Somerset Weather · Hyper-local Forecast

Claude's lab notebook

A working notebook for trying to improve the Somerset forecast. Ideas, plans, raw results and conclusions live here, including the ones that don't work. Nothing moves into the real forecast until it has passed a test it couldn't have known the answer to, and Brian has agreed.

How I work here
  1. Write down the plan and the passing mark before running the test.
  2. Grade only on days set aside: the model is built from earlier months and graded on later ones it was never built from (with an 8-day gap between them, so nothing leaks across).
  3. Compare against the forecast as it is today, on exactly the same days, and publish the result either way.
  4. Keep failures. A clean 'no' is a result.
  5. Promote nothing into the forecast without a solid result and Brian's OK.
49experiments3promoted to the forecast 15ideas waiting12research threads last updated November 18, 2026

Every experiment and case study

Newest first. Tap one to open it below. Experiments (L) test a change to the forecast against a bar written before the test; case studies (C) look back at one notable day or night.

Show:

Research

Long-running investigations, from the ground to the upper air. Each thread grows over many sessions and produces ideas; only a formal experiment below can change the forecast.

PROGRAMResearch program activeupdated October 16, 2026

Brian's charge (October 16, 2026): do novel research on every part of the forecast, from the ground to the upper atmosphere. Chase any connection that looks real. Research can run across many sessions and wait for data to build up. The goal is a more accurate model, and where possible a narrower +/- range, even if only for the nearest days.

This folder is where that research lives. Each thread is one text file that grows over time: what we know, what the literature says, what was explored, what is waiting for data, and what to try next. The Lab Notebook entries (L-00N) are the formal tests. A thread produces ideas; only a pre-registered test can turn an idea into a change.

How research becomes a forecast change

  1. Explore on the discovery years (walk-forward residuals for March 2024 to September 2025). Look freely. Nothing found here counts as evidence by itself, because looking at many ideas on the same days will always turn up some that work by luck.
  2. Pre-register the most promising idea as a Lab Notebook entry: question, plan and pass/fail bar written and published before the test.
  3. Test walk-forward on the test year (October 2025 onward), against today's model on identical days.
  4. Confirm on months nobody has looked at (from October 2026 on). The test year has now been used by many experiments, so a pass there is necessary but not enough for promotion.
  5. Promote only after Brian and Claude agree in a session.

The honest ceiling

Our day-1 errors are already small: about 1.6 F for lows and 1.8 F for highs on average (discovery years). Typical national guidance runs 2.5 to 3 F. What is left is partly weather nobody can see coming a day ahead, and partly the station's own noise. Gains from here will be tenths of a degree, hard won. Narrower ranges are more within reach than big accuracy jumps, because a range can shrink on the days we can tell are easy.

Data on hand (Tower: /lab/data, Mac: hyperlocal/data/research)

FileWhatPeriodUse it for
runs_gfs/ecmwf/icon/gem.parquetReal model runs (00Z, 12Z), 7 days, every level from 1000 to 300 mb plus CAPE, lifted index, freezing level, water vapour2026-04-02 on, grows dailyHonest tests at true lead times
upper_hist_*.parquetThe models' first hours stitched together, same variables2023 onLearning relationships only ("perfect prog"). NEVER score a forecast with it: it is far better than anything known at issue time
era5.parquetERA5 reanalysis at the station: soil moisture at 3 depths, soil temperature, boundary layer height, water vapour, evapotranspiration, vapour pressure deficit2023 to about 6 days agoLand surface and boundary layer research
asos.parquetHourly airport observations: Marion, Kokomo, Grissom, Muncie, Huntington, Fort Wayne, Logansport, Anderson (temperature, dew point, wind, gusts, pressure, cloud layers and heights, visibility, weather)2023 onNeighbours, upstream clouds, siting checks
trainer data/cumulus_minutes.parquet and data/truth_cumulus.parquet (Mac: hyperlocal/data)The station's own MINUTE record before Weather Display, from Cumulus: October 14, 2017 to March 8, 2023 (every 10 minutes until January 2018, then every minute), temperature, humidity, dew point, pressure, wind, 10-minute gust, direction, rain, solar. Matches Weather Display minute for minute in the February 2023 overlap. truth_cumulus has the same daily windows as truth.parquet, with 14 frozen stretches and 16 frozen or copied days removed (qc_report_cumulus.json)Fog history (R-012), cold-air drainage and wind by direction (R-004), more nights for the physics (R-010), WxSim grading on our own day and night windows (R-009). Not in the forecast model's training (its model archives start 2022 and 2024)
data/research/nws_products_point.json (Mac; copy on WorkingDrive/hyperlocal-lab/data)Every NWS watch, warning and advisory for the station's point, October 2017 to October 20, 2026 (689; IEM vtec_events_bypoint; zone INZ024): dense fog, winter weather, wind, frost and freeze, severeTruth for fog (R-012), winter weather and the significant-weather rules; refresh with the same IEM call
WorkingDrive "4. Archives/Meteorology" (NOAAMO*, NOAAYR*)Weather Display's NOAA monthly and yearly climate reports: the station's daily high, low and rainFeb 2018 to Jun 2020 (the full set to March 7, 2023 is in WorkingDrive "shardir/noaa reports/Reports", found 2026-10-20)Grading WxSim's 2018-2020 forecasts (archive: WorkingDrive "4. Archives/WxSim/wxsim/f*.txt", 5,651 runs 2018 to March 2023)
/trainer/data/*Station truth, 1-minute station history, model forecasts by lead (surface), sunshine, CAPE, wind direction, pressure2023 onEverything
/hl/hyperlocal.dbLive collector: observations, model snapshots, the 122-run ensembles (saved since mid-October 2026), NWS discussions and alertsliveRecent and live studies

python /lab/research_data.py runs adds the newest model runs (the morning watcher does this daily); era5, asos and upper top up the others.

Threads

ThreadQuestionStatus
R-001Can we tell easy days from hard days, so the +/- is narrow when it can be?exploring: small lead found
R-002Night cooling: what controls how cold it gets, and what do we miss?exploring: busts follow calm (decoupling) nights the models can't see
R-003The atmosphere above: mixing, inversions and the upper air for highs and lowsL-004 not passed: right direction, too small in summer; winter check scheduled
R-004The land: soil moisture, crops, leaf-out and leaf-off, snowexploring: dry soil lines up with highs running too warm (unexpected sign)
R-005Rain: moisture depth, lift, storm type, timing, and the gauge itselfnot started
R-006Neighbours and upstream: what nearby stations see before we doexploring: upstream low cloud is weak
R-007Frost and freeze: are the chances honest near the threshold?exploring: slight warm bias near freezing
R-008Hour by hour: our own station-corrected forecast for every hour of the next two daysL-003 not passed (9.6% vs 10%, better every month); L-005 confirms on fresh months
R-009WxSim: how good has it been, and could it help our forecast?WxSim adds about 0.01-0.02 F on discovery years; L-006 tests it on fresh months
R-010Our own physics: a night-cooling model built for this stationversion 1 live (L-007 passed); 1G evening layer on trial (L-026), grass start (L-028); every open problem traces to the grass's cooling at dusk: fit it on probe evenings (Nov 2026)
R-011What the experiments teach, read together: patterns across all testsseven rules from the first read-through (new information wins; small-effect ceiling; alternate-month splits; near-analysis flatters; fix roots; live trials for narrow misses; physics for rare regimes and timing). Re-read monthly. November 2 re-read: rules 1-8 hold, rules 9-13 added (bundles hide parts; regime claims need the other regimes; calm-versus-breezy misses; entries tied to replaced structures; thin historical checks)
R-012Fog at the station: knowing it, forecasting it, and the morning commuteexploring: the station saturates on 28% of nights (discovery days), the airports see fog on 11%; Brian's fog and frost log started October 20, 2026

Library (literature behind the threads)

Questions for the library

The unattended lab has no web browsing (a safety choice: see the brief). When a session needs literature, it adds a question here and Claude looks it up in the next session with Brian.

  • (none yet)
R-001Easy days and hard days: a +/- that knows how confident to be exploringupdated October 18, 2026

The question

Every forecast carries an 80% range (the +/-). Today it widens only when the four models disagree (expected miss = a + b x model spread). Can we tell easy days from hard days better than that, so the range is narrow when the weather is predictable and honestly wide when it is not? Brian asked for a smaller +/-; this is the most direct route.

Why it might work (physics and literature)

  • The spread of an ensemble carries information about the size of the error (the spread-skill relationship). EMOS (Gneiting et al. 2005) builds the forecast spread from a constant plus the ensemble variance and gives better calibrated ranges than a fixed width. We use 4 models; the 122-run ensembles we now save are a richer signal.
  • Some weather is simply easier: a sunny afternoon with a steady wind mixes the air predictably; an overcast day whose clouds may or may not break, a front arriving mid-day, or a snow-covered night are hard.
  • Our discovery data already shows it for highs (below).

What has been explored

2026-10-18, cold and warm spells (from Brian's review of a long-running physics forecaster's history). That program found its forecasts erred differently in unusually cold or warm spells, and warned that measuring it against what HAPPENED instead of what was FORECAST manufactures a false pattern. Checked on our forecast, discovery years, out-of-sample predictions, leads 1 to 7, with 1991-2020 normals for daily highs and lows (ERA5 at the station, shifted to the station by month; research/normals.py, research/spells.py).

Error (observed - forecast) per degree the FORECAST is from normalDay 1Day 3Day 5Day 7
Lows+0.02+0.00+0.01+0.08
Highs+0.01+0.00+0.01-0.03
Lows, measured against what happened instead (the false pattern)+0.07+0.10+0.18+0.36

No spell bias: our forecast is neither too timid nor too bold in big cold or warm spells, at any lead, and a cross-validated correction made every lead slightly worse (0.02 to 0.06 F). Spells are HARDER (day-1 lows 2.0 F off when forecast 12+ F below normal, 1.4 F near normal) but not biased. The false pattern is real and grows with lead: calibrating on observed departures would have pushed our day-7 forecasts a third of a degree further from normal per degree, the wrong way. Every correction we test is built on what is known at issue time, which avoids it. One thing to watch: day-1 lows forecast 12+ F below normal ran 0.7 F too warm (36 nights, within noise). The deep cold this winter will tell.

2026-10-16, discovery days (Mar 2024 to Sep 2025), leave-one-month-out. Expected miss from model spread alone vs. spread plus weather regime (clear-and-calm index, model cloud and wind, season, recent error size, snow on the ground, cold nights). Both scaled so 80% of misses fall inside.

today's rule: avg +/-regime: avg +/-regime: narrowest 10% of daysCRPS change
Highs, day 12.82.91.8-0.7%
Highs, day 23.73.52.2-0.3%
Lows, day 12.72.72.2+0.2%
Lows, day 23.43.32.8+0.3%
Highs, day 77.77.74.8-0.4%

(CRPS: the standard score for a whole forecast distribution, lower is better.)

What it means so far: the regime version spreads the ranges out (easy days get much narrower ranges, hard days wider) while keeping 80% coverage, but the overall score barely moves. Sunny light-wind days have smaller misses in the highs (1.7 F vs 2.4 F on day 1-2). Snow on the ground makes lows harder (2.4 F vs 1.9 F, 40 nights). A lead, not a finding.

Scripts: hyperlocal/research/r001_explore.py, results r001_explore.json.

2026-10-18, the miss review look-back (Brian asked what we do with big misses, after the October 18 high came in 68.4 F against 63.3 F). Our day-1 high and low as forecast walk-forward (each day from data before it), with the conditions of each day: the models' forecast (disagreement, cloud, sunshine, dryness) and what happened (pyranometer sunshine, wind, dew point swing, rain). Patterns looked for on March 2024 to July 15, 2026 only; July 16 to October 12, 2026 held back to check them. Script: hyperlocal/research/miss_lookback.py (output miss_lookback.out).

DaysTypical missOutside the 80% range (should be 20%)5 F+ misses
Highs, exploration8341.71 F24%31 (3.7%)
Highs, held back791.74 F28%3
Lows, exploration8381.68 F21%20 (2.4%)
Lows, held back781.47 F19%1
Highs: share outside the 80% rangeExplorationHeld back
Forecast cloud 80%+30% (207 days)40% (15)
0.1 in of rain or more in the window34% (89)55% (11)
All other days21% (596)22% (59)

What it shows:

  • The +/- for the high is honest on ordinary days and too narrow on cloudy and wet days. That is idea 3 below, answered for highs: the range is right on average but wrong on the days that matter most. Lows are close to 80% in every group checked.
  • Big misses are mostly misses the four models made too: on 23 of the 31 highs off by 5 F or more, the models' average was off by 4 F or more as well. Our correction cannot fix what every model gets wrong. 42% of the big misses came on days whose sunshine differed from the forecast by 15% of clear sky or more (21% of all days): cloud timing again (R-010).
  • Days like October 18 (September to November, 80%+ sunshine, dry air): 76 days, typical miss 1.36 F, came in 0.41 F warmer than forecast on average, 18% outside the range. No sign of a systematic problem; October 18 was a rare day (the held-back 12 such days: +0.76 F, 33% outside, too few to say more).
  • Lows ran 0.9 F too cold on nights the dew point held steady (306 nights; held back +1.27 F, 14 nights) and 0.9 F too warm when it swung 12 F or more (an air-mass change; held back -0.58 F). The swing is measured after the fact, so it cannot be used as it is; the question is whether the models' FORECAST of the swing carries the same signal.

Ideas to chase

0b. ~~Lows: does the models' forecast overnight dew point change predict the warm/cold lean above?~~ Checked

  1. Wider ranges for the high on days forecast cloudy (80%+) or wet (rain chance high): fit the extra width on the exploration days, test coverage on fresh days (the held-back months are too few cloudy days to judge). 2026-10-18 (research/airmass_lows.py, airmass_lows.out): no. The day-before forecast change matches the real swing well (correlation 0.75 with the observed overnight dew-point range, 838 exploration nights) but not our miss (0.01); a leave-one-month-out correction made the lows slightly worse (1.68 to 1.70 F, worse in every resample). The after-the-fact pattern comes from the UNFORECAST part of the change (fronts early or late, air masses the models missed), which only shows once the night is under way: a lead for the live adjustment (L-024: the station's dew point running against the forecast as a second signal), not for the forecast made the day before. Closed; the check months were not used. Live lead checked the same evening (research/livetrack_dew.py, livetrack_dew.out): adding the station's dew point gap to the tested live adjustment (same walk-forward, each month fitted on earlier days only) changed tonight's low by 0.002 F (exploration, 263 days) and 0.003 F (check, 85 days); even on hours with the dew point 5 F or more off the forecast, 1.54 to 1.52 F. For the high: -0.005 F (exploration), +0.018 F (check), inconsistent. When an unforecast air mass arrives the station's TEMPERATURE leaves the forecast too, and the live adjustment already follows that. Closed.
  1. The 122-run ensemble spread as the confidence signal (needs about 6 months saved: spring 2027).
  2. Upper-air stability and mixing as a confidence signal for highs: strong inversions and shallow mixing make highs hard (R-003 data).
  3. Check that the narrower ranges are honest within each regime (80% coverage on sunny days and on cloudy days separately), not just overall. A range that is right on average but wrong on the days it shrinks is worse.
  4. Different shapes: our misses may not be bell-shaped near freezing or on snow nights (see R-007).

Next step

When the run archive covers a winter, pre-register: "regime-aware ranges, highs days 1-2: mean +/- at least 5% narrower at 80% +/- 2% coverage on the test year, and coverage within 75-85% in each regime tercile." Then confirm on fresh months.

Log

  • 2026-10-16: thread opened by Claude with Brian. First exploration above.
  • 2026-10-18: cold and warm spells checked: no spell bias at any lead; the observed-departure false pattern confirmed.
  • 2026-10-18: miss review look-back: highs' range too narrow on cloudy and wet days; big misses are shared with the models; the live miss review (hyperlocal/misses.py) now records every finished period's conditions.
  • 2026-10-18: idea 0b checked and closed: the forecast air-mass change does not explain the lows' lean; the unforecast part might, live (L-024 lead).
  • 2026-10-18: the dew point as a second live signal checked and closed: the temperature gap already carries it.
R-002Night cooling: what decides how cold it gets exploringupdated October 17, 2026

The question

On a clear calm night the ground radiates heat to space and the air near it cools fast; wind stirs warmer air down; clouds send heat back; moisture puts a floor under the temperature because dew and frost release heat as they form. Our lows are already good (1.6 F average miss on day 1), but about 7% of nights miss by 7-9 F even when every model agreed. What are we missing, and can anything we can measure warn us?

Literature

  • Radiation-night minimum from the temperature and dew point about two hours after sunset (UC Davis), valid only on clear nights with wind under about 5 mph.
  • NWS TM 102: the dew point acts as a floor because condensation offsets radiative cooling; clouds or wind of any duration usually keep minima well above the radiative value.
  • NWS TM 106: raise a MOS minimum when it is 5 F or more below the afternoon dew point (fog and dew limit cooling).
  • The boundary layer can "decouple" on calm nights (wind under about 1 m/s near the ground): the surface layer cools on its own. Models handle this poorly and the sign of their error varies by model and surface scheme.

What has been explored

2026-10-16, discovery nights (Mar 2024 to Sep 2025, 560 nights, day-1 forecast). At 8-9 PM: temperature vs our low, dew point vs our low (the condensation floor), dew point depression, wind, the 2-hour cooling rate, pressure change. Leave-one-month-out correction of our low:

CorrectionMAE (F)
Our day-ahead low1.618
+ how the evening temperature is running (like today's live rule)1.634
+ dew point floor1.628
+ floor, wind, cooling rate, pressure1.624
+ calm-night physics interactions1.618

Nothing helps. The strongest single signal is the evening cooling rate (correlation 0.14 with the miss) and the dew point floor (0.11), both tiny. The models (through our corrections, which already include a clear-and-calm index and the dew point depression) already know the evening state. What causes the big misses happens later in the night (the next entry looks at what).

Scripts: hyperlocal/research/r002_explore.py, results r002_explore.json.

2026-10-16, bust night anatomy (same 560 discovery nights, station minute data). Nights grouped by our miss:

Night endedNightsModels' wind (mph)Station's wind overnight (mph)Share of calm minutesDew point change overnight (F)
4 F or more colder than forecast178.43.849%0.0
2-4 F colder727.23.050%-2.8
within 2 F3877.53.339%-2.2
2-4 F warmer757.83.925%-1.0
4 F or more warmer147.64.423%+0.7

The clearest signal so far: how much of the night goes calm. Nights that ended colder than forecast were calm half the time; nights that ended warmer were calm a quarter of the time (correlation with the miss -0.22, the strongest of anything tried). Within these groups the models' wind barely differs (7.2 to 8.4 mph), so the extra calm on bust nights is beyond what the wind forecast shows (across all nights the models' wind does track calm nights well; see R-003). This is boundary layer decoupling, as the literature describes. Air masses arriving overnight (temperature rising 3 F or more after 10 PM, or the dew point dropping 5 F or more) explain only 2 of 14 warm busts and 2 of 17 cold busts. Cloud surprises explain little (see R-006).

Script: hyperlocal/research/r002_bust_anatomy.py, results r002_bust_anatomy.json.

Ideas to chase

  1. Predict decoupling (top priority, with R-003). What could tell us in advance that a night will go calm at the ground? Candidates: the forecast wind at 925 mb (about 2,500 ft; a weak flow aloft lets the surface decouple, a low-level jet keeps it stirred), the strength of the evening inversion (925 mb minus 2 m temperature), the pressure gradient. Explore on upper_hist (perfect prog), then test on the real runs. A decoupling chance could also feed R-001: calm-prone nights deserve a wider range on the cold side.
  2. Soil and ground heat. Wet, warm soil in autumn releases heat at night (R-004 data: ERA5 soil temperature and moisture). Does it explain warm misses in fall?
  3. Done 2026-10-16: evening snapshot (null), upstream clouds (weak, R-006), bust anatomy (decoupling).
  4. Decoupling, measured (new 2026-10-17). The ground sensor (R-007, "Ground sensor data") gives the grass-minus-air gap and the soil heat profile every 5 minutes. Ideas: (a) an evening nowcast: does the gap at, say, 9 PM tell how far below our forecast the night's low will land? (b) does the 8-inch minus 2-inch soil difference (heat flowing up) explain warm misses on autumn nights (idea 2, now with our own soil instead of ERA5)? Exploration on the nights as they come; any correction is tested only on months after it is written down.

Log

  • 2026-10-16: thread opened by Claude with Brian (from the backlog idea "evening frost nowcast"). Null result above closes the simple version of that idea.
  • 2026-10-16: bust anatomy: decoupling (calm share) is the leading mechanism. Next: predict it from the air above.
  • 2026-10-16: R-003 perfect prog: the inversion strength lines up with the miss (strong inversion = colder than forecast), matching the known too-much-mixing bias of models on stable nights. Formal test planned in R-003.
  • 2026-10-17: our ground sensor went in (R-007). It measures decoupling directly; idea 4 added.
  • 2026-10-19 (the lab assistant's Tower session; written that morning, lost online to a Mac publish the next day and restored from the Tower's backup on 2026-10-26): live pair (see C-002, now C-004): breezy night to Oct 17 ran 4.8 F warmer than our forecast (36.0 vs 31.2); calm clear night to Oct 18 ran 3.4 F colder (31.5 vs 34.9). Models, NWS and ours all missed in the same direction each night. Two nights only; they fit the calm-versus-breezy switch.
  • 2026-10-19: from C-002: forecast winds aloft (925/850/1000 hPa) do not predict the night's inversion breaking in a usable way (strengthening overnight: weak, right direction); calm-and-clear-looking nights run 0.35 F colder than our forecast on average (L-008's blend leans right).
R-003The air above: mixing, inversions and the upper atmosphere L-004 not passed (right direction, too small in spring and summer); winter check scheduledupdated October 16, 2026

The question

Everything our model uses today is at the surface (2 m temperature, dew point, cloud, wind, sunshine, CAPE, pressure). The weather at the ground is decided partly aloft: how deep the afternoon mixing goes, how warm the air is at 850 mb (about 5,000 ft), whether a temperature inversion traps cold air, how moist the column is. Does the upper air explain misses that the surface fields cannot?

Physics and literature

  • Mix-down for highs: on a sunny day the boundary layer mixes, and the high approaches the 850 mb temperature brought down dry-adiabatically (about +13 C from 850 mb to the surface). It fails with cloud, rain, fronts or strong temperature advection, and in winter inversions (Haby hints 296; OU METR 2413).
  • Inversions and snow: in winter, a shallow cold layer under warm air aloft keeps highs below model values, especially with snow cover or low cloud.
  • Wind aloft at night: a strong 925 mb wind (low-level jet) keeps nights stirred and warm; a weak one lets the surface decouple (R-002).
  • Freezing level and the column's moisture matter for rain type and amount (R-005).
  • Models mix stable night air too much, on purpose. Operational models (ECMWF documented it) keep extra turbulent mixing in stable conditions because it helps the large-scale forecast, though it degrades near-surface temperatures on stable nights (Sandu, Beljaars, Bechtold, Mauritsen and Balsamo 2013, JAMES 5:117-133: https://www.ecmwf.int/sites/default/files/elibrary/2012/12086-why-it-so-difficult-represent-stably-stratified-conditions-nwp-models.pdf). Prediction: under strong inversions the models' nights run too warm.

Data

  • upper_hist_<model>.parquet: 2023 on, stitched first hours of each run (close to an analysis). Learning only.
  • runs_<model>.parquet: real runs from April 2, 2026 at true lead times. The only fair data for testing a forecast input. Grows by two runs a day per model.
  • Observed soundings (ILN Wilmington OH, ILX Lincoln IL, DTX White Lake MI) are available from the IGRA archive if needed to check the models aloft.

What has been explored

2026-10-16, decoupling and inversions, PERFECT PROG (GFS near-analysis aloft, 565 discovery nights, day-1 lows). An upper bound only: these values are known in hindsight.

What predicts a calm night at the station (correlation with the share of calm minutes, 11 PM to 7 AM): the models' 10 m night wind -0.75, our clear-and-calm index 0.79, the 925 mb wind -0.67, the inversion 0.44. So the models do know which nights will be calm in general; the 925 mb wind adds little to that.

But the inversion strength (925 mb temperature minus 2 m temperature, overnight) lines up with our miss, steadily:

Inversion overnightNightsShare calmOur miss (obs - forecast)
none (warmer at the ground)23922%+0.45 F
0-3 F13944%-0.05 F
3-6 F10050%-0.23 F
6-10 F6358%-0.58 F
over 10 F2470%-0.85 F

Under strong inversions the night ends colder than we forecast; with none it ends warmer. Exactly what the too-much-mixing bias predicts. Leave-one-month-out correction of our low: 1.618 -> 1.601 F with 925 mb wind and inversion, 1.587 F adding wind x clear sky and the evening inversion (2% better, an upper bound).

Caution: the inversion here uses the near-analysis's own 2 m temperature, which may partly stand in for knowing the night's outcome. Only the real runs (forecast inversion at true lead times) can say whether this survives.

Script: hyperlocal/research/r003_decoupling_pp.py, results r003_decoupling_pp.json.

2026-10-17, L-004 on the real runs (pre-registered). 177 nights, April to September 2026, ECMWF and ICON 00Z runs. Lows days 1-2: 1.541 -> 1.528 F (0.013 better; bar 0.02), positive in 74% of resamples (bar 90%). Strong forecast inversions (3 F or more, 72 forecasts): correction -0.42 F on average, error 1.181 -> 1.170 F (physics check passed). Day 1 +0.031 better, day 2 0.005 worse. Not passed. Winter confirmation (October 2026 to March 2027) still scheduled with the frozen correction; a pass there would justify a new pre-registered version, not promotion.

Next step (as pre-registered, now run: see L-004 in the notebook)

Formal test, written into the notebook before any run-period data is examined: forecast inversion strength (925 mb minus 2 m, 11 PM to 5 AM) and 925 mb wind from the real runs, the run issued closest to our issue time for each lead, as a correction to our day 1-2 lows. Perfect-prog design: learn the correction on upper_hist (2023 to September 2025), then apply it unchanged to the real runs' forecast values (the runs start too late to train on). Test months April to September 2026 (test-year months never used with these inputs), confirm on October 2026 to March 2027 (winter, when inversions are strongest). Draft bar: lows days 1-2 MAE at least 0.02 F better on both periods, and the gain coming from the strong-inversion nights (physics, not luck).

Ideas to chase

  1. Mix-down potential for highs: 850 and 925 mb temperatures brought down dry-adiabatically, compared with our forecast high, on sunny days. Does the gap predict our miss?
  2. Inversion strength (925 mb minus 2 m temperature; 850 minus 925) as a confidence signal (R-001) and as a correction in winter.
  3. Low-level jet at night: 925 mb wind speed after midnight vs our low's miss.
  4. Thickness and air mass: 1000-500 mb thickness (from geopotential heights) as an air-mass signal when the surface models disagree.
  5. Perfect-prog route: learn relationships on upper_hist (2023 on), then apply them to the real runs and test on the run period only.
  6. Aircraft profiles (Brian's question 2026-10-16). Airliners report temperature and wind (some also humidity) on every climb and descent; Indianapolis (FedEx hub, busy overnight) gives many evening and night profiles. They observe the low-level inversion directly, hours before the night plays out. Access: NOAA MADIS keeps aircraft data since 2001; most is public after 48 hours, real time only for the humidity-sensor aircraft (mainly UPS and Southwest) or with an approved MADIS application (https://amdar.noaa.gov, https://madis-data.cprk.ncep.noaa.gov/wmo_gdc.shtml). Research test with the delayed data: does the evening inversion measured by aircraft at IND explain our lows' miss better than the models' inversion? If clearly yes, that justifies applying for real-time access.

Log

  • 2026-10-16: thread opened. Backfill of upper-air history (4 models, 2023 on) and real runs (April 2026 on) started.
  • 2026-10-16: perfect-prog inversion signal found (above). Next: pre-register the real-run test.
  • 2026-10-17: L-004 pre-registered and run: right direction, too small in spring and summer. Not passed.
R-004The land: soil, crops, leaves and snow exploring: puzzling first resultupdated October 20, 2026

The question

The station sits in north-central Indiana farm country. The ground changes through the year: bare fields in spring, corn and soybeans pumping water into the air in July, harvest in October, leaves on and off, snow cover. Wet soil spends the sun's energy evaporating water (cooler highs, higher dew points); dry soil heats the air (hotter highs, bigger day-night swings). Do our forecast misses follow the state of the land?

Physics and literature

  • Mueller et al. (2016): cropland intensification across the Midwest cooled the hottest summer days through evapotranspiration; no cooling during droughts, because transpiration needs soil water.
  • Soil moisture errors in models shift 2 m temperature forecasts by up to about 2 C (Geophysical Research Letters 2024), with the biggest effects when soils are dry.
  • "Corn sweat": peak-season corn raises dew points in July, a known Corn Belt effect on heat index forecasts.

Data

  • era5.parquet: soil moisture at 0-7, 7-28 and 28-100 cm; soil temperature; evapotranspiration; vapour pressure deficit. Known with about a 6-day delay, so a live version would use the station's own rain history or a model's soil analysis instead.
  • The station's rain history (minutes.parquet) gives antecedent rain for free, live.
  • USDA weekly crop progress for Indiana (planting, silking, harvest dates) would date the crop calendar each year.

What has been explored

2026-10-16, soil moisture vs. our day-1 highs (561 discovery days). ERA5 soil moisture anomaly (vs. the same time of year) at 7 AM; the station's rain over the previous 7 and 14 days.

Correlation of our miss (observed minus forecast) with the soil moisture anomaly, sunny days in brackets: spring +0.13 [+0.24], summer +0.15 [+0.14], fall +0.17 [+0.27 at 0-7 cm, +0.37 at 7-28 cm], winter -0.11 [-0.15].

Spring and summer highs by soil moisture (7-28 cm): driest fifth of days -0.89 F (we ran too warm), dry -0.16, normal -0.57, wet -0.18, wettest fifth -0.11.

The sign is the opposite of the textbook. Dry soil should make the real high hotter than a model expects; here, on dry-soil days, the day came in cooler than our forecast. Possible readings: (1) the models already over-react to dry topsoil (their soil schemes dry the surface fast, while deep-rooted corn and soybeans keep transpiring from water a metre down); (2) a confound: the dry spells of 2024 and 2025 fell in particular weather patterns. As a correction it is worthless: leave-one-month-out MAE 1.805 -> 1.805 (soil), 1.817 (station rain, worse).

Script: hyperlocal/research/r004_soil_explore.py, results r004_soil_explore.json.

2026-10-18, the field south of the mast: corn or soybeans (research/crop_check.py, crop_check.out). Brian's field log: strict rotation, corn in even years (2024, 2026), soybeans in odd years (2023, 2025); corn harvest usually late October to early November, 2024's on October 14. Against the four-model day-1 blend (which knows nothing about the field):

July-AugustDew point, station minus modelsDaytime wind, station / modelsOur high came inOur low came in
2024 corn+3.8 F0.56+0.08 F-0.37 F
2025 soybeans+3.6 F0.64-0.10 F+0.15 F
2026 corn+3.2 F0.49+0.52 F-0.28 F
Mid-September to mid-OctoberCropOur high came in (all days / sunny)
2024corn, standing until October 14+1.43 / +1.50 F
2025soybeans+0.02 / -0.28 F
2026corn, still standing+1.18 / +1.28 F
  • The station is about 3.5 F more humid than the models every summer, corn or beans: a station or regional offset, not the crop.
  • Tall corn shelters the daytime wind about 15% more than beans; after the 2024 harvest the daytime wind ratio rose from 0.57 to 0.67 while the same weeks of 2025 fell slightly (0.77 to 0.72).
  • The clearest signal: in both corn falls, the highs ran about 1.3 F warmer than our forecast from mid-September until harvest (sunny days most), against none in the soybean fall. A plausible mechanism: a dry, senescent standing canopy evaporates almost nothing and shelters the mast, so the sun heats the air; beans come off earlier. After the 2024 harvest the lean fell to +0.84 F. October 18, 2026 (corn standing; high 68.4 against 63.3 forecast) fits it.
  • Caveats: two corn falls and one bean fall; fall 2024 was also a dry fall in Indiana, which by itself warms highs. Nothing is changed. The model already learns an AVERAGE field; it cannot know which crop is in.

The site and its surroundings (from Brian, 2026-10-20)

Distances are from the mast, by Brian's estimate.

  • South: a corn and soybean field begins a few feet from the mast base. The ground is hilly and falls toward bottom ground at least 30 ft below the base of the mast.
  • North: an alfalfa field from about 15 ft out.
  • East: a line of trees (a tree line, not a woods) at about 500 ft; another corn and soybean field beyond it.
  • West: an alfalfa field from about 500 ft, with a shed and a grain bin (about 25 ft tall).
  • Northwest: the house, about 1,500 ft away, on about 2 acres of yard. About 200 ft west of the house, a 2-acre pond with heavy woods on its west and north sides.
  • North, about 3/4 mile: a reservoir, with heavy woods on its south face.

Brian's answers (2026-10-20): the mast is on a LOCAL HIGH POINT, the ground falling away mainly to the south; the tree line to the east is about 40 ft tall; the reservoir is LOWER than the mast; the fog camera looks EAST and does not see the bottom ground. So: the mast should shed cold air on calm nights (a thermal belt site), the trees block the sun below about 4.6 degrees in the east (roughly the first 20 to 30 minutes after an eastern sunrise in spring and fall), and the fog log describes fog at the mast's level to the east, not in the bottom, where fog likely forms more often.

What this could mean (hypotheses to test, not findings):

  1. Cold-air drainage. On clear, calm nights the coldest air slides downhill. The mast sits above bottom ground 30 ft or more lower, so the coldest air may drain away from it and pool in the bottom: the mast's low would run warmer than flat ground, and the bottom colder, with frost and fog forming there first. That fits the fog study's note that local fog forms in low spots the airports miss, and possibly the October 16-17 night (C-001), when the mast stayed at 36.0 F against a 31.2 F forecast. It does not explain October 17-18 (C-004), which came in colder than forecast, so drainage may depend on wind direction or on how calm the night stays.
  2. Water upwind. With a light north wind the air crosses the reservoir (3/4 mile) and the pond and woods; in fall, water warmer than the night air adds moisture and can make steam fog. Fog nights at the mast may favor a north or northwest drift.
  3. Wind shelter by direction. The tree line about 500 ft east and the bin and shed about 500 ft west may slow the wind the anemometer (12 ft) sees from those directions; standing corn to the south slows a south wind in summer and fall. Alfalfa to the north is short, so a north wind should reach the mast least disturbed.
  4. Low sun. The trees (about 40 ft) block the sun below about 4.6 degrees above the eastern horizon, delaying the first direct sun on the solar sensor and the grass probe by some minutes after sunrise; the bin does the same, less, at sunset.

Field log

YearCropHarvest
2023soybeansunknown
2024cornOctober 14
2025soybeansunknown
2026corn(standing on October 18)

Ideas to chase

  • (Site, 2026-10-20) Station wind against the models' wind by direction (8 sectors, discovery days): is the east and west sheltering, and the south corn effect, visible? A direction-aware wind scaling would help the spray windows and the physics.
  • (Site, 2026-10-20) Cold-air drainage: on clear, calm nights, does the mast's low run warmer than our forecast more than on breezy nights, and does it depend on the drift direction (downslope toward the south bottom versus upslope)? Discovery days only.
  • (Site, 2026-10-20) Morning and evening shading: solar sensor against clear sky by sun azimuth and elevation in the first and last hour of the day.
  • Standing corn in fall: pre-registered as L-027. Its historical part (WxSim against the 2018 NOAA reports; 2020's reports stop in June) went the other way: in 2018 the high ran 1.9 F COLDER relative to WxSim from mid-September than in late summer. Not supported so far; the fresh falls (2026 until harvest, 2027 beans, 2028 corn) decide it. Originally: pre-register "in corn years, mid-September until harvest, our high runs warm" and correct it as a crop-and-season term; the 2026 days until harvest are a first fresh check, 2028 the next full one. A field log on the site (crop, planting, harvest: a few taps a year) makes it usable.
  1. Untangle the dry-soil sign: does the models' own forecast high on dry days run hotter than ours (i.e. are the models over-reacting and our correction only part-way)? Compare raw model highs vs. truth by soil anomaly.
  2. July-August dew point miss vs. crop stage: do models under-forecast dew points at peak corn?
  3. Leaf-off (late October) and harvest: does the low's bias shift in the weeks after?
  4. Snow cover: lows are harder with snow on the ground (R-001 found 2.4 F vs 1.9 F, 40 nights); is the miss mostly one sign (models too warm over snow)?
  5. Done 2026-10-16: soil anomaly by season and station antecedent rain (above).

Log

  • 2026-10-17: the station mast stands a few feet north of a corn field (Brian). The crop cycle (planting, canopy, harvest, corn/soybean rotation) may leave a seasonal mark on the station's own readings; log crop and dates yearly.
  • 2026-10-17: siting details (Brian): the Davis station is on the same mast; air sensor about 5 ft (below the top of standing corn), anemometer about 12 ft. The 2026 corn is standing and dried down, harvest date to come. 12 ft is only about 4 ft above the canopy, so winds from the south (over the corn) likely read slower and gustier than north winds until harvest; check wind speed by direction before vs after harvest in the minute data.
  • 2026-10-16: thread opened; ERA5 soil and boundary layer history (2023 on) saved.
  • 2026-10-16: first exploration: soil moisture lines up with our highs' miss with the unexpected sign; no gain.
  • 2026-10-18: field log from Brian (rotation, 2024 harvest Oct 14); corn falls run ~1.3 F warm on the high until harvest, beans not; summer wind sheltering by corn.
  • 2026-10-18: L-027 part A (2018 WxSim vs NOAA) not met: the corn-fall effect is not supported historically so far.
  • 2026-10-20: Brian described the land around the mast (new section "The site and its surroundings"); three site ideas added.
  • 2026-10-20: Brian's answers: mast on a local high point (ground falls mainly south), tree line about 40 ft, reservoir lower than the mast, fog camera faces east (cannot see the bottom).
R-005Rain: moisture, lift, storm type, timing, and the gauge not startedupdated October 16, 2026

The question

Our rain chance is a calibrated blend of the models' rain amounts, the ensemble and the NWS discussion. It uses surface fields only. Rain needs moisture through a deep layer and something to lift it; models are often right about rain but wrong about when, or get showers right in general but not over one gauge. Where do our rain chances go wrong, and which of those causes can we measure?

Physics and literature

  • Classic MOS rain chance equations lean on the mean relative humidity of the column, 850 and 700 mb humidity and the model's rain amount; precipitable water, K index and vertical motion are secondary (NWS ER TA 99-1; MDL TPB 487).
  • Convective showers are hit or miss at a single point: a 60% chance of storms means scattered cells, not 60% of the area wet everywhere. Stratiform rain from a large system is far more predictable.
  • Tipping-bucket gauges undercatch, most in wind and in light rain (splash, evaporation, wetting loss). A 0.01 in tip threshold means drizzle can fall without counting (Dunn et al. 2025 review).

Data

  • Real runs (April 2026 on): 850/700/500 mb humidity, vertical velocity, water vapour, CAPE, lifted index, freezing level at true lead times.
  • upper_hist (2023 on) for learning.
  • Airports (asos.parquet): hourly precipitation and weather codes (RA, DZ, TS) to separate drizzle from rain and to see whether rain fell nearby but missed the gauge.

Ideas to chase

  1. Miss anatomy: sort discovery-year misses (wet when we said under 30%, dry when we said over 70%) by storm type (CAPE, lifted index), season and timing.
  2. Deep moisture: mean 850-500 mb humidity from the models as an extra input (perfect-prog first, then on runs).
  3. Timing: for missed rain, did it fall within 3 hours of a period boundary? If so the miss is timing, not occurrence, and a probability spread across the boundary might help.
  4. Gauge check: days when two or more neighbouring airports had measurable rain and our gauge had none (or 0.01 in). How often, and in what wind? An honest estimate of what the gauge misses.
  5. Showers vs. steady rain: separate calibration for convective and stratiform regimes.

Log

  • 2026-10-16: thread opened.
R-006Neighbours and upstream: what nearby stations see first exploringupdated October 16, 2026

The question

Weather moves. Most of ours arrives from the west and southwest. Eight airports within about 70 miles report every hour: temperature, dew point, wind, gusts, pressure, cloud layers with their heights, visibility and present weather. They measure what is happening while the models can only predict it, and weather from the west reaches them hours before it reaches us. Can they correct our nearest forecasts (tonight, tomorrow) and the live track?

Stations

Marion (MZZ, closest, about 15 miles), Kokomo (OKK, west), Grissom AFB (GUS, northwest), Logansport (GGP, northwest), Muncie (MIE, southeast), Anderson (AID, south), Huntington (HHG, northeast), Fort Wayne (FWA, northeast), Rochester (RCR, north), Warsaw (ASW, north). Outer ring (50 to 105 miles): TYQ, LAF, IND, SBN, AOH, DNV, DAY, GYY.

Ideas to chase

  1. Upstream cloud for bust nights (with R-002): at 8 PM, cloud cover and ceiling at the stations upwind (chosen by the 850 mb wind direction) vs. what the models have for our sky after midnight.
  2. Neighbour consensus: the average of their evening temperature anomalies vs. ours: is our station running unusually cold or warm, a siting signal?
  3. Siting: compare our station with Marion on clear calm nights. A consistent difference measures our cold-air pooling, and its size may depend on wind and season.
  4. Live track input: neighbour temperatures as an extra signal in Today's Track.
  5. Rain nearby (with R-005): did rain fall at neighbours when our gauge recorded none?
  6. Aviation forecasts (TAFs, Brian's idea 2026-10-16). Forecasters write hour-by-hour cloud, visibility, wind and rain timing for Fort Wayne (KFWA), Indianapolis (KIND) and Lafayette (KLAF), 24-30 hours out. The Iowa Environmental Mesonet archives them, already parsed (cloud layers and heights, weather, timing). Questions: do TAF cloud groups for tonight separate the clear-calm bust nights better than the models' cloud? Does TAF rain timing sharpen our 12-hour rain chances near period boundaries? Note: the forecast discussion we already read includes an aviation section that summarises the same TAFs, so the new information is the exact timing and structure. Pilot reports (PIREPs) and AIRMETs: too sparse or too broad for a point forecast.

What has been explored

2026-10-16, upstream clouds and bust nights (560 discovery nights, day-1 lows). Cloud cover from airport reports (the highest of the two lowest layers) at 8-9 PM, and through the night at all eight airports, compared with the models' night cloud.

Night endedNightsModels' cloudAirports' cloud overnightSurprise
4 F or more colder1738%30%-9 points
2-4 F colder7048%43%-6
within 2 F38546%39%-6
2-4 F warmer7450%51%0
4 F or more warmer1452%58%+6

Cloud surprises lean the expected way (clearer than forecast = colder) but weakly (correlation 0.13). The west airports' evening cloud does not predict the miss (correlation 0.06; leave-one-month-out MAE 1.620 vs 1.612 for our low, no help). Caveat: airport ceilometers only see clouds below 12,000 ft, so high cloud is invisible here. Conclusion: upstream low cloud is not the bust mechanism at our station; decoupling is (R-002).

Script: hyperlocal/research/r006_upstream_clouds.py, results r006_upstream_clouds.json.

2026-10-18, the outer ring (Brian: should the ring cover 12 to 24 hours of advection?). Eight more airports about 50 to 105 miles out: Indianapolis Executive TYQ (49 mi, SSW), Lafayette LAF (60, WSW), Indianapolis IND (71, SSW), South Bend SBN (76, NNW), Lima AOH (94, E), Danville DNV (98, WSW), Dayton DAY (99, ESE), Gary GYY (106, NW). History from 2023 (backfill_outer.py). Same question as the inner ring: does the outer upwind airports' model error now predict the station's error hours later, beyond its own error and the inner upwind airports? Downwind is the control.

Hours aheadOwn error+ inner upwind+ outer upwind+ outer downwind (control)
Freshest-view errors, discovery hours, wind 5 mph+ (about 7,850 h)
30.9400.9300.9280.931
61.0381.0361.0351.036
121.0971.0971.0971.098
Real forecasts 6+ h old, April 2 to July 15, 2026 (about 2,290 h)
41.8491.8011.7961.802
61.9781.9481.9511.952
122.1022.1062.1072.112

On windy hours (12 mph or more) the outer upwind side does beat the downwind side by 0.004 to 0.009 F at 4 to 9 hours, so the direction effect is genuine, but it is a hundredth of a degree. Conclusion: no. Beyond about 6 hours the forecast runs themselves have already taken in the air mass that is upwind (they are rerun every 6 to 12 hours from the same observations), and what is left of the station's error is local (siting, the ground, the corn) rather than carried in. The useful upwind reach is the inner ring and the first few hours (L-013). Not pre-registered: nothing worth testing. Scripts: physmodel/outer.py, outer_error_hours.parquet.

Data

asos.parquet, 2023 on (Iowa Environmental Mesonet). Live use would need the collector to fetch the same hourly reports (public, free).

2026-10-18, the ring goes live (ring_now.py). Brian pointed out the station sits inside a ring of automated airports (Marion 14.6 mi SSE, Kokomo 15.5 SW, Grissom 17.7 W, Huntington 23.2 ENE, Logansport 28.6 W, Rochester 33.1 NW, Muncie 37.6 SE, Fort Wayne 38.7 ENE, Anderson 40.3 SSE, Warsaw 41.7 N; Wabash has no automated reports). The Tower now reads their latest reports every 15 minutes (aviationweather.gov, table ring_obs) and pushes an "around us" summary to the Briefing: sky and low cloud bases, rain or snow and where, fog, wind, the station against the ring, and the air arriving from upwind (the upwind airports' temperature and dew point against ours, and the travel time at the current wind). Rochester and Warsaw were backfilled into the archive (asos.parquet) to close the ring north. New idea (Brian, after WxSim's upwind METAR advection): OBSERVED ADVECTION OF MODEL ERROR. If the upwind airports are colder or drier right now than the models think, that error should drift here within an hour or two, before the next model run corrects it. Test: for each hour, the upwind ring stations' (observed minus model) temperature and dew point, carried by the wind, against the station's next 1 to 3 hours' (observed minus model). Needs the models' values at each ring airport (the Open-Meteo previous-runs archive by point) for discovery, then fresh months live.

2026-10-18, observed advection of model error: first look (physmodel/advect_error.py, fetch_ring_models.py). Model values (GFS + ICON stitched first hours, the freshest run's view of each hour) fetched at all ten ring airports (data/research/ring_models.parquet). Each hour: every airport's error (observed minus models) and the station's; the ring wind (1/distance vector mean); upwind = airports within 50 degrees of where the wind comes from, downwind = within 50 degrees of where it goes (the control). Discovery hours (March 2024 to September 2025), wind 5 mph or more, about 8,200 hours, cross-validated by 6 blocks, uncertainty by resampling whole days. Temperature error 1/2/3 h ahead, mean absolute error, own error only -> + upwind: 0.667 -> 0.656, 0.852 -> 0.836, 0.945 -> 0.934 F; gains 0.011 (95% 0.005 to 0.017), 0.016 (0.009 to 0.024), 0.011 (0.005 to 0.016) F; downwind gain 0.000 every time; the whole ring in between. The gain grows with wind (2 h: 0.016 at 5-10 mph, 0.018 at 10-15, 0.024 at 15+). That is the signature of air actually arriving: real, but small (1.5 to 2.5%). Dew point: no upwind edge (moisture errors look regional). Caveat: the "model" here is the freshest run's first hours, so the models have little error left to advect; the same idea against our issued forecast (hours old, less fresh) could matter more, and is the natural next step (a 1 to 6 hour nowcast of the station's temperature).

2026-10-18, L-013 PASSED: a 1 to 6 hour nowcast. From a real forecast (ECMWF + ICON, the newest run out at least 6 hours before), nudged by the station's own error and the upwind airports' error: test hours July 16 to October 11, 2026, averaged over 1 to 6 h: forecast 2.13 F, + own 1.76 F, + upwind 1.68 F (gain 0.083 F, 95% 0.056 to 0.111), + downwind instead 1.73 F. Against hours-old forecasts the upwind effect is about ten times what it is against the freshest runs. See the lab notebook, L-013.

Log

  • 2026-10-16: thread opened; eight airports' hourly history (2023 on) saved.
  • 2026-10-16: upstream clouds explored: weak. Siting and neighbour-consensus ideas still open.
  • 2026-10-18: the ring nowcast went live; idea added: observed advection of model error from upwind airports.
  • 2026-10-18: observed advection of model error: real (upwind beats downwind, grows with wind) but small (about 0.01-0.02 F).
  • 2026-10-18: L-013 passed: the upwind airports improve a 1 to 6 hour nowcast by 0.08 F on top of the station's own error.
  • 2026-10-18: outer ring (8 airports 50 to 105 miles out) explored: adds about 0.00 to 0.01 F at 4 to 12 hours; not worth carrying.
R-007Frost and freeze: are the chances honest near the threshold? exploringupdated October 17, 2026

The question

Our frost (36 F or lower) and freeze (32 F or lower) chances come from a bell curve centred on our forecast low, as wide as the +/- range. That assumes our misses are the same size and shape on cold nights as on any night. Near freezing, physics says they may not be: dew and frost forming release heat and slow the last few degrees of cooling; wet ground gives up heat; clear calm nights can plunge.

What has been explored

2026-10-16, discovery nights (Mar 2024 to Sep 2025), forecasts 1-3 days ahead (1,695 forecasts).

Frost (36 F or lower), 408 occurred. Brier score 0.028 vs 0.183 for climatology (much better).

We saidNightsSaid on averageHappened
0-10%595%12%
10-30%5019%20%
30-50%3238%50%
50-70%3261%66%
70-90%4080%63%
90-100%33899%97%

Freeze (32 F or lower), 309 occurred. Brier 0.018 vs 0.149.

We saidNightsSaid on averageHappened
0-10%475%6%
10-30%4219%7%
30-50%2439%29%
50-70%3661%67%
70-90%4281%69%
90-100%24599%99%

Nights forecast within 6 F of freezing ended on average 0.55 F warmer than forecast (frost: 0.42 F warmer).

What it means so far: broadly reliable. Two hints, both from small samples: low freeze chances (10-30%) came true less often than we said, and high ones (70-90%) for both frost and freeze also fell short. Both fit a slight warm tendency near the threshold, as the condensation-floor physics predicts. A lead, not a finding: the bins hold only 24 to 59 nights each.

Script: hyperlocal/research/r007_frost_check.py, results r007_frost_check.json.

Ideas to chase

  1. Is the near-freezing warm tendency tied to the dew point (nights where the dew point is near or above the forecast low)? That would be a physical, testable correction.
  2. Ground heat: soil temperature (ERA5) on early-season frost nights.
  3. A skewed curve instead of a bell curve for cold nights.
  4. Test on this autumn and winter (fresh months) before anything changes: there will be many frost nights.
  5. Our own ground sensor: BUILT 2026-10-17. See "Ground sensor data" below. It measures real grass frost, so frost (grass at or below 32 F) can now be scored directly, not just freeze (station air). Once a season of nights is in: forecast the grass low as our air low minus a predicted grass-below-air gap (from clouds, wind, dew point). Exploration only until then; the first fair test is autumn 2027.

Ground sensor data (from 2026-10-17)

  • Table ground_obs in /hl/hyperlocal.db: t_utc, seq, probe, position, temp_f, vbat, rssi, snr; one row per probe every 5 minutes. Positions: grass (2 in above short grass), soil_2in, soil_4in, soil_8in, all at the station mast, beside the station's own air sensor (about 5 ft up). Rows with position test are readings taken while testing the probes (hand-warmed) and must never be used.
  • Site: a corn field begins a few feet SOUTH of the mast base (Brian, 2026-10-17). In October the sun is low (about 40 degrees at noon), so standing corn about 8 ft tall would shade a spot within roughly 10 ft north of it for much of the day: the likely reason the grass read 8-10 F below the air an hour before sunset. The corn also blocks part of the southern sky and breaks the wind at grass height. HARVEST IS A BREAK IN THE SERIES (full sun, more open sky, more wind): log its date, and the crop each year (corn and soybeans usually alternate).
  • Daily summaries and records are pushed to the site as fc/hl/ground.json (hyperlocal/ground.py).
  • Probe history, which changes the readings: installed 2026-10-17 about 5 PM EDT with a bare stainless grass probe (low emissivity, so it likely reads a little WARM on clear nights). Brian painted it white on 2026-10-18 at 2:03 PM EDT (18:03 UTC). Expect clear-night grass readings a degree or two colder after that, and sunny-afternoon readings lower too (the bare probe soaked up sun). Treat 18:03 UTC as a break in the series: grass records and the grass-minus-air fits use before and after separately. The 14:02 and 14:07 EDT readings (69 F, against 57 F just before) are the probe being handled, not weather. At the same visit Brian found the probe had FALLEN and was lying on the ground; it is now held in the printed cradle mount 2 in above the grass. When it fell is unknown: the 5-minute readings show no clear step (the only candidate is 1:16 AM EDT on 2026-10-18, +1.3 F), and the first evening (grass 12 F under the air, then held at the dew point) looks like a probe in the grass. Today's daytime readings (5 to 8 F BELOW the sunny air) fit Brian's report: the old spot was in the mast's shade (and on the ground). The new spot is in the sun, so daytime grass readings will now run ABOVE the air (sun on the probe and the grass top, normal for a grass-minimum probe; only its NIGHT readings are the grass temperature that matters for frost). So: the bare-probe stretch (2026-10-17 5 PM to 2026-10-18 2 PM EDT) is "first mount, position uncertain"; the clean series starts with the mount, 18:03 UTC. Battery/solar move into a weatherproof box around the same time.
  • First evening (2026-10-17, clear and calm): the grass ran 12 F below the station's air by 7:10 PM, then its fall stalled at 35.5 F, within 0.1 F of the station's dew point (35.4 F): the dew-formation brake, seen directly. Ground truth: Brian touched the grass about 7:45 PM and found dew had formed (grass 34.6 F, station dew point 35.4 F). The stall marks dew onset.

Log

  • 2026-10-18: grass probe painted white at 2:03 PM EDT and put in the printed mount 2 in above the grass (it had fallen to the ground at some point); series break.
  • 2026-10-17: the detailed-run page (fc/detail.php) shows a night grass ESTIMATE from a starting rule: grass = air - 12 F x clear x exp(-wind/8 mph), held no lower than 1.5 F under the dew point (insight.py GRASS_*). Fit these on the ground sensor's nights; the rule is labelled experimental on the page.
  • 2026-10-17: our ground sensor went in (grass + soil 2/4/8 in at the mast); data notes above. First evening: grass 12 F below the air, stalled at the dew point.
  • 2026-10-17: Freeze Warning night, case study C-001 (outcome pending at 2 AM; station cooling pace pointed to a low above our 31 F forecast). Score it next session as the first fresh-month night.
  • 2026-10-16: thread opened by Claude with Brian. Reliability check above.
  • 2026-10-18 evening: first sunny-mount evening: grass 18 F under the air at sunset, levelling at the dew point (35.7-36.8 F) from 8 PM; Brian saw dew at 9 PM. Mounting intent clarified from the design (ground/mount/grass_mount.py): probe level, JUST ABOVE THE BLADE TIPS like a grass-minimum thermometer (not 2 in above the tips); Brian to check. Automatic calibration check added (ground.py calibration(): the probe within 2 F of the air in well-mixed air = an OVERCAST DAY (sun up, solar under 30% of clear sky, wind 3+ mph) or a WINDY NIGHT (8+ mph at the sheltered anemometer); a first version also counted light-breeze nights, which its first evening showed can carry a real 13 F gap, so they are excluded).
  • 2026-10-19: Brian checked the grass probe: placement correct, at the blade tips as designed. No series break; the clean series from 2026-10-18 18:03 UTC stands.
  • 2026-10-20: dew onset checked by hand (Brian): the slightest hint of dew, patchy, at about 6:45 PM; dew everywhere at 8:21 PM. The grass probe crossed below the air's dew point between 6:30 (48.4 F against 47.7 F) and 6:40 PM (45.6 F against 46.8 F), around sunset: the probe's dew onset matches the eye within about 10 minutes. The hourly grass estimate said 46.3 F at 8 PM (probe 40.6 F) and no dew until 9 PM: the forecast's evening grass runs about 6 F warm and dew about 2.5 hours late (L-031's target).
R-008Hour by hour: our own forecast for every hour of the next two days scorecard done (October 26, 2026); L-045 on trial, the station-shaped curve beside oursupdated October 26, 2026

The question

Our tested forecast works in 12-hour periods (the day's high, the night's low, rain chance per period), refreshed every hour. Everything finer on the pages is the raw model blend: the 3-hour strip's sky, wind and rain come straight from the models, and its temperatures follow a curve shaped to our high and low. Can we make our own hour-by-hour forecast for the first two days, corrected to the station the same way the highs and lows are? Brian asked for this on October 16, 2026.

Why it might work

  • The models have station-specific habits by hour, not only by day: how fast the evening cools, when the afternoon peaks, when the wind drops after sunset, how the dew point behaves at dawn. A daily correction can't fix timing.
  • The station's 1-minute history (2023 on) gives about 30,000 hourly truths to learn from.
  • The NWS does this operationally: its hourly guidance (LAMP, the Localized Aviation MOS Program, and the hourly National Blend) corrects models hour by hour and, for the first hours, leans on the latest observations. That last part is what our live track already does for the high and low.

Data

  • archive_<model>.parquet (/trainer/data): hourly model temperature, dew point, rain, cloud and wind at lead days 1-7, 2023 on. This is the training data: true lead times.
  • runs_<model>.parquet (/lab/data): real runs at exact lead hours, April 2026 on, for the honest recent test.
  • minutes.parquet: the station, every minute. Truth for every hour.

Plan (draft, to be pre-registered before testing)

  1. Targets: hourly temperature, dew point and wind speed for lead hours 1-48.
  2. Method: for each lead hour, correct the model blend with a ridge regression on the models' own values, the hour of day, season, and our daily high/low forecast (so the hourly path agrees with the daily numbers). For the first few hours, blend in the latest station reading (persistence), as LAMP does.
  3. Compare with what the page shows today (the curve plus the raw blend) on identical hours, walk-forward.
  4. Measures: hourly error by lead hour; the timing of the day's peak and the evening drop; wind speed error; how often the hourly path stays inside its range.
  5. Draft bar: at least 10% lower hourly temperature error than today's strip over lead hours 1-24, with the daily high and low no worse.

What has been explored

2026-10-16, discovery issues (March 2024 to September 2025, 415 issues scored, walk-forward). Mean hourly temperature error (F):

Hours aheadRaw model blendToday's stripHourly correction
1-62.062.021.64
7-241.841.951.76
1-241.891.971.73
25-482.212.192.12

Hour 1: 2.17 (blend), 1.82 (strip), 0.86 (correction): knowing how far off the blend is at 6 PM matters most early. Side finding: the strip is slightly worse than the raw blend away from the peak and the low, because it shifts every hour of a window by the high/low correction, which mostly belongs near the extreme. Formal test pre-registered as L-003 before any test-year run. Script: hyperlocal/research/r008_hourly.py.

2026-10-16, L-003 on the test year (358 issues). Hours 1-24: strip 1.946 F, correction 1.760 F, 9.57% better (bar 10%: fail). Hours 25-48: 2.326 vs 2.196 (pass). Better in 13 of 13 months (pass). 90% interval of the gain 7.7% to 11.4%. Hours 1-6 21% better, hours 7-24 5% better. Not passed; see the notebook entry for next steps.

The scorecard, October 26, 2026: where the page's curve misses the station

Brian: the highs and lows are close now, but the hour-by-hour curve does not fit what the station then does, above all the plunge after sunset, the overnight stirs and the morning rise. Before changing anything, the curve was measured three ways (hyperlocal/research/r008_scorecard.py; results in r008_scorecard_live.json, _replay.json, _seasons.json and _shape.json). Truth everywhere is the station's readings within 5 minutes of each hour. Errors are the mean absolute error in F with the bias in brackets. Regimes are from the models' own inputs at issue, never from what happened (L-038's definitions: a night is clear and calm under 30 percent cloud and 5 mph, cloudy at 70 percent or more, otherwise "other"; days on cloud alone).

  1. The live record: every hourly build the Tower made from October 14 to 26, 2026 (401 builds; the Tower's database holds nothing older), the page's curve rebuilt with the page's own code (insight.build_detail) on the snapshots, issued values, readings and airport ring as they stood at each build, so it is the curve the page showed.
  2. The seasons: the two forecasts of record a day (11Z and 23Z) rebuilt from the model archives at lead 1 and our walk-forward daily forecasts, March 2024 to October 2026 (1,835 issues), so winter is in the picture. The archive's runs are at least a day old at each hour, so this strip is a little worse than the live one; the shape of its errors is what matters here.
  3. The replay: the main model's replay on the real forecast runs, April 2 to October 13, 2026 (373 record issues), with the strip built the Tower's way from the same runs, for the sizing of the physics-shaped candidate below.

By hour after the issue. The live curve is 0.87 F off at hour 1, 1.5 at hour 2, 2.0 at hour 3, and 2.1 to 2.3 F from hour 4 to 12 (bias +0.3 to +1.3 F, warm). The seasons strip: 1.07, 1.41, 1.64, then 1.8 to 1.9 F. The first three hours owe most of their accuracy to the L-013 nowcast: without it, hour 1 is 2.3 to 2.8 F off.

Where the misses sit (the seasons record, by regime and hours since the sun event).

regime, hours since sunset or sunriseerror (bias)n hours
clear calm night, Sep to Oct, 1 to 2 h after sunset4.59 (+4.51)39
clear calm night, Sep to Oct, 2 to 3 h4.35 (+4.07)39
clear calm night, Sep to Oct, 3 to 5 h3.17 (+2.65)78
clear calm night, Jun to Aug, 1 to 3 h2.08 (+1.65)112
clear calm night, all seasons, 8 h or more1.68 (+0.35)378
cloudy night, Nov to Feb, 8 h or more1.81 (-1.03)544
other night, Nov to Feb, 5 to 8 h2.31 (-1.21)333
other night, Nov to Feb, 8 h or more2.64 (-1.65)530
cloudy day, Mar to May, 5 to 8 h after sunrise2.64 (+1.46)309
cloudy day, all seasons, 3 to 8 h2.1 (+0.7)1,650
clear day, all seasons, 0 to 1 h after sunrise1.06 (+0.51)152

The live October record says the same with fewer nights: clear calm nights 2.0 F too warm 2 to 3 hours after sunset and 1.6 F too warm at 3 to 5 hours, then 1.3 F too cold at 5 to 8 hours (the stirs: the station warms when the air mixes near midnight and the curve keeps falling); by local hour the curve runs 1.6 F cold at 1 AM, 1.9 to 2.4 F warm from 9 to 11 PM, and 2.1 to 3.1 F warm at 8 and 9 AM (the hour after sunrise, 3.7 F too warm on clear mornings: the models warm the morning faster than the station does). So the misses Brian sees are real and sit at fixed places in the day: the plunge after sunset on clear calm nights in fall (the largest, 4 to 4.5 F), the late-night cold bias of cloudy and windy winter nights, the hour after sunrise, and cloudy afternoons.

The physics-shaped curve, sized and rejected (R-011 rule 8). The change asked for first: a curve whose shape and timing come from the main model's run (column.py, restarted from the station's latest reading at the issue as column_live does) and whose endpoints are pinned to our corrected high and low, so the numbers the site shows cannot change and only the path between them does (the shift is zero at the issue, equals our value minus the physics' at each window's extreme, and runs in straight lines between). On the replay, over hours 1 to 12 after the forecasts of record, it is 0.58 F per hour farther from the station than the current curve (95 percent interval 0.47 to 0.68 F, 186 dates), and farther in every regime: clear calm nights +1.37, other nights +0.98, cloudy nights +0.55, clear days +0.51, cloudy days +0.47, other days +0.13 F per hour. The reason is the physics' own shape, not the pin: on clear calm nights the night physics is 5.0 F too warm 1 to 2 hours after sunset, 4.8 F at 2 to 3 hours and 3.6 F at 3 to 5 hours (the dusk lag known from R-010, here on real forecast runs and with the restart), and the day physics warms too slowly through the morning (0.4 to 0.6 F cold at 1 to 5 hours after sunrise on clear days in the replay, 2.4 to 3.0 F in the live October record). The Tower's own stored record curves agree: 4 to 7 F warm for the first four hours after the 23:05 builds of October 22, 23 and 25, and 5 F cold at hours 3 to 6 after the 11:05 builds of October 23, 25 and 26. On the live October record the pinned physics is 1.44 F per hour farther (9 dates). A variant that keeps the page's nowcast for the first six hours instead of the physics' restart is worse still (+1.17 F per hour on the replay). The physics-shaped curve was therefore not frozen: it cannot pass any bar until the dusk lag is fixed in the physics itself (R-010: the grass's evening cooling, L-031, and the soil column on the Studio), and this sizing is the measure of what that fix must deliver.

The station-shaped curve (frozen as L-045). The alternative uses the station's own record (rule 1): a table of the curve's mean residual by the hour's regime (the models' mean cloud and wind over the half, sunrise to sunset or sunset to sunrise) and the whole hours since the half's sun event, 87 cells, each shrunk toward zero by n over n plus 30 hours. The correction ramps to zero over the two hours before the curve's own extreme inside the half's scoreboard window and is zero from then on, and inside the window the curve is clamped so it never goes below the night's low or above the day's high. Fitted on March 2024 to September 2025 and sized on October 2025 to October 2026 (the test year, 360 dates, the two records a day, hours 1 to 12):

current curvestation-shapedgain per hour (95 percent interval)
hours 1 to 121.76 F1.73 F0.031 F closer (0.015 to 0.047)
hours 1 to 31.421.390.030 closer (0.010 to 0.050)
hours 4 to 61.851.820.027 closer (0.058 closer to 0.004 farther)
hours 7 to 121.891.860.034 closer (0.015 to 0.053)
clear calm nights (46)1.841.690.114 closer (0.265 closer to 0.034 farther)
cloudy nights (157)1.601.560.033 closer (0.009 to 0.057)
other nights (215)1.921.890.026 closer (0.000 to 0.051)
days (all three regimes)0.010 to 0.020 closer, intervals across zero

On clear calm nights the bias 1 to 2 hours after sunset falls from +2.0 to +0.4 F, at 2 to 3 hours from +1.8 to +0.2 F and at 3 to 5 hours from +1.2 to -0.1 F; late on cloudy nights from -0.6 to -0.4 F; through cloudy afternoons from +0.5 to +0.1 F. The window extremes do not move (by construction). A small gain, but resolved on 360 dates, in every lead group, and in the places Brian sees the curve miss. Power (rule 2): 0.031 F per hour against a per-date spread of 0.159 F needs about 207 records for 80 percent power; November to April provides about 360. The clear-calm-night gain alone would need about 160 such nights, more than a winter holds, so it is reported and not part of the verdict. Frozen as L-045 with the table refit on the whole record to October 11, 2026 (hourly_shape_frozen.json); on the Tower from the install of October 26, 2026, shown beside our curve on the Hour by hour page and logged at every build (hourly_issued), graded by l045_score.py on the forecasts of record through April 30, 2027.

What this does not fix: the physics' dusk lag (above), the hour after sunrise beyond what the table carries (the live October bias of 3.7 F there comes from 76 hours and the table's cell from the whole record says +0.5 F), and the stirs themselves, which no mean curve can place on a given night; those need the measured wind and the airports (R-002, R-006).

Ideas to chase

  1. The evening drop: does a station-corrected cooling curve catch calm-night decoupling earlier (with R-002/R-003)?
  2. Hourly rain timing from the ensemble, calibrated to the gauge (with R-005).
  3. If it works, the briefing's 24-hour strip and Today's Track would both use it.
  4. When the high and low happen (added 2026-10-17, Brian). Our daily forecasts give how warm or cold, never when; the hourly curves borrow the models' timing unchecked. Score the models' time of the day's high and low against the station (1-minute history), by season and cloud, and on the odd days: (a) a strong cold front late morning or early afternoon (the high comes early, then the temperature falls like a rock); (b) the calendar-day high at 4 AM ahead of a front or with warm air overnight; (c) warm-air nights where the low comes in the evening and the temperature rises toward dawn. Fronts: WPC passages (data/fronts/passages.csv) and the station's own frontal signature (fronts.py / frontlib.py). Questions: are the models early or late on the front's arrival here, by how many hours, and does a timing fix make the hourly track (and the calendar-day high vs the 8 AM-7 PM high) more accurate? First look below (station climatology only, no forecasts scored).

2026-10-17, timing first look (station only, Feb 2023 to Oct 2026, 1,271 days). Script: hyperlocal/research/r008_timing_count.py, results r008_timing_count.json. Calendar-day high: 1-5 PM on 54% of days, 5-8 PM 28%, but midnight-6 AM on 10% (6-11 AM only 1.5%). Off-hours highs (before 11 AM or after 8 PM) run 35-38% of days in December and January, 22% in November and February, under 4% June to September. On the 128 days whose high came before 9 AM, the calendar high averaged 4.8 F above the 8 AM-7 PM high. Lows: 4-9 AM on 59% of days, but 9 PM-midnight on 22% (the low at the end of the day: cold air arriving), 46% of December days away from dawn. Daytime cold fronts (WPC passage 9 AM-6 PM, 60 days): the high came before 1 PM on 25%; 3-hour drop median only 2.3 F (many fronts are weak or mistimed by the 3-hourly analyses), but 7 dropped 10 F or more, up to 21.7 F (2026-03-11). Takeaway: off-hours timing is a cold-season matter, and big front drops are rare but real; the scoring of model timing is the next step.

Log

  • 2026-10-16: thread opened at Brian's request.
  • 2026-10-16: discovery exploration (above); L-003 pre-registered and published.
  • 2026-10-16: L-003 result: 9.57% vs a 10% bar. Not passed.
  • 2026-10-17: idea 4 (timing of the high and low, front days) added at Brian's request; station-only first look.
  • 2026-10-26: the scorecard (live builds, the seasons, the replay); the physics-shaped curve sized and rejected before any freeze; the station-shaped curve sized and frozen as L-045.
R-009WxSim: how good has it been, and could it help our forecast? L-006 pre-registered on fresh months (October 2026 to March 2027)updated October 17, 2026

The question

Brian runs WxSim, a local physics model of his site. How accurate have its forecasts been against the station, and could it add something our forecast lacks?

Data

  • WxSim forecasts: 5,651 runs, January 2018 to March 2023 (gaps mid-2019 to mid-2020, late 2020, April 2021 to January 2022), six runs a day, about 7 days ahead in half-hour steps. WorkingDrive "4. Archives/WxSim/wxsim/f*.txt".
  • Station truth for those years: Weather Display's NOAA monthly reports, February 2018 to June 2020 (calendar-day high and low). WorkingDrive "4. Archives/Meteorology". No station record for 2021-2022 has turned up.
  • Recent WxSim: the forecast-vs-actual database on Hostinger, hourly since June 4, 2026 (needs an export).

What has been explored

2026-10-17, WxSim graded against the station, 424 days (February 2018 to June 2020). The 6 PM run each day (10 PM if missing); calendar-day highs and lows. Benchmarks on the same days: persistence (tomorrow like today) and the NWS GFS guidance for Marion (MOS, the run out around 6 PM). Mean error (F), with average bias (+ = too warm):

WxSimNWS guidance (MOS)Persistence
Highs, day 14.46 (-1.1)3.29 (+0.2)6.13
Highs, day 24.89 (-1.4)3.73 (+0.8)8.55
Highs, day 35.40 (-1.7)not available9.51
Lows, day 14.00 (+1.0)3.00 (+0.2)5.41
Lows, day 23.58 (+1.0)3.12 (+0.6)7.42
Lows, day 33.75 (+1.1)3.54 (+0.5)8.36
  • WxSim beat persistence easily but trailed the NWS guidance by about 1 F on day-1 highs and lows, less for lows at days 2-3.
  • Within 3 F of the actual: about 47% of day-1 highs and lows. Misses of 8 F or more: 61 day-1 highs (14%) and 46 day-1 lows (11%).
  • Season pattern, day 1: in summer WxSim ran 3.2 F too cool by day and 3.8 F too warm at night: its day-night swing was about 7 F too small. In winter its lows ran 1.5 F too cold. A damped summer swing is what a model does when it treats the ground as too moist or the air too humid or cloudy, so it is worth checking WxSim's soil moisture and ground settings for summer.
  • Day-1 lows (4.00) were worse than day-2 lows (3.58), which is unusual. Unexplained; one possibility is how the 6 PM run handles its first night.
  • Caveats: 2018-2020 settings and equipment may differ from today; WxSim keeps standard time year round while Weather Display uses the PC clock (one hour at midnight in summer); MOS's day and night windows differ slightly from calendar days. For rough context only (different years and day definitions): our forecast today misses by about 1.6 F (lows) and 1.8 F (highs) on day 1.

Script: hyperlocal/research/wxsim_grade.py, results wxsim_grade.json.

2026-10-17, WxSim's saved forecasts decoded. The current WxSim (Weather PC) keeps every run as f<YYMMDDHH>.wxf: 4,010 readable runs, June 2023 to now. Format: each half-hour step starts with a line "Mon DD",hours (standard time), then the air temperature in Kelvin, converted as (K - 273.0) x 1.8 + 32 (matches latest.csv exactly). Script: hyperlocal/research/wxsim_extract.py.

2026-10-17, WxSim as an input, discovery years (March 2024 to September 2025). WxSim value = newest run finished by our issue time, max/min over our windows. Mean error (F):

OursWxSim aloneError correlationBest WxSim share (information only)
Highs, day 1 (289)1.8233.1790.3911%
Highs, day 2 (272)2.3843.7060.4218%
Lows, day 1 (276)1.6493.4360.367%
Lows, day 2 (258)2.0983.7490.398%

WxSim improved since 2018-2020 (highs 4.5 -> 3.2, lows 4.0 -> 3.4) and its errors only partly overlap ours. Leave- one-month-out, WxSim's information alone moved our error by: highs day 1 +0.009 (worse), day 2 -0.021; lows day 1 -0.009, day 2 -0.012, day 3 -0.007. A fixed bias term looked helpful for day 2-3 highs but that was our own forecast's bias on those days, not WxSim's information, so it was left out. The test year had WxSim on only about 70 days (it was down most of the year), too few to detect 0.01-0.02 F, so the formal test (L-006) uses fresh months. Script: hyperlocal/research/wxsim_input_test.py.

Ideas to chase

  1. WxSim as an input: L-006 (fresh months). WxSim's own WXSIM-Lite blend has not been fed since August 21, 2026 (fdata.txt stale; its loop is being fixed); if Brian turns the blend on, WxSim changes, so log the date.
  2. Grade the live WxSim (June 2026 on) the same way, to see whether today's setup does better than 2018-2020.
  3. If Brian adjusts WxSim's summer ground settings, the live grading would show the effect.

Log

  • 2026-10-17: thread opened at Brian's request; 2018-2020 grading above.
  • 2026-10-17: saved forecasts decoded; discovery exploration; L-006 pre-registered on fresh months.
R-010Our own physics: a night-cooling model built for this station night model live (L-007 passed); day model live for display (L-009 failed on forecast clouds; physics itself holds up)updated November 18, 2026

The question

Our forecast is statistical: it corrects the global models with years of the station's record. Its biggest misses are nights that go calm and cool on their own (R-002), linked to the inversion near the ground (R-003), which the global models are known to mix too much (Sandu et al. 2013). That is a physical process with well-understood ingredients. Can a small physics model of the night at this one spot, calibrated on the station's own minute data, predict those nights better than statistics alone, and feed our forecast as one more input? Brian asked for this on October 17, 2026: build our own physics of the station's air and ground, piece by piece, each piece earning its place against the station.

Physics (version 1: sunset to sunrise)

  • The ground radiates to the sky. Outgoing longwave from the surface minus what the sky sends back. Clear-sky sky emission from air temperature and humidity (Brunt, Swinbank, Brutsaert-type formulas); cloud adds back radiation depending on cloud height and cover.
  • Wind mixes warmer air down. Turbulent heat exchange between the ground and the air, scaled by wind speed and cut off as the air near the ground becomes stable. Below a minimum wind the layer decouples and the ground cools on its own (van de Wiel et al. 2012, minimum wind speed for sustainable turbulence).
  • The ground gives back stored heat. A simple two-layer soil (the force-restore method, Deardorff 1978) started from the soil temperature: warm wet autumn soil slows cooling.
  • Moisture puts a floor under the temperature. Dew and frost forming release heat near the dew point (the condensation floor, NWS TM 102).
  • The air aloft caps it. The 925 mb temperature and wind bound how far mixing can reach.

Version 1 design

  • Start: the station's own 6 PM readings (temperature, dew point, wind). Soil: ERA5 for history; the models' soil temperature live.
  • Drive each hour overnight with the forecast at issue time: cloud (low, mid, high), wind at 10 m and 925 mb, humidity, 925 mb temperature (from the real runs since April 2026; near-analysis upper air for learning).
  • A handful of free parameters (exchange strength, soil heat capacity, cloud back-radiation, decoupling wind), calibrated on discovery nights (March 2024 to September 2025) against the station's minute-by-minute cooling curves.
  • Output: a physics low and a physics cooling curve for tonight. Use: (1) "physics low minus our low" as an input to the statistical model, (2) a decoupling indicator for the +/- range (R-001), (3) the evening path for the hourly forecast (R-008).
  • Test like every other input: pre-registered plan and bar, the test year, then a winter confirmation on fresh months.

What has been explored

2026-10-18, version 1 built (hyperlocal/physmodel/: scm.py the model, nights.py inputs, calibrate.py, crossval.py). Brian asked to start building our own physics, to run alongside the statistical model, never replacing it. Version 1 is the night model designed above: a grass/soil skin radiating to the sky (Brutsaert clear sky, cloud by layer), turbulent exchange with the 2 m air (cut back by stability), dew and frost latent heat, force-restore soil, mixing toward the 925 mb air brought down dry-adiabatically, plus (added after the first fit) a share BETA of the air-mass change aloft (advection). 13 constants, fitted by optimisation on 474 dry DISCOVERY nights (March 2024 to September 2025), sunset to sunrise in 5-minute steps, against the station's minute record. Inputs: the models' stitched first hours (upper_hist, 4-model mean) and ERA5 soil: perfect prog, so learning only.

Results (low of the night, F, mean absolute error and bias):

  • In-sample fit: physics 2.01 (-0.01); clear and calm nights 1.94 (-0.18). The models' own near-analysis 2 m temperature: 1.46 (+0.76); clear and calm 1.93 (+1.61). Adding BETA (advection) cut the curve error from 2.44 to 2.18 F and removed biases of about +/-1.2 F on nights when the air aloft cooled or warmed.
  • Cross-validated (6 blocks of months, constants refit without each block): physics alone 2.13; a blend with the models 1.40 vs 1.46; physics only on clear, calm nights 1.33 overall, 1.56 vs 1.93 on clear, calm nights.
  • BUT the fair rival is a plain statistical correction of the models on clear, calm nights (what our forecast already does with its clear-and-calm feature): 1.51. Adding physics to that rival made it slightly WORSE (-0.09 F, 95% interval -0.13 to -0.06). So version 1's apparent gain is only the calm-night warm bias, which statistics already removes. Version 1 adds no new information. Results: physmodel/crossval_v1.json, constants physmodel/params_v1.json.

Reading the fitted constants: the optimiser chose a deep cooling layer (Z_EFF about 160 m), strong coupling and almost no stability cut-off, i.e. it gave up on the shallow, decoupled 2 m layer and modelled the night as one bulk layer. That is the physics that matters most on our worst nights (R-002), so version 1 cannot capture it by construction. And the comparison is harsh: the baseline here is the models' near-analysis, which has already seen the night.

2026-10-18, version 2 (two air layers: a shallow 2 m layer that can decouple under a bulk night layer; physmodel/scm2.py, calibrate2.py). Fitted freely it reached 1.98 F in-sample (version 1: 2.01) by abandoning the shallow layer again (a 180 m "surface layer", a wind floor that erases calm). Held to physical ranges it was worse (2.15 F) and pushed against every limit. Driven by the STATION's measured wind instead of the models' it improved only to 1.92 F and still would not decouple. So the second layer is not where version 1 loses; shelved for now.

2026-10-18, the fair comparison and L-007. Against OUR day-ahead low (walk-forward, discovery nights), adding the cross-validated version 1 physics low (started at sunset) cut the error 1.63 to 1.56 F, while the same sunset readings used statistically added nothing (1.66). Pre-registered as L-007 and run on real forecast runs (ECMWF and ICON 12Z, April to October 2026, 149 dry nights): ours 1.52 F, ours + 0.25 x (physics - ours) 1.38 F, control with the fresh runs' own 2 m low 1.54 F, gain positive in 99.9% of resamples: PASS. See the lab notebook, L-007.

2026-10-18, the DAY model (physmodel/days.py, dayscm.py, calibrate_day.py; live in physlive.py). Brian asked for daytime physics with the same rules (published, runs on the Tower). Sunrise to sunset: clear-sky sunshine from the sun's height dimmed by low, mid and high cloud; net longwave; a share into the soil; an evaporative fraction from topsoil moisture and the growing season; the rest heats the air, which first burns off the night's inversion and then builds a mixed layer by encroachment (Carson 1973) through the morning's 925/850/700 mb profile; the 2 m air follows the mixed layer plus a hot layer at the ground on sunny, calm days; advection from the 850 mb change. Fitted on 455 dry DISCOVERY days. With all constants free the fit went unphysical (a negative albedo), so albedo 0.20, soil share 0.15 and entrainment 0.2 are fixed and the rest bounded. Version 1.1 added snow albedo (from 5 cm of snow), an uneven growing season (crops dry down in the fall), a calm-day hot layer and a morning inversion depth; that took the in-sample error on the high from 2.22 to 2.03 F (sunny days 1.75 F, better than the models' near-analysis 1.87 F; cloudy days 2.43 F are its weak spot). Fitted constants: low cloud blocks 79% of the sun, mid 62%, high 19%; evaporative fraction 0.49 + 0.52 x (moisture - 0.30) + 0.36 x (season - 0.5); inversion 334 m. Cross-validated against OUR walk-forward high: ours 1.68 F, ours + sunrise readings (statistical) 1.67, + day physics 1.63: gain 0.05 F, 95% interval -0.02 to 0.11. Not proven, so no real-run test yet and nothing blended; the day run is shown on the Physics page for interest. Driving it with the models' own sunshine instead of our cloud dimming was worse (stuck at a +2.8 F bias). Its highs come about 40 minutes late on average (mixing keeps going until the sun is low).

2026-10-18, the station's solar sensor (physmodel/solar.py, cloud_skill.py). The Davis pyranometer, logged by Weather Display every minute since January 2023, as 5-minute sunshine (370,076 values to October 14, 2026; data/research/solar5.parquet). Calibration is steady: the clearest middays read 0.98 to 1.10 of the clear-sky formula in summer and about 1.10 to 1.15 in winter (the formula runs low at low sun), with no drift year to year. (Earlier uses: tested as a statistical temperature input on October 14, no gain; shipped as front evidence, the sky clearing behind fronts, on October 15.)

  • Cloud forecasts against the sensor, next-day daytime clearness (8 AM to 5 PM), 12Z runs, April 3 to October 13, 2026, 183 days (error on a 0 to 1 scale / correlation): ECMWF + ICON own radiation averaged 0.097 / 0.86; ICON 0.101 / 0.84; ECMWF 0.116 / 0.82; our cloud-layer dimming of their clouds 0.136 / 0.81 (ICON's clouds through our dimming: -0.19 bias); yesterday's clearness 0.226 / 0.18; the month's average 0.198 / 0.05. The models' radiation schemes know cloud thickness; three cloud layers do not. For sunshine, use the models' radiation.
  • CORRECTION (same evening): the day fits with a switch on (measured sunshine, the models' sunshine) were wrong. A bug in calibrate_day.py scored every candidate with the DEFAULT model during fitting, so those constants were tuned for cloud dimming and only the report used the switch. The night model (no switches; L-007) and the cloud-forecast scores above (no fitting) are unaffected. Rerun correctly (cross-validated high, gain over ours + sunrise info): measured sunshine 2.24 F, gain 0.05 F (95% -0.04 to 0.13); the models' sunshine 2.42 F, gain 0.00. So with the old advection, even perfect sunshine barely helped: the bottleneck was advection after all.

2026-10-18, day ADVECTION version 2. The models' 925 mb, 850 mb and 2 m temperatures change through the day for two reasons: air moving in (advection, what we want) and the models' own daily heating (which our physics does itself). Version 2 removes the typical daily cycle for the month and UTC hour (from the stitched archive, discovery years only) and adds the remaining "weather" change since sunrise, weighted per level (fitted: 925 mb 0.71, 850 mb 0.02, 2 m 0.16; the low-level flow about 2,500 ft up carries it). Cross-validated high, discovery days:

Day modelPhysics aloneGain over ours + sunrise info (95%)
Advection v1, our cloud dimming2.32 F0.05 F (-0.02 to 0.11)
Advection v2, our cloud dimming1.68 F0.14 F (0.03 to 0.25)
Advection v2, measured sunshine (perfect clouds)1.51 F0.20 F (0.08 to 0.32)

Our walk-forward high is 1.68 F on the same days. Rule on discovery (v2, cloud dimming): high = ours + 0.408 + 0.498 x (physics - ours). Perfect clouds would add more on top: cloud forecasting pays once advection is right. Pre-registered as L-009 (real 00Z runs, April to October 2026). Files: params_day_v1_adv2.json, rerun_*.txt.

2026-10-18, L-009 FAILED (day physics on real 00Z runs, 135 dry days). Ours 1.46 F, rule 1.55 F, fresh-run control 1.35 F, physics alone 2.08 F (bias -1.23). Diagnostic on the same days and runs with the MEASURED sunshine: physics alone 1.48 F, level with ours. The physics and advection hold up; the cloud-to-sunshine step failed (right on average, 96% of measured, wrong on individual cloudy days). Next: the models' own radiation as the sunshine input, explored on discovery, tested on fresh days only.

2026-10-18, the day model's sunshine rebuilt on the models' radiation (step 1 after L-009), discovery only. Sensor calibration on discovery daytimes: measured / models' sunshine 0.907. (a) The physics fitted on MEASURED sunshine, then driven by 0.907 x the models' sunshine, cross-validated (blocks refit): physics alone 1.63 F (ours 1.68), gain over ours + sunrise information 0.10 F (95% -0.01 to 0.21); rule ours + 0.30 + 0.54 x (physics - ours). (b) Everything refit with the models' sunshine: 1.73 F, gain 0.08 F (-0.02 to 0.19), and the fit ran to its limits (it wanted 1.3 x the models' sunshine while the sensor says 0.9 x: constants trading off, not physics). (a) is anchored to the sensor at both ends and is the version frozen for L-010 (fresh days from October 19) and put live. L-008 (the night model live, against our issued low at sunset) was pre-registered at the same time.

2026-10-18, PRECIPITATION TYPE (physmodel/ptype.py, ptype_eval.py, fetch_levels.py, l011.py). The energy method (Bourgouin 2000) on the profile above the station, verified against what 8 nearby airports reported (Marion, Kokomo, Grissom, Muncie, Huntington, Fort Wayne, Logansport, Anderson; freezing rain only where sensors can report it: Fort Wayne, Grissom, Muncie). With only 925/850/700 mb: rain vs snow worse than a 33 F surface rule (marginal hours 70.8% vs 80.9% at Marion); freezing rain caught 32% with 83% of ice calls right. Added 10 levels (1000 to 600 mb, GFS and ICON stitched first hours, cold seasons 2023 to 2026, data/research/levels_*.parquet): freezing rain caught 41% (temperature) to 50% (wet-bulb profile), calls about 80% right; rain vs snow still worse than the rules (too snowy just above freezing: rain hours it called snow had the station at 35.5 F). Hybrid (physics decides ice, station wet-bulb 32.5 F decides rain or snow) on the two exploration winters: 95.6% vs 95.1%. L-011 on the unseen winter 2025-26: hybrid = wet-bulb rule (97.0%, marginal 76.2%), both beating the 33 F rule (marginal 71.3%); ice: 18 reported hours only, 1 caught, 16 of 18 ice calls during snow. Lesson: the wet-bulb rule is the solid gain; our ice detection is not shown to work.

2026-10-18, FOG, first look (discovery nights; physmodel/fog_nights.csv). Truth: visibility under 1 mile at Marion or Huntington during the night (about 1 night in 15 at any one airport; the two agree on only 25 of their 82 fog nights: fog is very local), and the station's own saturation: its Davis humidity sensor tops out at 98% (night maxima, 99th percentile), so 97% or more for 3 hours stands for "saturated". Station 97%+ for 5 hours or more came with regional fog on 50% of nights (base rate 9%). Ranking skill (AUC) on 474 dry discovery nights, regional fog / station saturation: night model's closest air-to-dew-point gap 0.63 / 0.67; its saturated calm hours 0.60 / 0.72; crossover (physics low minus the sunset dew point) 0.66 / 0.70; the models' near-analysis 2 m humidity through the night 0.90 / 0.90 (flattered, but the gap is far beyond that). The night model was fitted for temperature; its moisture (dew deposition plus the models' drift) is too crude for fog, which needs temperature and dew point right to within a degree at the same hour. A fog-capable night model needs its own moisture budget (deposition, evaporation from wet ground, mixing of drier air from above), calibrated on the station's humidity and the grass.

2026-10-18, night version 3: a MOISTURE BUDGET (physmodel/scm3.py, calibrate3.py; live physlive.integrate_v3). Version 1's temperature physics (frozen) plus water: dew deposits on the grass (drying the air), dew and then a share of the soil evaporate, fog forms beyond saturation (latent heat; fog sends longwave back like low cloud). Version 3a mixed the 2 m air's moisture toward the 925 mb air, which is about 9 F drier in dew point at sunset: the dew point collapsed (7.1 F error) and the lows went too cold (4.5 F). Version 3b keeps the models' own 2 m moisture and lets our physics add only the LOCAL departure (dew, evaporation) that fades back over TAU_Q hours. Fitted (discovery dry nights, temperature constants frozen): moisture layer 86 m, TAU_Q 3.1 h, soil evaporation almost nil at night, fog longwave 0.49 of low cloud. Cross-validated: dew point curve error 1.67 F (version 1 1.77), low 2.03 F (2.01); fog ranking skill, regional / station saturation: version 1 0.63 / 0.67, version 3 0.67 / 0.73, the models' near-analysis humidity 0.90 / 0.90 (flattered). Our ~2 F night temperature error is too large for fog, which needs air and dew point within a degree. The fair test against the models' FORECAST humidity at sunset is pre-registered as L-012 on fresh fall-to-spring nights; the Tower logs both signals from October 18.

2026-10-18, the 24-HOUR CYCLE (physlive.chain). Day and night joined: from the most recent sunrise or sunset (the station's reading then) each half runs to its end and the next starts where it finished, day (the L-010 version) into night (version 1) and on, four halves (about two days), with ECMWF + ICON clouds, wind and air mass. Shown on the Physics page against our forecast and the station; never blended. Each start's predicted extremes are logged (phys_runs kind 'chain') so a later test can judge the physics' own next-day high and low; first live chain (from the October 18 sunrise): today 62.4 F (ours 63.2), tonight 33.4 (34.5), Monday 65.8 (67.8), Monday night 36.1 (38.6): cooler than ours, more so further out.

2026-10-18, FAIR-WEATHER CUMULUS (physmodel/cumulus.py, ring.py). Cumulus forms when the mixed layer's top reaches the condensation level of the mixed-layer air (125 m per deg C of temperature minus dew point; mixed-layer moisture from the models' 2 m, 925 and 850 mb humidity averaged over the depth mixed). Truth: days with a clear morning and no weather, cumulus (FEW/SCT/BKN, base 1,500 to 9,000 ft) 10 AM to 5 PM. First against Marion alone (242 such days, cumulus on 41%, onset mostly 10 AM to noon, first base median 4,600 ft): the L-010 day model reached the condensation level on only 2 of 160 exploration days (ranking skill 0.76). Its afternoon mixed layer is too shallow: median 3,370 to 3,500 ft vs ERA5's boundary layer 4,810 ft (day-to-day correlation 0.73); fitted to temperature only, it got its warmth partly from the hot layer at the ground instead of depth. Refit with ERA5's depth in the loss (params_day_v1_obs_adv2_blh.json): 3,960 ft, temperature skill unchanged (high 1.51 F cross-validated, gain over ours 0.21 F). THE RING (Brian's idea): the six airports within 35 miles (Marion 14.6 mi SE, Kokomo 15.5 SW, Grissom 17.0 W, Huntington 23.2 NE, Logansport 28.6 W, Rochester 33.0 NNW; Wabash has no automated reports), weighted by 1/distance, as a consensus for the station (ring.py; Rochester and Warsaw were backfilled into the airport archive to close the ring to the north). 255 clear-morning days, cumulus on 35%. Ranking skill: physics 0.84 (Marion-only truth: 0.76), the models' near-analysis low cloud 0.92 (flattered). As yes/no it needs a 300 to 600 m allowance (the depth bias): 300 m: 81% of calls right, 38% caught; 600 m: 58% right, 88% caught. Onset off by 2.2 to 2.6 h; base off by 1,500 to 1,900 ft, worse than the forecaster's 227 ft per F rule. Conclusion: the physics knows which days lean to cumulus, not when or how high; the mixed layer is too crude (too shallow, three upper levels).

A pattern across the session: our physics adds real information where the models carry systematic LOCAL biases in temperature (the night low, L-007; the day high with measured or model sunshine, L-010 pending); for clouds, fog and precipitation type the models' own fields are stronger, and the useful physics there is simple and local (the wet-bulb rule; the station's own humidity ceiling for fog).

Road map toward Brian's goal (anticipating clouds, rain, snow and storms ourselves, which drive the temperatures):

  1. Afternoon cumulus: when the mixed layer reaches its condensation level (from the dew point), clouds form; that would let the day model make its own fair-weather clouds and cap its own heating.
  2. Storms: lift a parcel from the mixed layer through the models' upper air (instability, a lid) for our own thunderstorm potential and timing.
  3. Rain and evaporative cooling, then precipitation type from the column (freezing level, a warm layer aloft).
  4. Fog and dew at night from the night model's dew point and the grass.
  5. Join night and day into one continuous 24-hour cycle, each half starting from where the other ended.

Do the small effects add up? (2026-10-18, L-014)

Brian's question: many tested effects were each worth about 0.02 F and set aside; together, do they become something big? Every small candidate (the other forecast program, the inversion at issue time, soil moisture, the ring's error and the upwind airports' error at issue time) was put next to our forecast in one penalised regression, with and without the physics, cross-validated by 6 blocks of months on discovery days (research/combined.py).

Lows (225 nights), mean absolute errorFHighs (232 days)F
Ours1.71Ours1.70
+ physics1.61+ physics1.47
+ the five small effects, no physics1.65+ the five small effects, no physics1.76
+ physics + inversion1.59+ physics + inversion1.49
+ all six1.59+ all six1.53 to 1.56

They do not add up. Two reasons: they overlap (the other program and the inversion carry part of what the physics already knows, correlation 0.2 to 0.3), and the ring and upwind errors fade within hours, so at the next day's low or high they are noise (they belong to the 1 to 6 hour nowcast, where they PASSED in L-013). The only small piece left on top of the physics was the 6 PM inversion for lows. L-014 tested it on the 149 real-run nights: 1.38 -> 1.37 F, gain 0.014 F, positive in 72% of resamples: FAIL, the size discovery predicted. The physics is the one large piece.

Wet weather (2026-10-18, L-015 and L-016)

Version 1 of the night physics ran only on dry nights. On wet nights it ran far too warm: +1.7 F on average and +4.1 F with a quarter inch of rain or more (discovery, perfect prog). The cause was physics, not tuning: it brings the 925 mb air down to the ground along the DRY adiabat (9.8 K per km). In rain the air below the cloud is saturated and the air follows the MOIST adiabat (about 6 K per km): 2.5 K (4.5 F) colder at the ground on heavy-rain nights, the size of the bias. physmodel/scmw.py adds, on top of version 1's dry constants (identical when it is dry):

  • the moist adiabat in rain (full at 0.15 mm/h smoothed over 2 hours: fitted);
  • falling rain cooling the air toward its wet-bulb (the fit made it small once the moist adiabat was in);
  • a wet lawn evaporating after rain (fitted: about 4 mm counts as fully wet);
  • the rain's own heat, snow melting in air above freezing (holding it at 32 F), and freezing rain giving up the heat of fusion on a sub-freezing lawn, with a "drip" share that runs off, ice tracked and melted later. The discovery years have about 30 hours of snow and 1 hour of freezing rain on our nights, so these keep their physical starting values; this winter's nights check them.
Wet nights, low (F)Discovery (130, CV)Test year (30, real runs)
Physics version 12.62 (bias +1.71)2.79 (bias +2.55)
Physics + wet weather1.99 (bias +0.49)1.94 (bias +0.66)
Ours1.571.38
Ours + 0.25 x (wet physics - ours)1.521.38 (L-016: FAIL)

The wet physics makes the night model right on rainy nights, but it brings our forecast nothing there: rain-night lows are set by the rain and the air mass, which the statistical forecast already carries.

The day model needed no add-on. A wet-weather version (rain cooling toward the wet-bulb, a wet-ground store) made highs worse (bias -2 F): the day model already sees a rain day through the runs' dark sunshine and the soil moisture, and the add-on counted the water twice. Unchanged, the day physics is worth MORE on rainy days than on dry ones:

Wet days, high (F)Discovery (166, perfect prog)Test year (43, real 00Z runs)
Ours2.142.08
Day physics alone1.622.06
Ours + 0.30 + 0.54 x (physics - ours) (the L-010 rule)1.63 (CV share)1.74 (L-015: PASS, gain 0.33 F, 99.7% of resamples)

Our forecast misses rain-day highs by more than dry-day highs (2.1 vs about 1.5 F), and the physics, started from the station's sunrise reading and fed the runs' sunshine, catches a third of the miss. Scripts: physmodel/scmw.py, daywet.py, calibrate_wet.py, calibrate_daywet.py, l015.py, l016.py.

Snow on the ground (2026-10-18, L-017)

Version 1 ignored snow cover and ran far too warm over snow: +5.8 F on last winter's nights with 2 inches or more. Snow conducts heat a fraction as well as soil, so the soil's warmth reaches the surface through the snow in series, k = 1 / (1/k_soil + depth/k_snow), and a snow surface cannot warm past 32 F (physmodel/scmw.py, item 7). Fitted on the learning winters (40 dry snow nights): snow conductivity 0.118 W/m/K, the textbook range for settled snow; the fit did not support a thin fast-chilling surface layer. Held-out winter 2025-26 (perfect prog, a physics check: no real runs exist for that winter):

Night low, MAE (F)Without snowWith snowModels' own
Snow 5 cm or more (23 nights)6.14 (bias +5.79)4.53 (bias +2.23)2.77
All snow nights (43)4.59 (bias +2.62)4.02 (bias +0.26)2.12

A large step (L-017 passed), not the end: deep snow is still about 2 F too warm, likely because the learning winters had only 3 nights that deep. The ground probes will sit under real snow this winter and measure the insulation directly. Snow depth live comes from Brian's reports (fc/snow.php), carried forward by the Tower (snowpack.py), with the grass probe as a burial detector and the models' snow depth as the fallback.

The calibration run (2026-10-18, L-018)

Idea from the notes review: start each run from a state tuned to the station's recent past. For the day model the uncertain state is how wet the ground behaves. For every past day, the evaporative share that best reproduces the station's warming curve with the measured sunshine was inferred (calib_run.py). One day's value is mostly that day's noise (correlation with the day before 0.13); a two-week average holds a slow signal (0.23), and calm clear days carry it best (0.30, half the noise). On discovery days, half the two-week estimate improved the day physics (1.465 -> 1.417 F) and the L-010 rule (gain 0.04 F). With real forecast runs on 178 test-year days it added nothing (gain -0.004 F, L-018 FAILED): the forecast sunshine and air mass carry far larger errors than the ground's slow wetness. The grid's soil moisture stays as the day model's input. By-product: the L-010 rule, unchanged, beat ours on all 178 days (1.61 -> 1.53 F, dry and wet); its live test (L-010, dry days) continues.

The physics ensemble (2026-10-18, L-019)

Running the night physics once per forecast model and averaging the results (as forecast ensembles do), instead of once on the averaged forcing, did not help on discovery nights (2.12 vs 2.08 F, four models), even on nights when the models disagreed about low cloud; the constants were fitted on averaged forcing, which favours it. Adding GFS to the averaged forcing looked better on discovery nights (ECMWF + ICON 2.14, + GFS 2.06 F) but FAILED with real forecasts (1.79 -> 1.85 F, L-019): its help came from its first hours, which are close to an analysis. Real-run members alone: GFS 2.43, ECMWF 2.18, ICON 1.86 F; the ECMWF + ICON average beats each. Lesson for this program: an input's value in the stitched first hours does not transfer to forecasts; test every input on real runs. The spread of the members is a modest confidence signal (tighter half of nights 1.6 F, wider half 1.95 F).

The day model's cold bias and the crop calendar (2026-10-18, L-020)

With real forecasts the day physics ran about 1 F too cold. Swapping inputs on the exploration months (April 2 to July 15, 2026) split it: forecast inputs about 0.6 F (sunshine timing more than amount: forecast sunshine runs 11.7% over the pyranometer, close to the 0.907 calibration; plus the morning upper air and the air moving in), the model itself the rest. And the model's own error repeats every year by season: too cold in spring and fall, too warm in July-August and winter. Its evaporation followed a smooth season peaking in mid-July; real fields here are bare in spring, a closed transpiring canopy in July-August, and drying in the fall. A second seasonal harmonic (a crop calendar, refitted on discovery days with measured sunshine) fixed much of it (discovery 1.28 -> 1.19 F, 2023 1.59 -> 1.41 F), and a +0.84 F offset for real forecasts (fitted on the exploration months) the rest. On held-back months (L-020 passed): 1.79 -> 1.59 F, bias -1.22 -> -0.37 F, now better than ours alone (1.71 F). Midsummer (August) is still about 1.4 F cold: the next piece, probably the corn canopy's peak transpiration or the hot layer.

Clouds: what they cost, and the midday sunshine update (2026-10-18, L-021)

Neither model forecasts clouds itself; both inherit ECMWF and ICON's. With real forecasts (2026), the sunshine reaching the station differs from the forecast by 10% of clear-sky on a typical day; on the 10 much-darker days our high missed by 2.5 F (1.3 F too warm), on days the sunshine was about right 1.4 F. At night, against the cloud the airports reported, the low missed by 1.3 F when the cloud forecast was right and 1.4 to 1.7 F when not. Cloud surprises explain about 10% of our high's error and 20% of the day physics'; with the measured sunshine the day physics would improve from 1.63 to 1.43 F. The morning's sunshine surprise alone says little about the afternoon (correlation 0.09: morning clouds burn off, afternoon cumulus is its own). Restarting the day physics with the sunshine measured so far (and the last two hours' measured/forecast ratio, fading) and blending it with ours and the station's error improved the high on held-back months at 11 AM, noon and 1 PM (gains 0.07 to 0.11 F over the station-error update), but at the pre-registered 1 PM not consistently enough (86% of resamples): L-021 failed narrowly. A live trial would settle it. The night version (inferring the real cloudiness from the first hours of cooling after sunset, with the grass probe) waits for a few weeks of the mounted probe.

2026-10-18, the shape of the night (Brian's question from the October 18 chart). Version 1 was tested on the low; does it get the PATH right? The live dry-night model (params_v1.json) against the station, hour by hour after sunset, dry nights (physmodel/night_shape.py, output night_shape.out and night_shape_held.out).

Physics minus station (F), hours after sunset1 h2 h3 h5 h7 h9 h
A near-analysis, Mar 2024 - Sep 2025, all 474 dry nights+0.2+0.6+0.6+0.20.0-0.1
A, clear and calm (164)+1.2+1.8+1.7+0.9+0.5+0.1
B real runs, Apr 2 - Jul 15 2026, clear and calm (25)+0.5+1.1+1.2+0.60.0-0.7
C HELD BACK real runs, Jul 16 - Oct 11 2026, all 69+1.8+2.4+2.2+1.8+1.3+1.0
C held back, clear and calm (34)+2.4+3.0+2.6+2.1+1.5+0.9

(Clear and calm: models' low/mid cloud under 25% and 10 m wind under 3 m/s through the night.)

  • The physics cools too slowly in the first hours after sunset, worst on clear calm nights: the station did 54-57% of the night's cooling in the first 3 hours, the physics 41%. It catches up overnight, which is why the low tested well.
  • It also starts with a jolt: in its first 5-minute step the air drops 0.9 F (the grass starts a fixed 7.6 F under the air and the air is pulled toward it at once); the station drops 0.2 F in those 5 minutes.
  • The held-back late-summer and fall nights are worse than the spring ones and stay about 1 F too warm by morning (longer, drier nights). Seen in all three sets, so not chance.
  • Why it matters: the low is right but the timing is not: frost and dew onset, spray-window inversions and the grass curve all come too late on clear nights. A fit judged on the whole night's path (not only the low) is the natural next step, and the two-layer design below (a 2 m layer that decouples at dusk) is the physical fix it points to.

2026-10-18, version 1G: an evening layer that starts shallow and deepens (physmodel/scm_g.py, calibrate_g.py, params_g.json). Version 1's fitted cooling layer is a fixed 159 m block with a heavy skin (76,500 J/m2/K), so it cools almost evenly all night. On clear nights the real air near the ground goes still in a shallow layer just after sunset and cools fast, then slows as the layer deepens (the early evening transition: Acevedo and Fitzjarrald 2001; the nocturnal boundary layer's growth toward an equilibrium depth: Nieuwstadt and Tennekes 1981). 1G: depth = Z_EFF x (Z_F + (1 - Z_F) x (1 - exp(-t/Z_TAU))), two more constants, everything else version 1, fitted the same way on the same discovery nights. Fitted: full depth 247 m, 34% of it at sunset, deepening over about 1.6 h; the skin's heat capacity fell to a physical 4,500 J/m2/K and the starting grass-air gap to 2 F, which removes the 0.9 F first-step jolt (now 0.3 F, station 0.4 F).

Dry nightsHours 1-6 after sunsetWhole nightLow
Real runs Apr 2 - Jul 15 2026 (not fitted on), v1 / 1G1.64 / 1.54 F2.02 / 1.93 F1.69 / 1.69 F
Held back Jul 16 - Oct 11 2026 (indication only), v1 / 1G2.34 / 2.15 F2.34 / 2.20 F1.96 / 1.95 F

On clear, calm nights the lag at 2 hours fell from +1.1 to +0.5 F (spring and summer real runs); a lag remains. A wind-aware variant (1GW: the layer also shallower when calm) fitted slightly better overall but worse on the early evening, so it was set aside by the rule fixed before looking. Pre-registered as L-026, running live beside version 1 from October 19, 2026.

2026-10-18, late evening: which clear-sky formula? Brian asked whether the published formulas inside the model could be improved for this site. Version 1's fit turned Brutsaert's coefficient down from the textbook 1.24 to about 1.0, a sky about 20% dimmer than published. 1G was refitted (all constants, same nights, same loss) with Prata 1996 and with Dilley and O'Brien 1998 (forms checked against the MeteoIO library's code), each with one fitted strength factor.

Real-run exploration nights not fitted on (81)Hours 1-6Whole nightLowStrength vs published
1G, Brutsaert 19751.54 F1.93 F1.69 F~0.80
1G, Prata 19961.53 F1.93 F1.68 F0.80
1G, Dilley and O'Brien 19981.49 F1.88 F1.66 F0.88

Dilley won on all three (gain over Brutsaert positive in 100% of resamples) and replaced Brutsaert in 1G before L-026's first scored night (amendment recorded there). All three formulas wanted the sky turned DOWN, two of them by the same 20%: the model is probably using the sky to make up for something else, most likely too much mixing on calm nights (the stability formula's fitted constant is tiny, so stability barely cuts the turbulence). That is the next piece to examine. Dilley's fit also chose a very light grass surface (767 J m-2 K-1); plausible for a grass canopy, worth checking against the grass probe once it has a few weeks of readings.

2026-10-18, night: which stability shape, and the air's own radiative cooling? Following the sky clue, 1G (Dilley sky) was refitted with four published shapes for how stability cuts turbulence, plus a new term: the near-ground air's own longwave cooling (radiative flux divergence: a share K_RA of the clear sky's deficit spread through the evening layer; version 1 can cool the air only through contact with the grass). Same nights, same loss; judged on the real-run nights not fitted on (81).

1G (Dilley sky) with...Fitted nights: whole / lowUnseen 2026: whole / lowSky vs publishedCutoff
quadratic 1/(1+bRi)^2 (current)2.09 / 1.96 F1.87 / 1.66 F0.88b 0.05: barely cuts
Louis 1979 long tail2.08 / 1.95 F1.90 / 1.67 F0.91
exponential2.08 / 1.96 F2.06 / 1.82 F0.92
sharp, critical Ri2.09 / 1.99 F2.16 / 1.87 F0.94Ri_c 2.3
quadratic + air radiative cooling2.10 / 1.95 F1.90 / 1.69 F0.90K_RA 0.19
sharp + air radiative cooling2.09 / 1.95 F (best loss)2.01 / 1.79 F0.92Ri_c 2.5, K_RA 0.20
  • None beat the current 1G on unseen nights, so it stays (L-026 unchanged).
  • The physically fuller versions (a real cutoff, air radiative cooling) fit the near-analysis nights best, with the most physical constants (sky nearer its published strength, a realistic grass surface, the evening lag halved on clear calm nights), and lost on nights driven by REAL forecasts. A first guess was the reason operational models give for long-tailed mixing (a sharp switch turns small forecast-wind errors into large temperature errors); the measured-wind check below does not support it for this station. Robustness to imperfect inputs, not truer physics, is still what the dim sky and weak cutoff appear to buy; which input is the open question.
  • Our Richardson number is a bulk value across the grass-to-air gap, so its critical value is not directly comparable with the 0.20-0.25 measured for gradient Ri (Grachev et al., SHEBA).
  • Next: the sharp + radiative version could win where the inputs are good: driven by the station's own measured wind (a nowcast use, hours ahead), or once forecast night winds are corrected toward the station's.
  • CHECKED the same night, and the wind explanation above is NOT supported (physmodel/wind_diag.py, wind_diag.out): refitted and driven by the station's measured wind (12 ft anemometer, 5-minute means, scaled to 10 m), both versions fit the discovery nights worse (loss 4.77 and 4.40 against 3.81 and 3.79 with forecast wind) with absurd constants, and on the unseen 2026 nights did worse than the live 1G with forecast wind (whole night 1.99 and 2.30 F against 1.87 F; low 1.61 and 1.80 against 1.66 F). On 33 clear, calm nights (by the station's own wind) the fuller version beat the quadratic when both used station wind (1.83 against 1.99 F) but not the live version (1.55 F). The station's sheltered 12 ft wind, beside an 8 ft cornfield half the year, is not the wind that mixes the night air; the models' 10 m wind is the better driver. So the near-analysis / real-forecast gap is more likely CLOUDS (forecast night cloud is often wrong; the cloud-cost check found clouds the biggest single cost). A perfect-cloud diagnostic (airport-observed cloud) is the next check.
  • Cloud CHECKED too (physmodel/cloud_diag.py, cloud_diag.out): with airport-observed LOW cloud (Marion and Huntington; their ceilometers see nothing above 12,000 ft, so mid and high stayed the forecast's), refitted, both versions still lost to the live 1G on the unseen 2026 nights (1.93 and 2.04 F against 1.87 F), and on the 20 nights with a low-cloud surprise the airport cloud made it worse (1.64 against 1.36 F): the sky 10-20 miles away is often not the sky over the field.
  • DECISIVE (no fitting): on the same 2026 nights with NEAR-ANALYSIS inputs, the sharp + radiative version only ties 1G (Apr-Jul 1.71 against 1.67 F whole night; Jul-Oct 1.93 against 1.97 F). Its win on the 2024-25 nights it was fitted on does not carry to 2026 even with perfect inputs: it was mostly OVERFITTING (an extra constant and a sharp switch fit those years' quirks). Real forecast inputs cost both versions 0.15-0.2 F, the fuller one slightly more. Conclusion: the simple 1G (Dilley sky, quadratic) is the honest choice; the dim sky and weak cutoff are not mainly hiding an input problem. Files: physmodel/scm_g.py (stabf, k_ra), calibrate_g.py (louis|sharp|exp, rad), params_g_*.json, calibrate_g_*.out.

2026-10-19, the dew point after sunset (C-002 question 1; physmodel/dew_rise.py, dew_rise.out, dew_rise_check.out). On October 18 the station's dew point rose 4 F in the hour after sunset and the physics' did not. Checked on the dry nights:

  • That rise is rare: 2 F or more in the first hour on 14 of 895 dry nights (1-2%), 4 F once.
  • The usual evening goes the other way, strongest in late summer and fall. July 16 - October 11, 2026 (held back; looked at here for this question), clear calm nights: station dew point -2.3 F after 1 h, -3.6 F after 2 h, -7.1 F after 6 h; all three physics versions -0.6, -1.4, about -4.4 F. Dew forming on the grass and crops dries the air near the ground faster than the physics allows, because the physics spreads that drying through its whole cooling layer (85-250 m deep).
  • The dew miss explains much of the physics' fall low error: on those clear calm nights correlation between (station minus physics dew point, hours 1-2) and (physics low minus actual) -0.65 to -0.72; the physics' lows ran 0.7-1.1 F warm (1.3 F all dry nights): its dew brake engages too high. Spring and summer show it weakly (March 2024 - July 2026: correlation -0.15 to -0.35, dew miss under 1 F).
  • Next: a shallow MOISTURE layer (the few metres dew actually dries; version 3's Z_Q idea, refitted for this) fitted on the dew point's path as well as the temperature, explored on discovery nights and tested on fresh nights (the held-back fall nights have now been looked at for this pattern).

2026-10-19, a shallow moisture layer: tried and set aside (physmodel/moist_fit.py, moist_fit_report.out). 1G with dew drying only the bottom Z_Q metres (refilled from above, or not), fitted on the discovery nights with the dew point's path in the loss. Unseen 2026 spring-summer nights: 1G 1.48 / 1.87 / 1.66 F (hours 1-6 / whole night / low), dew path 1.59 F; with the layer 1.60 / 1.94 / 1.67 F, dew 1.56 F; without refill 1.58 / 1.91 / 1.67 F, dew 1.53 F. The dew point improved a little, the temperature got worse; the no-refill fit chose a 187 m layer (rejecting "shallow"), and neither fit made the grass start colder. Set aside. The reason: in the model the air dries only once the grass reaches the dew point, and the model's grass is far too warm at dusk (October 18: 54 F against the probe's 38 F), so dew forms late whatever the moisture layer. The evening lag, the too-cold probe-started restarts and the dew point that will not drop all come back to how the grass cools around sunset, which the historical nights cannot teach (no grass sensor before October 17, 2026). NEXT, the experiment only this station can do: after a few weeks of clean probe evenings (clear and cloudy, calm and windy), fit the grass's evening cooling and dew directly on the probe, then re-check the evening, the dew point and the fall lows together. L-028 (start from the measured grass) runs meanwhile.

Thin decks: when the models' sun disagrees with their cloud (2026-11-01)

Saturday October 31 asked it: the models called the sky 93% cloudy all day and still forecast a third of a clear sky's sun at midday, and the station got that sun (forecast 292 W/m2 at 1 PM, measured 377). The day physics turned it into 67.8 F against a station high of 66.7; the four-model average said 62.4, ours 62.3, and the main model's line pulled the physics down to 64.5. Question: on such days, does the physics beat the blend, and by enough that the line should trust it more on them? Sized on the main model's replay (physmodel/thin_deck.py, 367 day highs April 3 to October 12, both leads), with each day's forecast sun from the same archive runs (ECMWF and ICON mean shortwave 7 AM to 7 PM over the clear-sky shortwave) beside the replay's mean cloud.

Thin decks (cloud 80% or more and sun 30% or more of clear sky) are common: 100 of the 367 rows. On them the physics is off by 2.32 F with no bias (+0.26), the four-model average by 3.01 F and 2.2 F cold, ours by 2.49 F and 0.9 F cold. So the physics does beat the blend there, by 0.7 F. But it beats the blend by about as much on every kind of day (overcast 1.89 against 2.61, broken 1.93 against 2.41, mostly clear 1.69 against 2.40): the four-model average runs 2 F cold on highs at this station whatever the sky, and the thin deck is not special to it. What the thin deck is special to is ours: 2.49 F on thin-deck days against 1.83 F on all days, the statistical forecast's worst kind of day for the high, where the physics alone is the better of the two.

The main model's line already knows. Leave one month out, the one line (-0.04 + 0.134 physics + 0.889 average) is off by 2.13 F on thin-deck days with no bias (-0.09); a line fitted to thin-deck days alone (-1.21 + 0.363 physics + 0.670 average) does 2.17, and a line with a sun term (the physics-minus-average gap weighted by the sun ratio) 2.13. Nothing to gain out of sample, so no thin-deck rule: the line stays as it is. The finding is for the main model's case rather than for a treatment: on the days where ours is weakest, the main model's line is the best of the lot (2.13 against ours 2.49 and the physics 2.32). Saturday's 4.4 F miss by ours was about twice the typical thin-deck miss. The eight thin-deck days the average missed most (by 4 to 8 F) all ran warm, the physics closer on every one.

The jolt at sunset: a balanced start, sized and not adopted (2026-11-01)

Brian's question from the How it was made page: the night column starts the surface a fitted constant under the air (ts0, 4.2 K in version 1), so its first five-minute step drops the air about 1 F toward that cold surface, and on a cloudy evening like October 31 (sky 100%, the station 59.4 to 59.2 in the five minutes to sunset) that looks like a model that cannot even match the present. A candidate (treatments/cand_balanced_start.py, hook init added to column.night_init): start the surface where the column's first step reproduces the station's own fall over the ten minutes before sunset, by bisection on one trial step. Sized on 187 replay nights from sunset (physmodel/balanced_start.py, the night started from the station's sunset reading as the Tower's sunset run is), against version 1's start.

The jolt is real on average. Over the 187 nights the station itself falls 1.00 F in the first fifteen minutes after sunset, 1.64 F on clear calm nights; version 1 falls 1.17 and 1.47. So the constant-gap start reproduces the sunset plunge on the typical night and is wrong only night by night: too fast under cloud (October 31), too slow on the clearest evenings. The balanced start falls 0.52 F (1.23 on clear calm nights), too gentle, because the fall before sunset understates the fall after it; and it makes the known evening warmth worse, bias at 1, 2, 3 hours 1.7, 2.2, 2.1 F against version 1's 1.1, 1.7, 1.7, with the low 0.09 F worse on all nights (2.18 against 2.09) and 0.13 F worse on clear calm nights. Not adopted; no entry.

What the sizing says instead: the right start is a sky-aware one, and L-028 (the measured grass temperature at sunset, running since October 19) is already that. Under October 31's overcast the grass probe read 3 F under the air at sunset against the constant's 7.6 F, which would have cut the first step to roughly 0.4 F; on a clear evening the probe reads 14 F under and the start would plunge as the station does. The restart at each hourly issue (L-034's rule, in the main model) starts from the station's state and showed a 0.3 F first step on October 31. The evening warmth after the plunge, version 1 cooling too slowly from one to three hours after sunset, is untouched by any start rule and remains the night model's first fault.

A rain memory for the surface, sized and not needed (2026-11-05)

Brian's question from WXSIMATE's panel: WxSim is fed the last day's, week's and month's rain and the hours since rain ended, to set how wet the surface and soil are; our column starts every run with a dry surface and takes soil moisture from the models' top layer. Does the physics run warm after rain, as a missing evaporation would make it? Sized on the main model's replay (lead 1, April 2 to October 13, 2026) against the station's daily rain: the physics' error on the high and on the low, grouped by the rain of the day before and of the three days before.

No signal at this precision. The day high: physics bias -0.06 F after dry days, +0.50 after 0.25 to 1 inch in the three days before, -0.31 after an inch or more; wet (0.25 inch or more in three days) minus dry +0.38 F, about +/- 0.84. The night low: +1.48 dry, +1.84 wet, +0.33 after an inch or more; wet minus dry +0.01 F, about +/- 0.99. The four-model average does move with rain (its high bias -2.1 F dry, -1.3 F wet), which is its own soil moisture responding, and our day model reads that same moisture from the models' top layer, so the rain's memory already reaches the column by that road. WxSim needs the rain history because it has no soil moisture input of its own. Nothing to build; the August warmth of the day model (L-020) stays with the crop calendar, not the rain.

Night tuning from the station's own cooling, sized and not adopted (2026-11-18)

The question. After the clear calm nights of November 14 and 15 (every source 3 to 4 F warm; the first killing freeze, 27.3 F, against our 31.6), Brian asked for the night tuning that has been on the list since the jolt work: read the sky from how fast the station has cooled, and run the rest of the night with that sky rather than the models' cloud. Rules re-read first (R-011): size before freezing (8), report the other regimes beside the one claimed (10), the calm-versus-breezy switch is the repeated big miss and is to be tested on other nights, not fixed from the notable ones (11).

What was sized (physmodel/night_cooling.py and night_fade.py, the main model's 183 replay nights, evening issues April 2 to October 13, 2026). Two signals, because the forecast of record and the hourly updates do not see the same thing:

A. *Before the forecast of record.* On every replay night the 23:05 UTC issue falls before sunset, so "cooling since sunset" does not exist at the record; what exists is the station's change over the 1, 2 and 3 hours before the issue against the models' own 2 m change over the same hours (each model's latest run on hand, averaged). The station fell less than the models said (1 h: -1.25 F against -1.59; 3 h: -1.71 against -2.79), and that surprise has no relation to the night's error: its correlation with the physics' low error is +0.02, -0.07 and -0.01 at 1, 2 and 3 h, with the four-model average's -0.07, +0.05, +0.03. On the 37 clear calm nights alone +0.19 and +0.23 at 1 and 3 h, which a line cannot use: the main model's line with the surprise added, out of sample by month, is 1.543 F against 1.543 without it at 1 h and worse at 2 and 3 h (1.557, 1.554); a surprise term for clear calm nights only is worse on them (1.325 to 1.365 against 1.295). Nothing to take.

B. *Two hours after sunset*, the hourly update's situation and, in winter, the forecast of record's (the 6:05 PM EST issue is 35 to 80 minutes after sunset). The night column is run from the station's reading at sunset as the Tower's sunset run; two hours in, the station is under the run by 1.57 F on average (sd 2.77; it fell 5.19 F in those two hours, the run 3.62), by 2.38 F on clear calm nights and 1.37 on the rest: the evening lag again, in the restart's own words. The column is then restarted from the station at two hours as the main model restarts each hour, with the cloud changed by a gain times the departure (cand_cooling_sky, cloud fraction per F, 0 to 0.3 and -0.05 as a sign check). The plain restart (gain 0) takes the sunset run's low from 2.09 F (bias +1.30) to 1.98 (+1.13), and on clear calm nights to 1.58 with no bias (+0.04). A gain of 0.03 is best on every month fold and gives 1.873 out of sample against 1.976, but every bit of that is on the cloudy or breezy nights (1.940 against 2.077, bias +1.06 left); on clear calm nights it is worse (1.605 against 1.578, now cold, -0.29). Larger gains cool the cloudy nights further and the clear ones more. By rule 10 this is not a clear-night fix: it is a general warm bias of the physics on cloudy nights, which less cloud happens to offset, and the main model's calibration line already carries a bias term for that.

The eight largest departures (6 to 24 F under the run at two hours, five of them clear nights in late September and October) showed a third thing: after the restart the forecast still ended 3 to 5 F warm, because the main model lets the station's departure fade back to the run with a 3-hour e-folding time (L-038's restart rule). So the restart's memory was sized too (night_fade.py: the fade is a post-processing of the sunset run, so no re-run), at 1, 2 and 3 hours after sunset, fading over 1, 3, 6, 12 hours or persisting:

restartfade 3 h (live)best memoryout of sampleclear calm, 3 h to bestthe rest, 3 h to best
1 h after sunset2.051 (+1.28)12 h: 1.934 (+0.89)1.934 against 2.0511.598 to 1.6302.166 to 2.012
2 h after sunset1.998 (+1.20)6 h: 1.902 (+0.92)1.902 against 1.9981.604 to 1.6422.097 to 1.968
3 h after sunset1.881 (+1.08)6 h: 1.823 (+0.81)1.823 against 1.8811.606 to 1.6151.951 to 1.876

A longer memory helps by about 0.1 F, again only on cloudy or breezy nights, and the departure persisting all night is 2 F too cold on clear calm nights (bias -2.1 at 2 h): the station's early departure on a clear night does not carry to the low, because the run catches up as the surface cools. The 3-hour fade is right on clear calm nights (bias +0.14) and only too short where the physics is warm anyway.

Verdict. Neither the sky read from the cooling nor a longer memory is a treatment: both find about 0.1 F on the raw physics, all of it on the nights the calibration already corrects, and nothing on clear calm nights, which on the replay are the physics' best regime (1.3 F, unbiased after the restart). The November 14 and 15 misses are winter nights, and the replay has no winter (rule 3): the first real check of the clear-calm question is the Tower's own record from October 14 on, as a case study and then as the fresh nights accrue; nothing is widened or shifted from two notable nights (rule 11). Not adopted; the candidate hook stays in treatments/ as cand_cooling_sky for the record.

A first look at the November nights (the Tower's own sunset runs, hl/physics_runs.json; the Backups share was unreachable). Three clear calm nights, the station against the sunset run hour by hour:

nightmodels' 10 m windphysics aloneoursstationstation minus the run at +2 h, +5 h, +9 h
November 13 (the freeze)2.1 to 2.5 mph26.329.029.3+2.4, +2.6, +4.8
November 15 (the 4.3 F miss)0.7 to 0.9 mph26.231.627.3+2.3, -2.5, -1.2
November 161.5 to 1.7 mph27.429.629.3-3.4, -3.9, -1.7

On the night that hurt, the physics alone was right within 1.1 F and the statistical side was 4.3 F warm; on the two nights either side the physics ran 2 to 3 F cold and the statistical side was right. The two errors have opposite signs, so no weighting fixes both, and the physics' own night-to-night scatter on clear calm nights is about 3 F here as on the replay. And on November 15 the station was still 2.3 F warmer than the run two hours after sunset; the plunge came after 8 PM. A sky read from the cooling at the forecast of record (6:05 PM, 40 minutes after sunset) or at the first hourly updates would have pointed the wrong way on the one night that mattered. The question stays open on the fresh nights, as a count of clear calm winter nights where the physics alone beats the statistics, not as a treatment.

What would make it better (version 2)

  • Two air layers: the 2 m layer that decouples and a bulk layer above it (the "air mass"), so a calm night can run away at the ground while the layer above cools slowly. Calibrate the 2 m layer on the ground sensor's grass and station curves, and the stability constants on the nights the grass and air separate.
  • Start from the station's evening state (sunset temperature, dew point, the first hour's cooling rate) and the ground sensor's soil temperatures, then judge it where it could actually help: a 6 PM forecast of tonight's low with the 12Z runs' forecast clouds and winds, against our own issued low at the same hour. That needs cloud layers and 10 m wind from the real runs (to add to research_data.py), and a pre-registered plan before any test-year night.
  • Rain nights (evaporative cooling, wet ground), then daytime (sun, mixing depth: highs).
  • A grass-level temperature sensor at the station mast (and a soil probe a few inches down): the ground's actual temperature to start and check the model, and the ground-to-air difference in real time as decoupling begins. Tabled by Brian (see R-007); valuable here once it has a season of data.
  • An infrared sensor looking at the grass measures the radiating surface directly (check Meshtastic support first).

Later pieces

  1. Afternoon mixing for highs (with R-003): how deep the sun mixes the air, from the 850/925 mb temperatures.
  2. The shape of the hourly temperature curve (with R-008).

Questions for the library

  • Simple published nocturnal boundary-layer or surface energy-balance models tuned to single stations: which formulations worked best for minimum temperature?

Log

  • 2026-10-17: thread opened at Brian's request; version 1 designed.
  • 2026-10-17: the ground sensor went in the same day (R-007). Version 1 can now start from MEASURED soil temperature (2, 4, 8 in) instead of model soil, and each night's grass curve checks the radiation and exchange parameters directly (first night: grass 12 F under the air at dusk, then held at the dew point). Brian asked again for our own physics model; next step is building version 1 on the station's minute history.
  • 2026-10-18: version 1 built, fitted and cross-validated (above): matches the models on calm nights but adds nothing beyond a statistical correction. Version 2 plan: two air layers, evening start, real-run inputs.
  • 2026-10-18: version 2 explored and shelved; L-007 pre-registered and PASSED with version 1. Next: run live on the Tower and test against our live-adjusted low (L-008).
  • 2026-10-18: night shape checked: version 1 cools too slowly for the first 3 hours after sunset (up to 3 F too warm on clear calm fall nights) and starts with a 0.9 F jolt; the low is unaffected. Next: a path-based refit or the two-layer evening, explored on A/B, tested on fresh nights.
  • 2026-10-18: day model v1.1 built, fitted, cross-validated (gain 0.05 F, not proven) and put live for display; road map to clouds, storms and precipitation added.
  • 2026-10-18: solar sensor brought in; models' own radiation forecasts beat our cloud dimming.
  • 2026-10-18 evening: fitting bug found and corrected (switched day variants); advection v2 makes the day model add to our high (0.14 F cross-validated); L-009 pre-registered.
  • 2026-10-18: L-009 failed; the cause is the cloud-to-sunshine step (with measured sunshine the physics matches ours on the same days).
  • 2026-10-18: day sunshine rebuilt on the models' radiation (sensor-calibrated); L-008 and L-010 pre-registered; L-010 version live.
  • 2026-10-18: precipitation type explored and tested (L-011): the wet-bulb rule beats the 33 F rule; physics ice detection not confirmed.
  • 2026-10-18: fog first look: the night model as built is a poor fog forecaster (AUC 0.6-0.7 vs the models' humidity 0.9); needs a moisture budget.
  • 2026-10-18: night moisture budget (version 3) built: dew point slightly better, fog still weak; fog log live; L-012 pre-registered.
  • 2026-10-18: 24-hour cycle built and put live (display; extremes logged for a future test).
  • 2026-10-18: fair-weather cumulus explored with the ring of airports as truth: ranking yes, timing and base no; mixed layer refit with ERA5 depth.
  • 2026-10-18: the small effects combined (L-014): they overlap or fade by the next day; the physics is the one large piece.
  • 2026-10-18: wet weather: the moist adiabat fixes the night physics in rain (L-016 failed vs ours); the day physics passes on wet days (L-015).
  • 2026-10-18: snow cover added to the night physics (L-017 passed on the held-out winter); snow reports live.
  • 2026-10-18: the calibration run (ground wetness from the station's own curves) explored and tested: L-018 failed.
  • 2026-10-18: physics ensemble explored and tested: L-019 failed; live forcing stays ECMWF + ICON.
  • 2026-10-18: day cold bias diagnosed and fixed (crop calendar + real-forecast offset): L-020 passed.
  • 2026-10-18: cloud cost measured; midday sunshine update tested (L-021 failed narrowly, consistent direction).
  • 2026-10-18: observed clouds tried two ways at night (the ring airports' reports, L-023; MOS and National Blend sky cover at Marion, exploration): neither helps the physics; its cloud constants are tuned to the models' layered cloud. The night restart from the station's reading is what helped (L-023, narrow fail).
  • 2026-10-18: version 1G (growing evening layer) built and fitted; better evening and whole night on unseen nights, low unchanged; L-026 pre-registered, live beside v1 from October 19.
  • 2026-10-18: sky formula compared (Brutsaert, Prata, Dilley): Dilley adopted in 1G before L-026 starts; all three want a dimmer sky, pointing at the stability formula next.
  • 2026-10-18: four stability shapes and the air's own radiative cooling compared: none beat 1G on real-forecast nights; the physically fuller versions win only with near-perfect inputs (robustness lesson).
  • 2026-10-18: measured-wind diagnostic: station wind is a worse driver than forecast wind (sheltered 12 ft, cornfield); the wind explanation is withdrawn; clouds are the next suspect.
  • 2026-10-18: cloud diagnostic and near-analysis check: the fuller physics' advantage was overfitting to the fitting years, not an input problem; 1G stays.
  • 2026-10-19: evening dew point checked: the October 18 rise was rare; the usual fall pattern is fast drying the physics misses, which makes its fall lows run warm; next: a shallow moisture layer.
  • 2026-10-19: shallow moisture layer set aside (dew better, temperature worse); every thread points to the grass's cooling at dusk; next: fit the grass on a few weeks of probe evenings.
  • 2026-10-19: L-031 pre-registered: the grass's evening cooling fitted on the probe (5 coupling constants of 1G), fit and test on alternating ISO weeks, evaluate from November 16.
  • 2026-10-19: evening heat budget (dew night): 1G's formulas at the measured state bring 375 W/m2 into the grass, the real grass implies about 68; the weak stability damping compensates for the thick layer. L-031 amended to fit z_f and z_tau with the coupling constants.
  • 2026-10-20: C-003 (local radiation fog, low 35.5 F): physics family closest (measured-grass start 0.0 F, L-030 C2f +0.1 F); 1G 0.9 F warmer than version 1 on the fog-capped morning; grass probe 30.9 F but no frost seen near the house.
  • 2026-10-20: L-032 pre-registered (the physics race): the physics' own day-ahead forecast, logged live since October 20, blended into ours and graded like the scoreboard at leads 1 and 2 through April 30, 2027. The Cumulus minute record (2017-2023) could give the fit five more years with ERA5 inputs (backlog idea).
  • 2026-10-21: L-033 PASSED: restarting version 3 in the right state (saturated when the station is) cut the saturated-restart low error 1.04 -> 0.87 F and the bias -0.41 -> +0.09 F on the test months; also beat the live version 1 restart (0.19 F, 99.8%). Data assimilation, not a fog switch (Brian's question: rules or a model of the place?). Live trial L-034 through April 2027.
  • 2026-10-21: initialization run explored (Brian's WxSim question): afternoon cloud surprise into the night 1.75 -> 1.74 F (set aside); evening sky fitted to sunset-8 PM worse with version 1 (2.58 -> 2.68 F, knob absorbs the dusk lag), neutral with 1G. Not identifiable from 2 h; assimilate measured state instead (L-033).
  • 2026-10-21: slow recalibration explored (one constant refitted to the last 7-28 nights): best 1.75 -> 1.71 F, below the floor; a rolling bias correction (statistics) did better (1.65 F); blend unchanged (1.605 -> 1.602). Constants stable; errors are the weather inputs. Initialization idea closed: assimilate measured state, do not fit hidden knobs.
  • 2026-10-21: the physics ensemble built (51 ECMWF members, version 1 + version 3 per member): odds of the low, grass frost, freeze, saturation, split by sky and by wind thirds; graded as L-035. First night: calmest members warmest (version 1 has no radiative cooling of the air itself): a structural piece still to add.
  • 2026-10-21: version 1R (the air's own radiative cooling) built and refit on the discovery nights: eps_L 0.091, no measurable change on history (the near-calm forecast regime appears on 1 of 1,046 nights); in tonight's ensemble it narrows the calm-member warm bias by about half. Graded in the ensemble as L-036; the shallow calm layer is the remaining piece.
  • 2026-10-21: the day ensemble built (the L-020 day physics on each ECMWF member's sunshine, clouds, wind, profile and advection); graded as L-037 against the members' own highs, both centered on our forecast (only the spread can win).
  • 2026-10-21: version 1RW (shallow layer in calm air) fit 1.988 F in sample but made the ensemble's calm members worse (the fit weakened the grass-air exchange for the 3-5 mph nights); set aside. Calm-air constants cannot be fitted from forecasts that are never calm: use the station's observed calm nights next.
  • 2026-10-22 (Fable's session): the conversions sized on the main model's replay (full table in R-011). Research step opened: a bounded refit of the deepening layer's z_f and z_tau inside the main model (every other constant the main model's), then a sizing as a single change. On the replay 1G's layer, as fitted jointly with the Dilley sky, helped cloudy nights by 0.36 to 0.45 F a night and hurt clear calm nights by 0.31 F (bias -0.42 to -1.37 F); the Dilley sky was rejected at both scales. A layer whose growth depends on the wind and the cloud would be a second change and is sized as such, after the refit.

2026-10-22, the bounded refit with a bootstrap, run on the Mac Studio (studio/refit_layer.py; reports in deploy/reports). The research step above, done twice on the 474 dry discovery nights with 200 block-bootstrap refits over resampled months each, checked out of sample on the 81 dry real-run nights of April 2 to July 15, 2026. The loss is version 1's (curve plus low).

  1. *The layer alone, the other 13 constants held at the main model's values.* The data does not want it: z_f fits at 0.92 (5 to 95 percent 0.72 to 0.99; 57% of resamples above 0.9, which is no deepening to speak of), z_tau is unidentifiable (loss within 1% of its minimum from 10 minutes to 6 hours) and trades against z_f at -0.85. Loss 4.092 -> 4.087, nothing out of sample. 1G's layer values dropped into the main model unchanged make every night about 0.9 F too cold, in and out of sample: the mechanism behind the sizing result (clear calm nights colder). Closed as a single change.
  2. *All 14 together (the 13 of version 1 plus the layer), as 1G was fitted.* The optimiser found a second solution, not 1G's: loss 3.871 against 1G's 3.915, and a tie out of sample (low MAE 1.69 F for both; curve 1.96 against 1.93 F). Its constants: the skin starts 3.6 K WARMER than the air at sunset (3.3 to 4.1), the air layer starts at 3% of its full depth (about 8 m) and deepens in about an hour, the skin's heat capacity 6,239 (1,275 to 9,715), the wind floor driven to zero. A grass surface 6 F warmer than the air at sunset is not what the mast sees; the fit uses a warm skin over a thin layer to reproduce the first hour (clear calm bias at 1 h +0.9 -> +0.5 F), imitating heat stored in the ground under the grass coming back up through the early evening. A single slab can only say that as "the skin starts warm". The data cannot tell this branch from 1G's.

What the data pins down (loss within 1% of its minimum): the layer depth 231 to 267 m, the mixing aloft, the deep soil conductance 2.8 to 3.3, the exchange coefficient, both cloud factors, the clear-sky coefficient 0.995 to 1.016, the stability constant. What it does not: the skin heat capacity (1,560 to 10,397, a seven-fold spread, and 76,500 in version 1 against 4,457 in 1G), the sunset skin offset, the wind floor. Strongest trade-offs: layer depth against the clear-sky coefficient (-0.79), soil conductance against deepening time (-0.64). The three unpinned constants are all the ground.

Two limits of the job: every bootstrap refit started from the full fit, so the ranges describe the branch found, not the choice between branches; and the loss profiles for the sunset offset and z_f did not reach the fitted values (grids fixed for the next run). A bootstrap of version 1's own 13 constants, no layer, is the control and runs next.

Conclusion: the deepening layer is closed as a single change (alone it does nothing; jointly it is one of two skins the data cannot separate). The weakness it exposes is the ground: one slab with a fitted heat capacity, one conductance to one deep temperature, and a fitted sunset offset. Next research step: a grass node over a soil column (studio/soilcol.py). The grass gets a small heat capacity and takes the radiation; a fitted insulating conductance couples it to a column of levels to 1 m whose diffusivity comes from the soil's texture and water, not a fit. The soil survey (USDA SSURGO, queried October 22 at the mast's coordinates) gives Miami silt loam: plough layer 0 to 21 cm with 18% sand, 64% silt, 18% clay, 2% organic matter, bulk density 1.45; clay rising to 31% at 33 to 70 cm; dense till (bulk density 1.85) below 84 cm. The column is fitted first against the ground probes (grass at the blade tips, soil at 2, 4 and 8 inches, every 5 minutes since October 17, 2026: the grass probe as the top boundary, so the ground is settled on its own truth), and only then through the 2 m air on the discovery nights. The sunset state the slab had to fit (ts0) becomes a measurement: the probes' profile at sunset. Six probe nights so far; the job is built to re-run as the record grows and reports its ranges each time.

2026-10-22, later: the grass node over a soil column, first Studio run (studio/soilcol.py, soilcol_fit.py; reports in deploy/reports). Stage A had one complete probe night in the site's 48-hour window (October 21; the whole record follows with the next Tower install): the column driven by the grass probe reproduced the 2, 4 and 8 inch probes to 0.2 F all night, with the soil's conductivity at 0.51 of the published value for Miami silt loam at field capacity and the turf's conductance at 3.6 W/m2/K. Measured that evening: grass 16.5 F below the air at sunset, the soil at 2 inches 9 F below it; two hours later the soil 3 to 4 F above the air, the grass 8 F below. Stage B (the air refitted through the column on the 474 discovery nights, the two ground constants fitted too since stage A had one night; 100 month-bootstrap refits; converged):

Dry nightsversion 11Gcolumn
Loss on the discovery nights4.0923.9153.639
Out of sample Apr 2 to Jul 15 2026, whole night2.02 F1.93 F1.90 F
Out of sample, low MAE (all 81 / clear calm 25)1.69 / 1.411.69 / 1.411.54 / 1.25
Out of sample clear calm, bias at 1, 2, 3 h+0.5, +1.1, +1.2+0.3, +0.5, +0.7+0.8, +1.0, +1.0

The column has no deepening layer and beats 1G on the low out of sample by 0.15 F; the early-evening lag on clear calm nights remains (1G's layer handled that part). Cloudy nights run 0.8 F warm by morning in all three (a shared problem).

What changed in the air once the ground was a column: the cooling layer halved (z_eff 73 m, pinned 68 to 79; version 1 159, 1G 247), mixing aloft nearly doubled (trade-off with z_eff -0.95), the stability constant rose (0.136 against 0.081), the exchange coefficient fell (0.041 against 0.066), and the clear-sky coefficient moved UP to 1.066 (version 1 had turned it down to 1.008; every sky formula tried on October 18 had wanted the sky dimmer). So the slab was absorbing error that the sky and the layer were then fitted to cover. The ground constants: soil surface minus air at sunset -0.2 K (pinned -0.8 to -0.1; version 1 fitted its skin 4.2 K under the air), the grass node's heat capacity 532 (133 to 2,130, small as it should be), the grass's own sunset offset unidentifiable (the node forgets its start within minutes: a constant that disappears), conductivity factor 0.82 (0.76 to 0.88), turf conductance 50 W/m2/K (40 to 62). That last one disagrees with stage A's 3.6 from the probe night: through the air alone the fit ties the grass to the soil surface, which the measured grass (16 F under the air at sunset) says is wrong. Stage B lacks that truth; when stage A has 10 probe nights or more, its two ground constants are held in stage B by design, and that run decides whether the gain survives with the measured turf.

The control, version 1's own 12 constants bootstrapped (no layer): the refit lands exactly on version 1 (a true optimum), and its ranges show why the ground had to change: skin heat capacity 591 to 110,019 (200-fold), sunset offset -6.6 to +2.6 K, wind floor 0.02 to 1.26 m/s, skin capacity against sunset offset -0.89. The column's corresponding constants are measured or pinned.

Next: the probe archive (hourly from the Tower after the next install) feeds stage A with every night; then stage B with the ground held; if the out-of-sample gain holds with the measured turf, the column goes into column.py as a candidate, is sized on the replay as one change (R-011 rule 8) and frozen as a treatment with a bar. The dusk lag is a separate change (the layer or the stability shape) and is sized after the column, not with it.

2026-10-23, what six probe nights say about the evening (studio/ground_store.json; the first look, no fit). The grass decouples from the air BEFORE sunset, and that is where the gap peaks. Grass minus the station's air on the six clear or mostly clear evenings of October 17 to 22: within a degree three hours before sunset; about -4 F two hours before; -8 to -11 F one hour before; -13 to -17 F half an hour before; -12 to -19 F at sunset; then the gap SHRINKS all night, to 4 to 6 F at the morning low (grass low 29.5 to 34.6 F, air low 31.5 to 38.8 F). The 2 inch soil sits 2 to 9 F below the air at sunset and 3 to 8 F above it two hours later, as the air falls past it. So the evening transition is a grass event that begins two hours before sunset, while the sun is low and the air is still warm; the night model starts at sunset with a fitted grass offset (the fits say the data cannot identify it; the probes say it is about -16 F) and cannot make the first hour's fast drop any other way, which is the warm-skin branch the Studio found. The grass-probe start (L-028) uses the measured value at the issue; the model itself should start the grass's cooling about two hours before sunset from the day's state. Two other first looks: in the morning sun the grass stays 3 to 4 F BELOW the air until about 10 AM (dew drying; a truth for the moisture budget's evaporation), and the mast-top probe reads 1 to 3 F above the station's air in late-morning sunshine, which on a sunny day should be the other way round (a shield warming in the sun is the first suspect; the calibration waits for mixed air). Stage A with six nights: conductivity factor 0.47 (0.41 to 0.51), turf conductance 3.4 (3.3 to 3.6), probes to 0.2 F.

2026-10-23, the long record (studio/longrecord.py, first full run on the Studio). March 23, 2021 to March 8, 2023 built from the Cumulus minutes with the GFS historical-forecast archive above (the earliest date it holds 925 mb) and ERA5 soil: 706 nights, 588 dry, 235 of them November to March (two winters). The grass node over the soil column, fitted the same way as on the discovery nights, with 100 month-bootstrap refits:

version 1column fitgain
Long record, 588 dry nights (loss)6.995.8516%
Long winters, 235 nights (loss)7.375.9619%
Discovery 2024-25 for comparison (loss)4.093.6411%

The column wins on a second, independent record by a larger margin than on the first, and on winters it had never seen. On GFS inputs the whole error is larger (curve RMSE 2.36 F, low MAE 2.37 F against 2.05 and 1.89 F on near-analysis), which is the inputs. The cross test is the lesson: the long-record constants scored on the discovery nights give 4.43 (worse than version 1's 4.09; their own fit 3.64), and the discovery constants on the long record give 6.81 (barely better than version 1's 6.99; own fit 5.85). The structure carries over; the constants do not. The constants that differ most are the ones that touch the inputs: the wind floor 0.73 m/s on the long record against 0.04 (the GFS 10 m wind is a different wind from the near-analysis one), the stability constant 0.025 against 0.136, the grass heat capacity pinned at its floor. So a fitted constant is partly a calibration to the inputs it was fitted with, which is why the live model's constants must be fitted on the inputs it actually gets, and why the long record is for testing structure (does a change help on both records?) and not for lending constants. Two winters say one physical thing: in winter the turf's conductances halve (grass to air 0.019 against 0.036; grass to soil 28 against 64), and the soil surface starts 2.5 K under the air at sunset against 1.8. A dormant, insulating turf over a cold ground: the winter version of the insulating effect, and the candidate for a seasonal treatment once the probes have seen a winter. Stage A on six probe nights, with 200 bootstrap refits: conductivity factor 0.46 (0.41 to 0.51), turf conductance 3.4 (3.3 to 3.6), probes reproduced to 0.2 F. The multi-level air column is fitting now.

2026-10-24, the multi-level air column's first full fit (studio/sbl_fit.py; 7 hours on the Studio) and the night of October 23. The air resolved in 17 levels from 0.5 to 300 m over the grass node and the soil column, 15 constants, three starts of 4,000 evaluations and 50 month-bootstrap refits. Bottom line: it fits the discovery nights WORSE than the one-layer physics (loss 4.335 against version 1's 4.092 and 1G's 3.915), and on the 81 out-of-sample nights its low error is 1.71 F against 1.69 for version 1 and 1.54 for the one-layer column. The fit ran into its bounds on the stability constant and the exchange coefficient (both at their upper limits in every resample) and chose a rough surface (z0 0.13 to 0.28 m, a field of corn rather than grass), which says the closure as written cannot shut the exchange down hard enough on calm nights: the model's grass sits only 1.5 to 1.8 F under the station on clear calm nights where the probes show 8 to 16. Read: resolving the air in layers has not paid off with this closure; the next version needs a stability cut-off that goes to near zero in calm air (the long-tail and sharp forms were tried on the single layer in October 18's work), and the mast-top probe as a second truth. The one prediction worth keeping: on clear calm nights it puts the 2.92 m air 0.2 to 0.3 F above the station's 1.52 m and 10 m air 0.8 to 1.0 F above it. The probe's first night measured 0.5 to 2.0 F across the same 1.4 m for most of the night and 5.0 F at its widest (23:40), so the real inversion in the lowest metres is several times stronger than the column makes it: the resolved air is too well mixed.

The night of October 23 to 24 on the probes (clear, calm, the first with the mast-top probe). The grass gap peaked at 16 F within minutes of sunset and shrank to 6 to 8 F for the rest of the night; dew from about 7:30 PM. The air up the mast ran colder than the station for the first hour of fast cooling (lag of the station's shield, or the corn's stored warmth at canopy height), crossed at 7:40 PM, and held a 0.5 to 2 F inversion until it widened to 5.0 F at 11:40 PM; then, between 11:40 PM and midnight, the station's air jumped from 46.7 to 51.4 F in half an hour as the cups stirred to 1.2 mph and the dew point rose 2.8 F: the inversion broke and the warmer air above mixed down (the mast read 52.8, the inversion fell to -0.8 F). Cooling resumed, the inversion re-formed at 0.5 to 1.7 F, and the station's low came at 8:00 AM, 41.4 F, with the grass low 34.0 F at 6:00 AM. That midnight mixing event cost the night 2 to 3 F of cooling: ours said 39.5, the physics 38.2, the National Weather Service 42; the station gave 41.4. The models' wind was 4 mph all night against cups at zero, with one puff at midnight that nothing forecast. The morning: the grass stayed 6 to 8 F under the air until 8:30 AM and flipped above it by 9:00 AM once the dew dried, a 9 F jump in half an hour.

A grading bug, found October 24. Since the October 22 change that routed the main model's grading through the one truth table (review H-M3), station_extreme looked the date up as a string in a table keyed by date objects, so every period came back ungraded: the Engine page's graded counts read zero for the main model and every treatment, the Physics page's monthly table was empty, and the family's scorer saw no nights. Fixed in column_live.py; nothing was lost, since grading is recomputed from the issued and obs tables each time. The frozen entries' code hashes are re-recorded by code amendment (the replay is unchanged).

2026-10-29, the soil column's gain came from soil constants the probes contradict (studio/soilcol_fit.py; the morning runs of October 26 to 28, relayed to the site).

The Studio's morning run (9:30, every day since October 26) refits the soil column on the probe record as it grows. Stage A fits two ground constants on the buried probes alone, driven by the measured grass temperature: the soil's conductivity as a share of the textbook value for this silt loam (f_d) and the turf's grass-to-soil coupling (k_gs). Stage B then fits the air constants on the 474 training nights with the station's low as truth. Until the probe record reached ten nights, stage B fitted f_d and k_gs too; from ten nights on it holds them at stage A's values. That switch happened on October 27, and it reversed the column's result.

runprobe nightsf_d, k_gs in stage Bloss on the 474 nights (version 1 4.092, 1G 3.915)held-out 81 nights, low MAE (version 1 1.69)
October 269free: f_d 0.82, k_gs 49.63.639 (11 percent better than version 1)1.54
October 2811held: f_d 0.56, k_gs 3.574.149 (1 percent worse)1.76

Stage A is sharp and stable: f_d 0.556 (0.52 to 0.60 within 1 percent of the minimum; bootstrap 0.47 to 0.67), k_gs 3.57 W/m2/K (3.3 to 3.8), reproducing all three probes within about 0.3 F over 11 nights. The night-low fit wants the opposite ground: a turf coupling of 50, which welds the grass to the soil and lets the soil's heat capacity feed the air all night, and a conductivity near the textbook value. Held at the measured ground, the air constants compensate: the layer depth goes to 281 m (version 1: 159; the free column fit: 73), the soil starts 6 F below the air (ts0 -6.0, at its bound), the grass heat capacity triples, and the fit is still worse than version 1 everywhere, with the early-evening warm bias on clear calm nights grown to +2.4 F at two hours.

Reading. The probes measure the ground under the grass at one spot by the mast; the station's low is the air at 5 ft, which integrates the surface around it: the field, now bare, the road at grade, the house. The night-low fit is not finding "the soil"; it is finding the effective surface the station's air feels, and that surface stores and returns heat far more readily than the turf under the probe. Two readings fit this: the probe spot is not the station's surface (a two-surface model, turf and bare, with the bare share fitted, would say so), or the one-layer air model lacks a heat source it needs on clear calm nights (the air above, mixing down) and borrows it from the soil when allowed to. The second is the same early-evening problem the air column (October 23) and the evening transition (October 22 to 28) point at.

What follows. The soil column is not a candidate for the replay in this form: with the measured ground it does not beat version 1, and with the free ground it is fitting the wrong thing well. The long record result (October 23: 16 percent better than version 1 over 706 nights, 19 percent on winters) stands as a structural finding and was made with free ground constants, so it carries the same caveat. The probe constants are kept (they are the measurement); the next candidate is the evening transition fix sized on the replay, then a two-surface ground, each as its own numbered entry.

Bottom line, plainly: the buried probes now say what the ground under the grass is, and the night model, given that ground, gets no better than the current physics. The gain it showed last week came from pretending the ground was something else.

2026-10-29, the time step (physmodel/main_replay.py at 300, 60 and 20 s; studio/refit_layer.py --dt). Brian asked whether the iteration count a desktop model offers has a counterpart here. It does. The night and day physics step every 300 s, with the surface and the air advanced implicitly (stable at any step), and the constants were fitted at that step. The main model's replay (373 forecasts of record, April 2 to October 13, 2026) was run again at 60 s and at 20 s, nothing else changed:

stepnight physics, low MAE (bias)clear calm nightsday physics, high MAE
300 s (live)2.54 (+1.47)1.31 (-0.42)1.69
60 s2.41 (+0.80)1.44 (-1.09)1.69
20 s2.40 (+0.68)1.50 (-1.20)1.69

The finer step cools the night by 0.68 F on average (60 s against 300 s, largest 2.6 F); 20 s against 60 s moves it 0.12 F more, so the arithmetic has nearly converged by 60 s. The coarse step under-cools the fast part of the evening, and the constants absorbed that: at 60 s the published constants fit the discovery nights 6 percent worse than at their own step. So the step is not a free change; it is the step and its constants together. refit_layer.py gained --dt (the forcing held on its 5-minute bins, the model stepped finer, the output read back on the bins), and a first quick refit of version 1 at 60 s on the Mac fits the discovery nights 1 percent better than the published fit and lands the held-out 81 lows at 1.67 F against 1.69 (a wash on a quick fit; the curve RMSE 1.96 against 2.02). The full refit with its bootstrap is queued on the Studio (refit_layer.py --fit v1 --boot 100 --dt 60); if it beats the published model on the held-out nights, the step and its constants go in as one numbered entry. The day is untouched by the step.

The full refit at 60 s ran on the Studio the same evening (100 resamples, converged): discovery loss 4.0355 against the published 4.0921 (1.4 percent better); on the held-out 81 nights the low is off by 1.68 F against 1.69 at 300 s, the curve RMSE 1.96 against 2.02, clear calm nights 1.46 against 1.41. Refitted for the fine step, the constants move little (c_h 0.206, u_min 0.023, z_eff 193, ts0 -4.7 against version 1's 0.066, 0.033, 159, -4.2 at 300 s), and the forecast is the same to a hundredth of a degree. Verdict: the step error was real but the fitted constants had already absorbed it; a finer step with its own constants buys nothing on nights it never saw. The 300 s step stays, and the iteration count is not where the night model's error lives.

  • 2026-11-16: L-031 readiness check (counts only, nothing fitted or scored; experiments/R-010/l031_ready.py). Dry evenings with the probe and station covering the night, October 18 to November 15: 9 fit and 9 test, against the bar's 12 in each set. Test evenings by the first-4-hour station wind: 7 calm, 2 in between, 0 breezy (6 mph or more); the breezy evenings so far (November 2 to 6) fell in a fit week and were mostly wet. Cloud tags not checked. The bar is therefore not met and L-031 waits for its December 15 date. Also, the frozen fit script (hyperlocal/grass_fit.py) is not on the Tower, so the fit has to be run where it lives, in the collector container, or the file copied into /lab. The even-week split leaves the test set almost all calm, so a pass or fail would say little about breezy evenings.
R-011What the experiments teach, read together exploringupdated November 2, 2026

The question

Brian asked whether the successes and the failures, read as a whole, show patterns. Each test is judged on its own; this thread reads all of them side by side (L-001 to L-028, C-001, C-002, and the explorations set aside on October 18-19) and turns what they share into working rules. To be re-read and updated every month or so.

What has been explored

2026-10-19, the first read-through (28 notebook entries and 12 set-aside explorations).

  1. The winners bring in NEW information; the losers rearrange information we already have.
  • Won: the live adjustment from the station's own running temperature (L-024, 21% better than the so-far rule); the nowcast from the station's error and the upwind airports right now (L-013, 2.13 to 1.68 F); the physics where the statistics rarely see the weather (wet days L-015, +0.33 F; snow nights L-017, +1.6 F).
  • Lost: adding more model information to our correction: the NWS blend and wind direction (L-001), the models' inversions (L-004), a third model in the physics (L-019), forecast winds aloft and forecast dew-point change (C-002 follow-ups). Our correction of the four-model average already extracts most of what the models know.
  • Rule: prefer ideas that add something the models do not have (the station, the ground sensor, the airports now, the owner's observations).
  1. A small-effect ceiling. Many failures were too small to prove rather than wrong: L-004 +0.013 F, L-014 +0.014 F, L-018 -0.004 F, L-021 +0.09 F (86% of resamples), L-023 +0.07 F (88%), L-003 9.6% against a 10% bar. With 80-180 test days, gains under about 0.1 F are close to undetectable.
  • Rule: before writing a bar, estimate whether the test can detect the size of effect the exploration suggests (a power check from the exploration's day-to-day spread); do not run tests that cannot succeed; do not chase tweaks under about 0.1 F unless they can be pooled or run live for a long time.
  1. One test season. Every held-out test since October 2026 was judged on July 16 - October 11, 2026, and the explorations used spring and early summer (and the discovery years). Several effects turned out to be much larger in that late-summer-and-fall stretch (the physics' evening lag 3 F; the evening dew drying; the physics' lows 1.3 F warm).
  • Rule: future splits alternate months (or weeks) so every season is on both sides; live trials already do this.
  1. Near-perfect inputs flatter. Gains found with near-analysis inputs repeatedly shrank or vanished with real forecasts (the day physics L-009: +0.14 F discovered, -0.09 F tested; the sharp + radiative night physics; the advection pieces).
  • Rule: any gain found on near-analysis inputs is provisional until shown with real forecast runs.
  1. Fix the root, not the symptom. The shallow moisture layer failed because its root cause (the grass's cooling at dusk) is not observable in the history; 1G worked because it changed the structure that caused the evening lag.
  • Rule: when a fix does not take, trace the chain back to the first thing the model gets wrong, and ask whether we can observe it.
  1. Narrow misses deserve live trials. L-003 (9.6%) and L-021 (86%) went to live trials (L-005, L-022); L-023 (88%) did not.
  • Rule: a narrow miss with the right sign in every period goes to a live trial, written down before it starts.
  1. Where the physics earns its keep. It helps most in the regimes the statistics rarely see (rain, snow) and in explaining the night (timing, frost, dew, the shape), and least as a competitor on ordinary days (L-009, L-016, L-019).
  • Rule: aim the physics at rare regimes and at timing; judge it there.

2026-10-19, what pattern 3 means for past results (Brian asked whether anything must be re-run). Nothing can be re-run usefully: the real forecast-run archive starts April 2, 2026, so there is no winter to test on yet. Instead:

  • Unaffected (tested on a full year or on winter): L-001, L-003, L-011, L-017, L-024.
  • Promoted or in use but proven only April-October: L-007 (live check L-008 runs through winter), L-020 (live L-010), L-013 the nowcast on the Forecast page (NO winter check): winter confirmation pre-registered as L-029 (fresh months November 2026 - February 2027; if it fails, the nowcast goes off for the cold months).
  • Failures judged only in fall: L-021 (live trial L-022 runs through winter), L-023 (narrow, 88%): live trial L-030 pre-registered October 19 (October 20, 2026 to April 30, 2027). The large or long-span failures (L-004, L-014, L-018, L-019) are unlikely to flip with season.
  • October 18-19 physics work (1G L-026, grass start L-028): judged on fresh fall and winter nights, so fine.
  1. Size the candidates before one is frozen (October 21, 2026, Brian and Claude). L-039 (the sky term) was frozen and only then sized: it explains about a tenth of the miss it was aimed at. The sizing pass belongs in the exploration, for every candidate term a miss points at, before choosing which to freeze; L-040 (the moist lapse) was sized first and explains 3.6 times as much.
  • Rule: for each documented miss, replay every candidate change as a single changed parameter against the current main model, report the gain by regime and the physics share, and freeze in order of size.

Documented misses (October 21, 2026 on)

Brian's operating model: treatments of the main model come only from documented misses. Each miss names what was seen, where, on which nights or days (those are then not used to score its treatment), and the term it points at.

M-1. The physics runs warm on cloudy and rainy nights (the first treatment, L-039). Seen in the main model's replay on the real forecast runs (physmodel/main_replay.csv, April 2 to October 12, 2026; these nights are not used to score the treatment). The physics alone, against the station's low, by the models' own mean total cloud for the night:

nightlead 1lead 2
clear (under 30%)-0.1 F (51 nights)-0.4 F (54)
partly cloudy (30 to 70%), dry+1.3 F (48)+0.5 F (49)
cloudy (70% or more), dry+2.5 F (46)+2.7 F (48)
cloudy, with 0.10 in or more of rain+3.9 F (35)+5.3 F (26)

The error grows with the cloud, which points at the sky term: in version 1 the cloud part of the downwelling longwave is emitted at the model's own 2 m air temperature, the warmest level in the column at night, while a real cloud deck emits at its base, hundreds of meters up and colder; under overcast this is several W/m2 too much heat into the grass every step. The calibration hides part of it (the low's physics weight is 0.30), which is why the miss counts against the physics. Sizing of the rest of M-1 (rule 8, October 21, 2026; single parameters on the replay, physmodel/main_replay.py --change):

candidatecloudy rainy nightscloudy dry nightsper forecast-cloudy nighton top of L-040physics share, low / high
L-039, cloud at the 925 mb airbias +4.5 to +4.1 F, error -0.34-0.10-0.18 F0.318 / 0.153
L-040, moist lapse in the saturated share+4.5 to +3.1 F, -1.07-0.44-0.63 F0.342 / 0.184
rain cooling on every night+4.5 to +3.0 F, -1.17-0.11-0.49 F-0.31 F0.345 / 0.169
reference air from the cloud base (below 925 mb)+4.5 to +3.3 F, -0.96-0.36-0.53 F-0.35 F0.331 / 0.154
rain cooling re-derived, no fitted constant (October 22)+4.5 to +2.0 F, -1.68-0.42-0.87 F0.334 / 0.188

(main model: share 0.305 / 0.134.) Brian broke the first tie on physics and caveats (October 22): the cloud-base depth is frozen as L-041 (no constant; its 0.2 F shift on forecast-clear nights comes only from the low cloud such nights can still carry, exactly 0.000 F where there is none); the first rain cooling is withdrawn because it used the constants fitted for the failed L-016, and its constant-free re-derivation (treatments/rain_cooling.py) is the largest candidate so far and returns as a diff from the main model after the May merges. Rain cooling and the cloud-base depth were tied within what the replay resolves (gap 0.04 F a night, 95% interval -0.12 to +0.20 F), so neither is frozen yet. They act on different nights: rain cooling only where rain is forecast, the cloud-base depth on every cloudy night (and it cools clear nights' bias from -0.23 to -0.45 F with no change in their error). Rain cooling uses the wet-weather constants fitted for L-016, which failed as a whole.

Related, not treated with it: the day model's cloud weights in its longwave (0.8, 0.5, 0.2) are set by hand while the night's are fitted (0.72, 0.47, 0.20); a later treatment if a miss in the day points there. Rain nights add the missing rain cooling (version 1 runs dry on rainy nights since L-016 failed); that part is a separate question.

M-2. The station's summer wind is sheltered by the crop (wind check, research/wind_ring.py). In airport wind of 7 to 18 mph the station reads about 0.7 of the ring airports in winter and 0.44 to 0.57 in July to September every year; in 2026 the extra drop is in winds from the southeast to southwest, over the corn field (0.49, against 0.62 to 0.64 in 2023 to 2025), more than from other directions (0.48, against 0.50 to 0.56). The 12 ft cups sit in the wake of a crop about 8 ft tall in season, so the station's wind is not the open 10 m wind the physics expects then. Consequences: calm-night fits on station wind from June to harvest are biased toward calm; the calm-night depth treatment (the calm-night miss, L-038's first documented miss) waits for this to settle. What settles the bearings question: the ratio after harvest against November of earlier years (0.60 to 0.66), and the hand spin test of the cups.

Ideas to chase

  • Done: L-030, the live trial of L-023. Its power check (rule 2) set the length: about 150 to 200 nights for a 0.07 F gain, and the main measure became the average of the three restart times (correlated, so averaging cuts the spread from 0.50 to 0.43 F).
  • Re-split the next exploration/test periods by alternating months (rule 3), starting with the grass-fit work in November.
  • A power check template in the lab tools: from an exploration's per-day gains, the number of test days needed.

Log

  • 2026-10-19: thread opened at Brian's question; first read-through of all entries; seven patterns, each with a rule.
  • 2026-10-19: pattern 3 audit: no re-runs possible; L-029 (winter check of the nowcast) pre-registered; L-023 live trial proposed.
  • 2026-10-19: L-030 pre-registered (L-023's live trial), with the first written power check: six months of nights.
  • 2026-10-21: documented misses M-1 (warm on cloudy and rainy nights; first treatment L-039) and M-2 (the crop shelters the station's summer wind) recorded at Brian's direction.
  • 2026-10-21: rule 8 (size the candidates before freezing) after L-039; L-040 sized and frozen.
  • 2026-11-02: monthly re-read; first eight rules hold; rules 9 to 13 added (bundles hide parts, regime claims need the other regimes, calm-versus-breezy misses, entries tied to replaced structures, thin historical checks).

Sizing of the conversions, October 22, 2026 (rule 8; Fable's session; physmodel/size_candidates.py and size_report.py against main_replay.csv, April 2 to October 13, 2026, main model v0.1)

Each candidate is a single change against the main model, as a module in treatments/ (cand_layer, cand_dilley, cand_dilley_g, cand_budget); the main model's own replay with the new hook points hashes identically to the frozen one. Gains are per night, physics alone, positive when the candidate is closer; intervals are 95 percent bootstraps over dates.

Candidateclear calm (72)cloudy dry (107)cloudy rainy (68)all nights (364)calibrated low, months held out
The deepening layer alone (1G's z_f, z_tau)-0.31 [-0.55, -0.09]; bias -0.42 to -1.37+0.36 [+0.14, +0.52]+0.45 [+0.33, +0.65]+0.13 [+0.02, +0.23]1.688 to 1.680 F, weight 0.318 to 0.311
The Dilley sky, textbook scale 1.0-4.57-2.66-1.42-3.041.825 F
The Dilley sky, 1G's fitted scale 0.877-0.34 [-0.69, +0.09]-0.46 [-0.57, -0.26]-0.25 [-0.36, -0.16]-0.36 [-0.46, -0.23]1.701 F
The moisture budget (version 3's)-0.02 [-0.05, +0.01]+0.09 [+0.05, +0.12]+0.17 [+0.08, +0.22]+0.06 [+0.03, +0.08]1.686 F, weight 0.323
The budget with a saturated state restartthe same as the budget alone except on 3 saturated issues (4.30 to 3.70 F)
Version 1 with a state restart (saturated)identical to the main model

What it says. (1) 1G's result was carried by its layer, not its sky: the Dilley formula is worse than the fitted Brutsaert sky at either scale, by whole degrees at the textbook scale. The sky component of L-026 is rejected on the replay. (2) The deepening layer helps the cloudy nights by a third to a half of a degree and hurts the clear calm nights by the same, turning a small cold bias into a larger one: as fitted jointly with 1G's other constants it is not a single improvement. The honest next step is a bounded refit of z_f and z_tau inside the main model (research, then a sizing), or a layer whose growth depends on the wind and the cloud, which is a second change and must be sized as such. (3) The moisture budget is a small, well-resolved gain on cloudy nights (a tenth of a degree, interval clear of zero) and nothing on clear nights; its value is in the fog words and odds, which this sizing does not score. (4) The saturated state restart cannot be sized on this replay: from April to mid-September the night half is not in progress at either forecast of record (sunset is after 23:05 UTC and sunrise before 11:05 UTC), so the restart rule for the night acts on only a handful of October issues. The winter is where it acts, and L-034 is logging exactly those restarts. (5) The day halves are untouched by every night candidate, as they should be.

Monthly re-read, November 2, 2026 (49 entries: L-001 to L-045, C-001 to C-005; the session read the notebook's titles, statuses and results, and the logs of October 22 to November 1)

Honest limit: no new held-out result has arrived since October 22 (L-038 and the treatments L-039 to L-044 start scoring only after the install and evaluate in May 2027; L-029 evaluates in March 2027; L-031 from November 16). So this re-read adds patterns from process and case studies, and checks the first eight rules against what happened.

Do the first eight rules still hold? Yes, none was contradicted. Rule 2 (small effects) was confirmed again by two sizings that found nothing: thin-deck days (October 31: the main model's one line is already unbiased on them, a special line gains nothing) and the jolt at sunset (the constant-gap start is right on average, 1.0 F fall in the station against 1.2 F in version 1). Rule 8 (size before freezing) paid off at once: it showed that L-026's result came from its layer, not its sky.

  1. Bundled changes hide which part works; single changes show it. L-026 changed the sky and the layer together and looked like a gain; replayed one at a time (October 22) the sky was worse by whole degrees and only the layer helped, and only on cloudy nights (it hurt clear calm ones). The soil column (October 29) likewise gained from constants the probes contradict.
  • Rule: never read a gain from a multi-change fork as support for each of its parts; replay the parts separately, and check any fitted constant against a direct measurement (the probes) before believing it.
  1. A gain that is the same on every regime is not evidence for a regime. The physics beat the four-model average on thin-deck days by 0.7 F, but by as much on every sky (October 31). So that result says something about the physics in general, nothing about thin decks.
  • Rule: when a method is claimed for a regime, report the gain on the other regimes beside it; keep the claim only if the regime stands out.
  1. The calm-versus-breezy switch is the repeated big miss, and the measured-state physics tends to be closest. C-004 (two opposite misses of 4.8 F cold and 3.4 F warm, the four-model average wrong in both directions), C-003 (radiation fog: physics started from the measured grass 0.0 F off, statistics 2 F warm, global models 3 to 6 F warm) and C-005 (the post-frontal night: every source 1.5 to 6 F warm, ours 2 F in the window, the physics sunset run closest at 43.4 F against the morning's 39.6 F; the 9 F miss first written there was a misread of the next date's forecasts, corrected November 2) all point the same way, C-005 least. Three nights chosen because they were notable are a discovery set, not a result.
  • Rule: test it on other nights (the backlog idea on post-frontal calm clear nights, plus L-028 and L-034's logged nights); do not widen or shift anything from these cases alone. Keep Brian's fog and frost log as the truth.
  1. Entries tied to a structure that is replaced stall. On October 22 ten entries (L-008, L-010, L-012, L-022, L-026, L-030, L-032, L-035 to L-037) were closed without a verdict because the main model replaced the structure they tested, and their questions moved to L-038 to L-044. That costs the record some verdicts.
  • Rule: before pre-registering a live trial, check that the structure under test will still exist at evaluation time; if a rebuild is planned, register the question against the rebuild.
  1. Historical checks can reverse a suspected effect, but one year is thin. L-027 part A (2018 only, WxSim as the forecast, no soybean control) found the opposite sign of the standing-corn effect. It neither rules the effect out nor supports it.
  • Rule: for a seasonal land effect, a historical check needs at least two years and a control crop; otherwise wait for the fresh falls (parts B and C).

Open from this read: the nowcast (L-013) is the promoted piece with new information in it that has no winter test yet (L-029 evaluates in March 2027); the live adjustment (L-024) was tested on a full year. R-011 should be re-read next month against whatever the winter checks have logged.

R-012Fog at the station exploringupdated October 20, 2026

The question

Brian (October 20, 2026): fog causes real trouble here for commuting and school buses, so make it a major area of study. This is hyper-local on purpose: what happens over the field at the station, not in a town a few miles away. Three parts:

  1. Knowing: when is there fog at the station, how dense, from when to when? The airports often miss it (C-003).
  2. Forecasting: from the evening before, will fog form tonight, when, how dense, and when will it clear?
  3. The commute: the chance of fog between 6:00 and 8:30 AM, and the time it burns off.

Physics and literature

  • How radiation fog forms: on clear, calm nights the ground loses heat to the sky, and the air touching it cools.
  • Dew first: dew forms first and dries the lowest air.
  • Then fog: when the air a few feet up reaches saturation, fog forms, starting shallow and deepening through the night.
  • What stops it: a breeze above a few mph mixes it away; dry air, or dew taking the moisture out first, can prevent it altogether.
  • What it does to the low: once formed, fog gives off heat as it condenses and sends heat back down to the ground, so the cooling slows (C-003: the air crept from 40 to 37 F over five hours).
  • Clearing: after sunrise, sunshine warms the ground and mixes the layer, and the fog lifts or burns off; deeper fog lasts longer.
  • Other kinds are rarer here: fog carried in on the wind, and fog after rain.
  • Questions for the library:
  • the classic field studies of radiation fog (Roach and colleagues, 1976, at Cardington)
  • reviews of fog forecasting (Gultepe and colleagues, 2007)

Both are for the onset conditions and the dew-against-fog competition.

Data

  • Truth, at the station:
  • Brian's fog and frost log (fc/fog.php, from October 20, 2026, seeded with his October 18 to 20 sightings). Fog: none, light, moderate or dense; the grass: dry, dew, heavy dew, patchy frost or frost; where; when it started and cleared; notes. Clear mornings count too.
  • The station: humidity, wind, air and dew point every minute (data/minutes.parquet, February 2023 on), and the grass probe (October 17, 2026 on).
  • Truth, regional: the airports' visibility (Marion MZZ and Huntington HHG nearest; asos.parquet).
  • Instrument check (October 20): the station's humidity sensor reaches 98 to 99% in fully saturated air in every month from 2023 to 2026, so it has not drifted. Light fog on October 20 came with 96 to 97% at 5 ft. Saturated air at 5 ft and visible fog are related but not the same thing.
  • Possible forecast signals:
  • the models' closest approach of the 2 m air to its dew point overnight (the fog log, L-012)
  • the physics' version (version 1 and version 3)
  • the models' own visibility forecast (in the physics snapshot)
  • wind and cloud
  • the evening dew point and its trend
  • rain in the past two days and soil moisture
  • the crop beside the station

What has been explored

2026-10-20, first look (research/fog_explore.py, discovery days March 2024 to September 2025, 566 nights with both records). Station saturation means humidity of 95% or more with wind at 1 mph or less, for 2 hours or more between 8 PM and 9 AM. Airport fog means Marion or Huntington at 1 mile or less.

No airport fogAirport fog
Station not saturated38720
Station saturated11445
  • The station saturates far more often than the airports see fog: 159 nights (28%) against 65 (11%). On 114 nights it saturated with no airport fog at all, so October 20 was the usual pattern. Some of those nights were surely only dew, or fog too shallow to see. Brian's log is what will separate them.
  • When airports did have fog, the station was saturated on 45 of 65 nights. The other 20 were probably fog at the airports that never reached the field.
  • The season: strongly late summer and early fall. Saturated nights by month: July 39, August 42, September 28, May 17; winter only 1 to 4 a month.
  • The commute: on 246 nights the station sat at 95% or more for at least 30 minutes between 6:00 and 8:30 AM. That's a first, generous count of the mornings fog was possible during the commute.

2026-10-20, case study C-003: light radiation fog over the field, which no airport saw.

  • The night: dead calm; the air cooled 13 F in 2 hours after sunset, then stalled once nearly saturated.
  • The fog signals at sunset: physics version 1 brought the air within 0.4 F of its dew point, the models within 3.9 F.
  • The grass: below the air's dew point all night. Heavy dew; possibly one frosted blade.

2026-10-20, Brian's camera review (the camera facing the mast, September 30 to October 19). Two mornings with light, wispy fog, both logged:

MorningStation: humidity 95%+ and calmCamera saw fogLowNearest airports' lowest visibility
October 4from 11:18 PM, 10.6 hours, peak 98%7:10 to 7:50 AM43.5 F at 7:50Marion 2 mi, Anderson 1 mi, Huntington 7 mi
October 5from 1:58 AM, 7.2 hours, peak 98%6:50 to 8:10 AM42.1 F at 5:56Marion 2.5 mi, Anderson 1.5 mi, Huntington 10 mi
  • The fog likely formed hours before the camera could show it: the camera's first sightings come at first light, after long saturated, calm stretches.
  • Burn-off came 30 to 40 minutes after first light, with the air then warming 6 to 7 F in the next hour.
  • The airports had haze or mist, not fog, on both mornings.
  • Camera review is a practical way to backfill the log. Mornings with no fog on the camera are worth recording too.

2026-10-20, night footage (Brian reviewed the seven October nights when the station sat saturated and calm for 4 to 11 hours): on every one "hard to tell, there might have been light fog". The mast camera confirms fog in daylight (October 4, 5 and 20) but cannot show light fog at night with confidence. All seven were clear by 7:00 to 8:30 AM. Logged as "can't tell", which the fog detector's calibration will count as neither fog nor clear. October in the log so far: 3 fog mornings, 18 clear mornings, 7 uncertain nights.

  • For the commute question this is still useful: no light fog that may have formed on those nights lasted into 7:00 to 8:30.
  • Night onset times need another source:
  • the station's saturation as a stand-in, calibrated on daylight-confirmed cases
  • a small light near the mast for the camera (fog glows in a beam)
  • the planned air probes
  • later, perhaps a simple homemade visibility sensor: an LED and a light sensor reading what fog scatters back

2026-10-20, burn-off: the last real wisp at 8:57 AM (Brian), about an hour after the 7:58 sunrise, still hazy after. The air held at its low until about 8:25 and then rose 5 F in 35 minutes. Humidity rose from 96 to 98% and the dew point from 34.4 to 40.0 F as the heavy dew evaporated. October 4 and 5 cleared within 10 to 30 minutes of sunrise. A first hint that burn-off timing varies a lot from morning to morning, and that the heavy dew feeds the morning haze.

2026-10-20, a better camera (Brian): firm answers for most of the saturated October nights. Fog: October 11-12 (light, 10:08 PM to about 6:30 AM) and October 4 (a light wisp from 5:45 to 8 AM). No fog: October 5-6, 7-8 and 8-9. Still can't tell: October 17-18. First look at the saturated, calm nights with a firm answer:

NightFogHours saturated and calm (95%+, wind 1 mph or less)Saturated fromAir minus dew point at 7 PM (F)
October 3-4yes, 5:45 to 8 AM10.611:18 PM15.1
October 4-5yes, about 4 to 8:10 AM7.21:58 AM17.5
October 11-12yes, 10:08 PM to 6:30 AM11.310:33 PM6.0
October 19-20yes, seen 7:15 to 8:57 AM7.51:43 AM14.8
October 5-6no4.94:41 AM11.1
October 7-8no4.21:41 AM14.4
October 8-9no4.91:20 AM15.7
October 17-18 (can't tell)?7.51:40 AM12.9
  • Duration separates them: every fog night had 7 or more hours of calm saturation, every no-fog night under 5. Peak humidity did not separate them (97 to 99% for both).
  • An early start: on October 11 fog showed about 25 minutes before the station reached 95%, after a sunset only 6 F above the dew point.
  • Caveats:
  • Seven nights: a lead, not a rule.
  • Fog keeps the air saturated, so duration is partly a result of fog. That's fine for a detector judging the night afterward, not a forecast signal; the forecast needs evening signals, such as the sunset gap.
  • To test on other nights: the duration rule as a station fog detector, checked against Brian's log going forward (November on).

2026-10-20, the first test of the duration pattern: a miss. Before anyone looked, the pattern called fog for October 17-18 (7.8 hours of calm saturation from 1:40 AM). Brian checked it on a third camera with decent vision of the area: no fog that he could see. October 18-19 never saturated overnight, and also had no fog, as expected.

  • The one thing that set October 17-18 apart: it was at or below freezing, 32 F from about 1 AM and 31.4 F at 5:25 AM. Every fog night was above freezing (lows 35 to 51 F).
  • There is a physical reason freezing nights could favor frost over fog. Water vapor deposits onto ice more readily than it condenses into droplets: air saturated with respect to ice is still short of saturation with respect to water. So frost on the grass and roofs can keep pulling vapor out of the air before fog droplets can form. That night ended in frost on the grass and the roof.
  • What it means: a hypothesis, from one night. On nights near or below freezing, frost may take the moisture first and fog becomes less likely.
  • Caveat: the third camera might miss light fog.
  • Running count: fog on 4 of the 4 above-freezing nights with 7+ hours of calm saturation. No fog on 3 nights with under 5 hours, nor on the one freezing night with 7.8 hours.

2026-10-20, eight years back (research/fog_history.py, nights in research/fog_history_nights.csv). The station's own minute record (one Davis on the same mast throughout, Brian: Cumulus October 2017 to January 2023, Weather Display since), explored on October 2017 to September 2025 (October 2025 on kept back for testing). "Fog signs" = the October camera pattern: 7+ hours of humidity 95%+ with wind 1 mph or less, the low above freezing. The wind sensor was broken from about February 2019 to June 2020 (zero on 40 to 76% of daytime minutes), so those 476 nights are left out of everything below (2,221 nights kept). Airports: Marion or Huntington at 1 mile or less (fetched back to 2017 from the Iowa Environmental Mesonet).

  • The season: fog signs on 28% of July nights, 30% of August, 22% of September, 5% of October, 0 to 2% from November to June. The October camera checks calibrated the pattern in October only: in midsummer, heavy dew under saturated air may pass for fog. Summer camera checks (July to September 2027) are needed before trusting the summer numbers.
  • The station and the airports often disagree: on fog-sign nights the airports had fog 41% of the time (80 of 194), against 15% on all nights; but on 248 of the 328 nights the airports had fog, the station showed no fog signs. The mast sits on a local high point (R-004), so fog in the low ground and at the airports can leave the mast clear.
  • Calm in the evening is the strongest sign. Mean wind from 7 to 10 PM: 0.5 mph or less, fog signs on 32% of nights; 0.5 to 1 mph, 21%; 1 to 2, 11%; 2 to 3, 5%; over 3, 1% or less. (Partly built in, since the definition needs calm overnight.)
  • Evening dryness hardly matters, short of very dry air. In July to October the rate is 19 to 26% for any 7 PM gap between the air and the dew point from 0 to 20 F, and 8% above 20 F: the night's cooling closes a moderate gap.
  • Recent rain: a small lift (July to October: 19% after two dry days, 23 to 24% after rain).
  • The reservoir and pond (north drift): no sign of an effect. July to October by evening drift direction: southwest 34%, east and southeast 30%, west 29%, northeast 27%, north 26%, northwest 24%, south 22%.
  • Corn against soybean years: August to October fog signs 23% in corn years and 15% in soybean years, but the airports also had more fog in corn years (22% against 16%): mostly the years' weather, not the crop.
  • Year to year: August to November fog-sign nights ranged from 11 (2021) to 33 (2018) and track the airports' fog (2018: 33 against 34; 2021: 11 against 15).
  • Humidity cannot time the burn-off: after the fog lifts, evaporating dew keeps the humidity high (October 20: 96 to 98% after the fog cleared), so "humidity below 90%" comes about 3 hours after sunrise whatever the fog did. Burn-off times need the cameras or a visibility sensor.
  • What it points to: the forecast needs the night's wind (and cloud) more than the evening's dryness. Next exploration: the models' forecast night wind and cloud against the station's fog signs on the discovery days (March 2024 to September 2025, real forecast inputs), then a pre-registered test on fresh months.
  • Found on the way, the wind sensor: the station usually reads 70 to 80% of Marion's daytime wind (it is sheltered and at 12 ft). Since summer 2026 that fell to 62% (July), 57% (August), 61% (September) and 53% (October so far); the last corn year, 2024, stayed at 68 to 77%. Standing corn explains some; worn anemometer bearings are the usual cause on a Davis. Brian to check that the cups spin freely.

2026-10-20, the National Weather Service's dense fog advisories (data/research/nws_products_point.json: every watch, warning and advisory issued for the station's point since October 2017, 689 in all, from the Iowa Environmental Mesonet's archive; zone INZ024). 59 Dense Fog Advisories; by month: January 10, February 4, March 4, April 3, May 3, June 1, July 1, August 4, September 9, October 7, November 5, December 8. On the 40 advisory nights in the explored record (the broken wind sensor months left out): the station saturated (95%+, calm) for 2 hours or more on 82%, the airports reached 1 mile or less on 95%, but the full fog-sign pattern (7+ calm saturated hours) on only 32%.

  • Two kinds of fog: the forecasters' dense fog peaks in winter, the station's long calm saturation in late summer. Winter fog here often comes with some breeze or moist air moving in, so the October pattern (7 calm hours) is too narrow for it.
  • A second truth source, for free: the advisories mark regional dense fog that mattered (the county-scale decision); Brian's log marks fog at the field; the airports measure visibility. A fog detector should be judged against all three.
  • School delays (Brian's idea, October 20): a record of travel impact rather than of weather (a snow day can bring delays for days after, from drifting). There is no public archive; the district's own log or its social-media history is the likely source. If one turns up it joins the advisories as an "impact" record.

Ideas to chase

  1. A station fog detector, calibrated on Brian's log once there are 20 to 30 logged mornings with and without fog. Candidates: humidity and how long it stays high, calm, the gap between the grass and the dew point, and the cooling stalling.
  2. Evening signals: on the discovery days, with station saturation as the stand-in truth, find what separates saturated nights from the rest. Candidates:
  • the gap between air and dew point at sunset, and the first hour's cooling
  • the models' and the physics' closest approach to the dew point
  • forecast wind and cloud
  • recent rain and soil moisture
  • the crop

Then pre-register a test on fresh months, alternating them (R-011 rule 3).

  1. A commute product: the chance of fog from 6:00 to 8:30 AM, and the burn-off time. Only after 1 and 2 are tested.
  2. Brian's cameras as a fog detector: how sharp the far tree line looks over the night. That's new information the models do not have (R-011 rule 1), and it could backfill nights Brian was not out.
  3. Freezing nights: does the duration rule hold only above freezing (frost taking the moisture first below it)? Check on this winter's nights.
  4. The air profile probes being planned (5, 11 and 25 ft) would show when the fog layer forms and how deep it grows.
  5. The models' own visibility forecast: does it mean anything at this spot?

Log

  • 2026-10-20: thread opened at Brian's request ("a major area of study"); first look on the discovery days; the humidity sensor checked (no drift); the fog and frost log page built (fc/fog.php) and seeded with his sightings from October 18 to 20.
  • 2026-10-20: Brian's camera review added two light-fog mornings (October 4 and 5); the station had been saturated and calm for 7 to 11 hours before the fog showed on camera.
  • 2026-10-20: night footage of seven saturated October nights: light fog cannot be confirmed at night on this camera; logged as can't tell (a "Can't tell" choice added to the log page).
  • 2026-10-20: burn-off at 8:57 AM logged (about an hour after sunrise; dew evaporating raised the dew point 5.6 F).
  • 2026-10-20: a better camera resolved most October nights; first pattern: fog on all four firm fog nights with 7+ hours of calm saturation, none on the three with under 5.
  • 2026-10-20: the better camera (and two others on the same network switch) went offline on October 17, likely a failed switch, so October 17-18 and 19-20 cannot be checked on it. Once it is back, it is the best fog record we have for nights.
  • 2026-10-20: first test of the duration pattern missed: October 17-18 (7.8 calm saturated hours, at or below freezing) had no fog on a third camera; hypothesis: below freezing, frost takes the moisture first.
  • 2026-10-20: Brian described the land around the mast (R-004, "The site and its surroundings"): bottom ground 30 ft or more below the mast to the south, a 2-acre pond and a reservoir (3/4 mile north) upwind on a north drift. Worth testing on the logged fog nights: wind direction in the hours before fog, and whether fog shows on the cameras in the bottom before the mast.
  • 2026-10-20: eight years of the station's own record (Cumulus and Weather Display, one Davis) explored for fog signs; the season, the airports, evening calm as the main sign, dryness and the reservoir drift not mattering, burn-off not measurable from humidity; the wind sensor's 2019-2020 failure and its 2026 slide found.
  • 2026-10-20: the NWS's advisories for the point (2017 on, IEM) added as a second truth source; dense fog advisories peak in winter, only 32% met the October fog-sign pattern.
  • 2026-10-21: fog night October 20-21 logged by Brian (4-8 AM); version 1 kept cooling under the fog. L-033 (restart saturated when the station is) passed; L-034 live through the fog season.
C-005case study

The October 29 cold front: the high was nailed (78.9 vs 79.3 F); the night low was missed by 2 to 3 F, not the 9 F first written (corrected November 2: the night's forecasts had been read from the next date)

opened October 30, 2026 · by the lab assistant on the Tower
The question

A strong front crossed on the evening of October 29 (temperature down 11 F, dew point down 16 F, pressure up 4.6 mb). How did the forecast do for the day before it and for the night after it, and why was the night so far off?

The plan, written before the test

Read the station's readings through the front (obs), every source's forecast for the October 29 high and the October 30 low as they changed (issued table), the physics model's sunset run (phys_runs), and the neighbouring airports' temperature, dew point, wind and sky (ring_obs). Scripts in /lab/experiments/C-005/. One front, so everything here is a pointer, not evidence. The days looked at (Oct 29 and Oct 30) are now discovery days.

What counts as a pass, written before the test: Case study: no pass/fail. Findings become ideas for proper tests on other days.
Results

The day before: the high on October 29 was 79.3 F at 4 PM, and our last forecast was 78.9 F (live-adjusted 79.3 F). NWS said 78, WxSim 79.3, four-model average 76.5. A good day for us. The front arrived at the station about 5:30 PM: the temperature fell from 77.4 F at 5 PM to 65.5 F at 6 PM with 0.21 inch of rain, and the airports around us show the same drop within the hour (Marion 77.9 to 65.8 F, Huntington 73.9 to 64.4 F between 5 and 6 PM). The night after: the wind went to 0.0 mph by 7 PM and stayed near zero all night. The station fell from 65 F at 7 PM to 54 F at midnight, 46.6 F at 2 AM, 40.8 F at 6 AM, and the low was 39.6 F around 8:30 AM (dew point 38.7 F, so near saturated). Our forecast of the low was 48.5 F (+/- 2.7), too warm by 8.9 F. Others: NWS 46 (6.4 warm), WxSim 46.8 (7.2 warm), four-model average 49.5 (9.9 warm), the main model 50.0 (10.4 warm). Every source missed in the same direction, ours was in the middle of the pack. The one that came close was our own night-cooling physics: its sunset run at 7:05 PM said 43.4 F (skin 42.4 F), only 3.8 F warm, and was still falling when the night ended. The published number never moved toward it: ours stayed at 48.6 to 48.8 F through the evening and even rose to 49.9 F at midnight, when the station was already at 54 F and falling about 4 F an hour. The stated +/- of 2.7 F missed by far. The airports around us show the same night: calm winds from about 10 PM, temperature falling to meet the dew point, and very low cloud or fog (ceilings of 100 to 500 feet) at Marion, Anderson, Huntington and Muncie by 1 to 5 AM. Marion reached 37.2 F by 8 AM. The station itself stayed 1 to 3 F above its dew point all night. CORRECTION (2026-11-02, Claude on the Mac, from the issued table in the Tower's October 31 snapshot): the night's forecasts above were read from the wrong date. The figures 48.5 (ours), 46 (NWS), 46.8 (WxSim), 49.5 (four-model average) and 50.0 (main model) are the forecasts for the NEXT night, October 30 to 31, as issued on the morning of October 30. For the post-frontal night itself (date 2026-10-29) the forecasts of record were: ours 42.7 at 7:05 PM (42.1 at 8:05 PM, the live-adjusted 42.7), the NWS 42, WxSim 45.8, the four-model average 43.9, the main model 42.7. The station's low in the grading window (8 PM to 8 AM) was 40.6; it reached 39.6 at about 8:30 AM, after the window. So every source was 1.5 to 6 F warm, not 6 to 10, and ours was 2 F warm in the window (3.1 against the later 39.6 that the Scoreboard shows). The claim that the published number never moved toward the cooling is also withdrawn: it was October 30's forecast that stayed at 48.6 to 49.9 through that evening, correctly, since Friday night's low was 52.3. The physics sunset run at 43.4 stands as read (phys_runs, night 2026-10-29) and was still the closest to the morning's 39.6; its lead over the others is 1 to 2 F, not 5. The case remains a pointer, now a smaller one.

Final forecasts vs station, October 29 high and October 30 low (F)
TargetStationOurs (+/-)NWSWxSimFour-model avgMain modelPhysics at sunset
Oct 29 high 79.3 78.9 (2.7) 78 79.3 76.5 78.9 n/a
Oct 30 low 39.6 48.5 (2.7) 46 46.8 49.5 50.0 43.4
  • The low of 39.6 F is the station's lowest reading to about 9:30 AM on October 30; midnight was 54 F, so no earlier minimum hides in the night. It was still falling at 8:30 AM, so the true dawn low may be a hair lower.
  • Rule 3: these two days are now looked-at days. Any test of an idea from here needs other days.
  • The Oct 30 daytime high (forecast 66 F) is not scored here: the day was not over when this was written.
  • One front. The physics model coming closer on one night proves nothing; L-007 and L-026 are the real tests of it.
What it means

The daytime half was easy: a strong front arriving late in the afternoon did not hurt the high because it arrived after the high. The night was the repeat of C-004's calm clear night: the models, NWS and WxSim included, kept the night warm (45 to 50 F) while a dry air mass, a calm wind and clear sky let the station drop 25 F in 14 hours. What stands out is not that the models missed but that our own physics model had most of it at sunset (43.4 F) while the published low sat at 48.7 F, and the evening live adjustment did not move toward the falling temperature.

Next

Ideas for tests on OTHER nights (backlog): (1) post-frontal calm clear nights: after a cold front with the wind below 3 mph and the sky clearing by dusk, does the published low run warm against the physics sunset run? Check how the physics run and the blend perform on every such night in L-007, L-026 and L-028 (and the discovery-year nights in R-010) before opening a new test. (2) Does the evening live adjustment ever move up while the station is falling steeply (it did at midnight, 48.7 to 49.9 F)? Look at that in the replay logs. (3) Front nights at night: widen the +/- when a front passes within the previous 12 hours and the models disagree with the dew point behind it. Fog at the airports is a thread for R-012.

L-045pre-registered (fresh builds from the Tower install that carries it to April 30, 2027; evaluate in May 2027)

The hour-by-hour curve shaped by the station's own record: our curve minus its typical miss by regime and hours since sunrise or sunset (research R-008)

opened October 26, 2026 · by the Claude with Brian (session of October 26, 2026)
The question

Over the 12 hours after each forecast of record, is the Hour by hour page's temperature curve closer to the station when the station's own mean residual for the hour's regime and the hours since the half's sunrise or sunset is taken out of it, with the high and low left exactly where they are?

Why it might work

Brian (October 26, 2026): the highs and lows are close now, but the hour-by-hour curve does not fit what the station then does, above all the plunge after sunset, the overnight stirs and the morning rise. The R-008 scorecard of October 26, 2026 (research/r008_scorecard.py) measured the page's curve against the station three ways: every hourly build the Tower made from October 14 to 26, 2026, rebuilt with the page's own code (401 builds); the two forecasts of record a day rebuilt from the model archives, March 2024 to October 2026 (1,835 issues); and the main model's replay on the real forecast runs, April 2 to October 13, 2026 (373 issues). The misses sit at fixed places in the day: on clear calm nights in September and October the curve is 4.5 F too warm 1 to 2 hours after sunset and still 2.7 F too warm at 3 to 5 hours (it misses the plunge); in the live October record it is 1.6 to 2.0 F too warm from 2 to 5 hours after sunset and 1.3 F too cold 5 to 8 hours after (the stirs); on cloudy and other nights from November to February it runs 0.8 to 1.7 F too cold from 5 hours after sunset on; on cloudy days it runs 0.7 F too warm from 3 to 8 hours after sunrise (1.3 to 1.5 F in spring); in the live October record it is 3.7 F too warm in the hour after sunrise. The change asked for first, a curve shaped by the main model's physics run and pinned to our high and low, was sized on the same hours before any freeze (R-011 rule 8) and rejected: on the replay it is 0.58 F per hour farther from the station over hours 1 to 12 (95 percent interval 0.47 to 0.68 F, 186 dates) and farther in every regime, because the night physics cools 4 to 5 F too slowly for the first three hours after sunset and the day physics warms too slowly through the morning; on the live October record it is 1.44 F per hour farther. That shape waits for the dusk-lag work in R-010. The candidate here uses the station's own record instead (R-011 rule 1: new information), aimed at timing (rule 7): a table of the curve's mean residual by regime (the models' mean cloud and wind over the half) and whole hours since the half's sunrise or sunset, 87 cells, each shrunk toward zero by n over n plus 30. Fitted on March 2024 to September 2025 and sized on October 2025 to October 2026 (the test year): 0.031 F per hour closer over hours 1 to 12 (95 percent interval 0.015 to 0.047 F closer, 360 dates), closer in every lead group, and on clear calm nights 0.11 F per hour closer (interval 0.27 F closer to 0.03 F farther, 46 nights) with the bias 1 to 2 hours after sunset cut from +2.0 to +0.4 F and at 2 to 3 hours from +1.8 to +0.2 F. A small gain, but resolved, and the first that moves the shape where Brian sees it miss.

The plan, written before the test

The candidate (hyperlocal/hourly_shape.py, standard library) runs in every hourly build right after the page's curve is built (insight.build_detail) and is shown beside it on the Hour by hour page as a dashed line and a table column marked on trial, never in its place: the page's current curve (the four models' hourly blend moved by our high and low correction and the 1 to 6 hour nowcast, L-013) minus the frozen table's residual (hourly_shape_frozen.json, fitted on March 1, 2024 to October 11, 2026, the model archives at lead 1 against the station) for the hour's regime and whole hours since the half's sun event. The regime of each half (sunrise to sunset, sunset to sunrise) comes from the four core models' mean total cloud and 10 m wind over the half as downloaded at the build, L-038's definitions: a night is clear and calm under 30 percent cloud and 5 mph, cloudy at 70 percent or more, otherwise other; days on cloud alone. The correction ramps to zero over the two hours before the current curve's own extreme inside the half's scoreboard window (the high's window 8 AM to 8 PM, the low's 8 PM to 9 AM) and is zero from that hour on, and inside the window the curve is clamped so it never goes below the night's low or above the day's high: the high and low the site shows are the issued values and the candidate issues none. Both curves are logged at every build for hours 1 to 24 in the table hourly_issued (sources ours_hourly and L-045, with the regime) and graded by l045_score.py: the forecasts of record (the 11:05 and 23:05 UTC builds), hours 1 to 12, against the station's readings within 5 minutes of the hour, from the first record after the Tower install that carries the candidate through April 30, 2027; the hourly updates are reported beside them and are not part of the verdict. The builds of October 14 to 26, 2026 are the exploration set and are never scored. Frozen by the hashes of hourly_shape.py, hourly_shape_frozen.json and l045_score.py; the table is never refit during the trial.

What counts as a pass, written before the test: Evaluated in May 2027 with l045_score.py (--verify). No verdict below 150 records; then the window runs to May 31, 2027 and is evaluated whatever the count, saying so. The measure is, per local date, the mean over that date's records and their 12 hours of the candidate's absolute error minus the current curve's (negative: the candidate is closer); intervals are bootstraps over dates. Pass (the station-shaped curve replaces the current one on the page, with the table refit on the trial months added) only if all three hold: (1) the per-date mean over hours 1 to 12 is below zero at one-sided 2.5 percent; (2) no lead group (hours 1 to 3, 4 to 6, 7 to 12) is worse than 0.05 F per hour at the upper end of its 95 percent interval; (3) the high and low are unchanged: the candidate issues no high or low, and on every record its extreme inside each scoreboard window its logged hours cover equals the current curve's within 0.05 F (the scorer counts the windows that fail; the count must be zero). Otherwise fail: the candidate leaves the page and the table goes to R-008 as a description of the miss. Reported without a bar: the gain on clear calm nights, the bias by hours since sunset and sunrise, and the same measures over every hourly build. Power (R-011 rule 2): on the test year the per-date gain is 0.031 F per hour with a per-date spread of 0.159 F, so about 207 records give 80 percent power at one-sided 2.5 percent; November to April provides about 360. The clear-calm-night gain alone (0.11 F per hour, per-night spread 0.51 F) would need about 160 such nights, more than a winter holds, which is why it is reported and not part of the verdict.
Next

Starts with the Tower install that carries hourly_shape.py (the bundle of October 26, 2026). If it passes, the station-shaped curve becomes the page's curve and the 3-hour strip on the Today page follows it; the next shape question is the physics' own dusk lag (R-010), which the sizing above measured at 4 to 5 F for the first three hours after sunset on clear calm nights.

L-044pre-registered (nights and days from the first forecast of record after the next Tower install to April 30, 2027; evaluate in May 2027)

The main model on every ECMWF ensemble member: are its odds honest? (the questions of L-035, L-036 and L-037 carried forward)

opened October 22, 2026 · by the Claude, with Brian, in a session
The question

When the main model (L-038, the current version with its restart rule, no treatment) is run once on each of the ECMWF ensemble's members through its two days, and each member's departure from the members' middle is laid on the corrected forecast, are the odds honest: does the station's low land inside the 8-in-10 range about 8 nights in 10; is the chance of frost on the grass better than knowing how often frost happens; is the spread wider on the nights the middle run misses by more; and by day, does the physics' spread catch the station's high better than the members' own 2 m highs do, both centred on their forecasts?

Why it might work

The odds on the Physics page came from version 1 and version 1R on the members from the sunset start (L-035, L-036) and from the day model on the members (L-037). Under the main-model approach the odds should come from the main model itself, so that every change to it reaches the odds, and so that one set of code is graded. The old three entries are closed with their questions carried here (the one-time closing pass, APPROACH-2026-10-21.md section 4a); their code leaves the live physics once this entry has issued its first run on the Tower. Version 1R's question (the air's own radiative cooling) becomes a sizing candidate for a treatment, not a second ensemble.

The plan, written before the test

Frozen before any graded night by the hashes of its code and scripts (column_ensemble.py, l044_score.py, column.py, column_live.py, physlive.py, hlcore.py). Every hourly build runs column_ensemble.build: for each member the two downloads are re-based onto the ensemble's time axis with the member's own values wherever the ensemble carries that field (cloud by layer, 10 m wind, 2 m air and dew point, the 925, 850 and 700 mb air and heights, sunshine; precipitation, soil moisture and snow stay the deterministic downloads', as the single run has them), through column.segments_for and column.run with the same station start, soil, pressure and the reset to the station's latest reading as the main model. Recorded per period at leads 1 and 2: the members' physics values, p10, p50 and p90 of those values re-centred on the corrected main forecast (centre plus each member's departure from the members' median), the raw p10, p50, p90, the share of members whose grass surface (the night half's lowest skin from the issue on) reaches 32 F and whose air reaches 32 F, and for a day the members' own 2 m highs in the window and the chance of reaching each 5 F mark; and a fan of the air at every third step. At the two forecasts of record (11:05 and 23:05 UTC) the record is logged in phys_runs (kind 'main_ens'). Truths: the station's low and high from the one truth table the Scoreboard uses (column_live.station_truth); the grass probe's lowest reading in the night window (frost if 32 F or colder). Nights and days: every one from the first forecast of record after the Tower install that carries this entry through April 30, 2027; degraded periods (the physics could not run) are left out and counted. Scored by l044_score.py on the Tower in May 2027.

What counts as a pass, written before the test: Evaluated in May 2027; no verdict below 100 graded nights (lead 1) for the night items, 100 graded days (lead 1) for the day item; then the window runs to May 31, 2027 and is evaluated whatever the count, saying so. The night odds pass only if all hold at lead 1: (1) the station's low inside the re-centred p10 to p90 on 70% to 90% of nights; (2) the Brier skill of the grass-frost chance against the trial's own frost frequency above zero in at least 95% of 10,000 bootstrap resamples by night; (3) the spread (p90 minus p10) larger on the nights the middle run missed by more than its median miss than on the rest, in at least 95% of 10,000 resamples. The day odds pass only if, at lead 1, the per-day CRPS of the re-centred physics members is lower than the per-day CRPS of the members' own 2 m highs re-centred on our forecast in at least 95% of 10,000 resamples by day. Night and day are verdicts of their own: a pass puts that side's odds on the Today page with a link here; a fail keeps that side's odds on the Physics page as being graded and says what failed. Reported, no bar: lead 2 the same way, the freeze chance, reliability by tenths of the frost chance, the calm-member bias (the calmest third against the rest), and the fog odds once L-043 is in the main model.
Next

Starts with the next Tower install. Once its first run on the Tower is confirmed, the old ensemble code (physlive.ensemble, r_summary, day_ensemble and their parameter file phys_params_r.json) leaves the live physics with the next install, and the Run charts draw the fan from this entry's record.

L-043pre-registered (treatment of the main model, with L-039, L-040, L-041 and L-042 on the same nights from its first issue after the next Tower install to April 30, 2027; evaluate in May 2027)

Fifth treatment of the main model: version 3's moisture budget inside the main model's night (from L-012 and L-034)

opened October 22, 2026 · by the Claude, with Brian, in a session
The question

On nights the models forecast cloudy, does the main model's night physics miss the station's low by less when the air's humidity is carried by a moisture budget (dew water on the grass, evaporation from it and from the soil, a local departure from the models' moisture fading over a few hours, fog's latent heat and its own longwave) instead of the drift-limited humidity, with nothing else changed? And, reported as a secondary outcome, does its saturation rank the fog nights in Brian's log better than the models' own humidity forecast?

Why it might work

Version 3's moisture budget (phys_params_v3.json, fitted on discovery nights) was the fog version of the night physics, logged beside the sunset run for L-012 and used by the saturated restart L-033 and L-034. Under the main-model approach it belongs in the main model as a single change or not at all. Sized against the main model v0.1 on the replay (R-011 rule 8, April 2 to October 13, 2026; physmodel/size_candidates.py with treatments/cand_budget.py, now treatments/L043.py): forecast-cloudy nights 0.11 F better a night (95 percent interval 0.07 to 0.14 F; nights with mean total cloud of 70 percent or more: 0.13 F, spread 0.20 F), cloudy rainy nights 0.17 F, cloudy dry nights 0.09 F, clear calm nights no change (-0.02 F, interval across zero), the calibrated low unchanged (1.688 to 1.686 F with the months held out), the physics weight of the low 0.318 to 0.323. A small gain, but a well-resolved one, and the physics the fog work needs: saturation, dew water and fog are states of the run rather than words read off the humidity.

The plan, written before the test

Frozen before anything is scored by its replay's output hash (physmodel/l043_replay.csv, main_replay.py --treatment L-043), with the code it ran (column.py, column_live.py, physlive.py, hlcore.py, treatments/__init__.py, treatments/L043.py, phys_params_v3.json and the main model's parameter files) and the scripts (l039_score.py, treatments.json, calibration_L-043.json, main_replay.py, main_calib.py) recorded beside it. The treatment is the main model (L-038, v0.1) with one change, in its own module (treatments/L043.py) through the hooks surface_moisture, cloud_extra, after_air and on_restart: at the surface, dew deposition and evaporation with a store of dew water on the grass and a share of evaporation from the soil set by the topsoil moisture (e0, e1); in the air, the humidity is the models' 2 m moisture plus a local departure fed by that evaporation and deposition over the layer depth z_q and fading over tau_q hours; when the air is pushed past saturation the excess condenses as fog, warming the air by half its latent heat, and while fog is present k_fog is added to the cloud the sky sends back. The constants are version 3's (phys_params_v3.json, fitted on discovery nights 2024 and 2025); every temperature constant is the main model's; nothing is refitted. It has its own calibration, fitted by the main model's procedure on its own replay (calibration_L-043.json) and refit monthly the same way, and is issued every hour beside the main model as source L-043 by the same build. Nights: the family's, from the first night on which L-039 to L-043 were all issued (the Tower install that carries L-043) through the night of April 30, 2027; never the replay nights. Pairs, forecast-cloudy nights (mean total cloud for the night at issue of 70 percent or more), truth and intervals exactly as the rest of the family, one scorer (l039_score.py, Holm over five). Secondary outcome, reported and not part of the verdict: on the nights in Brian's fog log (fc/fog.php, the truth set of R-012), the ranking skill (AUC) of this treatment's saturation (its fog state at the run's end, and the hours its air sat at saturation) against the models' own minimum dew-point depression for the night, as L-012 planned, valid only with at least 20 logged fog nights.

What counts as a pass, written before the test: Evaluated in May 2027 with l039_score.py (--verify), the same bar form as L-039 to L-042. No verdict below 40 forecast-cloudy nights with both physics runs; then the window runs to May 31, 2027 and is evaluated whatever the count, saying so. Pass (merge into the main model, in order of size within the family) only if all hold: (1) on forecast-cloudy nights, the physics alone misses the low by less with the change: the per-night difference (treatment minus main) is below zero at one-sided 2.5 percent after the Holm adjustment over the family (L-039 to L-043); (2) the calibrated low, every night, is not worse with the change: the upper end of the 95 percent interval of the per-night difference is under +0.10 F; (3) the physics share of the low in the treatment's own monthly refit at evaluation is at least the main model's. Otherwise fail: the change is deleted and its lesson written here and in R-011. Power (R-011 rule 2): item 1, by resampling the replay's 90 forecast-cloudy nights (per-night mean -0.13 F, spread 0.20 F), passes the strictest Holm step over five about 99 percent of the time with 40 nights and essentially always with 70. Item 2: the calibrated difference is -0.002 F a night on the replay; the upper end of the interval should sit near zero, under the 0.10 F margin. Item 3: 0.323 against 0.318 in the replay. The secondary fog outcome has no bar: it is reported with its sample size, and a ranking gain of 0.02 or more (L-012's bar) on at least 20 logged fog nights is the sign that a fog-odds entry should follow.
Next

Starts with the next Tower install. If it passes, the moisture budget goes into the main model and the saturated start (L-034's question) is sized against it on the winter's logged restarts.

L-042pre-registered (treatment of the main model, with L-039, L-040 and L-041 on the same nights from its first issue after the next Tower install to April 30, 2027; evaluate in May 2027)

Fourth treatment of the main model: rain cooling from the physics, with no fitted constant (R-011 miss M-1)

opened October 21, 2026 · by the Claude, with Brian, in a session
The question

On nights the models forecast cloudy, does the main model's night physics miss the station's low by less when the forecast rain falling through the night layer evaporates into it, cooling it by the latent heat of that rain spread over the layer and never past its wet bulb, with nothing else changed?

Why it might work

R-011 miss M-1: the physics runs warm in proportion to the forecast cloud, most of all on rainy nights (+4.5 F on cloudy rainy nights). The main model does not model rain on nights without snow on the ground, because the wet-weather night version failed L-016. A first rain cooling candidate used that version's fitted constants (k_ev, e_p) and was withdrawn on Brian's caveat (October 22, 2026). Re-derived from the physics with no fitted constant (treatments/L042.py) and sized against the main model on the replay (R-011 rule 8, April 2 to October 12, 2026): forecast-cloudy nights 0.87 F better a night (95% interval 0.53 to 1.22 F, spread 1.68 F, 63% of nights better), the largest candidate so far; cloudy rainy nights bias +4.5 to +2.0 F and error down 1.68 F; cloudy dry nights error down 0.42 F (their forecast rain is under 0.10 in for the night but not zero); partly cloudy rainy nights bias +2.0 to +0.1 F; clear nights no change (+0.03 F); the calibrated low +0.023 F a night; the physics share of the low 0.305 to 0.334 and of the high 0.134 to 0.188.

The plan, written before the test

Frozen before anything is scored by its replay's output hash (physmodel/l042_replay.csv, main_replay.py --treatment L-042), with the code it ran (column.py, column_live.py, physlive.py, hlcore.py, treatments/__init__.py, treatments/L042.py) and the scripts (l039_score.py, treatments.json, calibration_L-042.json, main_replay.py, main_calib.py) recorded beside it. The treatment is the main model (L-038, v0) with one change, in its own module (treatments/L042.py): on nights without snow on the ground (snow nights already carry the wet-weather version tested in L-017), at each 5-minute step with forecast rain, the air of the night layer is cooled by the latent heat of that step's rain, rain x Lv x dt / (rho cp z), and never by more than its distance from its wet bulb; the moisture the rain gives the air is the matching cp / Lv per degree. The layer depth z is the main model's own; no constant is fitted or refitted. It has its own calibration, fitted by the main model's procedure on its own replay (calibration_L-042.json) and refit monthly the same way, and is issued every hour beside the main model as source L-042 by the same build. Nights: the family's, from the first night on which L-039, L-040, L-041 and L-042 were all issued (the next Tower install) through the night of April 30, 2027; never the replay nights. Pairs, forecast-cloudy nights (mean total cloud for the night at issue of 70% or more), truth and intervals exactly as the rest of the family, one scorer (l039_score.py). Family: L-039, L-040, L-041 and L-042 test changes for the same miss on the same nights, item 1 Holm-adjusted over the four (one-sided 2.5%: the smallest p against 0.00625, then 0.0083, 0.0125, 0.025). At the verdicts the passes are merged in order of size; each later one is re-run as a diff from the new main model on the window's nights from the stored downloads and goes in only if its item 1 still holds there. A code change that alters the replay output is allowed only as a declared amendment (both output hashes, the replay nights it changes, the first live night it issues), and the evaluation reports the nights on each side of it separately. Holm power for the family of four (recomputed October 22, 2026, by resampling the replay's 90 forecast-cloudy nights jointly, so the correlation between the four, 0.59 to 0.81, is kept): with 40 nights L-039 96%, L-040 95%, L-041 97%, L-042 90%; with 70 nights 99 to 100% each; with 90, 100%. The evaporated fraction: each step evaporates the share of that step's forecast rain that the layer can take, all of it at most, so the cooling is limited at every step by the rain available and by the wet bulb; in the replay all of a step's rain evaporated on 24% of rainy night steps and the wet bulb limited the rest, so the evaporated fraction averaged 0.32 a step and 0.07 of the rain by volume (once rain begins the layer stays within a few hundredths of a degree of its wet bulb). Amendment of October 22, 2026, before anything was graded: the main model now starts the half in progress from the station's latest reading (L-038, amendment 12), so this treatment's replay and calibration were re-run on that main model and recorded as a declared output change. Against the v0.1 main model, on forecast-cloudy nights: -0.822 F a night; weight on the physics for the low 0.353 (the main model's 0.318). Holm power over the four with 40 forecast-cloudy nights: L-039 95%, L-040 93%, L-041 96%, L-042 87% (this one 87%); with 70 nights 98 to 100%. Amended 2026-10-22: L-043 (the version 3 moisture budget inside the main model's night, sized on the replay, R-011) joins the family; the family is L-039, L-040, L-041, L-042 and L-043, scored by one scorer (l039_score.py) with the Holm adjustment over five, and the family's nights run from the first night on which all five were issued (the Tower install that carries L-043).

What counts as a pass, written before the test: Evaluated in May 2027 with l039_score.py (--verify), the same bar form as L-039, L-040 and L-041. No verdict below 40 forecast-cloudy nights with both physics runs; then the window runs to May 31, 2027 and is evaluated whatever the count, saying so. Pass (merge into the main model, in order of size within the family) only if all hold: (1) on forecast-cloudy nights, the physics alone misses the low by less with the change: the per-night difference (treatment minus main) is below zero at one-sided 2.5% after the Holm adjustment over the family (L-039, L-040, L-041, L-042); (2) the calibrated low, every night, is not worse with the change: the upper end of the 95% interval of the per-night difference is under +0.10 F; (3) the physics share of the low in the treatment's own monthly refit at evaluation is at least the main model's. Otherwise fail: the change is deleted and its lesson written here and in R-011. Power (R-011 rule 2): item 1, by resampling the replay's nights jointly with the rest of the family, passes about 90% of the time with 40 forecast-cloudy nights and 99% with 70 (its per-night spread, 1.68 F, is the family's widest, because it acts mostly on the rainiest nights). Item 2: the calibrated difference is +0.023 F a night; over about 170 nights the upper end of the interval should sit near +0.05 F, under the 0.10 F margin. Item 3: 0.334 against 0.305 in the replay. Reported, not part of the verdict: cloudy dry and cloudy rainy nights separately, clear and partly cloudy nights, the calibrated high and low against ours and the main model, the share of the high, and any declared output change with the nights on each side. Amended 2026-10-22: the Holm adjustment in item 1 is over the family of five (L-039 to L-043).
Next

Starts with the next Tower install. Nothing else is frozen until it is issuing (Brian, October 22, 2026).

L-041running (treatment of the main model, with L-039, L-040 and L-042 on the same nights from their first issue after the next Tower install to April 30, 2027; evaluate in May 2027)

Third treatment of the main model: the reference air comes down from the models' cloud base when it is below 925 mb (R-011 miss M-1)

opened October 21, 2026 · by the Claude, with Brian, in a session
The question

On nights the models forecast cloudy, does the main model's night physics miss the station's low by less when, in the share of the column the models cloud (their low cloud fraction) and wherever their cloud base is below 925 mb, the reference air the night air mixes toward is the air at the cloud base brought to the surface, instead of the 925 mb air brought down from above the cloud, with nothing else changed?

Why it might work

R-011 miss M-1, the part L-039 and L-040 leave: after L-040 the physics is still about 3 F warm on cloudy rainy nights. In version 1 the reference air is the 925 mb air brought down dry-adiabatically from about 750 m; under a cloud deck whose base is lower, the air the surface layer exchanges with is the sub-cloud air, and the 925 mb air is inside or above the cloud. Sizing (R-011 rule 8; single parameters against the main model on the replay, April 2 to October 12, 2026): per forecast-cloudy night 0.53 F better (95% interval 0.36 to 0.70 F, spread 0.84 F, 72% of nights better); cloudy rainy nights bias +4.5 to +3.3 F, error down 0.96 F; cloudy dry nights error down 0.36 F; the calibrated low unchanged (+0.001 F a night); the physics share of the low 0.305 to 0.331 and of the high 0.134 to 0.154. On top of L-040 it adds 0.35 F a night. It tied the first rain cooling candidate within the replay's resolution; Brian broke the tie on physics and caveats (October 22, 2026): the cloud-base depth adds no constant, while that rain cooling used the constants fitted for the failed wet version (L-016). Rain cooling re-derived with no fitted constant (treatments/rain_cooling.py; sizing only) gains 0.87 F a night on the same nights and comes back later as a diff from whatever the main model is then. Why it moves forecast-clear nights by about 0.2 F (Brian asked before freezing; physmodel/diag_l041_clear.py): only because a night forecast clear (mean total cloud under 30%) can still carry some low cloud in the models. On the 8 clear nights with no low cloud at any step the change is exactly 0.000 F; on the 43 with some (mean 7.7%, up to 59% at some steps) it grows with the low cloud: -0.06 F below 5%, -0.45 F at 5 to 15%, -0.54 F above 15%. The cloud base is below 925 mb on 75% of clear-night steps (the night air near its dew point), so the low cloud weight is what limits the change, as designed; clear nights' error does not change (+0.00 F).

The plan, written before the test

Frozen before anything is scored by its replay's output hash (physmodel/l041_replay.csv, main_replay.py --treatment L-041), with the code it ran (column.py, column_live.py, physlive.py, hlcore.py, treatments/__init__.py, treatments/L041.py) and the scripts (l039_score.py, treatments.json, calibration_L-041.json, main_replay.py, main_calib.py) recorded beside it. The treatment is the main model (L-038, v0) with one change, in its own module (treatments/L041.py): at each night step where the models' cloud base, the lifting condensation level of their own 2 m air and dew point (125 m per degree C of spread), is below the 925 mb level, the reference air becomes, in the share of the column given by the models' low cloud fraction, the air at the cloud base (read off the models' profile between their 2 m air and their 925 mb air) brought to 2 m along the dry adiabat; elsewhere, and with no low cloud, nothing changes. No constant is fitted or refitted. It has its own calibration, fitted by the main model's procedure on its own replay (calibration_L-041.json) and refit monthly the same way, and is issued every hour beside the main model as source L-041 by the same build. Nights: the family's, from the first night on which L-039, L-040 and L-041 were all issued (the next Tower install) through the night of April 30, 2027; never the replay nights. Pairs, forecast-cloudy nights (mean total cloud for the night at issue of 70% or more), truth and intervals exactly as L-039 and L-040, one scorer (l039_score.py). Family: L-039, L-040 and L-041 test changes for the same miss on the same nights, item 1 Holm-adjusted over the three (one-sided 2.5%: the smallest p against 0.0083, the next against 0.0125, the last against 0.025). At the verdicts the passes are merged in order of size (the larger gain in item 1 first); each later one is re-run as a diff from the new main model on the window's nights from the stored downloads and goes in only if its item 1 still holds there. Family amended October 22, 2026, before anything was graded: L-042 (rain cooling from the physics, no fitted constant) joins, so the family is L-039, L-040, L-041 and L-042, scored from the first night on which all four were issued, with item 1 Holm-adjusted over the four (one-sided 2.5%: the smallest p against 0.00625, then 0.0083, 0.0125, 0.025); the merge rule covers all four. Holm power for the family of four (recomputed October 22, 2026, by resampling the replay's 90 forecast-cloudy nights jointly, so the correlation between the four, 0.59 to 0.81, is kept): with 40 nights L-039 96%, L-040 95%, L-041 97%, L-042 90%; with 70 nights 99 to 100% each; with 90, 100%. A code change that alters the replay output is allowed only as a declared amendment (lab_tools freeze --output-amend, recording both output hashes, the replay nights it changes and the first live night it issues), and the evaluation reports the nights on each side of it separately; an undeclared change is refused. Amendment of October 22, 2026, before anything was graded: the main model now starts the half in progress from the station's latest reading (L-038, amendment 12), so this treatment's replay and calibration were re-run on that main model and recorded as a declared output change. Against the v0.1 main model, on forecast-cloudy nights: -0.496 F a night; weight on the physics for the low 0.343 (the main model's 0.318). Holm power over the four with 40 forecast-cloudy nights: L-039 95%, L-040 93%, L-041 96%, L-042 87% (this one 96%); with 70 nights 98 to 100%. Amended 2026-10-22: L-043 (the version 3 moisture budget inside the main model's night, sized on the replay, R-011) joins the family; the family is L-039, L-040, L-041, L-042 and L-043, scored by one scorer (l039_score.py) with the Holm adjustment over five, and the family's nights run from the first night on which all five were issued (the Tower install that carries L-043).

What counts as a pass, written before the test: Evaluated in May 2027 with l039_score.py (--verify), the same bar form as L-039 and L-040. No verdict below 40 forecast-cloudy nights with both physics runs; then the window runs to May 31, 2027 and is evaluated whatever the count, saying so. Pass (merge into the main model, in order of size within the family) only if all hold: (1) on forecast-cloudy nights, the physics alone misses the low by less with the change: the per-night difference (treatment minus main) is below zero at one-sided 2.5% after the Holm adjustment over the family (L-039, L-040, L-041, L-042); (2) the calibrated low, every night, is not worse with the change: the upper end of the 95% interval of the per-night difference is under +0.10 F; (3) the physics share of the low in the treatment's own monthly refit at evaluation is at least the main model's. Otherwise fail: the change is deleted and its lesson written here and in R-011. Power (R-011 rule 2), from the sizing: item 1's per-night spread is 0.84 F; with 40 forecast-cloudy nights its standard error is 0.13 F and the replay's 0.53 F gain passes the family-of-four Holm test about 97% of the time; with 90 nights, nearly always. Item 2: the calibrated difference is +0.001 F a night with a spread of 0.16 F; over about 170 nights the upper end of the interval sits near +0.03 F. Item 3: 0.331 against 0.305 in the replay. Reported, not part of the verdict: cloudy dry and cloudy rainy nights separately, clear and partly cloudy nights (with their low cloud), the calibrated high and low against ours and the main model, and the share of the high. Amended 2026-10-22: the Holm adjustment in item 1 is over the family of five (L-039 to L-043).
Next

Starts with the next Tower install, together with L-039 and L-040. Rain cooling (re-derived, no fitted constant) returns as a diff from the main model after the May merges.

L-040running (treatment of the main model, with L-039, L-041 and L-042 on the same nights from their first issue after the next Tower install to April 30, 2027; evaluate in May 2027)

Second treatment of the main model: the 925 mb air comes down the moist adiabat where the models saturate the low column (R-011 miss M-1)

opened October 21, 2026 · by the Claude, with Brian, in a session
The question

On nights the models forecast cloudy, does the main model's night physics miss the station's low by less when the 925 mb air it relaxes toward is brought to the surface along the moist adiabat in the share of the low column the models saturate (their low cloud fraction), and along the dry adiabat elsewhere, with nothing else changed?

Why it might work

R-011 miss M-1 (the physics runs warm in proportion to the forecast cloud). The first treatment, L-039 (the sky term), explains about a tenth of it. In version 1 the 925 mb air is brought to 2 m along the dry adiabat (9.8 C per km); in a cloudy, saturated column the lapse is moist (about 5 to 6 C per km), so over the 600 to 800 m below 925 mb the reference air is about 2 to 3 C too warm and the night air is pulled toward it. Sized on the replay before freezing (Brian's direction, October 21, 2026; physmodel/main_replay.py --change, l040_replay.csv): the physics alone's warm bias falls from +2.6 to +1.8 F on cloudy dry nights and from +4.5 to +3.1 F on cloudy rainy nights (error down 0.44 and 1.07 F); forecast-cloudy nights improve by 0.63 F a night on average (spread 1.02 F, 77% of nights better), 3.6 times L-039; clear nights do not move (-0.02 F); the calibrated low is unchanged within 0.013 F; the physics share of the low rises from 0.305 to 0.342 and of the high from 0.134 to 0.184. With L-039 the two nearly add (0.75 F a night together). Whole-degree on cloudy rainy nights, about half a degree on cloudy dry nights.

The plan, written before the test

Frozen before anything is scored (code: column.py night_step with ref = moist, treatments.json, calibration_L-040.json, l039_score.py, physmodel/l040_replay.csv, main_replay.py, main_calib.py). The treatment is the main model (L-038, v0) with one change: the weight on the moist adiabat when bringing the 925 mb air to the surface (the reference the night air mixes toward) becomes the larger of the main model's rain weight (snow nights only) and the models' low cloud fraction at that step; everything else, every constant (none refitted) and the day are the main model's. It has its own calibration, fitted by the main model's procedure on its own replay (calibration_L-040.json) and refit monthly the same way, and is issued every hour beside the main model as source L-040 by the same build. Nights: the same as L-039's, from the first night on which both were issued (the next Tower install) through the night of April 30, 2027; never the replay nights (April 2 to October 12, 2026). Pairs, forecast-cloudy nights, truth and intervals exactly as L-039. Family: L-039 and L-040 test two changes for the same miss (R-011 M-1) and are scored on the same nights, from the first night on which both were issued, by one scorer (l039_score.py), with item 1 adjusted for the family (Holm, one-sided 2.5%: the smaller p against 0.0125, the other against 0.025). At the verdicts the passes are merged in order of size (the larger gain in item 1 first); the second is re-run as a diff from the new main model on the window's nights from the stored downloads and goes in only if its item 1 still holds there. (Amendment) Code change before anything was graded (October 21, 2026): column.py gained two inert parameters for sizing the next candidates for miss M-1 (rain cooling on every night; the reference air brought down from the models' cloud base when it is below 925 mb), and the night forcing now carries the models' 2 m air and dew point for them. With neither set, the main model's replay and the replays of L-039 and L-040 are byte for byte the same, and parity with the tested functions is 0.0 K. Family amended October 22, 2026, before anything was graded: L-041 (the reference air from the models' cloud base) joins, so the family is L-039, L-040 and L-041, scored from the first night on which all three were issued, with item 1 Holm-adjusted over the three (one-sided 2.5%: the smallest p against 0.0083, the next against 0.0125, the last against 0.025); the merge rule covers all three (in order of size, each later one re-run as a diff from the new main model before it goes in). Family amended October 22, 2026, before anything was graded: L-042 (rain cooling from the physics, no fitted constant) joins, so the family is L-039, L-040, L-041 and L-042, scored from the first night on which all four were issued, with item 1 Holm-adjusted over the four (one-sided 2.5%: the smallest p against 0.00625, then 0.0083, 0.0125, 0.025); the merge rule covers all four. Holm power for the family of four (recomputed October 22, 2026, by resampling the replay's 90 forecast-cloudy nights jointly, so the correlation between the four, 0.59 to 0.81, is kept): with 40 nights L-039 96%, L-040 95%, L-041 97%, L-042 90%; with 70 nights 99 to 100% each; with 90, 100%. A code change that alters the replay output is allowed only as a declared amendment (lab_tools freeze --output-amend, recording both output hashes, the replay nights it changes and the first live night it issues), and the evaluation reports the nights on each side of it separately; an undeclared change is refused. Amendment of October 22, 2026, before anything was graded: the main model now starts the half in progress from the station's latest reading (L-038, amendment 12), so this treatment's replay and calibration were re-run on that main model and recorded as a declared output change. Against the v0.1 main model, on forecast-cloudy nights: -0.591 F a night; weight on the physics for the low 0.359 (the main model's 0.318). Holm power over the four with 40 forecast-cloudy nights: L-039 95%, L-040 93%, L-041 96%, L-042 87% (this one 93%); with 70 nights 98 to 100%. Amended 2026-10-22: L-043 (the version 3 moisture budget inside the main model's night, sized on the replay, R-011) joins the family; the family is L-039, L-040, L-041, L-042 and L-043, scored by one scorer (l039_score.py) with the Holm adjustment over five, and the family's nights run from the first night on which all five were issued (the Tower install that carries L-043).

What counts as a pass, written before the test: Evaluated in May 2027 with l039_score.py (--verify), the same bar form as L-039. No verdict below 40 forecast-cloudy nights with both physics runs; then the window runs to May 31, 2027 and is evaluated whatever the count, saying so. Pass (merge into the main model, in order of size with L-039) only if all hold: (1) on forecast-cloudy nights, the physics alone misses the low by less with the change: the per-night difference (treatment minus main) is below zero at one-sided 2.5% after the Holm adjustment over the family (L-039, L-040, L-041, L-042); (2) the calibrated low, every night, is not worse with the change: the upper end of the 95% interval of the per-night difference is under +0.10 F; (3) the physics share of the low in the treatment's own monthly refit at evaluation is at least the main model's. Otherwise fail: the change is deleted and its lesson written here and in R-011. Power (R-011 rule 2), from the sizing: item 1's per-night spread is 1.02 F; with 40 forecast-cloudy nights its standard error is 0.16 F and the replay's 0.63 F gain passes the family-of-four Holm test about 95% of the time; with 90 nights, nearly always. Item 2: the calibrated difference is +0.013 F a night with a spread of 0.18 F; over about 170 nights the upper end of the interval sits near +0.04 F, so the margin holds nearly always, about 90% if the winter spread doubles. Item 3: 0.342 against 0.305 in the replay. Reported, not part of the verdict: cloudy dry and cloudy rainy nights separately, clear and partly cloudy nights, the calibrated high and low against ours and the main model, and the share of the high. Amended 2026-10-22: the Holm adjustment in item 1 is over the family of five (L-039 to L-043).
Next

Starts with the next Tower install, together with L-039. Left of miss M-1 after both: the missing rain cooling on rainy nights without snow, and the day model's hand-set cloud weights.

L-039running (treatment of the main model, with L-040, L-041 and L-042 on the same nights from their first issue after the next Tower install to April 30, 2027; evaluate in May 2027)

First treatment of the main model: the cloud deck emits at the 925 mb air temperature, not the 2 m air (R-011 miss M-1)

opened October 21, 2026 · by the Claude, with Brian, in a session
The question

On nights the models forecast cloudy, does the main model's night physics miss the station's low by less when the cloud part of the downwelling longwave is emitted at the 925 mb air temperature instead of the model's own 2 m air, with nothing else changed?

Why it might work

R-011 miss M-1, found on the main model's replay (April 2 to October 12, 2026): the physics alone runs warm in proportion to the forecast cloud, about 0 F on clear nights, +1.3 F partly cloudy, +2.5 to 2.7 F cloudy and dry, +3.9 to 5.3 F cloudy with rain. In version 1 the whole sky, cloud included, radiates at the 2 m air temperature, the warmest level in the night column; a cloud deck radiates at its base, hundreds of meters up and colder. Brian chose this as the first treatment (October 21, 2026), ahead of the calm-night layer depth, because it is a larger effect and because the calm-night work waits for the wind question (miss M-2). Exploration on the same replay (main_replay.py --change, l039_replay.csv): the change moves the physics the right way on every cloudy class but explains only part of the miss: forecast-cloudy nights' physics-alone error falls by 0.175 F per night on average (spread 0.29 F; 68% of nights better; dry cloudy nights about 0.10 F, rainy cloudy nights 0.30 to 0.39 F), clear nights do not move, the calibrated low does not change (1.561 against 1.564 F at lead 1), and the physics share of the low rises from 0.305 to 0.318 (the high from 0.134 to 0.153). So this is the right sign and a real structural correction, but most of the cloudy-night warm bias lies elsewhere.

The plan, written before the test

Frozen before anything is scored (code: column.py night_step with cloud_t = 925, treatments.json, calibration_L-039.json, l039_score.py, physmodel/l039_replay.csv, main_replay.py, main_calib.py). The treatment is the main model (L-038, v0) with one change: in the night step the cloud part of the downwelling longwave, (1 - clear-sky emissivity) x cloud x sigma T^4, uses T = the 925 mb air temperature (the models' 925 mb temperature at that step) instead of the 2 m air; the clear-sky part, every constant (none refitted), the day and everything else are the main model's. It has its own calibration, fitted by the main model's procedure on its own replay (calibration_L-039.json) and refit monthly the same way, and is issued every hour beside the main model as source L-039 by the same build. Nights: from the treatment's first logged issue (after the next Tower install) through the night of April 30, 2027; never the replay nights (April 2 to October 12, 2026) on which the miss was found and the change explored. Pairs: the main model and the treatment from the same build, at the cut of each night's lead 1 and lead 2 (forecaster.score's rule); a night's difference is the mean over the leads it has. Forecast-cloudy night: the models' mean total cloud for the night, at issue, 70% or more. Truth: the station's low in the scoreboard's window. Intervals by resampling nights (10,000). Merged into the main model on May 1, 2027 if it passes, deleted if it fails. Family (amended October 21, 2026, before the treatment issued): L-039 and L-040 test two changes for the same miss (R-011 M-1) and are scored on the same nights, from the first night on which both were issued, by one scorer (l039_score.py), with item 1 adjusted for the family (Holm, one-sided 2.5%: the smaller p against 0.0125, the other against 0.025). At the verdicts the passes are merged in order of size (the larger gain in item 1 first); the second is re-run as a diff from the new main model on the window's nights from the stored downloads and goes in only if its item 1 still holds there. (Amendment) Code change before anything was graded (October 21, 2026): column.py gained two inert parameters for sizing the next candidates for miss M-1 (rain cooling on every night; the reference air brought down from the models' cloud base when it is below 925 mb), and the night forcing now carries the models' 2 m air and dew point for them. With neither set, the main model's replay and the replays of L-039 and L-040 are byte for byte the same, and parity with the tested functions is 0.0 K. Family amended October 22, 2026, before anything was graded: L-041 (the reference air from the models' cloud base) joins, so the family is L-039, L-040 and L-041, scored from the first night on which all three were issued, with item 1 Holm-adjusted over the three (one-sided 2.5%: the smallest p against 0.0083, the next against 0.0125, the last against 0.025); the merge rule covers all three (in order of size, each later one re-run as a diff from the new main model before it goes in). Family amended October 22, 2026, before anything was graded: L-042 (rain cooling from the physics, no fitted constant) joins, so the family is L-039, L-040, L-041 and L-042, scored from the first night on which all four were issued, with item 1 Holm-adjusted over the four (one-sided 2.5%: the smallest p against 0.00625, then 0.0083, 0.0125, 0.025); the merge rule covers all four. Holm power for the family of four (recomputed October 22, 2026, by resampling the replay's 90 forecast-cloudy nights jointly, so the correlation between the four, 0.59 to 0.81, is kept): with 40 nights L-039 96%, L-040 95%, L-041 97%, L-042 90%; with 70 nights 99 to 100% each; with 90, 100%. A code change that alters the replay output is allowed only as a declared amendment (lab_tools freeze --output-amend, recording both output hashes, the replay nights it changes and the first live night it issues), and the evaluation reports the nights on each side of it separately; an undeclared change is refused. Amendment of October 22, 2026, before anything was graded: the main model now starts the half in progress from the station's latest reading (L-038, amendment 12), so this treatment's replay and calibration were re-run on that main model and recorded as a declared output change. Against the v0.1 main model, on forecast-cloudy nights: -0.163 F a night; weight on the physics for the low 0.334 (the main model's 0.318). Holm power over the four with 40 forecast-cloudy nights: L-039 95%, L-040 93%, L-041 96%, L-042 87% (this one 95%); with 70 nights 98 to 100%. Amended 2026-10-22: L-043 (the version 3 moisture budget inside the main model's night, sized on the replay, R-011) joins the family; the family is L-039, L-040, L-041, L-042 and L-043, scored by one scorer (l039_score.py) with the Holm adjustment over five, and the family's nights run from the first night on which all five were issued (the Tower install that carries L-043).

What counts as a pass, written before the test: Evaluated in May 2027 with l039_score.py (--verify). No verdict below 40 forecast-cloudy nights with both physics runs; then the window runs to May 31, 2027 and is evaluated whatever the count, saying so. Pass (merge into the main model) only if all hold: (1) on forecast-cloudy nights, the physics alone misses the low by less with the change: the per-night difference (treatment minus main) is below zero at one-sided 2.5% after the Holm adjustment over the family (L-039, L-040, L-041, L-042); (2) the calibrated low, every night (about 170 in the window), is not worse with the change: the upper end of the 95% interval of the per-night difference (treatment minus main) is under +0.10 F (the same interval test as L-038's item 1, with this margin); (3) the physics share of the low in the treatment's own monthly refit at evaluation is at least the main model's (Brian's progress meter). Otherwise fail: the change is deleted and its lesson written here and in R-011. Power (R-011 rule 2), from the exploration: the per-night difference has a spread of 0.29 F; the replay found about 13 forecast-cloudy nights a month, so the window should hold 70 to 90; with 40 nights it passes the family-of-four Holm test about 96% of the time; with 90 nights, 0.09 F. Item 2 (margin set by the power check, amended from 0.05 F): the exploration's calibrated difference is +0.002 F a night with a spread of 0.05 F; over about 170 nights the interval's half-width is under 0.01 F, so the margin holds even if the winter spread is several times the replay's. Item 3: 0.318 against 0.305 in the replay. Plain reading: a pass is likely, and it would remove about a tenth of the cloudy-night warm bias, not a whole degree; the rest of miss M-1 waits for its own treatments. Reported, not part of the verdict: dry and rainy cloudy nights separately, clear and partly cloudy nights, the treatment's calibrated high and low against ours and the main model, and the share of the high. Amended 2026-10-22: the Holm adjustment in item 1 is over the family of five (L-039 to L-043).
Next

The next Tower install starts it (treatments.json). The candidates the exploration leaves for the rest of miss M-1, each a later treatment of its own: the 925 mb air brought to the surface along the dry adiabat on nights the column is cloudy and near saturation (a moist lapse would be colder by roughly 2 to 3 F over that depth; not yet measured); the missing rain cooling on rainy nights; the day model's hand-set cloud weights against the night's fitted ones.

L-038pre-registered (first forecast of record November 1, 2026, 23:05 UTC; evaluate in May 2027)

The main model, alongside: does the physics forecast, corrected with the station's record, beat our statistical forecast?

opened October 21, 2026 · by the Claude, with Brian, in a session
The question

From November 1, 2026 the Tower issues, every hour beside our forecast, a second forecast of today's and tomorrow's high and low: the main model, one physics integrator for the column over the mast (day and night in one loop, started from the station's latest sunrise or sunset reading, driven by the weather models), corrected by a calibration learned from the station's record. Its forecasts of record (the 11:05 and 23:05 UTC issues) are graded beside ours at the same cut-offs on the same scoreboard. Is it at least as good as ours overall, and better in a regime where the physics should matter? If so it replaces our forecast; if not, it keeps running alongside and is judged again on fresh periods.

Why it might work

Brian's operating model (October 21, 2026; hyperlocal/APPROACH-2026-10-21.md section 4a): the physics is the engine, the statistical forecast becomes the benchmark and the calibration, and the physics forecast is the candidate main model, not a treatment. This is a new entry, not an extension of L-032: L-032 (frozen October 20, graded since October 21) blends our forecast toward the physics, with ours as the engine; this candidate is the reverse, with its own bar and grading rule. Replay on the real forecast runs, April 2 to October 13, 2026, at the two record times each day (physmodel/main_replay.py; the integrator gives the same numbers as the tested functions to 0.0 K on 216 halves), each month held out in turn (physmodel/main_calib.py): the physics alone missed by 1.69 F (lead 1) and 2.28 F (lead 2) on the high against ours 1.68 and 1.97, and by 2.53 and 2.75 F on the low against 1.40 and 1.70 (it runs about 1.5 to 1.8 F warm at night, almost all of it on cloudy and rainy nights). Calibrated on the physics alone it stays 0.5 to 0.65 F a day behind ours at lead 2. Calibrated on the physics and the four-model average together it ties ours: high 1.59 and 1.94 F, low 1.56 and 1.84 F, per-day difference at lead 2 +0.05 F (spread 0.92 F). The fitted weights put 0.13 on the physics for the high and 0.30 for the low (the rest on the four-model average): today the physics carries real information for the low (the four-model average alone, calibrated, misses by 2.03 F at lead 2; with the physics, 1.84) and little for the high. The physics weight is published every month; as treatments improve the physics it should rise.

The plan, written before the test

Frozen before the first forecast of record (code: column.py, l038_score.py, main_calibration.json, physmodel/main_replay.csv, main_replay.py, main_calib.py, fingerprints in this entry's freeze record). (1) The candidate: column.py version v0, every term the current best answer from the tests that passed: night = version 1 with the L-007 constants, with the L-017 snow insulation and the wet-weather terms it was tested with only when snow is on the ground (L-016 failed, so rain is not modelled on other nights); day = the L-020 version with its offset; the air temperature carries across sunrise and sunset; start from the station's reading at the latest sunrise or sunset at or before the issue (readings whose station clock did not move are not stored); deep soil from the 4 and 8 inch probes (the last hour, else the last 24 hours). Nothing is substituted when an input is missing: a period the physics cannot reach is issued from our forecast, flagged degraded, and graded as issued. (2) The calibration: per target, value = a + b x physics + c x four-model average, least squares, fitted on the replay (April 2 to October 13, 2026); refit by the same code on the first build of each local month on the replay plus every graded live period the physics reached (column.refit). (3) Issue: every hour at 5 minutes past by the wall clock, as source 'main' in the issued table, leads 1 and 2 only (days 3 to 7, rain and the hour-by-hour stay with our forecast); the forecast of record is the last issue at or before the cut (the 11:05 UTC build for the high, 23:05 UTC for the low), exactly forecaster.score's rule. (4) Grading (l038_score.py): periods whose record issue is at or after November 1, 2026, 23:00 UTC, through the night of April 30, 2027; truth = the station's high or low in the scoreboard's windows, 90% of the 5-minute slots present; the per-day difference = the mean, over that date's high and low, of |main - truth| - |ours - truth|; intervals by resampling days (10,000). (5) Regimes, defined at issue time from the models' own forecast for the half (never from what happened): clear and calm night (mean total cloud of the three layers combined under 30% and mean 10 m wind under 5 mph, sunset to sunrise); rain (forecast precipitation of at least 0.10 in in the half); snow (snow on the ground at issue of at least 1 in, or at least 0.05 in of precipitation with a mean 2 m temperature at or below 34 F). (6) The main model is fixed inside the window apart from the monthly refit and bug fixes recorded here with their dates; treatments of the main model run alongside it as their own entries and are merged only at a window boundary. (7) The first documented miss, which points at the first treatment of the main model: on L-007's 149 dry nights (April to October 2026, started at sunset; exploration), the physics with the models' own low, calibrated, missed the 77 clear and calm nights by 0.22 F more than ours (the physics alone by 0.42 F more); in this replay the miss concentrates in the calmest nights (forecast wind under 2.5 mph: 5 nights, the physics 0.64 F too warm), which is the structural stop in cooling found in the code review (4.5): with a fixed 159 m layer the 2 m air decouples once the wind dies. The first treatment is the night layer's depth as a state (a stable layer that collapses when the wind dies), fitted on the station's own calm nights once the wind check (the ratio after harvest and the spin test) shows the anemometer can be trusted. The second documented miss, recorded and not yet treated: the physics runs about 2 F warm on cloudy and rainy nights. (8) The progress meter (amended before anything was graded, at Brian's direction): the physics share of the calibration (the weight b on the physics, per target, in each monthly refit) is the project's progress meter. Every treatment of the main model is judged partly by whether it raises that share and lowers the physics-alone error (written into each treatment's bar). The share and the physics-alone error, beside the corrected forecast and ours, are published monthly on the Physics page. (9) Code changes in the same amendment, none of which changes the main model's numbers: column.py gained the treatment hook (a treatment is the main model's parameters with one change, issued beside it as its own source with its own calibration); main_replay.py and main_calib.py gained optional arguments to replay and calibrate a treatment by the same procedure. With no change given, the integrator again matches the tested functions to 0.0 K on 306 halves and the replay file is byte for byte the same. (10) A second code change before anything was graded: column.py gained the parameter for treatment L-040 (the moist-adiabat weight); with no change given the main model's replay is again byte for byte the same. (11) Code change before anything was graded (October 21, 2026): column.py gained two inert parameters for sizing the next candidates for miss M-1 (rain cooling on every night; the reference air brought down from the models' cloud base when it is below 925 mb), and the night forcing now carries the models' 2 m air and dew point for them. With neither set, the main model's replay and the replays of L-039 and L-040 are byte for byte the same, and parity with the tested functions is 0.0 K. (12) Amendment of October 22, 2026, before anything was graded (finding H1 of the October 22 review): the half in progress starts from the station's latest reading at the issue. Its air is set to the station's reading and the station's departure from the run fades with an e-folding time of 3 hours, by night and by day (column.run; version v0.1, carried in every run record). The exploration set (Brian's choice) was the Tower's own logged restarts: on 17 restart times over 3 nights (L-030 and L-034), restarting the night integrator from the station with the grass keeping its gap missed the low by 4.46 F against 1.93 F with no restart (3 to 6 F too cold every time, because the night physics cools late in the evening) and was not adopted; the fading departure missed by 2.06 F. On L-022's 9 afternoon times the fading departure missed the rest of the day's high by 1.02 F, against 1.77 F with no restart and 1.34 F for L-022's measured-sunshine re-run. In the replay the rule takes the physics alone's low at lead 1 from 2.52 to 2.34 F and leaves the high (1.69 F) and lead 2 unchanged. The replay and the calibration were re-run and recorded as a declared output change; the scorer also reports the days on each side, by the version that issued them. Weight on the physics after the refit: low 0.318, high 0.136 (b over b plus c: 0.324 and 0.133). Power: the per-day spread at lead 2 is 0.93 F; with about 170 graded days item 1 passes about 94% of the time if the two are truly equal and 79% at the replay's +0.05 F; item 2 about 92% at the replay's +0.01 F. The same amendment: every failure inside the issue becomes a degraded issue with its reason, and the Engine page counts periods with no main row at the cut; the replay runner accepts only a fixed command template and runs without a shell; the six parameter files are part of the recorded code; the run curves are published for the Physics page.

What counts as a pass, written before the test: Evaluated in May 2027 (l038_score.py on the Tower's database, with --verify). No verdict below 120 graded days at lead 2: the window then runs to May 31, 2027 and is evaluated whatever the count, saying so. The verdict is 'replace' only if all hold: (1) at lead 2, highs and lows pooled by day, the upper end of the 95% interval of the per-day difference (main minus ours) is under +0.25 F; (2) the same at lead 1; (3) in at least one of the three regimes at lead 2, the main model is better than ours with the 95% interval wholly below zero after a Holm adjustment over the three regimes (a regime with fewer than 10 graded periods counts as not passed); (4) degraded issues are at most 10% of graded periods. Otherwise the verdict is 'not yet': the candidate is not deleted; it keeps running alongside and is judged next, with this bar, on fresh periods only (May 1 to October 31, 2027, evaluated in November 2027); no look in between counts. Power (R-011 rule 2), from the replay's held-out months: the per-day difference at lead 2 has a spread of 0.92 F; with about 170 graded days its standard error is about 0.07 F, so item 1 passes about 94% of the time if the two are truly equal, about 81% at the replay's +0.05 F, and rarely at +0.25 F. A plain 'better on the season' bar would have 80% power only for a true gain of about 0.2 F. Regimes at lead 2 (80% power after the Holm step): clear and calm nights (about 19% of nights, so about 35 in the window; spread 0.66 F) for a gain of about 0.37 F, and the replay shows +0.04 F; rain (about 70 periods; spread 1.45 F) for about 0.56 F, and the replay shows -0.11 F; snow depends on the winter (about 30 periods would detect about 0.9 F; L-017 found 1.6 F for the physics alone, which the calibration shrinks). Plain reading: items 1, 2 and 4 are likely to hold; item 3 probably holds only through snow, so a mild winter will most likely end in 'not yet'. Reported, not part of the verdict: the physics alone, the four-model average and ours at both leads; each target separately; each regime; fall, winter and spring; the calibration's weights each month; the count of degraded issues and why.
Next

Deploy B puts column.py on the Tower with the wall-clock build at 5 past the hour; the issues before November 1, 23:00 UTC are a shake-down and are not graded. Scoreboard: the main model's line beside ours from the first graded morning (November 2).

L-037Closed continues as L-044

The day ensemble: does the physics spread the high better than the ECMWF ensemble's own members do?

opened October 21, 2026 · closed October 22, 2026 · by the Claude, with Brian, in a session
The question

Centered on our forecast high, does the spread of the day physics run on each ECMWF member catch the station's high better than the spread of the members' own 2 m highs?

Why it might work

Brian (October 21): an ensemble for the day model too. The day's uncertainty is mostly sunshine and clouds (L-009 failed on forecast clouds), which an ensemble expresses. The rival is strong: the members already forecast the high, and by day the air is well mixed and the models do better than at night (R-011 rule 7). So the test isolates the SPREAD: both sets of members are re-centered on our own forecast high, so only the shape of the spread can win. First day (October 22, sunny and settled): physics 70.4 to 71.3 F in 8 runs of 10, members' own highs 68.6 to 69.8 F; ensembles are known to be too narrow on settled days.

The plan, written before the test

Frozen October 21, 2026, before any graded day (Tower: physlive.day_ensemble, physlive.py sha256 bf8a7b6c4514402a...; the night code is unchanged since L-034 to L-036 were frozen, only the day ensemble was added). Each day's sunrise run (kind 'day') stores 'ens': the L-020 day physics (crop calendar and real-forecast offset) on each ECMWF member's sunshine, cloud by layer, 10 m wind, 925/850/700 mb profile at sunrise and air-mass change, through day_forcing; soil moisture and snow from the main runs; from the station's sunrise reading. Also stored: the members' own 2 m highs (sunrise to sunset) and our forecast high at that moment. Truth: the station's highest reading, sunrise to sunset. Dry days only (0.02 in or less between sunrise and sunset), October 22, 2026 to April 30, 2027. (Amended the same day, before any graded day: a display-only fill-in for a day run logged before the ensemble existed; nothing logged changes.)

What counts as a pass, written before the test: Evaluated in May 2027; no verdict below 100 dry days. Primary: the per-day CRPS of {ours + (physics member - physics median)} lower than that of {ours + (member's own high - their median)} in at least 95% of 10,000 bootstrap resamples by day. Reported, no bar: the coverage of the physics' 8-in-10 range as issued and re-centered; the chances of reaching each 5 F mark (Brier against the station); the sunshine thirds. If the primary fails, the day odds stay off the Today page and the physics' day spread is not used.
Results

October 21, 2026. Bug fix installed October 21, 2026, 10:02 PM EDT (02:02 UTC October 22), plan, constants and scoring unchanged: the Tower no longer stores a station reading whose station clock has not moved (a stalled Weather Display or upload), so no run starts from a repeated reading that looks current (code review 4.1). Before this, six such repeats were stored in the week of October 14 to 21.

October 22, 2026. Closed on 2026-10-22 without a verdict (the one-time closing pass, APPROACH-2026-10-21.md section 4a): its question (does the day physics spread the high better than the members' own 2 m highs) continues as L-044, the main model on every member through the day halves, with the same CRPS bar; the day ensemble code leaves the live physics with the install after L-044's first run is confirmed. Nothing was graded and no pass or fail is claimed; whatever this trial logged is exploration only and is never used to write a bar on the same nights. Its question continues as L-044.

Next

First graded day October 22, 2026 (the sunrise run).

L-036Closed continues as L-044

Version 1R (the air's own radiative cooling) in the physics ensemble: more honest odds than version 1?

opened October 21, 2026 · closed October 22, 2026 · by the Claude, with Brian, in a session
The question

Does the ensemble run with version 1R give more honest odds of the low and of frost on the grass than the same ensemble run with version 1?

Why it might work

The physics ensemble's first night showed version 1's calmest members coming out warmest (median 39.9 F against 35.8 F): version 1 cools the 2 m air only through the grass, so in near-calm air it stops cooling. Version 1R (physmodel/scm_r.py) adds the layer's own longwave: Q = eps_L (2 sigma Ta^4 - eps_s sigma Ts^4 - L_down), cooling the layer by Q / (rho cp z_eff). Refit of all 13 constants on version 1's 474 dry discovery nights (calibrate_r.py): eps_L = 0.091, the fit's loss 4.092 -> 4.087, low error 2.011 -> 2.010 F, no change by wind: the regime is absent from history (forecast wind under 1.5 mph on 1 night of 1,046), so no test on deterministic runs can succeed (R-011 rule 2). The ensemble's members visit that regime: on October 21 version 1R narrowed the calm members (calmest third 39.9 -> 38.1 F, top member 47.3 -> 43.0 F, 8-in-10 range 34.5-40.9 -> 34.6-38.7 F) but did not remove the effect: the fitted layer (about 184 m) is too deep to cool fast; the shallow calm layer (version 1G's idea) is the remaining piece.

The plan, written before the test

Frozen October 21, 2026, before any graded night (Tower: physlive.ensemble with phys_params_r.json = physmodel/params_r.json; physlive.py sha256 e6f8ad2a27a5de11...; the eps_L term enters physlive.integrate only when the constants carry eps_l, and version 1's runs are unchanged to 1e-13 F, so L-034 and L-035 still test what they froze). Each dry night's sunset run stores ens.r: the same 51 members, the same start, version 1R. Truths and nights as L-035. Amended October 21, 2026, 7:45 PM, before any night was graded: a bug stored the sunset run before its ensemble was attached, so the stored record lacked 'ens' (the page showed it). Fixed (physlive.py sha256 3f0e2711c9e718e6...): the run is stored again with its ensemble; a run stored without it is repaired once, from the ensemble and main-run downloads the Tower had when the run was made (the same inputs; the soil probes and the pressure as of the repair, a few tenths at most), marked ens_repaired.

What counts as a pass, written before the test: Evaluated in May 2027 with L-035's nights. Primary: the Brier score of the grass-frost chance lower with version 1R than with version 1 in at least 95% of 10,000 bootstrap resamples by night, AND the 1R range's coverage of the station's low at least as close to 80% as version 1's. Reported, no bar: the calm-third bias against the station on nights the station's wind averaged under 2 mph. Honest power note: the two ensembles differ mostly on near-calm nights, so the difference may be too small to prove in one season; if so, the trial runs a second winter rather than being judged on noise.
Results

October 21, 2026. Bug fix installed October 21, 2026, 10:02 PM EDT (02:02 UTC October 22), plan, constants and scoring unchanged: the Tower no longer stores a station reading whose station clock has not moved (a stalled Weather Display or upload), so no run starts from a repeated reading that looks current (code review 4.1). Before this, six such repeats were stored in the week of October 14 to 21.

October 22, 2026. Closed on 2026-10-22 without a verdict (the one-time closing pass, APPROACH-2026-10-21.md section 4a): version 1R's question (does the air's own radiative cooling make the odds more honest) is not carried as a second ensemble: it becomes a sizing candidate for a single-change treatment of the main model (R-011 rule 8), to be sized on the replay before any freeze; the 1R ensemble code leaves the live physics with the install after L-044's first run is confirmed. Nothing was graded and no pass or fail is claimed; whatever this trial logged is exploration only and is never used to write a bar on the same nights. Its question continues as L-044.

Next

Live from the night of October 21-22, 2026.

L-035Closed continues as L-044

The physics ensemble: are its odds honest? (51 runs of the night physics on the ECMWF ensemble)

opened October 21, 2026 · closed October 22, 2026 · by the Claude, with Brian, in a session
The question

When the night physics is run on each of the ECMWF ensemble's 51 members from the sunset start, do its odds hold up: does the station's low land inside its 8-in-10 range about 8 nights in 10, and are its chances of frost on the grass, a freeze and saturation better than knowing only how often those happen?

Why it might work

Brian (October 21): a local ensemble. L-019 showed averaging per-model physics runs does not improve the single low, so the ensemble is judged on its odds, not its average. The night-to-night errors come from the weather inputs (the initialization and recalibration explorations of October 21), which an ensemble of inputs is built to express. First live night: the 51 members spanned 34 to 47 F on a clear, calm night, all from wind between 0.6 and 2.3 mph; and the calmest runs came out WARMEST (version 1 cools the air only through the grass), a structural limit recorded here before grading starts.

The plan, written before the test

Frozen October 21, 2026, before any graded night (Tower: physlive.ensemble, physlive.py sha256 eecfa3b1700ec1ae...; collector source ens_phys = Open-Meteo ensemble API, ecmwf_ifs025, every 6 hours). Each night's sunset run (physlive.build, kind 'sunset') stores 'ens': version 1 on each member's cloud by layer, 10 m wind, 925 mb air (brought down with the main runs' 925 mb height) and 2 m dew-point drift, from the station's sunset reading; version 3 on the same for saturation; the newest ensemble snapshot at that hour. Recorded: the members' lows (p10, p50, p90), the chance of the grass surface reaching 32 F, of the air reaching 32 F, of version 3's air minus dew point reaching 0.5 F or less. Truths: the station's lowest reading sunset to sunrise; the grass probe's lowest (frost if 32 F or colder); the air at 32 F or colder; saturation = humidity 95% or more with temperature minus dew point 1.5 F or less for 60 minutes or more. Dry nights only (0.02 in or less, no snow on the ground), October 22, 2026 to April 30, 2027. Amended October 21, 2026, 7:45 PM, before any night was graded: a bug stored the sunset run before its ensemble was attached, so the stored record lacked 'ens' (the page showed it). Fixed (physlive.py sha256 3f0e2711c9e718e6...): the run is stored again with its ensemble; a run stored without it is repaired once, from the ensemble and main-run downloads the Tower had when the run was made (the same inputs; the soil probes and the pressure as of the repair, a few tenths at most), marked ens_repaired.

What counts as a pass, written before the test: Evaluated in May 2027; no verdict below 100 dry nights. Pass needs all of: (1) the station's low inside p10-p90 on 70% to 90% of nights; (2) the Brier skill of the grass-frost chance against the trial's own frost frequency above zero in at least 95% of 10,000 bootstrap resamples by night; (3) the spread (p90 - p10) larger on the nights the middle run missed by more than its median miss than on the rest (a one-sided bootstrap, 95%). Reported, no bar: freeze and saturation chances, reliability by tenths, the calm-member bias. If it passes, the odds go on the Today page with a link to the evidence; if (1) fails, the range is widened or narrowed by a factor fitted on the trial and tested on the next season.
Results

October 21, 2026. Bug fix installed October 21, 2026, 10:02 PM EDT (02:02 UTC October 22), plan, constants and scoring unchanged: the Tower no longer stores a station reading whose station clock has not moved (a stalled Weather Display or upload), so no run starts from a repeated reading that looks current (code review 4.1). Before this, six such repeats were stored in the week of October 14 to 21.

October 22, 2026. Closed on 2026-10-22 without a verdict (the one-time closing pass, APPROACH-2026-10-21.md section 4a): its question (are the night physics' odds on the ECMWF members honest) continues as L-044, the main model on every member, re-centred on the corrected forecast and graded by the same bar; the version 1 ensemble code leaves the live physics with the install after L-044's first run is confirmed. Nothing was graded and no pass or fail is claimed; whatever this trial logged is exploration only and is never used to write a bar on the same nights. Its question continues as L-044.

Next

First graded night October 22-23, 2026.

L-034running (live trial, nights October 21, 2026 to April 30, 2027; evaluate in May 2027)

Live trial: restart the night in the right state (follows L-033)

opened October 21, 2026 · by the Claude, with Brian, in a session
The question

Through the fog season, does the version 3 restart started saturated when the station is saturated (C) beat the live version 1 restart (A) and version 3 from the measured dew point (B) on the night's remaining low?

Why it might work

L-033 passed on the April-October test months (saturated restarts: C 0.87 F against A 1.02 and B 1.04; bias +0.09 against -0.46 and -0.41), but those months hold few true fog nights; November to March is the fog season (research R-012). The comparison that would change the live forecast is C against A.

The plan, written before the test

Frozen October 21, 2026, before any live night (Tower: physlive.restart_state, physlive.py sha256 cf17f401abc51d04...; phys_params_v3.json unchanged). Each dry night, at every whole hour T from 10 PM to 5 AM local (logged within two hours after T, kind 'restart3'): the forecasts the Tower had at sunset; A = version 1 and B = version 3 restarted from the station's temperature and dew point at T, C = version 3 with the dew point set to the temperature when the station is saturated (humidity 95% or more and temperature minus dew point 1.5 F or less in every reading of the 30 minutes before T), else C = B; the grass keeps the gap each version's own sunset run built by T. Truth: the station's lowest reading from T to sunrise. Nights October 21, 2026 to April 30, 2027.

What counts as a pass, written before the test: Evaluated in May 2027. Primary: on saturated restarts, the per-night mean of (|A error| - |C error|) above zero in at least 95% of 10,000 bootstrap resamples by night, and C's mean bias closer to zero than A's. No verdict below 60 saturated nights (L-033's power check: C over A needs roughly 150 for a 0.13 F gain at its spread; 60 detects about 0.2 F, the size L-033's test found): if fewer, the trial runs to the next April. Reported, no bar: C over B, all restarts, the nights Brian logged as fog. If it passes, C replaces A as the physics inside the live restart (L-030's blend weights refit only through a new pre-registered test).
Results

October 21, 2026. Bug fix installed October 21, 2026, 10:02 PM EDT (02:02 UTC October 22), plan, constants and scoring unchanged: the Tower no longer stores a station reading whose station clock has not moved (a stalled Weather Display or upload), so no run starts from a repeated reading that looks current (code review 4.1). Before this, six such repeats were stored in the week of October 14 to 21.

Next

Live logging starts with the night of October 21-22.

L-033passed (October 21, 2026); live trial L-034 next

Restart the night in the right state: start the physics saturated when the station is saturated (research R-010, R-012)

opened October 21, 2026 · by the Claude, with Brian, in a session
The question

When the station shows saturation during the night, does restarting the fog-capable night physics (version 3) WITH THE AIR SATURATED improve the low and the rest of the night's curve, compared with restarting it from the dew point the sensor reports?

Why it might work

Brian (October 21): 'are we just creating a bunch of if-this-then-that rules, or are we ultimately building algorithms that model the atmosphere of this place?' The fog night of October 20-21 (the station stopped cooling at 41 F at 4 AM under fog; version 1 restarts ran to about 34 F) suggested a fog switch. Instead, data assimilation: a Davis humidity sensor never reads 100% (it tops out near 95-98% in fog and heavy dew), so every restart begins slightly DRY and the model has to rediscover saturation. Correct the starting state and let version 3's own moisture budget decide whether fog forms, holds or clears. No new constant, no switch: only the starting dew point changes. Saturation is common (95% or more on 107 of the April-October nights, mostly heavy dew), so the model, not a rule, must tell dew nights from fog nights.

The plan, written before the test

Frozen October 21, 2026, before the test months were run (physmodel/l033.py, sha256 cb04689c064f7ec8...). Restarts at each whole hour T from 10 PM to 5 AM local; dry nights (0.02 in or less, sunset to sunrise); the 12Z ECMWF + ICON runs (l019.night_inputs, the forecasts the Tower would have had); the grass keeps the gap its own sunset run had built by T (as L-030). A = version 1 restart (the live L-030 physics); B = version 3 (params_v3b.json) from the station's temperature and dew point; C = version 3 with the dew point set to the temperature when the station is saturated, else the same as B. Saturated: humidity 95% or more AND temperature minus dew point 1.5 F or less in every 5-minute reading of the 30 minutes before T (threshold chosen on the exploration months below). Truth: the station's lowest reading from T to sunrise; also the mean absolute error of the curve from T to sunrise. Months alternate (R-011 rule 3): EXPLORE April, June, August, October 2026 (done); TEST May, July, September 2026, excluding September 19-25 (gauge glitch). Exploration (77 nights, 149 saturated restarts on 37 nights): low error on saturated restarts A 1.12, B 1.20, C 1.02 F; bias A -0.85, B -0.61, C -0.18 F; curve A 1.03, B 1.00, C 0.86 F; per-night gain C over B 0.185 F (sd 0.377), C over A 0.131 F (sd 0.578). All restarts: A 1.47, B 1.41, C 1.37 F.

What counts as a pass, written before the test: Primary (the assimilation idea): on the test months' saturated restarts, the per-night mean of (|B error| - |C error|) of the low is above zero in at least 95% of 10,000 bootstrap resamples by night, AND C's mean bias is closer to zero than B's. Power check (R-011 rule 2): at the exploration's effect (0.185 F, sd 0.377) about 35 saturated test nights give roughly 85% power; the effect may shrink. Reported, no bar: C against A (the live restart) on saturated and on all restarts, and the curve error. C against A is the comparison that would change the live forecast, and it is underpowered here: at the exploration's 0.131 F (sd 0.578) it needs about 150 saturated nights, so if the primary passes, C is logged alongside every live restart from the next night as a separate pre-registered live trial (L-034) through April 30, 2027, the fog season.
Results

October 21, 2026. TEST months May, July, September 2026: 71 nights, 568 restarts; saturated: 152 restarts on 42 nights. Saturated restarts, low error A 1.02, B 1.04, C 0.87 F; bias A -0.46, B -0.41, C +0.09 F; curve A 0.93, B 0.91, C 0.74 F. Per-night gain C over B 0.204 F (100.0% of 10,000 resamples above zero); C over A 0.190 F (99.8%). All restarts: A 1.46, B 1.38, C 1.34 F; gain C over A 0.118 F (93%). Files: physmodel/l033_test.csv, l033_test_result.json (scored by l033_eval.py).

What it means

PASS. Starting the restart in the right state (the air saturated when the station is) removed nearly all of the physics' cold bias on saturated nights (-0.41 F to +0.09 F) and cut the low's error by about a sixth, with no new constant and no switch: version 3's moisture budget decides on its own whether fog forms. It also beat the live version 1 restart, by more than the exploration suggested (0.19 F, 99.8%), though the plan treats that comparison as unproven until the live trial: real-run test months were all April-October, and the fog season is November-March.

Next

L-034: log C alongside every live restart (L-030's 10 PM, midnight and 2 AM, plus hourly to 5 AM) from the next night through April 30, 2027, written down before it starts.

L-032Closed continues as L-038

The physics race: does the physics' own day-ahead forecast improve ours one and two days out?

opened October 20, 2026 · closed October 22, 2026 · by the Claude, with Brian, in a session
The question

Since October 20, 2026 the Tower records the physics model's own forecast of each high and low, made the same way ours is: by chaining day into night into day from the station's last sunrise or sunset reading, with the weather models' clouds, wind and air mass, and logged every hour for today and tomorrow only. Graded on the same cut-offs as the scoreboard (lead 1: made just before the period began, before 8 AM for the high and before 8 PM for the low; lead 2: made the day before), does moving our forecast toward the physics make it better? And how does the physics alone compare with ours, the National Weather Service and WxSim?

Why it might work

Brian's goal (October 20): the best forecast for Somerset, a forecast and not a nowcast, with the physics doing the work one to two days out and the statistics taking over further out; added alongside, changing nothing else. The physics' sunset-start blend already passed (L-007, a quarter of the way toward the physics cut the dry-night miss from 1.52 to 1.38 F) and the sunrise-start day blend is on trial (L-010), but neither was a true day-ahead forecast: this is the first test of the physics as a forecast made a day before. R-011 rules applied: rule 2 (power): on L-007's 149 nights the per-night gain from the quarter blend averaged 0.14 F with a night-to-night spread of 0.58 F; the day-ahead physics strays further from ours, so the plan assumes a spread of 0.8 F. With highs and lows pooled (about 380 periods by April 30) a true 0.10 F gain clears a 90% resample bar about 88% of the time (0.05 F: about 47%); with one target alone (about 190) about 67%. So the main measure pools highs and lows, and the trial runs into spring. Rule 3: the window spans late fall, winter and early spring. Rule 7: the physics should earn its keep in the regimes the statistics see least, so those are reported separately.

The plan, written before the test

Frozen October 20, 2026, before any graded period (Tower: physlive.py chain() and log_forecast(), sha256 2659ff7d60454e0a...; the issued table, source 'physics', leads 1 and 2). Grading exactly as forecaster.score: for each period and lead, the last issue at or before the cut-off, the same cut-off for every source; truth = the station's high (8 AM to 8 PM) or low (8 PM to 9 AM), the periods fixed in UTC. Blends computed from the logged values at the same issue: B = ours + w x (physics - ours), w = 0.25 for lows (the L-007 weight) and 0.5 for highs (the L-010 weight, without its offset, which belonged to the sunrise run). 'Ours' is the forecast as issued (the scoreboard's), never the live-updated number. Periods without a physics value at the cut-off (Tower down, physics error) are left out for every source and counted. Part B, exploration only (cannot change anything in part A): the lab checks whether the real forecast-run archive (runs_*, true lead times, April 2, 2026 on) carries the inputs the chain needs (clouds by layer, wind, the temperature about 2,000 to 5,000 ft up, sunshine, precipitation); if it does, it replays the chain day-ahead for April 2 to October 20, 2026 and reports the same measures, labeled as exploration on months already looked at.

What counts as a pass, written before the test: Evaluated in May 2027 on October 21, 2026 to April 30, 2027. No verdict below 150 graded highs and 150 graded lows at lead 2: if fewer, the trial runs to May 31, 2027 and is then evaluated whatever the count, saying so. Pass only if all hold: (1) at lead 2, highs and lows pooled, the blend's mean absolute error is at least 0.05 F lower than ours; (2) that gain is positive in at least 90% of 5,000 resamples of days (a day's high and low resampled together); (3) at lead 2, neither highs nor lows alone is worse with the blend than ours by more than 0.05 F; (4) at lead 1, pooled, the blend is not worse than ours. Reported alongside, not part of the verdict: the physics alone against ours, the National Weather Service, WxSim and the four-model average at both leads; each target and lead separately; fall (October-November), winter (December-February) and spring (March-April); clear calm nights, nights and days with rain or snow, and nights with snow on the ground; the share inside our 80% range; other weights (0.25 and 0.5 for both targets) as exploration only. A monthly look is published on the Physics page and the scoreboard as a running count, with no verdict before May 2027. If it passes, the blend at leads 1 and 2 is proposed for promotion (Brian and Claude decide together); leads 3 and beyond stay with the statistics.
Results

October 21, 2026. Bug fix installed October 21, 2026, 10:02 PM EDT (02:02 UTC October 22), plan, constants and scoring unchanged: the Tower no longer stores a station reading whose station clock has not moved (a stalled Weather Display or upload), so no run starts from a repeated reading that looks current (code review 4.1). Before this, six such repeats were stored in the week of October 14 to 21.

October 21, 2026. Bug fix installed October 21, 2026, 10:02 PM EDT (02:02 UTC October 22): one daylight-time rule (second Sunday of March to first Sunday of November) replaces the month-only rule in the Tower code (code review 4.2). For this trial it affects only the local date used to number the leads of rows written in the weeks of a clock change.

October 21, 2026. Schedule fixed October 21, 2026, 10:02 PM EDT (02:02 UTC October 22): the forecast is built at 5 minutes past every hour by the wall clock (it drifted by 10 to 20 minutes a day and reset after each restart), so the last issue before each cut is now always the 11:05 or 23:05 UTC build. The grading rule (the last issue at or before the cut, the same for every source) is unchanged; the installer's check build now runs on a copy of the database, so no row is logged off the hour.

October 22, 2026. Closed on 2026-10-22 without a verdict (the one-time closing pass, APPROACH-2026-10-21.md section 4a): replaced outright by L-038, which issues the same chained physics corrected with the station's record, at the same cut-offs, with a plan and a bar; L-038 reports the physics-alone line this entry was. Nothing was graded and no pass or fail is claimed; whatever this trial logged is exploration only and is never used to write a bar on the same nights. Its question continues as L-038.

Next

Part B (the replay on the forecast-run archive) is the lab's to check first. Live grading starts with the night of October 20-21 (graded the morning of October 21) at lead 1 and October 21-22 at lead 2.

C-004case study

Two freeze-watch mornings, two opposite misses: October 17 (we said 31.2 F, it was 36.0 F) and October 18 (we said 34.9 F, it was 31.5 F) [written as C-002 by the lab assistant on October 19; renumbered C-004]

opened October 19, 2026 · by the lab assistant on the Tower
The question

The held-back trigger for October 18 (first freeze, 31 F, Freeze Warning) sent me to the data, and it also showed that C-001's scoring had the wrong morning. What did each morning do, what did we and the others forecast, and why did the two misses go in opposite directions?

The plan, written before the test

Read the final forecast of every source before each sunrise (issued table), the station's hourly readings for both nights (obs), and the neighbouring airports' sky and wind (asos.parquet, which only reaches the evening of October 17 so the second night has no airport data). Scripts in /lab/experiments/C-002/. Only two nights, so everything here is a pointer, not evidence.

What counts as a pass, written before the test: Case study: no pass/fail. Findings become ideas for proper tests on other days.
Results

Morning of Oct 17: low 36.0 F (still 36.1 F at 8 AM). Ours 31.2 F, 4.8 too cold; NWS 32, WxSim 33.5, four-model average 34.0, all too cold too. The night had a light north wind of 2 to 4 mph at the station and 6 to 10 knots at the airports around us, with scattered cloud at several of them, so the air never settled the way a calm clear night does. The temperature fell a steady 1.2 F an hour. Morning of Oct 18: low 31.5 F (dew point 30.2 F at the low). Ours 34.9 F, 3.4 too warm; NWS 36, WxSim 34.7, four-model average 37.2 F. This night was dead calm (0.0 mph) and clear: the station fell from 66 F at 6 PM to 42 F at 10 PM (11 F in the first hour after sunset) and then crept down to 31.5 F. Our live adjustment at 9 PM said 34.4 F, closer but still 2.9 F warm, and the forecast jumped back up to 35.6 F at 10:33 PM. So the same correction that was 4.8 F too cold on a breezy night was 3.4 F too warm on a calm clear night. The four-model average was warm by 5.7 F on the calm night and too cold by 2.0 F on the breezy one: the models miss the wind-and-calm switch in both directions, which is the R-002 decoupling finding showing up live. Over the two nights our mean miss was 4.1 F (against the usual 1.6 F for lows) and the NWS's was 4.25 F (4.0 and 4.5): no better than us.

Final forecast before sunrise vs the station low (F)
MorningStation lowOurs (+/-)NWSWxSimFour-model avg
Oct 17 (breezy, scattered cloud) 36.0 31.2 (3.1) 32 33.5 34.0
Oct 18 (calm, clear) 31.5 34.9 (3.1) 36 34.7 37.2
  • C-001 had scored the wrong morning (corrected in that entry). Both mornings count as looked-at days, so any test of an idea from here must use other days.
  • The stated +/- of 3.1 F covered neither miss (4.8 and 3.4), though 3.1 is meant to be an 80 percent range, so two misses in a row is odd but two nights cannot show it is wrong.
  • Oct 19 morning is not scored: the night had not finished when this was written (forecast 38.1 F at 1:34 AM, station 42.8 F at 2 AM, calm and slowly rising).
  • Renumbered October 20, 2026: a session on the Mac had opened a different C-002 (the October 18-19 Frost Advisory night) on October 18 but published it only on October 19, after this entry; that publish replaced this one online. Restored from the Tower's copy under the next free number.
What it means

Not a model failure that points one way: the breezy night ran warmer than guided and the calm clear night ran colder than guided, and the models, ours included, cannot tell beforehand which kind of night it will be. That matches R-002 (decoupling) and R-007. The one thing the evening data did show: on the calm night the 11 F drop in the first hour and the 66 F to 42 F fall by 10 PM were already far steeper than the forecast curve, and the live adjustment moved the right way (35.6 to 34.4 F) but only a third of the way.

Next

Ideas for tests on OTHER nights (backlog): (1) a calm-night flag from the evening: if the wind is 0 and the first-hour fall after sunset is steep, does the low come in colder than our forecast? L-028 and L-026 already collect the fresh calm nights, so check them first instead of opening a new test. (2) Does the +/- need widening for freeze-risk nights when the wind forecast is near the calm threshold? Explore on discovery years in R-002 before any pre-registration. Asos.parquet stops at Oct 17: run research_data.py asos next session.

C-003case study, scored

Light radiation fog over the field that no airport saw (night of October 19-20, 2026)

opened October 20, 2026 · closed October 20, 2026 · by the Claude, with Brian, in a session
The question

A dead-calm, clear night after a 71 F day. Brian saw dew on the grass at 8:45 PM and light, wispy fog at 7:15 AM (camera footage suggests light fog from about 2:15 AM, unconfirmed); heavy dew but no frost visible at 8 AM. The statistical forecasts said 38 F, the physics 35.6 to 37.3 F. What happened, which forecasts were closest, and why?

Why it might work

The physics is meant to earn its keep in regimes the statistics rarely see (R-011, pattern 7): calm, saturated nights with fog are one. It was also the first live night of the overnight restart (L-030), 1G (L-026) and the grass-start runs (L-028).

The plan, written before the test

Read the station's minute log (Weather Display: air, humidity, dew point, wind), the grass and soil probes, the six nearest airports' visibility reports (Iowa Environmental Mesonet: Marion, Huntington, Kokomo, Fort Wayne, Anderson, Muncie), every forecast and trial logged on the Tower, the sunset run's heat budget (lab/checks/2026-10-19_dew.json) and Brian's observations. This night is also a fresh night of the running fog test L-012, looked at here at Brian's request; L-012's method is frozen with no fitted weights, so nothing in it can change, and the night stays in it.

What counts as a pass, written before the test: Case study: no pass/fail. Findings become ideas for proper tests on other nights.
Results

Calm all night (not one minute of wind after 7:53 PM). The air fell 13.4 F in the 2 hours after sunset (60.3 to 47.7 F), then slowed as it closed on a falling dew point: humidity 95% from about 2 AM and 96-97% from 3:37 AM (no higher that night), the air creeping from 40 to 37 F between 1 and 6 AM, then the low of 35.5 F at 7:48 AM, at sunrise. None of the six airports reported fog (visibility above 5 miles all night): the fog was local to the field. The grass probe sat 4 to 6 F under the air's dew point from 7 PM to morning (dew all night), at or below 32 F from 4:47 to 8:12 AM, lowest 30.9 F at 6:57 AM; Brian saw heavy dew and no frost at 8 AM (checked on the shaded east side of the house, not at the mast). Closest to the low: physics version 1 started from the measured grass (L-028) 35.6 F, the overnight restart with the physics (L-030 C2f) 35.7 F at 10 PM and 35.3 F at 2 AM, our live adjustment (L-024) 35.9 F at 10 PM; ours and the NWS 37.6 to 38 F, the global models 38.4 to 41.3 F.

The night's low (35.5 F at 7:48 AM, minute log; 35.6 F in the Tower's 5-minute record)
ForecastLow (F)Miss (F)
Physics version 1 from the measured grass at sunset (L-028) 35.6 0.0
Overnight restart with the physics, 10 PM (L-030 C2f) 35.7 +0.1
Overnight restart with the physics, 2 AM (L-030 C2f) 35.3 -0.3
Our live adjustment at 10 PM (L-024) 35.9 +0.3
WxSim 36.0 +0.4
Physics version 1 from the station at sunset 36.4 +0.8
Overnight restart without the physics, 10 PM (L-030 C1) 36.9 +1.3
Physics 1G (from the station or the grass) 37.2 to 37.3 +1.7
Ours a quarter toward the physics (L-007) 37.3 +1.7
Ours at sunset 37.6 +2.0
Ours (midnight issue), NWS 38.0 +2.4
GFS, HRRR, GEM 38.4 to 38.5 about +3
ICON, NBM 40.2 to 40.3 about +4.7
ECMWF 41.3 +5.7
The night at the station (every 2 hours)
TimeAir (F)Dew point (F)HumidityGrass probe (F)
7 PM 59.5 44.7 58% 42.7
9 PM 47.7 42.8 83% 38.0
11 PM 42.2 39.8 91% 35.6
1 AM 40.2 38.6 94% 34.2
3 AM 38.8 37.5 95% 33.1
5 AM 38.4 37.4 96% 32.3
7 AM 37.3 36.2 96% 31.1
7:48 AM 35.5 34.6 96% about 31
  • The physics' heat budget at 8:52 PM (lab/checks/2026-10-19_dew.json): at the measured state 1G's formulas would bring about 375 W/m2 into the grass from the air and dew, the steady real grass about 68; the stability damping barely bites. Led to the L-031 amendment.
  • Fog signals at sunset (the fog log): version 1's air came within 0.4 F of its dew point overnight, the models' within 3.9 F. L-012 scores version 3's signal, which is in the Tower's log and is not quoted here.
  • By L-012's own truths this night counts as no fog: no airport fog, and the station's humidity reached 97% only briefly (L-012's station truth needs 97% for 3 hours). Brian saw light fog. Corrected October 20: the sensor does read 98 to 99% in fully saturated air (every month 2023-2026); this night it stayed at 96 to 97%, so the 5 ft air stayed just short of saturation while light fog formed, likely shallow or patchy. A fog truth that needs 97% for 3 hours misses light fog like this.
  • Brian, about 8:20 AM: one blade on the southwest side of the house, about 250 ft from the mast, looked like frost (maybe); everywhere else heavy dew. Consistent with borderline, patchy frost at a grass probe reading of 30.9 F: frost forms first on the most exposed blade tips, while blades already wet with the evening's dew are slow to turn white.
  • Burn-off (Brian): the last real wisp at 8:57 AM, still hazy after. Sunrise about 7:58. The air held at its low (35.4 F at 8:20) until about 8:25, then rose 5 F to 40.5 F by 9:00; humidity rose from 96 to 98% and the dew point from 34.4 to 40.0 F as the heavy dew evaporated off the warming grass. The fog lasted about an hour after sunrise, against 10 to 30 minutes on October 4 and 5.
What it means

A textbook local radiation-fog night: dead calm, dew from early evening, the air pressing down onto a falling dew point until it saturated after 2 AM, with the cooling then held back by condensation until a last drop at sunrise. The fog stayed in the field; the airports never saw it. The physics family was closest (the measured-grass start 0.0 F, the restart with physics 0.1 F), the statistics 2 F warm and the global models 3 to 6 F warm, which is the regime where the physics should help. 1G ran 0.9 F warmer than version 1: 1G is built for the evening, and a fog-capped morning may be its weak side. The grass probe read 30.9 F while Brian saw no frost; with one look, away from the mast, this cannot be resolved: frost may have formed at the mast only, the probe at the blade tips may run colder than the grass within the canopy, or dew already on the blades froze clear rather than white.

Next

Ideas, each to be tested on other nights: (1) a fog and frost observation log at the mast from Brian (time, light or dense, frost or dew), the only truth for local fog and frost the airports cannot give; (2) when the grass probe's reading means visible frost, built from that log; (3) a fog truth from the station for a future test, checked against Brian's log, since 97% for 3 hours missed this light fog; (4) Brian's cameras as a fog detector (new information the models do not have); (5) 1G against version 1 by night type (saturated or fog nights) when L-026 is scored.

L-031scheduled (probe evenings from October 18, 2026; evaluate from November 16, 2026 once the conditions in the bar are met, at the latest December 15)

The grass's evening cooling, fitted on the grass probe (the root of the physics' evening lag)

opened October 19, 2026 · by the Claude, with Brian, in a session
The question

Every evening problem in the night physics traces back to how fast the grass cools around sunset (research R-010): the evening lag in the air, dew forming late, the dew point that will not drop. The history could not teach it (no grass sensor before October 17, 2026). With the grass probe, can the constants that couple the grass to the air and the soil be fitted directly on measured evenings, so that night version 1G gets the evening right on evenings it never saw?

Why it might work

R-011: rule 5 (fix the root, not the symptom: the grass is the first thing the model gets wrong, and the probe makes it observable), rule 1 (the probe is information the models do not have), rule 3 (fit and test weeks alternate, and the test must include clear and calm, cloudy and breezy evenings), rule 4 (real forecasts only: the inputs are the forecasts the Tower actually had at sunset). Power (rule 2): no probe history exists yet to measure the night-to-night spread of the gain; assuming 0.8 F, a 0.5 F gain on 12 test evenings would clear a 90% resample bar about 4 times in 5. The observed spread and the achieved power are reported with the result.

The plan, written before the test

Frozen October 19, 2026 (hyperlocal/grass_fit.py, run in the collector container on the live database, read-only). Model: night version 1G (phys_params_g.json, unchanged) started at sunset from the station's air and dew point and the probe's grass (as L-028), deep soil from the 4 and 8 inch probes, driven by the newest ECMWF + ICON snapshots before sunset. Fitted, on the fit evenings only: the five constants that couple the grass to the air and the soil (grass heat capacity c_surf, exchange coefficient c_h, stability damping ri_b, calm-wind floor u_min, soil conduction k_soil) and, by the amendment below, the two that set the night layer (z_f, its share of full depth at sunset, and z_tau, how fast it deepens); everything else as frozen for 1G. Loss: squared error of the air (station) and the grass (probe) every 10 minutes from sunset to sunrise, the first 4 hours counted double. Control: 1G as frozen, same start and inputs. Evenings: sunsets after the probe went into its mount (October 18, 2026 18:03 UTC); dry (models' rain under 0.02 in overnight and no measured rain); no snow (models' depth under 0.2 in); probe and station covering 80% of sunset to sunrise. Split fixed now: October 18 and 19 (already looked at) and every evening in an odd ISO week fit; every evening in an even ISO week tests (first test week October 26 - November 1). Weather tags (first 4 hours): breezy (station wind 6 mph or more), cloudy (models' cloud 60% or more), clear and calm (cloud under 30%, wind under 3 mph), other. The probe calibration check (ground.py) must read 'ok' before evaluation; if it finds an offset instead, the probe readings are corrected by that offset (stated with the result). AMENDED October 19, 2026, 9 PM, before any test evening (the first test week starts October 26; only fit evenings, October 18 and 19, had been looked at): z_f and z_tau added to the fitted constants, seven in all; the bar is unchanged. Reason, from that evening's heat budget (lab/checks/2026-10-19_dew.json): at the measured state (air 47.8 F, grass 38.5 F, dew point 42.8 F, the forecast's 1.9 mph) 1G's formulas would bring 375 W/m2 into the grass from the air and dew, while the steady real grass implies about 68 W/m2 (sky -88, soil +20); with the station's calm they bring about 1 W/m2. The stability damping barely bites (90% of normal stirring at a bulk stability of 1.0). The weak damping appears to compensate for the thick night layer, the only way the model cools the 5 ft air fast enough; freeing the damping without the layer could improve the grass while worsening the air.

What counts as a pass, written before the test: Evaluated from November 16, 2026 when all hold: at least 12 evenings in each set, at least 2 test evenings each clear and calm, cloudy and breezy, and the probe calibration ok; otherwise on December 15, 2026 with whatever exists (said so). The fit is run once, then scored once on the test evenings. Pass only if all hold: (1) the evening air (hourly, 1 to 4 hours after sunset) mean absolute error at least 0.5 F lower than 1G's; (2) that gain positive in at least 90% of 5,000 resamples of evenings; (3) the night's low no worse than 1G's by more than 0.1 F; (4) the evening grass error lower than 1G's (the root fixed, not only the symptom). Reported alongside: each weather tag, version 1 from the same start, the fitted constants against 1G's. If it passes, the proposal is a live run alongside 1G through the winter, not promotion.
L-030Closed continues as L-038

Live trial of the night restart from the station's reading (L-023 missed narrowly)

opened October 19, 2026 · closed October 22, 2026 · by the Claude, with Brian, in a session
The question

L-023 restarted the night physics at 10 PM, midnight and 2 AM from the station's own reading (the forecast clouds unchanged) and blended it with our low and the station's error. On the held-back nights it beat the station-error update alone at all three times, but at midnight only in 88% of resamples (90% needed). Does the restart hold up on fresh nights through late fall, winter and early spring, run live on the Tower?

Why it might work

Research R-011: rule 6 (a narrow miss with the right sign everywhere goes to a live trial) and rule 3 (L-023 was judged only on July to October). Power check (rule 2), from L-023's held-back nights: the per-night gain over the station-error update averaged 0.07 F with a night-to-night spread of 0.50 F at midnight, and 0.43 F for the average of the three restart times (they are correlated 0.43 to 0.76, so averaging helps). For a 0.07 F gain to clear a 90% resample bar with a fair chance needs about 150 to 200 nights: at 150 nights about 65% for midnight alone and 78% for the three-time average; at 175 nights 70% and 83%. So the main measure is the three-time average and the trial runs about six months. Even then a real 0.07 F gain fails about one time in five; a fail here means 'not shown', not 'no effect'.

The plan, written before the test

Frozen October 19, 2026, before any live night (Tower: physlive.restart; weights unchanged from physmodel/nightcloud_frozen.json, the C1 and C2f rows). Each night: the forecasts the Tower had at sunset (the newest ECMWF + ICON snapshot before sunset); night physics version 1 (the weather version on nights with rain, snow or snow on the ground) run from the station's sunset reading. At T = 10 PM, midnight and 2 AM local the air restarts from the station's reading at T (temperature and dew point), the grass keeps the gap to the air that the sunset run had built by T, and the run continues to sunrise with the same forecast clouds and winds. C1 = ours + a + b x e_s; C2f = ours + a + b x e_s + c x (restarted physics low - ours); each never above the lowest reading since sunset. ours = our forecast low as logged with the sunset run; e_s = the station minus the two models' 2 m temperature at T. Logged once per restart time, within 2 hours after T (phys_runs, kind 'restart'), together with the Forecast page's live low at that moment (L-024) for comparison. Truth: the station's lowest reading from sunset to sunrise. Nights: October 20, 2026 to April 30, 2027, every night with all three restarts logged. Dropped from L-023: the airports' reported clouds (C2), which added nothing. Differences from L-023, all fixed in advance: live snapshots instead of exactly the 12 UTC runs, ours as of sunset instead of the walk-forward day-1 low, the live soil probes for the deep soil.

What counts as a pass, written before the test: Evaluated in May 2027. No verdict below 150 nights: if fewer by April 30, the trial runs to May 31, 2027 and is then evaluated whatever the count, saying so. Main measure, the per-night average over the three restart times: pass only if all hold: (1) C2f at least 0.04 F lower mean absolute error than C1; (2) C2f lower than ours at each of the three times; (3) the gain in (1) positive in at least 90% of 5,000 resamples of nights. Reported alongside, not part of the verdict: midnight alone (L-023's measure), each time separately, fall (October-November), winter (December-February) and spring (March-April) separately, and C2f against the Forecast page's live low (L-024) at the same moments. If it passes, the proposal is to show the restarted low as the overnight update on the Physics page; it reaches the Forecast page only after a separate pre-registered test against the L-024 live adjustment, which already updates tonight's low from the station.
Results

October 21, 2026. Bug fix installed October 21, 2026, 10:02 PM EDT (02:02 UTC October 22), plan, constants and scoring unchanged: the Tower no longer stores a station reading whose station clock has not moved (a stalled Weather Display or upload), so no run starts from a repeated reading that looks current (code review 4.1). Before this, six such repeats were stored in the week of October 14 to 21.

October 21, 2026. Bug fix installed October 21, 2026, 10:02 PM EDT (02:02 UTC October 22): one daylight-time rule (second Sunday of March to first Sunday of November) replaces the month-only rule in the Tower code (code review 4.2). For this trial it affects only the replay check script (l030_check.py); the restarts themselves already used the correct rule.

October 22, 2026. Closed on 2026-10-22 without a verdict (the one-time closing pass, APPROACH-2026-10-21.md section 4a): the night restart from the station's reading is now part of the main model (v0.1), chosen on this trial's logged nights, which were its exploration set; the remaining question, whether the hourly updates after the forecast of record beat it, goes to the update-hours entry. Nothing was graded and no pass or fail is claimed; whatever this trial logged is exploration only and is never used to write a bar on the same nights. Its question continues as L-038.

L-029scheduled (fresh months November 2026 - February 2027; evaluate in March 2027)

Winter confirmation of the 6-hour nowcast (L-013), which was tested only from July to October

opened October 19, 2026 · by the Claude, with Brian, in a session
The question

The nowcast on the Forecast page (the station's own error now plus the upwind airports') passed on July 16 - October 11, 2026 and was promoted. Every held-out test so far used that same late-summer-and-fall stretch (research R-011, pattern 3). Winter changes the station's error (snow cover, frozen ground, long nights, fog). Does the frozen nowcast still beat the forecast alone, and do the upwind airports still add, in winter?

Why it might work

R-011 (October 19, 2026): all held-out tests were judged in one season. L-013 is the one promoted feature on the main Forecast page with no check outside July-October and no live trial running.

The plan, written before the test

Frozen October 19, 2026: the same replay as L-013 (physmodel/nowcast.py with nowcast_frozen.json, unchanged weights): forecast F from the newest ECMWF + ICON run out at least 6 hours before, the station's error now e_s, the upwind and downwind ring airports' errors now; N1 = own error, N2 = own + upwind (the version on the page), N3 = own + downwind (control). Test hours: November 1, 2026 to February 28, 2027, every hour with all inputs (archives topped up daily by the lab's research_data.py). Truth: the station's temperature at t + h. Reported separately: hours with snow on the ground (snow estimate) and without.

What counts as a pass, written before the test: Evaluated in March 2027 (no verdict below 60 days). Averaged over leads 1 to 6 h, the page's version must hold its place: (1) N2's mean absolute error at least 0.2 F lower than the forecast alone (in July-October it was 0.45 F lower); (2) N2's error no worse than N1's (the upwind airports do not hurt in winter); (3) the gain of (1) positive in at least 90% of 3,000 resamples of whole days. If (1) or (3) fails, the nowcast is switched off for the cold months (the page shows the forecast alone, with a note) until a winter version is tested.
Results

October 21, 2026. Before this trial starts (November 1): the nowcast's early-morning base was corrected October 21, 2026, 10:02 PM EDT (02:02 UTC October 22) (code review 4.6, see L-013), so the whole trial runs on the corrected base.

L-028running (fresh dry nights from October 19, 2026)

Starting the night physics from the MEASURED grass temperature at sunset (research R-010, R-007)

opened October 18, 2026 · by the Claude, with Brian, in a session
The question

The night physics starts the grass a fixed amount below the air at sunset (fitted before the grass probe existed). On October 18, 2026, 30 minutes after sunset, the grass probe read 37.8 F, 16 F under the air, while the physics had the grass at 55.9 F; the station fell 11.6 F in the first hour, the physics 4.2 F. Does starting from the probe's reading make the evening and the night better, for version 1 and for 1G?

Why it might work

First live look at the evening (October 18, 2026, the probe's first evening in its sunny mount, painted white, 2 in above the grass): at 7:29 PM station 53.9 F, physics 57.8 F, probe 37.8 F, physics grass 55.9 F; at 7:59 PM 48.8 / 56.2 / 36.8 / 54.6 F. Brian found dew on the grass at 9 PM; the forecast had it at about 10 PM. One night only: why this is a test and not a change.

The plan, written before the test

Frozen October 18, 2026, evening (physlive.py, kind 'sunset' record, field 'gs'). Each sunset from October 19, 2026 with a grass-probe reading within 15 minutes of sunset, the Tower also runs version 1 and 1G from the station's air and dew point AND the probe's grass temperature as the starting grass (the hourly restarts already do this; the tested sunset runs do not). Constants unchanged (fitted with the fixed offset). Compared with the same night's runs without the probe (version 1's sunset run; 1G's 'g'). Nights: dry (0.02 in of rain or less), bare ground (no snow on the probe). Truth: the station's 5-minute readings; for the grass, the probe itself. Scores per night: hours 1-6 after sunset (mean absolute error of the hourly values), the whole night (root mean square error), the low (absolute error).

What counts as a pass, written before the test: Evaluated at 40 such nights (or on April 30, 2027 with whatever is there; no verdict below 25). For each version, pass only if all hold: (1) hours 1-6 at least 0.3 F better than the same version without the probe; (2) the whole night better; (3) the low no more than 0.1 F worse; (4) the hours 1-6 gain positive in at least 90% of 5,000 resamples of nights. If it fails narrowly, the next step is a refit of the starting constants on probe-started nights (the constants were fitted with the fixed offset), as a new test.
Results

October 21, 2026. Bug fix installed October 21, 2026, 10:02 PM EDT (02:02 UTC October 22), plan, constants and scoring unchanged: the Tower no longer stores a station reading whose station clock has not moved (a stalled Weather Display or upload), so no run starts from a repeated reading that looks current (code review 4.1). Before this, six such repeats were stored in the week of October 14 to 21.

October 22, 2026. Note, 2026-10-22: version 1G was closed as a trial (L-026) and its code leaves the live physics, so from the next Tower install this entry's 1G comparison stops; its version 1 comparison (the night started from the measured grass against the fixed offset) continues unchanged, and the evaluation reports the 1G half on the nights it has.

L-027running (part A done: not met; fresh parts fall 2026, 2027 and 2028)

Standing corn in fall: does the daytime high run warm until harvest? (research R-004)

opened October 18, 2026 · by the Claude, with Brian, in a session
The question

The field just south of the mast rotates corn (even years) and soybeans (odd years). In both corn falls on record (2024, 2026) our day-1 high came in about 1.3 F warmer than forecast from mid-September until harvest, and in the soybean fall (2025) it did not. Is that a real effect of the standing, drying corn, so it can be corrected?

Why it might work

Crop check (R-004, October 18, 2026): mid-September to mid-October our high came in +1.43 F (2024 corn), +0.02 F (2025 beans), +1.18 F (2026 corn), against +0.58, +0.07, +0.18 F from mid-August to mid-September. A plausible mechanism: a dry, senescent standing canopy evaporates almost nothing and shelters the mast. Caveats: two corn falls, one bean fall; fall 2024 was also dry.

The plan, written before the test

Frozen October 18, 2026, BEFORE any 2018-2020 data was read. Proposed correction: our day-1 high + 1.3 F on corn-year days from September 15 until harvest (harvest from Brian's field log). Three parts. (A) HISTORICAL, independent forecasts: WxSim's archived runs (Brian's archive, 2018-2023) against the station's NOAA monthly reports (2018-2020); the corn falls with both are 2018 and 2020 (the archive has no soybean fall, so there is no crop control; each year is compared with itself). WxSim day-1 high = the last run issued before 7 AM local on the day; truth = the report's daily high. Lean = mean of (actual - WxSim). Per corn year: jump = lean(September 15 - October 13) - lean(August 15 - September 14). (B) FRESH 2026: our graded day-1 high from October 19, 2026 until harvest (miss review), mean of (actual - ours), reported, no verdict (too few days). (C) FRESH full test: fall 2027 (soybeans) and fall 2028 (corn), our graded day-1 high.

What counts as a pass, written before the test: (A) consistent with the effect only if the jump is at least +0.5 F in both 2018 and 2020 and the pooled jump is positive in at least 90% of 5,000 resamples of days; it can support but not prove the effect (no soybean control). (C) the verdict: pass only if all hold: fall 2027 (beans) lean between -0.5 and +0.5 F over September 15 - October 13; fall 2028 (corn) September 15 until harvest: the +1.3 F correction lowers our high's mean absolute error by at least 0.2 F, positive in at least 90% of 5,000 day resamples. Evaluated after the 2028 harvest. If it passes, the correction goes into the forecast, switched by the field log.
Results

Part A (historical), run after the plan was published: only 2018 could be scored (the station's NOAA reports in the archive stop in June 2020). 2018, corn: actual minus WxSim's day-1 high +1.70 F from August 15 to September 14 (30 days), -0.25 F from September 15 to October 13 (28 days): a jump of -1.94 F, the OPPOSITE of the effect; October 14-31 -1.66 F, November -3.21 F. Bar (A) not met. One year, with no soybean control and WxSim's own seasonal habits, so it cannot rule the effect out, but it does not support it: the 2024 and 2026 pattern may be a quirk of our own forecast in early fall or the dry fall of 2024. Parts B and C continue as planned (research/l027_historical.py, l027_historical.out).

L-026Closed

Night physics 1G: an evening layer that starts shallow and deepens, for the shape of the night (research R-010)

opened October 18, 2026 · closed October 22, 2026 · by the Claude, with Brian, in a session
The question

Version 1 gets the low right but cools too slowly for the first hours after sunset (up to 3 F too warm on clear, calm fall nights) because its cooling layer is a fixed, deep block. Version 1G lets the layer start shallow at sunset and deepen through the night. On fresh nights, does it follow the evening better without losing the low?

Why it might work

Night shape check (R-010, October 18, 2026): version 1 did 41% of the night's cooling in the first 3 hours after sunset, the station 54-57%. The early evening transition (Acevedo and Fitzjarrald 2001: changes in the first 1-2 hours after sunset many times larger than later, as turbulence dies and the mixed layer collapses into a shallow stable layer) is what a fixed-depth layer cannot do. 1G fitted on dry discovery nights (March 2024 to September 2025, near-analysis inputs) with version 1's loss; on the real-run exploration nights (April 2 to July 15, 2026, not fitted on) hours 1-6 after sunset: 1.64 F (v1) to 1.54 F (1G), whole-night 2.02 to 1.93 F, low 1.69 to 1.69 F. A wind-aware variant (1GW) was also fitted and set aside: on those nights it did worse than 1G on the early evening, the point of the change.

The plan, written before the test

Frozen October 18, 2026 (physmodel/scm_g.py; constants phys_params_g.json = physmodel/params_g.json; z(t) = Z_EFF x (Z_F + (1 - Z_F)(1 - exp(-t/Z_TAU))), Z_EFF 247 m, Z_F 0.343, Z_TAU 1.6 h). On the Tower, every sunset run also runs 1G from the same station reading with the same forecast inputs (physlive.py, logged in the same phys_runs record as 'g'). Nights: dry (0.02 in of rain or less sunset to sunrise) from October 19, 2026. Truth: the station's 5-minute readings. Scores per night: hours 1-6 after sunset, mean absolute error of the hourly values; the whole night, root mean square error; the low, absolute error. AMENDED October 18, 2026, evening, before any scored night (the first is October 19): the clear-sky longwave formula in 1G changed from Brutsaert 1975 to Dilley and O'Brien 1998, chosen by a rule fixed before the comparison (replace Brutsaert only if, on the real-run exploration nights not fitted on, it beats it on hours 1-6 and the whole night without worsening the low). On those 81 nights: hours 1-6 Brutsaert 1.54, Prata 1996 1.53, Dilley 1.49 F; whole night 1.93, 1.93, 1.88 F; low 1.69, 1.68, 1.66 F; Dilley's gain over Brutsaert positive in 100% of 5,000 resamples on both. Constants now physmodel/params_g_dilley.json (= phys_params_g.json): sky strength 0.88 of the published formula. The bar is unchanged.

What counts as a pass, written before the test: Evaluated at 60 fresh dry nights (or on April 30, 2027 with whatever is there; no verdict below 40). Pass only if all hold: (1) the hours 1-6 error at least 0.05 F lower than version 1's; (2) the whole-night error lower than version 1's; (3) the low no more than 0.1 F worse than version 1's; (4) the hours 1-6 gain positive in at least 90% of 5,000 resamples of nights. Reported alongside, no verdict: the July 16 to October 11, 2026 nights (held back, but already looked at for version 1's shape, which is what prompted this change). If 1G passes, it replaces version 1 on the Physics page and in the night blend's next test; the blend tested in L-007/L-008 is not changed by this test.
Results

Early indication only, no verdict: on the held-back real-run nights, July 16 to October 11, 2026 (69 dry nights, already looked at for version 1's shape), the amended 1G (Dilley sky): hours 1-6 after sunset 2.34 F (v1) to 2.10 F, gain 0.25 F positive in 100% of 5,000 resamples; whole night 2.34 to 2.15 F; low 1.96 to 1.88 F. (The first 1G, Brutsaert sky: 2.15, 2.20, 1.95 F.) All four bar items would be met. The verdict comes from fresh nights (physmodel/l026_heldback.py, l026_heldback_dilley.out).

  • Bug fix installed October 21, 2026, 10:02 PM EDT (02:02 UTC October 22), plan, constants and scoring unchanged: the Tower no longer stores a station reading whose station clock has not moved (a stalled Weather Display or upload), so no run starts from a repeated reading that looks current (code review 4.1). Before this, six such repeats were stored in the week of October 14 to 21.
L-025running (fresh days from October 19, 2026)

A wider +/- for the daytime high on days forecast cloudy or wet (research R-001)

opened October 18, 2026 · by the Claude, with Brian, in a session
The question

The 80% range on our daytime high is honest on ordinary days but too narrow on cloudy and wet ones. If the range is widened on days FORECAST cloudy or wet, by one factor fitted on past days, does it hold 80% on fresh days and score better than today's range?

Why it might work

Miss review look-back (R-001, October 18, 2026): on exploration days the high landed outside the range on 30% of days forecast 80%+ cloudy and 34% of days with 0.1 in of rain or more, against 21% on other days; the held-back months pointed the same way (40% and 55%, small samples). Grouping by what is known when the forecast is made: day 1, exploration days March 2024 to July 15, 2026, 320 of 834 days forecast cloudy or wet, inside the range 69% (other days 81%).

The plan, written before the test

Frozen October 18, 2026 (hyperlocal/research/r001_range_fit.py, r001_range_frozen.json). Group: the four core models' average day cloud 80% or more, OR 2 or more of GFS, ECMWF, ICON and GEM with 0.01 in or more of rain in the day window (7 AM to 7 PM standard time), from the model run in hand before the period began. Widened range = today's +/- x 1.253 on day 1 (x 1.179 on day 2) for days in the group; other days unchanged. Fresh days: daytime highs from October 19, 2026, day 1, graded on the forecast issued before the period began, from the Tower's miss review (hyperlocal/misses.py: ours, pm, obs, forecast cloud, wet_votes). Nothing on the pages changes during the test.

What counts as a pass, written before the test: With at least 60 fresh days in the group (evaluated then, or on April 30, 2027 with whatever is there; no verdict below 40): pass only if both hold in the group: (1) the share of highs inside the widened range is between 74% and 86%; (2) the 80% interval score (Winkler: width plus 10 x the distance outside when it misses) is lower with the widened range than with today's in at least 90% of 5,000 resamples of days. Other days are reported alongside (unchanged by the rule). If it passes, the pages use the widened range.
L-024passed (October 15, 2026); promoted October 18, 2026; live check running

Updating today's high and tonight's low with how the station is running, once the period has begun

opened October 15, 2026 · by the Claude, with Brian, in a session
The question

During the day, does the gap between the station's current temperature and our forecast curve for that hour tell us how today's high will end up (and after the evening start, tonight's low)?

Why it might work

Written into the notebook on October 18, 2026: the test itself was pre-registered and run on October 15 (hyperlocal/data/livetrack_prereg.txt, livetrack_results.txt) and its result drove the Today page's live tracker, but the Forecast table kept showing the hourly rebuild, which never looks at the station. On October 18 at 5:33 PM it still showed a high of 64 F after the station had reached 68 F. Brian asked for the forecast to be calibrated to the station once a period is underway, and for proof on real days.

The plan, written before the test

Pre-registered October 15, 2026 (hyperlocal/livetrack.py). Forecast curve for an hour = the 4-model hourly blend (lead 1) shifted so its peak matches our forecast high (trough for the low). Running gap e = station temperature now minus the curve now. Adjusted high at hour h = max(highest so far, our forecast from before the window + a_h + b_h x e), a_h and b_h fitted per hour (lows: the mirror). Walk-forward: a_h, b_h fitted only on days before each test month (8-day gap); test October 1, 2025 onward. Baselines: (1) our morning forecast; (2) the so-far rule, max(highest so far, our forecast). Promoted October 18, 2026 with Brian's OK: the Forecast table, Today cards and discussion show the adjusted number once a period has begun (with its tested 80% range), the forecast made before the period began shown under it and still the one the main scoreboard grades. Live check: every hour the adjusted value (ours_live) and the so-far rule (ours_floor) are logged; How it works grades them at local 11 AM and 2 PM (highs) and 9 PM and 11 PM (lows), same days in every column, beside the hourly rebuild and the NWS at the same moment.

What counts as a pass, written before the test: Test bar (October 15): show it only if, averaged over 10 AM to 4 PM (highs) or 9 PM to 5 AM (lows), it cuts mean absolute error by at least 10% against the so-far rule, and is never worse than that rule by more than 0.1 F at any single hour. Live check (October 18): after 60 scored days, if the adjusted value is worse than the so-far rule at 2 PM or at 11 PM, the pages go back to the so-far rule and the result is written here.
Results

Test, October 1, 2025 to October 2026, walk-forward: highs 356 days, bar hours: so-far rule 1.25 F to adjusted 0.99 F (21% better), worst single hour 0.07 F better than the rule: PASS. Lows 353 nights: so-far rule 1.41 F to adjusted 1.17 F (17% better), worst hour 0.06 F better: PASS. The gain grows through the day as more of the running gap carries into the final number (high: 21% of it at 8 AM, 64% at 2 PM, 75% at 4 PM).

Today's high, mean absolute error (F) by local standard hour, test days
HourMorning forecastSo-far ruleAdjusted
10 AM 1.57 1.52 1.45
11 AM 1.57 1.49 1.32
12 PM 1.57 1.43 1.14
1 PM 1.57 1.32 1.00
2 PM 1.57 1.17 0.85
3 PM 1.57 0.97 0.63
4 PM 1.57 0.85 0.52
Tonight's low, mean absolute error (F) by local standard hour, test nights
HourEvening forecastSo-far ruleAdjusted
9 PM 1.70 1.61 1.54
11 PM 1.70 1.53 1.40
1 AM 1.70 1.45 1.18
3 AM 1.70 1.31 0.97
5 AM 1.70 1.12 0.72
  • Bug fix installed October 21, 2026, 10:02 PM EDT (02:02 UTC October 22) (code review 4.7): a high or low counts as the station's truth only after its window has ended (before, from 90% of its readings, so the night's low could be final about 15 minutes early and feed the recent-error feature in the hour before 9 AM). The scored checkpoints already waited for the window's end.
L-023failed

The night restart: does restarting the night physics from the station's reading during the night improve the low? With the airports' observed clouds? (research R-010)

opened October 18, 2026 · by the Claude, with Brian, in a session
The question

Brian asked for the night physics to use the clouds the ring airports actually report. The test design: run from sunset with the airports' reported low and mid cloud up to the restart time (or the forecast clouds, the control), restart the air from the station's reading, carry the cloud surprise forward (fading over 3 h), and blend with ours and the station's error now.

Why it might work

Exploration months (April 2 to July 15, 2026, real 12Z runs, 101 nights, cross-validated): station-error update C1 at 10 PM / midnight / 2 AM 1.40 / 1.31 / 1.12 F; + restarted physics with airport clouds C2 1.36 / 1.27 / 1.06; + restarted physics with forecast clouds C2f 1.36 / 1.25 / 1.04. The restart from the station helps; the airport clouds as substituted do not add (an airport's cover over one ceilometer is not the model grid box's cloud fraction the physics was tuned on).

The plan, written before the test

Frozen October 18, 2026 (physmodel/nightcloud.py, nightcloud_frozen.json fitted on all exploration nights). Night physics scmw (v1 + wet + snow) from sunset with the 12Z ECMWF + ICON runs; C2f: forecast clouds, air restarted at T from the station (dew point too; the grass keeps the run's gap); C2: the same with the ring airports' two lowest reported layers (low below 6,500 ft, mid to 12,000 ft, 1/distance weighted) up to T and the surprise fading after. C1 = ours + a + b e_s; C2/C2f add c (physics - ours); final = min(lowest so far, model). T = 10 PM, midnight, 2 AM. Nights: July 16 to October 11, 2026, excluding September 19 to 25.

What counts as a pass, written before the test: At MIDNIGHT, pass only if all hold: (1) C2f at least 0.04 F lower mean absolute error than C1; (2) C2f lower than ours; (3) the gain positive in at least 90% of 5,000 resamples of nights. C2 (airport clouds), 10 PM and 2 AM reported alongside. No verdict below 60 nights.
Results

77 held-back nights (July 16 to October 11, 2026). At midnight: ours 1.400 F, station-error update C1 1.269, restarted physics with forecast clouds C2f 1.201 (gain 0.068 F, 95% -0.04 to +0.18, positive in 88.4% of resamples), with the airports' clouds C2 1.208. Bars (1) and (2) met, (3) missed (90% needed): FAIL, narrowly. Alongside: 10 PM 1.53 / 1.42 / 1.34 F (91%), 2 AM 1.28 / 1.14 / 1.06 F (90%). The night restart helped at every time on months it never saw; the airports' reported clouds added nothing over the forecast's. The same narrow pattern as the afternoon update (L-021).

Held-back nights, the night's low, mean absolute error (F)
RestartOurs (6 PM)C1 station errorC2f restarted physicsC2 + airport cloudsShare positive (C2f)
10 PM 1.531 1.417 1.344 1.338 91.3%
Midnight 1.400 1.269 1.201 1.208 88.4%
2 AM 1.281 1.136 1.057 1.062 90.3%
L-022Closed continues as L-038

Live trial of the midday sunshine update (follows L-021)

opened October 18, 2026 · closed October 22, 2026 · by the Claude, with Brian, in a session
The question

L-021 found the midday physics restart with measured sunshine improved the afternoon high at 11 AM, noon and 1 PM on held-back months, but not consistently enough to call proven (86% of resamples at 1 PM). Does it hold on fresh days?

Why it might work

Brian asked to run it live and show it on the Physics page as 'Afternoon update (on trial)'.

The plan, written before the test

Frozen October 18, 2026: physlive.py midday() with phys_midday_frozen.json (= physmodel/sun_update2_frozen.json). Each day at 11 AM, noon and 1 PM local (computed within two hours after), from the forecasts the Tower had at sunrise: the L-020 day physics restarted with the station's measured sunshine (clientraw field 127) since sunrise and the last two hours' measured/forecast ratio carried forward (fading over 3 h); B1 = our sunrise-issue high + a + b x station error now; B2 adds c x (physics - ours); each final = max(the highest so far, the model). Logged in phys_runs (kind 'midday'). Days from October 19, 2026 with the update logged. Truth: the station's highest reading, sunrise to sunset.

What counts as a pass, written before the test: Evaluated once 100 days are logged at 1 PM (or on June 30, 2027 with whatever is there; no verdict below 60). At 1 PM, pass only if all hold: (1) B2 at least 0.05 F lower mean absolute error than B1; (2) B2 lower than our sunrise-issue high; (3) the gain over B1 positive in at least 90% of 5,000 resamples of days. 11 AM and noon reported alongside.
Results

October 21, 2026. Bug fix installed October 21, 2026, 10:02 PM EDT (02:02 UTC October 22), plan, constants and scoring unchanged: the Tower no longer stores a station reading whose station clock has not moved (a stalled Weather Display or upload), so no run starts from a repeated reading that looks current (code review 4.1). Before this, six such repeats were stored in the week of October 14 to 21.

October 21, 2026. Bug fix installed October 21, 2026, 10:02 PM EDT (02:02 UTC October 22): one daylight-time rule (second Sunday of March to first Sunday of November) replaces the month-only rule in the Tower code (code review 4.2). The midday restarts were placed an hour early from March 1 to the second Sunday under the old rule; from March 2027 they run at 11 AM, noon and 1 PM as the plan says.

October 22, 2026. Closed on 2026-10-22 without a verdict (the one-time closing pass, APPROACH-2026-10-21.md section 4a): the midday restart from measured sunshine is now part of the main model (v0.1: the half in progress starts from the station's latest reading, the departure fading over 3 hours), chosen on this trial's logged afternoons; the remaining question, whether the hourly updates after the forecast of record beat it, goes to the update-hours entry. Nothing was graded and no pass or fail is claimed; whatever this trial logged is exploration only and is never used to write a bar on the same nights. Its question continues as L-038.

L-021failed

The midday sunshine update: does restarting the day physics with the sunshine actually measured so far improve the afternoon high? (research R-010)

opened October 18, 2026 · by the Claude, with Brian, in a session
The question

Cloud surprises cost our forecasts on about half of days (R-010: much-darker-than-forecast days miss the high by 2.5 F). By midday the station's pyranometer knows how much sun actually came. Simply nudging our high by the morning's sunshine surprise did not help (correlation 0.09: morning clouds and afternoon clouds differ). Restarting the day physics with the measured sunshine so far, and scaling the afternoon forecast sunshine by the last two hours' measured/forecast ratio (fading over 3 hours), and blending it with ours and the station's current error: does that beat the station-error update alone?

Why it might work

Exploration months (April 2 to July 15, 2026, real 00Z runs, 100 days, cross-validated): at 1 PM ours 1.41 F, + station error (B1) 1.34 F, + restarted physics (B2) 1.20 F; at noon 1.45 / 1.48 / 1.40 F; at 11 AM no gain.

The plan, written before the test

Frozen October 18, 2026 (physmodel/sun_update2.py, sun_update2_frozen.json, fitted on all exploration days): at update time T (11, 12, 13 local), e_s = station temperature at T minus the 00Z ECMWF + ICON 2 m temperature at T; physics_T = the L-020 day physics (crop calendar, + 0.84 F) with the station's measured sunshine up to T and after T 0.907 x the forecast sunshine x (1 + (r - 1) x exp(-(t - T)/3 h)), r = measured/forecast over the last 2 hours (clipped 0.3 to 2). B1 = ours + a + b e_s; B2 = ours + a + b e_s + c (physics_T - ours); each final = max(the highest so far, the model). Days: July 16 to October 11, 2026 with the pyranometer covering the day, excluding September 19 to 25. Truth: the station's highest reading, sunrise to sunset.

What counts as a pass, written before the test: At 1 PM, pass only if all hold: (1) B2's mean absolute error at least 0.05 F lower than B1's; (2) B2 lower than ours; (3) the gain over B1 positive in at least 90% of 5,000 resamples of days. 11 AM and noon reported alongside. No verdict below 60 days.
Results

77 held-back days (July 16 to October 11, 2026). At 1 PM: ours 1.537 F, station-error update B1 1.438 F, + restarted physics B2 1.350 F; gain over B1 0.088 F (95% -0.08 to +0.26), positive in 86% of resamples: bars (1) and (2) met, (3) missed (90% needed): FAIL. Alongside: 11 AM 1.66 / 1.63 / 1.52 F (gain 0.11, 98% positive), noon 1.61 / 1.57 / 1.50 F (gain 0.07, 84%). The direction held at all three times on months it never saw; the evidence is not strong enough to call it proven. A live trial would add days.

Held-back days, the day's high, mean absolute error (F)
UpdateOurs (7 AM)B1 station errorB2 + restarted physicsGain B2 - B1Share positive
11 AM 1.660 1.629 1.521 0.108 97.8%
Noon 1.614 1.570 1.499 0.071 83.5%
1 PM 1.537 1.438 1.350 0.088 85.7%
L-020passed (in use for the Physics page's day high)

The day physics' cold bias: a crop calendar for evaporation, and an offset for real forecasts (research R-010)

opened October 18, 2026 · by the Claude, with Brian, in a session
The question

With real forecast runs the day physics ran about 1 F too cold. Exploration (April 2 to July 15, 2026, real runs) split it: the forecast inputs (sunshine timing, morning upper air, air moving in) about 0.6 F, the model itself the rest; and the model's error is SEASONAL, the same every year (too cold in spring and fall, too warm in July-August and winter): the evaporation followed a smooth season peaking mid-July, while real fields here are bare in spring, a closed transpiring canopy in July-August, and drying in the fall. Do a sharper crop calendar plus an offset for real forecasts fix it on months never looked at?

Why it might work

Crop calendar (second seasonal harmonic of the evaporative fraction, refitted with ef_c and ef_s on discovery dry days with measured sunshine): discovery 1.28 -> 1.19 F, 2023 (never fitted) 1.59 -> 1.41 F, 2026 exploration with real runs 1.96 -> 1.83 F (bias -1.04 -> -0.84 F). Disclosure: the overall -1.1 F bias over all 178 test-year days (which include the test months below) was seen before this plan; the fix and the offset were fitted without the test months.

The plan, written before the test

Frozen October 18, 2026: physmodel/dayscm.py, params_day_v1_obs_adv2.json + params_day_season.json (ef_c 0.063, ef_s -0.021, ef_h2c 0.035, ef_h2s 0.006), sunshine 0.907 x the ECMWF + ICON forecast, advection v2, driven by the 00Z ECMWF + ICON runs, started at sunrise from the station (daybias.py); then + 0.84 F (day_offset.json: the mean bias on the exploration months). Compared with the L-010 version (same runs, no calendar, no offset). Days: July 16 to October 11, 2026, all days with the runs complete (wet and dry), excluding September 19 to 25. Reported alongside: both in a rule against ours refitted on the exploration months.

What counts as a pass, written before the test: Pass only if all hold: (1) physics alone at least 0.10 F lower mean absolute error than the L-010 version; (2) the size of the bias smaller; (3) the gain positive in at least 90% of 5,000 resamples of days. No verdict below 60 days.
Results

78 test-month days (July 16 to October 11, 2026, real 00Z runs, all days). Physics alone: L-010 version 1.786 F (bias -1.22), crop calendar + offset 1.591 F (bias -0.37); gain 0.194 F (95% -0.02 to 0.41), positive in 96.6% of resamples. All three bars met: PASS. The corrected physics alone now beats ours on these days (1.714 F). In a rule against ours (refitted on the exploration months) it adds nothing beyond the old one (1.61 vs 1.59 F): the old rule's intercept was already absorbing the bias. By month the bias went July -0.40 -> -0.28, August -1.84 -> -1.35, September -0.77 -> +0.53, October -1.61 -> +0.16 F: midsummer is still too cold.

Test-month highs, real runs, mean absolute error (F)
OursPhysics (L-010 version)Physics (calendar + offset)Rule oldRule new
1.714 1.786 1.591 1.589 1.610
L-019failed

The physics ensemble: does the night physics do better on three models' forcing (GFS + ECMWF + ICON) than on two? (research R-010)

opened October 18, 2026 · by the Claude, with Brian, in a session
The question

The night physics runs on the average of the ECMWF and ICON forecasts. On discovery nights (each model's stitched first hours), running it once per model and averaging the results did NOT beat averaging the forcing first (2.12 vs 2.08 F, 4 models), even when the models disagreed about low cloud; but adding GFS to the averaged forcing did help: ECMWF + ICON 2.141 F, + GFS 2.062 F, + GEM instead 2.182 F, all four 2.077 F (324 dry nights). GFS runs cold and ECMWF warm; together they cancel. Does that hold with real forecasts?

Why it might work

Brian asked for an ensemble; the useful part turned out to be which models go into the average, plus a confidence flag: the spread of the per-model lows was a modest warning (widest quarter of nights, about 7.5 F apart: 2.6 F misses, others about 1.95 F).

The plan, written before the test

Frozen October 18, 2026: physmodel/scmw.py with params_v1.json, params_wet.json, params_snow.json, started at sunset from the station as L-007. Forcing E2: the mean of the ECMWF and ICON 12Z runs (as live). Forcing G3: the mean of the GFS, ECMWF and ICON 12Z runs (GFS surface fields from the single-runs API, model gfs_global, fetched October 18 for this test; GFS upper air from the saved ncep_gfs025 runs). Ours: the walk-forward low, lead 1. Rules: ours + 0.25 x (physics - ours) with each. Nights: April 2 to October 11, 2026, dry (station 0.02 in or less), excluding September 19 to 25; nights where all three runs are complete. Reported alongside: the per-model physics lows, their spread against the size of the miss.

What counts as a pass, written before the test: Pass only if all hold: (1) the physics alone with G3 at least 0.05 F lower mean absolute error than with E2; (2) the rule with G3 no worse than the rule with E2; (3) the physics-alone gain positive in at least 90% of 5,000 resamples of nights. No verdict below 100 nights.
Results

148 dry test-year nights (real 12Z runs). Physics alone: ECMWF + ICON (E2, live) 1.785 F; + GFS (G3) 1.848 F; gain -0.062 F (95% -0.19 to +0.07), positive in 16% of resamples. Rule: 1.445 vs 1.455 F (ours 1.524). FAIL. GFS's help on the discovery years came from its first hours (near-analysis), not from its forecasts. The live forcing stays ECMWF + ICON. Per-model physics alone: GFS 2.43, ECMWF 2.18, ICON 1.86 F; the average beats every member. The spread of the three member lows is a modest warning: nights in the tighter half (spread under about 3 F) missed by 1.6 F, the wider half by about 1.95 F (correlation 0.24).

Dry test-year nights, low, mean absolute error (F)
OursPhysics E2 (live)Physics G3Rule E2Rule G3GFS aloneECMWF aloneICON alone
1.524 1.785 1.848 1.445 1.455 2.43 2.18 1.86
L-018failed

The calibration run: does the station's own recent days correct the ground's wetness for the day physics? (research R-010)

opened October 18, 2026 · by the Claude, with Brian, in a session
The question

The day physics takes the ground's wetness from a grid estimate of topsoil moisture (about 15 miles across). For each past day, the share of the sun's energy going into evaporation that best reproduces the station's actual warming curve (with the measured sunshine) says how wet the ground really behaved. Does a slow estimate from the previous two weeks improve today's high?

Why it might work

On 1,245 days with the pyranometer, a single day's inferred correction is mostly that day's own noise (correlation with the day before 0.13), but a two-week average holds a slow signal (0.23), and calm, clear days carry it best (0.30 from one quiet day to the next, half the noise). Discovery years, cross-validated: every block of months chose half the estimate (k = 0.5); physics alone 1.465 -> 1.417 F; in the L-010 rule against ours 1.404 -> 1.362 F (gain 0.042, 95% 0.022 to 0.062). Disclosure: the persistence statistics above were computed over all years, test year included; k and the method were chosen on discovery years only, and no test-year forecast has been scored.

The plan, written before the test

Frozen October 18, 2026 (physmodel/calib_run.py, calib_test.py, l018.py). Each day's inferred offset: the day physics (L-010 constants, advection v2) with the station's measured sunshine, offsets -0.45 to +0.45 in steps of 0.05 added to its evaporative fraction, scored on the 5-minute curve from sunrise to the observed high; days need the pyranometer for 90% of daylight. Today's estimate: the weighted mean of the previous 14 days' offsets (quiet days, cloud under 25%, 10 m wind under 4 m/s, 925 mb change under 1.5 K, weight 1; others 0.25; at least weight 2 in the window). Forecast: the L-010 day physics driven by the 00Z ECMWF + ICON runs (as l015.py), evaporative fraction + 0.5 x estimate. Rule C: ours + 0.30 + 0.54 x (calibrated physics - ours); rule U: the same with the usual physics. Days: April 2 to October 11, 2026, all days (wet and dry) with an estimate, excluding September 19 to 25, 2026.

What counts as a pass, written before the test: Pass only if all hold: (1) rule C's mean absolute error at least 0.02 F lower than rule U's; (2) the gain positive in at least 90% of 5,000 resamples of days; (3) at least 120 days.
Results

178 test-year days (April 2 to October 11, 2026, real 00Z runs). Rule U (usual physics) 1.526 F, rule C (calibrated) 1.531 F: gain -0.004 F (95% -0.03 to +0.02), positive in 37% of resamples. FAIL. With real forecasts, the day physics' errors come from the forecast sunshine and air mass; the ground's slow wetness signal (worth 0.04 F with near-perfect inputs) is lost under them. The grid's soil moisture stays as it is. By-product: the unchanged L-010 rule beat ours on all 178 days, 1.606 -> 1.526 F (dry and wet together); the live L-010 test decides that properly on fresh days.

Test-year highs, all days, mean absolute error (F)
OursPhysics (usual)Physics (calibrated)Rule URule C
1.606 1.886 1.889 1.526 1.531
L-017passed

Snow cover: does the night physics with snow insulation get snow nights right? (research R-010)

opened October 18, 2026 · by the Claude, with Brian, in a session
The question

Version 1 of the night physics knows nothing about snow on the ground and ran +5.5 F too warm on last winter's nights with 2 inches or more. Snow conducts heat a fraction as well as soil, cutting the ground off from the soil's warmth, and a snow surface cannot warm past 32 F. With those added, is the physics right on a winter it never saw?

Why it might work

Learning winters (March 2023 to September 2025, 40 dry snow nights, only 3 with 5 cm or more), each winter scored with constants fitted on the other: without snow 2.69 F (bias +1.71 F), with snow 2.35 F (bias -0.36 F). Fitted on all 40: snow conductivity 0.118 W/m/K (the textbook range for settled snow), snow-surface heat capacity about 40,000 J/m2/K (the fit did not support a thin, fast-chilling surface).

The plan, written before the test

Frozen October 18, 2026: physmodel/scmw.py with params_v1.json, params_wet.json and params_snow.json (k_snow 0.1177, c_snow 39,676), full cover from 2 cm. Nights: the held-out winter, October 1, 2025 to April 30, 2026, dry (station 0.02 in or less and the models' precipitation under 0.02 in), with the models' snow depth 1 cm or more at sunset. Inputs: the stitched models' first hours and their snow depth (perfect prog: no real runs exist for that winter, so this checks the PHYSICS, not a forecast; the forecast test is live from this winter with Brian's snow reports). Compared: the same physics without snow cover. Truth: the station's lowest reading, sunset to sunrise.

What counts as a pass, written before the test: Pass only if all hold, on nights with the models' snow 5 cm or more: (1) mean absolute error at least 0.5 F lower with snow cover than without; (2) the size of the bias smaller; (3) the gain positive in at least 90% of 5,000 resamples of nights. No verdict with fewer than 15 such nights. All snow nights (1 cm or more) are reported alongside.
Results

Held-out winter 2025-26, 23 dry nights with the models' snow 5 cm or more: without snow cover 6.14 F (bias +5.79 F), with snow cover 4.53 F (bias +2.23 F); gain 1.61 F (95% 0.06 to 3.15), positive in 97.8% of resamples. All three bars met: PASS. All 43 snow nights: 4.59 -> 4.02 F, bias +2.62 -> +0.26 F. The snow physics is a large step, but deep-snow nights stay the hardest: the models' own low was 2.77 F on the same nights, and a warm bias of about 2 F remains (the learning winters had only 3 nights this deep, so deeper snow's insulation is likely underestimated). This winter's ground probes, under real snow, will measure it.

Held-out winter 2025-26, night low, mean absolute error (F)
NightsWithout snow coverWith snow coverModels' own low
Snow 5 cm or more (23) 6.14 (bias +5.79) 4.53 (bias +2.23) 2.77
All snow 1 cm or more (43) 4.59 (bias +2.62) 4.02 (bias +0.26) 2.12
L-016failed

Wet nights: does the night physics with wet-weather physics improve our low on rainy nights? (research R-010)

opened October 18, 2026 · by the Claude, with Brian, in a session
The question

Version 1 of the night physics runs far too warm on rainy nights (+1.7 F on average, +4.1 F with a quarter inch or more, discovery). The new wet-weather physics fixes most of it. On rainy nights with real forecast runs, does the L-007 rule using it improve our low?

Why it might work

The main cause was physics, not tuning: version 1 brings the 925 mb air down to the ground along the DRY adiabat; in rain the air below the cloud is saturated and follows the MOIST adiabat, about 2.5 K (4.5 F) colder at the ground on heavy-rain nights. With that, plus rain cooling the air toward its wet-bulb and a wet lawn, on 130 discovery wet nights (cross-validated): physics 2.62 -> 1.99 F (bias +1.71 -> +0.49); heavy rain 4.08 -> 2.13 F. Against ours (1.57 F): the rule ours + 0.25 x (physics - ours) 1.52 F, the same as with version 1 overall, better with heavy rain (1.70 -> 1.55, 29 nights), worse on light-rain nights (1.40 -> 1.46).

The plan, written before the test

Frozen October 18, 2026: physmodel/scmw.py with params_v1.json (dry) and params_wet.json (k_ev 7.1e-5, e_p 0.85, w_full 4.2 mm, p_moist 0.15 mm/h; snow melting and freezing rain at their physical starting values), driven by the mean of the ECMWF and ICON 12Z runs and started at sunset from the station (physmodel/l016.py, as l007.py plus precipitation and the 850 mb air). Rule W: ours + 0.25 x (wet physics - ours). Reported alongside: rule V1 (version 1 physics in the same rule) and the runs' own 2 m low in the same rule (control). Ours: the walk-forward low, lead 1. Truth: the station's lowest reading, sunset to sunrise. Wet nights: more than 0.02 in at the station between sunset and sunrise. Test-year days April 2 to October 11, 2026 (the real-run archive), excluding September 19 to 25, 2026 (a rain-gauge glitch in the station archive).

What counts as a pass, written before the test: Pass only if all hold: (1) rule W's mean absolute error at least 0.05 F lower than ours; (2) lower than the control; (3) the gain over ours positive in at least 90% of 5,000 resamples of nights. No verdict with fewer than 30 wet nights.
Results

30 wet test-year nights (real 12Z runs). The wet-weather physics fixed the physics itself: version 1 2.79 F (bias +2.55 F), with wet physics 1.94 F (bias +0.66 F). But in the rule against ours it adds nothing: ours 1.38 F, rule W 1.38 F (gain -0.01, 95% -0.20 to +0.19), rule V1 1.42 F, control 1.42 F. Bars (1) and (3) missed: FAIL. On rainy nights our forecast is already good (the lows are set by the rain and the air mass, which the models carry); the physics now gets them about right but brings no new information.

Lows on wet nights, mean absolute error (F)
OursPhysics v1Physics + wetRule WRule V1Control
30 wet nights 1.38 2.79 1.94 1.38 1.42 1.42
L-015passed

Wet days: does the day physics, as it stands, improve our high on days with rain? (research R-010)

opened October 18, 2026 · by the Claude, with Brian, in a session
The question

The live day test (L-010) scores dry days only. On rainy days, does the same day physics, driven by real forecast runs, improve our high?

Why it might work

A wet-weather add-on for the day model (rain cooling toward the wet-bulb, wet ground) made highs WORSE on discovery wet days (bias -2.0 F): the model already sees rain days through the models' dark sunshine and the soil moisture, and the add-on counted the water twice. Without it, on 166 discovery wet days with the models' first-hours sunshine (which flatters it, since it already saw when the rain came): physics 1.62 F, ours 2.14 F, ours + a share of (physics - ours) 1.63 F. The fair version is real forecast runs.

The plan, written before the test

Frozen October 18, 2026, before any test-year wet day is looked at with this version: the L-010 day physics exactly (dayscm.py, params_day_v1_obs_adv2.json, advection v2, sunshine 0.907 x the ECMWF + ICON shortwave), driven by the 00Z ECMWF and ICON runs of that day (physmodel/l015.py), started at sunrise from the station. Rule: high = ours + 0.30 + 0.54 x (physics - ours) (the L-010 rule, unchanged). Control: ours + 0.85 + 0.50 x (the two runs' own 2 m high - ours). Ours: the walk-forward high, lead 1. Truth: the station's highest reading, sunrise to sunset. Wet days: more than 0.02 in at the station between sunrise and sunset. Test-year days April 2 to October 11, 2026 (the real-run archive), excluding September 19 to 25, 2026 (a rain-gauge glitch in the station archive).

What counts as a pass, written before the test: Pass only if all hold: (1) mean absolute error at least 0.05 F lower than ours; (2) lower than the control; (3) the gain over ours positive in at least 90% of 5,000 resamples of days. No verdict with fewer than 30 wet days.
Results

43 wet test-year days (April 2 to October 11, 2026, real 00Z runs). Ours 2.08 F, physics alone 2.06 F, rule 1.74 F, control 2.04 F. Gain over ours 0.33 F (95% 0.10 to 0.58), positive in 99.7% of 5,000 resamples. All three bars met: PASS. Every month improved (April to September). The day physics, unchanged, is worth more on rainy days than on dry ones: our forecast misses rain-day highs by more, and the physics, started from the station at sunrise and fed the runs' dark sunshine, catches much of it.

Highs on wet days, mean absolute error (F)
OursPhysics aloneRuleControl (runs' own high)
43 wet days 2.08 2.06 1.74 2.04
L-014failed

Do the small effects add up? The inversion at 6 PM on top of the night physics rule (research R-010)

opened October 18, 2026 · by the Claude, with Brian, in a session
The question

Brian's question: many tested effects were each worth only about 0.02 F and were set aside. Added together, do they become something big? On discovery days (lead 1, cross-validated by 6 blocks of months), every small candidate was combined with our forecast in one penalised regression. The only small effect that survived on top of the night physics was the inversion strength at 6 PM (925 mb minus 2 m temperature). Does it improve the L-007 rule on real-run nights?

Why it might work

Discovery, 225 lows and 232 highs with every candidate: ours 1.71 F (lows), 1.70 F (highs). Lows: physics alone gains 0.10 F; all five small effects together (the other forecast program, inversion, soil moisture, ring error, upwind error) gain 0.06 F, nearly all from the inversion; physics + inversion 0.125 F; all six 0.11 to 0.12 F. Highs: all five small effects together LOSE 0.06 F; physics alone gains 0.23 F and adding the rest only takes away. The small effects mostly overlap with each other or do not carry 12 to 30 hours ahead (the ring and upwind errors help 1 to 6 hours ahead, L-013, not the next day).

The plan, written before the test

Frozen October 18, 2026, before looking at the inversion on any test night. Rule P = the L-007 rule: ours + 0.25 x (physics - ours). Rule P+I = rule P + 1.716 + 0.1852 x inv, where inv = 925 mb temperature minus 2 m temperature (F) at 22Z (6 PM EDT) in the mean of the ECMWF and ICON 12Z real runs; the two numbers fitted on all 472 discovery nights (residual of rule P, perfect-prog physics). Nights: the L-007 real-run nights (April 2 to October 11, 2026, dry, 149 nights), then fresh nights from October 19, 2026 logged by the Tower.

What counts as a pass, written before the test: Pass only if both hold on the test nights: (1) P+I's mean absolute error at least 0.02 F lower than P's; (2) the gain positive in at least 80% of 5,000 resamples of nights. Expected gain from discovery is small (about 0.02 F), so a no-pass is the likely outcome; the test is underpowered and is reported either way.
Results

149 real-run test nights (April 2 to October 11, 2026). Ours 1.52 F, physics rule P 1.38 F, P + inversion 1.37 F. Gain 0.014 F (95% -0.03 to 0.06), positive in 72% of resamples. Neither bar met: FAIL. The size is what discovery predicted (about 0.02 F); it is real-looking but too small to prove or to be worth the complication. Answer to the question: the small effects do not add up to something big here. They overlap, and most of them belong to the next few hours rather than the next day.

Lows, mean absolute error (F)
SetDiscovery (CV, 225 nights)Test (149 real-run nights)
Ours 1.71 1.52
+ physics 1.61 1.38 (rule P)
+ physics + inversion 1.59 1.37
+ all five small effects, no physics 1.65 not tested
L-013

A 1 to 6 hour nowcast: do the upwind airports add to the station's own error? (research R-006)

L-012Closed continues as L-043

Fog: does our night physics add to the models' own humidity forecast? (research R-010)

opened October 18, 2026 · closed October 22, 2026 · by the Claude, with Brian, in a session
The question

Made at sunset each night: the models' forecast of how close the 2 m air comes to its dew point overnight, and the same from our night physics with its moisture budget (version 3). Does adding the physics improve the ranking of fog nights over the models alone, on fresh fall-to-spring nights?

Why it might work

On discovery nights the physics was a weak fog signal (ranking skill 0.67 regional, 0.73 station saturation, version 3) against the models' near-analysis humidity (0.90), but that comparison flatters the models (their first hours have already seen the night). The fair rival is the models' FORECAST humidity at sunset, which the Tower logs from October 18.

The plan, written before the test

Frozen October 18, 2026 (physlive.py fog_signals and integrate_v3; phys_params_v3.json = physmodel/params_v3b.json). Each night the Tower's sunset run logs: models_min_dd = the smallest 2 m temperature minus dew point overnight in the ECMWF + ICON forecast (F); v3_min_dd = the same from the night physics version 3 started at the station's sunset reading. Score M = -models_min_dd; score C = the average of the two scores' ranks (no fitted weights). Truths: regional fog = visibility under 1 mile at Marion or Huntington between sunset and sunrise (airport archive, research_data.py asos); station saturation = the station's humidity at 97% or more for 3 hours or more. Nights: October 19, 2026 to April 30, 2027, dry (0.02 in or less).

What counts as a pass, written before the test: Ranking skill (AUC). Pass only if C beats M by at least 0.02 on BOTH truths. Valid only with at least 20 regional-fog nights (otherwise report, no verdict). Fog outcomes on fresh nights are not looked at before the evaluation.
Results

October 22, 2026. Closed on 2026-10-22 without a verdict (the one-time closing pass, APPROACH-2026-10-21.md section 4a): its question (does the night physics with a moisture budget add to the models' own humidity forecast for fog) continues as a single-change treatment of the main model: the version 3 moisture budget inside the main model's night, sized on the replay (R-011, October 22, 2026) and frozen as L-043 with the fog ranking against Brian's log as its secondary outcome. Nothing was graded and no pass or fail is claimed; whatever this trial logged is exploration only and is never used to write a bar on the same nights. Its question continues as L-043.

L-011

Rain, snow or ice: does the physics of the profile help? (research R-010)

L-010Closed continues as L-038

Our day physics with the models' sunshine, on fresh days (research R-010)

opened October 18, 2026 · closed October 22, 2026 · by the Claude, with Brian, in a session
The question

After L-009 failed on its cloud-to-sunshine step: with the sunshine taken from the models' own radiation forecasts (calibrated against the station's sensor), does the day physics improve the high we issue at sunrise, on days that did not exist when this was written?

Why it might work

On L-009's days the physics driven by MEASURED sunshine matched our high (1.48 vs 1.46 F), so the physics holds up and the sunshine input failed. The models' own radiation forecasts beat our cloud dimming against the sensor (clearness error 0.097 vs 0.136). On discovery days, the physics fitted on measured sunshine and driven by 0.907 x the models' sunshine, cross-validated: physics alone 1.63 F (ours 1.68), gain over ours + sunrise information 0.10 F (95% -0.01 to 0.21). Those April to October days are not reused.

The plan, written before the test

Frozen October 18, 2026: physmodel/dayscm.py; constants params_day_v1_obs_adv2.json (fitted on discovery days with measured sunshine, advection v2: 925 mb 0.71, 850 mb 0.02, 2 m 0.14); sunshine = 0.907 x the mean of the ECMWF and ICON shortwave forecasts in the Tower's newest snapshot; advection v2 with the discovery-year daily-cycle composites (phys_composite.json); start: the station's reading at sunrise; profile, clouds, wind, snow and ECMWF topsoil moisture from the newest snapshots. Logged once per day by the Tower (phys_runs kind 'day'), with our high as issued at that hour and the two models' own 2 m high. Rule: high = ours + 0.30 + 0.54 x (physics - ours). Control: ours + 0.85 + 0.50 x (the models' 2 m high - ours). Truth: the station's highest reading, sunrise to sunset. Dry days only (0.02 in or less of rain at the station, sunrise to sunset). Days from October 19, 2026.

What counts as a pass, written before the test: Evaluated once 120 dry days are logged (or on June 30, 2027 with whatever is there, no verdict below 120). Pass only if all hold: (1) mean absolute error at least 0.05 F lower than ours; (2) lower than the control; (3) the gain over ours positive in at least 90% of 5,000 bootstrap resamples.
Results

October 21, 2026. Bug fix installed October 21, 2026, 10:02 PM EDT (02:02 UTC October 22), plan, constants and scoring unchanged: the Tower no longer stores a station reading whose station clock has not moved (a stalled Weather Display or upload), so no run starts from a repeated reading that looks current (code review 4.1). Before this, six such repeats were stored in the week of October 14 to 21.

October 22, 2026. Closed on 2026-10-22 without a verdict (the one-time closing pass, APPROACH-2026-10-21.md section 4a): the day blend at sunrise belongs to the structure the main model replaced; whether the day physics helps the high is answered inside L-038 by the main model's day half and its calibration weight. Nothing was graded and no pass or fail is claimed; whatever this trial logged is exploration only and is never used to write a bar on the same nights. Its question continues as L-038.

L-008Closed

Our night physics, live: does it improve the low we issue at sunset? (research R-010)

opened October 18, 2026 · closed October 22, 2026 · by the Claude, with Brian, in a session
The question

On nights after October 18, 2026, run live on the Tower: does L-007's rule (ours + 0.25 x (physics - ours)) beat the low our forecast had issued at the same hour?

Why it might work

L-007 passed against our day-ahead low on April to October nights. Live, the physics starts at sunset from the station and our forecast is refreshed hourly with the newest runs, a tougher rival, and the season turns to winter.

The plan, written before the test

Frozen: physlive.py as deployed October 18, 2026 (phys_params.json = physmodel/params_v1.json). Each night the Tower logs, in table phys_runs (kind 'sunset'), the physics low from the station's sunset reading and our forecast low at that hour (the newest issue). Rule: ours + 0.25 x (physics - ours). Truth: the station's lowest reading, sunset to sunrise. Dry nights only (0.02 in or less of rain at the station). Our live-adjusted low on Today's Track is reported alongside, no bar.

What counts as a pass, written before the test: Evaluated once 120 dry nights are logged (or on June 30, 2027 with whatever is there, no verdict below 120). Pass only if: (1) mean absolute error at least 0.05 F lower than ours; (2) the gain positive in at least 90% of 5,000 bootstrap resamples.
Results

October 21, 2026. Bug fix installed October 21, 2026, 10:02 PM EDT (02:02 UTC October 22), plan, constants and scoring unchanged: the Tower no longer stores a station reading whose station clock has not moved (a stalled Weather Display or upload), so no run starts from a repeated reading that looks current (code review 4.1). Before this, six such repeats were stored in the week of October 14 to 21.

October 22, 2026. Closed on 2026-10-22 without a verdict (the one-time closing pass, APPROACH-2026-10-21.md section 4a): the quarter-way blend of the night physics into our low belongs to the structure the main model replaced (the physics nudging the statistical forecast); no outcome it could reach changes what the project does, and the physics-alone and corrected numbers it compared are reported by L-038. Nothing was graded and no pass or fail is claimed; whatever this trial logged is exploration only and is never used to write a bar on the same nights.

L-009failed

Does our own day physics improve the day's high? (research R-010)

opened October 18, 2026 · by the Claude, with Brian, in a session
The question

Our day physics model (version 1.1 with advection version 2), started at sunrise from the station's own reading and driven by real forecast runs: does a fixed share of it make our high more accurate on days it has never seen, and is any gain more than just having a fresher model run?

Why it might work

On discovery days (March 2024 to September 2025), with near-analysis inputs, the cross-validated day physics added 0.14 F to our high beyond the same sunrise information used statistically (95% interval 0.03 to 0.25 F), once advection was handled as the weather part of the models' 925 mb, 850 mb and 2 m change (the typical daily cycle removed). Near-analysis inputs flatter the physics; the honest test needs forecast inputs.

The plan, written before the test

Frozen before scoring (hyperlocal/physmodel/: dayscm.py, params_day_v1_adv2.json fitted on discovery days only; days.py with the daily-cycle composites from the discovery years). Days: sunrises April 2 to October 11, 2026, dry (0.02 in or less of rain from sunrise to sunset), with both models' 00Z run of that date (out before sunrise) and station data. Forcing: the mean of the ECMWF and ICON 00Z runs (clouds low, mid and high through our cloud dimming, 10 m wind, the 925/850/700 mb profile at sunrise, the weather part of the 925 mb, 850 mb and 2 m change, snow depth); start: the station's 5-minute reading at sunrise; topsoil moisture and deep soil: ERA5 at sunrise. Physics high = its maximum, sunrise to sunset. Ours = our walk-forward day-1 high. Rule (fixed on discovery): high = ours + 0.41 + 0.50 x (physics - ours). Control (fresher run, no physics): ours + 0.85 + 0.50 x (the two 00Z runs' own 2 m high - ours), its constant fitted on discovery the same way. Truth: the station's highest 5-minute reading, sunrise to sunset.

What counts as a pass, written before the test: Pass only if all hold: (1) mean absolute error at least 0.05 F lower than ours; (2) lower than the control; (3) the gain over ours positive in at least 90% of 5,000 bootstrap resamples of days. Sunny and cloudier days reported separately, no bar. Fewer than 120 days: report, no verdict.
Results

135 dry days scored (185 built), April 2 to October 11, 2026. Our day-1 high missed by 1.46 F; the frozen rule (ours + 0.41 + 0.50 x (physics - ours)) by 1.55 F, worse, positive in only 13% of resamples. The control, the fresh 00Z runs' own 2 m high, improved ours to 1.35 F. The physics alone missed by 2.08 F with a -1.23 F bias (cloudier days 2.65 F). No bar met: FAIL.

Day high, mean absolute error (F), 135 dry days
ForecastAllSunny (77)Cloudier (58)
Ours (day 1) 1.46 1.37 1.57
Ours + 0.41 + 0.50 x (physics - ours) 1.55 1.38 1.78
Control: fresh 00Z runs' 2 m high 1.35 1.26 1.47
Physics alone 2.08 1.66 2.65
What it means

Failed, and the diagnostic shows why: on the same 135 days with the same runs, the same physics driven by the sunshine the station's sensor actually MEASURED missed the high by 1.48 F (bias -0.59), level with ours. So the physics core and the new advection hold up on real forecasts; the cloud-to-sunshine step (our dimming of the runs' low/mid/high cloud) is what failed. It gets the total right on average (96% of measured) but misses individual cloudy days badly. The models' own sunshine forecasts are better (sensor comparison, R-010). Also notable: the fresh 00Z runs alone improve our day-1 high by 0.1 F, a reminder that a fresher run is real information.

Next

Rebuild the day model's sunshine on the models' own radiation forecasts (calibrated against the sensor), explore on discovery days only, then test on FRESH days (October 18, 2026 on, from the live log), not again on these April to October days.

L-007passed (the night blend shown on the Physics page; live check L-008)

Does our own physics model improve tonight's low? (research R-010)

opened October 18, 2026 · by the Claude, with Brian, in a session
The question

Our physics model of the night (version 1), started at sunset from the station's own readings and driven by real forecast runs: does a fixed share of it make our low forecast more accurate on nights it has never seen, and is any gain more than just having a fresher model run?

Why it might work

On discovery nights (March 2024 to September 2025), with near-analysis inputs, adding the cross-validated physics low to our day-ahead low cut the error from 1.63 to 1.56 F, while the same sunset readings used statistically added nothing (1.66 F); gain 0.11 F, 95% interval 0.05 to 0.17 F. Near-analysis inputs flatter the physics, so the honest test needs forecast inputs, which the real-run archive (ECMWF and ICON since April 2, 2026) provides.

The plan, written before the test

Frozen before scoring (hyperlocal/physmodel/: scm.py version 1, params_v1.json fitted on discovery nights only on October 18). Nights: sunsets April 2 to October 11, 2026, dry (0.02 in or less of rain between sunset and sunrise), with both models' 12Z run of that date and station data. Physics forcing: the mean of the ECMWF and ICON 12Z runs (clouds low, mid and high, 10 m wind, 925 mb temperature and height brought down dry-adiabatically, the change in 2 m dew point); start: the station's 5-minute reading at sunset; deep soil: ERA5 7-28 cm at sunset. Physics low = its minimum, sunset to sunrise. Ours = our walk-forward day-1 low (trained only on earlier days). Rule (fixed on discovery): low = ours + 0.25 x (physics - ours). Control (same rule, fresher run, no physics): ours + 0.25 x (the two 12Z runs' own 2 m low - ours). Truth: the station's lowest 5-minute reading, sunset to sunrise.

What counts as a pass, written before the test: Pass only if all hold: (1) mean absolute error at least 0.05 F lower than ours; (2) lower than the control; (3) the gain over ours positive in at least 90% of 5,000 bootstrap resamples of nights. Clear and calm nights reported separately, no bar. Fewer than 120 nights: report, no verdict.
Results

149 dry nights scored (187 built; nights with rain or missing data dropped as planned), April 2 to October 11, 2026. Our day-1 low missed by 1.52 F on average; the frozen rule (ours plus a quarter of the physics difference) missed by 1.38 F, a 0.14 F (9%) gain, positive in 99.9% of resamples (95% interval 0.05 to 0.23 F). The control, the same rule using the two fresh 12Z runs' own 2 m low instead of physics, did not help (1.54 F): the gain is the physics, not just a fresher run. The physics alone missed by 1.80 F, against 2.40 F for the runs' own 2 m low. All three bars met: PASS.

Night low, mean absolute error (F), 149 dry nights
ForecastAll nightsClear and calm (69)Bias
Ours (day 1) 1.52 1.18 -0.41
Ours + 0.25 x (physics - ours) 1.38 1.15 -0.14
Control: ours + 0.25 x (12Z runs' 2 m low - ours) 1.54 1.37 -
Physics alone 1.80 1.70 +0.66
The 12Z runs' own 2 m low 2.40 - +1.85
By month (mean absolute error, F)
MonthNightsOursRule
April 21 1.82 1.68
May 26 1.24 1.13
June 23 1.77 1.47
July 25 1.34 1.24
August 23 1.50 1.46
September 21 1.43 1.24
October 10 1.72 1.65
What it means

Our own physics adds real information to tonight's low, on nights it never saw and with real forecast inputs. Not a bias effect: with our bias removed in hindsight the rule still gains (1.49 to 1.37 F), and it gained in every month. Limits: warm season only (no winter nights yet); the physics starts from the sunset reading while our day-1 low was issued the day before, so part of the gain may be that fresh reading (the same reading used statistically added nothing on discovery nights); deep soil came from ERA5. Nothing changes in the live forecast yet.

Next

Run the physics live on the Tower every evening and log its low (October 18 on). L-008, to pre-register: against our live-adjusted low at sunset (the real competitor, which already uses the evening readings), on fresh nights through winter, before any blend goes into the forecast.

C-002case study, scored

The October 18-19 Frost Advisory night: the dew-point brake against the physics (the low was 34.2 F; our forecast 35.6 F, live-updated 34.4 F; the physics 32.1 F)

opened October 18, 2026 · by the Claude, with Brian, in a session
The question

A clear, calm, dry night with a Frost Advisory. The statistical forecasts (ours, the NWS) said 34 to 36 F, the physics 30 to 32 F. Which was closer, and why?

Why it might work

The two ways of forecasting disagreed by about 5 F for the same sky: a clean case of the dew-point 'brake' (statistics) against pure energy accounting (physics). Frozen calls: lab/checks/2026-10-18_frost_dew.json.

The plan, written before the test

Read the station's minute log (Weather Display, air, dew point, wind), the ground sensor (grass probe, mounted and painted that afternoon) and the physics runs logged on the Tower; Brian's observations at 9 PM and 7:54 AM. Case study: no pass/fail; findings become questions for proper tests.

What counts as a pass, written before the test: Case study: no pass/fail. Findings become ideas for proper tests on other days.
Results

The station's low was 34.2 F at about 7:45 AM (air), the grass probe's 29.8 F at 7:30 AM: grass below 32 F from about 4 AM, a very light frost at dawn (Brian, 7:54 AM). Our forecast issued before the night began said 35.6 F (1.4 too warm); updated with the station through the evening it said 34.4 F by 10 PM (0.2 too warm). The NWS said 36 F. The physics' sunset run said 32.1 F (2.1 too cold); our low moved a quarter of the way toward it (the tested L-007 rule) gives 34.7 F, closer than our forecast alone. The physics restarts started from the grass probe at 9-10 PM said 29.3-30 F (4-5 too cold). Frost call at crop height (31.0 F, physics grass) right; frost chance 81% right; freeze (air 32 F) 9% right. Dew formed about 2 hours earlier than the frozen call (on by 7:30-8 PM, seen at 9 PM; called 10 PM). Three things happened that matter: (1) the evening fell 11 F in the first hour while the physics fell 4 F (the evening lag, R-010: 8 F too warm at 10 PM); (2) after sunset the station's dew point ROSE 4 F (34 to 38 F) as the shallow layer trapped moisture, while the physics' dew point stayed near 34.7 F, so its dew-point brake engaged about 3 F too low and its morning ran colder; (3) from 11:45 PM to 1:15 AM puffs of 0.6-1.2 mph (gusts to 4) and the wind backing from south to south-southeast warmed the air 3.2 F (40.7 to 43.9 F) with the dew point unchanged: an intermittent mixing event that stirred warmer air down from the inversion, which no forecast here can predict. The dew point then fell from 36.5 to 32.4 F by morning (dew and frost drying the air) and the cooling slowed as the air closed on it (35.3 to 34.2 F from 6 to 7:45 AM).

The night's low (air, 5 ft): 34.2 F
ForecastLowOff by
Ours, issued before the night began 35.6 F +1.4
Ours, updated with the station by 10 PM 34.4 F +0.2
Ours moved 1/4 toward the physics (L-007) 34.7 F +0.5
NWS 36 F +1.8
Physics, sunset run (version 1) 32.1 F -2.1
Physics restarts from the grass probe, 9-10 PM 29.3-30.0 F -4 to -5
Hour by hour
TimeStation air / dew pointPhysics air / dew point (sunset run)Grass probe
7 PM 60.4 / 34.0 59.5 / 34.0 40.7
8 PM 49.6 / 38.0 55.8 / 34.7 36.8
10 PM 42.1 / 37.0 50.5 / 34.7 34.6
1 AM (mixing) 43.6 / 36.5 44.4 / 33.5 34.7
4 AM 39.8 / 34.7 38.9 / 32.7 32.0
7 AM 34.8 / 33.0 33.5 / 31.1 30.7
C-001case study, scored

The October 16-17 Freeze Warning night: what we said, and how it went (the low was 36.0 F; our forecast 31.2 F, 4.8 F too cold; corrected October 19)

opened October 17, 2026 · by the lab assistant on the Tower
The question

The NWS issued a Freeze Warning for the night ending the morning of October 17, 2026. How did our low forecast evolve over the four days, how did it compare with the NWS, the four-model average and WxSim, and what is the station doing? This session ran at about 2:20 AM local, before the low happened, so the final result is not known yet.

Why it might work

Freeze nights are the ones that matter most to a gardener. Research thread R-007 found a slight warm tendency near freezing; this is a live example to compare against.

The plan, written before the test

Read the forecast trail in hyperlocal.db (issued table, night_min for Oct 17 and Oct 18) and the station readings through 2:20 AM. No scoring: the night is not over. A follow-up scoring on the finished night is left to the next session.

What counts as a pass, written before the test: Case study: no pass/fail. Findings become ideas for proper tests on other days.
Results

Our forecast low for the night held between 31 and 34 F for four days and ended at 31.1 to 31.2 F (plus or minus about 3.1 F). The NWS said 33 F four days out, 32 F two days out and 31 F the afternoon before. The four-model average was 2 to 3 F warmer than ours throughout (33.9 to 34.0 F at the last issue), and WxSim wandered between 30.4 and 36.5 F (33.5 F at its last run I saw). So we sat with the NWS, below the raw models. At 2:20 AM the station read 43.0 F with a dew point of 39.0 F, wind 1.5 to 2.9 mph and falling about 1.2 F per hour. That pace would bring it to roughly 36 to 38 F by sunrise, well above 31 F, unless cooling speeds up. The dew point is also falling about 0.7 F per hour. SCORED October 18: the night was the one ending the morning of October 18 (the text above called it the night ending October 17, a slip). The station's low was 31.5 F. Our final forecast 31.5 F, the NWS 31 F, WxSim 29.5 F, the four-model average 34.2 F. At 2 AM the cooling pace had pointed above 31 F; the night finished on our forecast. Frost was on the grass and the roof at dawn (Brian). CORRECTION October 19: the SCORED lines above mixed up two mornings. The morning of October 17 (the night this entry is about) finished at 36.0 F, not 31.5 F: our final forecast of 31.2 F was 4.8 F too cold, the NWS 32 F, WxSim 33.5 F, the four-model average 34.0 F (all too cold, ours the most). The 31.5 F low and the frost on the grass belong to the NEXT morning, October 18, when our forecast was 34.9 F (3.4 F too warm; see C-002). The 2 AM cooling-pace guess in this entry was right.

Forecast low for the night of Oct 16 to 17, by issue time (local)
IssuedOurs (+/-)Four-model avgNWSWxSim
Oct 14 5 PM (4 days out) 33.2 (4.1) 34.8 33 -
Oct 15 9 PM (about 2.5 days out) 31.9 (4.7) 34.1 - 34.6
Oct 16 2 PM 31.3 (4.0) 33.3 31 30.4
Oct 16 6 PM 32.0 (3.3) 34.0 - 32.5
Oct 17 2 AM 31.2 (3.1) 34.0 - -
Station readings overnight (local)
TimeTemp FDew point FWind mph
Oct 16 6 PM 51.3 49.3 7.9
Oct 16 10 PM 48.6 44.4 4.1
Oct 17 12 AM 46.0 41.4 6.6
Oct 17 2 AM 43.5 39.4 2.5
  • Not scored: the night's low had not happened when this was written. Our low for the following night (Oct 18) is 36.3 F, which is above the frost-free line of 36 F only barely; the models are 2 F warmer.
  • Restored October 20, 2026: the lab assistant's correction of October 19 was overwritten online by a later publish from a session on the Mac (publishing then replaced the whole notebook); brought back from the Tower's copy. The two lows were checked against the station's minute log: 36.0 F at 8:13 AM October 17, 31.4 F at 5:25 AM October 18.
What it means

Preliminary. Our forecast correction pulled the cold side of the raw models by 2 to 3 F, in agreement with the NWS. Whether that was right depends on the final low. The cooling pace at 2 AM suggests we may have been too cold, which would fit the warm tendency near freezing that R-007 found, but this is a guess from half a night and proves nothing.

Next

Next session: read the finished low for Oct 17 (tmin_day_f in the obs table after sunrise) and the Oct 18 night, and append the error to this entry and to R-007. If we were warm-missed again, that adds one night to the fresh-month evidence for R-007 idea 4 (tested only on October 2026 onward). No change to the forecast is suggested by one night.

L-006running (collecting forecasts to March 31, 2027)

Does WxSim's forecast improve our highs and lows? (research R-009)

opened October 17, 2026 · by the Claude, with Brian, in a session
The question

Blending a small, fixed share of WxSim's forecast into ours: does it make our day 1-2 highs and lows more accurate, on months nobody has looked at?

Why it might work

WxSim is a physics model of the column over the station, a different approach from the global models we correct, and its mistakes overlap ours only partly (error correlation 0.36 to 0.52 on the discovery years). On those years, WxSim's own information was worth about 0.01 to 0.02 F where it helped (day-2 highs, lows days 1-3) and slightly hurt next-day highs. Small, but WxSim now runs reliably for the first time, so a clean test on fresh months is possible.

The plan, written before the test

Frozen method (hyperlocal/research/wxsim_input_test.py and wxsim_frozen.json as of today): our forecast + b x (WxSim - our forecast), no extra bias term, b fitted on discovery days: highs 0.109 (day 1) and 0.176 (day 2), lows 0.074 and 0.084. WxSim value = the newest WxSim run finished by our issue time (7 AM for highs, 6 PM for lows; at most 12 hours old), its max over our day window or min over our night window. The test year was deliberately NOT used: WxSim ran on only about 70 of its days, too few to detect an effect this size. Fresh issues: October 17, 2026 to March 31, 2027, run once our walk-forward forecasts cover them. Brian keeps WxSim's settings unchanged in this window (any change is logged with its date).

What counts as a pass, written before the test: Valid only with at least 300 forecasts that have a WxSim value (otherwise extend to June 30, 2027). Pass only if all hold, days 1-2 highs and lows combined: (1) mean error at least 0.02 F lower; (2) gain positive in at least 90% of resamples by day; (3) neither highs nor lows more than 0.01 F worse on their own.
L-005running (collecting fresh months to January 31, 2027)

Confirming the hour-by-hour forecast on fresh months (follow-up to L-003)

opened October 17, 2026 · by the Claude, with Brian, in a session
The question

Does the hour-by-hour correction that just missed its bar in L-003 (9.6% better than the strip against a 10% bar, better every month) hold up on months nobody has looked at?

Why it might work

L-003 showed a steady gain but used the test year, which many experiments have now seen. A pass on fresh months is the honest way to decide; the bar is set now, before any of those months exist.

The plan, written before the test

Exactly the frozen L-003 method (hyperlocal/research/r008_hourly.py, only the dates changed): 6 PM issues, next 48 hours, compared with the strip and the raw blend on identical hours. Fresh issues: October 17, 2026 to January 31, 2027. Run by the lab once the training data covers January 2027.

What counts as a pass, written before the test: Pass only if all three hold: (1) hours 1-24 at least 5% lower error than the strip; (2) hours 25-48 lower than the strip; (3) hours 1-24 better than the strip in each full month (November, December, January). If it passes, promotion still needs Brian and Claude to agree, and the hourly path reconciled with our daily high and low.
L-004complete, not promoted

Do the models' forecast inversions correct our lows? (research R-003, R-002)

opened October 17, 2026 · closed October 17, 2026 · by the Claude, with Brian, in a session
The question

When the forecast shows warm air sitting over cold air near the ground overnight (an inversion), does a correction learned from that inversion make our day 1-2 lows more accurate, using real forecasts available at issue time?

Why it might work

Our lows bust mostly on nights that go calm and stable at the ground (R-002). With hindsight upper-air data, the inversion strength lined up with our miss steadily: strong inversion nights ended 0.6-0.9 F colder than forecast, nights with none 0.45 F warmer (R-003). Operational models are known to mix stable night air too much (Sandu et al. 2013). The real question is whether the FORECAST inversion, known at 6 PM, carries that signal.

The plan, written before the test

Frozen method: hyperlocal/research/r003_inversion_test.py as of today. Features per night (11 PM to 5 AM local), mean of ECMWF and ICON: inversion = 925 mb minus 2 m temperature (F), and 925 mb wind (mph). The correction (one small ridge regression per lead) is learned only on the discovery nights (March 2024 to September 2025) from near-analysis upper air, already fitted: lead 1 -0.25 F and lead 2 -0.63 F per 4.4 F of inversion, +0.07 and +0.09 F per 12 mph of 925 mb wind. Applied unchanged to the real model runs: the 00Z run of the issue day, available long before the 6 PM issue. Test: our walk-forward lows (leads 1-2) for issue days April 2 to September 30, 2026 (test-year months, never used with these inputs; the runs start April 2). Confirmation, a separate later run: issue days October 17, 2026 to March 31, 2027, same frozen correction, same bar.

What counts as a pass, written before the test: Pass only if all three hold: (1) lows days 1-2 combined, mean error at least 0.02 F lower with the correction; (2) resampling nights, the gain is positive in at least 90% of resamples; (3) physics check: on nights with a forecast inversion of 6 F or more (3 F if fewer than 20 such nights), the correction lowers the forecast on average and lowers the error there. Promotion also needs the same three on the winter confirmation months.
Results

177 nights (348 forecasts, leads 1-2), issue days April 2 to September 30, 2026, real ECMWF and ICON runs. The correction moved the right way but too little: 0.013 F better overall (bar 0.02), positive in 74% of resamples (bar 90%). The physics check passed: on the 72 forecasts with a forecast inversion of 3 F or more, the correction lowered the low by 0.42 F on average and cut the error from 1.181 to 1.170 F. Fewer than 20 forecasts reached 6 F, so the 3 F threshold applied, as pre-registered.

Mean error of our lows, test months (F)
ForecastNightsOursWith the inversion correction
Day 1 172 1.406° 1.375°
Day 2 176 1.673° 1.677°
Days 1-2 348 1.541° 1.528°
  • Bar 1 (at least 0.02 F better): 0.013. FAIL.
  • Bar 2 (positive in 90% of resamples): 74%, 90% interval -0.021 to +0.047 F. FAIL.
  • Bar 3 (strong-inversion nights lowered and improved): -0.42 F, 1.181 to 1.170 F. Pass.
  • Day 1 improved 0.031 F; day 2 got 0.005 F worse. Noted, not acted on: picking day 1 alone after seeing this would be a forking path.
  • These are spring and summer months; strong inversions are a winter phenomenon, and only 72 forecasts had one of 3 F or more.
What it means

The signal is real in direction and physics (strong-inversion nights do end colder, and the correction knows it), but in spring and summer it is too small to clear the bar. Not passed, not promoted.

Next

The pre-registered winter confirmation (issue days October 17, 2026 to March 31, 2027) still runs with the same frozen correction: winter is when inversions are strongest, so it is the fair season to look. Because this test failed, a winter pass cannot promote this version by itself; it would justify a new pre-registered version, tested on further fresh months.

L-003complete, not promoted

Our own hour-by-hour temperature forecast for the next 48 hours (research R-008)

opened October 16, 2026 · closed October 16, 2026 · by the Claude, with Brian, in a session
The question

Can a station-corrected hourly temperature forecast beat what the Weather Briefing's 24-hour strip shows today (the model blend, shifted by our high/low correction for windows starting within 30 hours)?

Why it might work

The models have station habits by hour (how fast evenings cool, when the afternoon peaks) that a daily correction can't fix, and the station's own reading at issue time says how far off the blend is right now. The NWS does this operationally (LAMP, hourly National Blend). Exploring on the discovery years (March 2024 to September 2025) the correction scored 1.73 F per hour over the first 24 hours against 1.97 F for the strip and 1.89 F for the raw blend.

The plan, written before the test

Frozen method: hyperlocal/research/r008_hourly.py as of today. Issue at 6 PM EDT (22Z) each day; targets every hour for the next 48 hours; truth = the station's temperature within 5 minutes of each hour. Only information available at issue: model hours from runs at least 24 h (hours 1-24) or 48 h (hours 25-48) older than the target hour, so always before the issue time; our walk-forward daily forecasts; the station reading at issue. Method: the 4-model blend plus a ridge correction (one per lead hour) on our daily correction for the hour's window, the blend's error at issue, model spread, cloud, wind, dew point, each model's temperature, time of day and season. Walk-forward by issue month with an 8-day gap, training on everything before (discovery years included). Test issues October 1, 2025 to October 9, 2026, compared on identical hours with (a) the strip as the page builds it and (b) the raw blend.

What counts as a pass, written before the test: Pass only if all three hold on the test year: (1) mean hourly error over hours 1-24 at least 10% lower than the strip; (2) lower than the strip over hours 25-48; (3) lower than the strip over hours 1-24 in at least 9 of the 12 test months. Then confirm on fresh months (October 2026 on) before promotion. Note: if promoted, the hourly path must be reconciled with our daily high and low so the page never shows two different answers.
Results

Test year, 358 issues (October 1, 2025 to October 9, 2026), 16,938 hourly targets. The hourly correction was better than the strip every month, but over hours 1-24 it was 9.57% better, just short of the 10% bar. Criterion 1 fails; 2 and 3 pass. By the rules this does not pass.

Mean hourly temperature error, test year (F)
Hours aheadRaw blendToday's stripHourly correction
1-24 2.003° 1.946° 1.759°
25-48 2.380° 2.326° 2.196°
By month, hours 1-24 (F)
MonthToday's stripHourly correction
2025-10 1.823° 1.650°
2025-11 1.827° 1.673°
2025-12 1.824° 1.641°
2026-01 1.953° 1.759°
2026-02 2.404° 2.196°
2026-03 2.001° 1.767°
2026-04 2.224° 1.952°
2026-05 1.894° 1.666°
2026-06 1.858° 1.701°
2026-07 1.802° 1.592°
2026-08 1.824° 1.757°
2026-09 1.927° 1.725°
2026-10 1.968° 1.925°
  • Criterion 1 (hours 1-24 at least 10% lower than the strip): 9.57% lower. FAIL.
  • Criterion 2 (hours 25-48 lower than the strip): 2.196 vs 2.326. Pass.
  • Criterion 3 (better in at least 9 of 12 months): better in 13 of 13. Pass.
  • Resampling issue days: improvement 90% interval 7.7% to 11.4%; 37% of resamples reach 10%. The gain is real and steady; its size sits right at the bar.
  • Where it comes from: hours 1-6 21% better (the station reading at issue), hours 7-24 5% better. Every season better.
  • Conservative setup: model hours came from runs at least 24 h older than the target hour; the live system has fresher runs.
  • The discovery side finding (strip worse than the raw blend) did not hold on the test year: strip 1.946, raw blend 2.003.
What it means

A real, steady improvement (every month, every season, strongest in the first six hours), but smaller than the 10% we asked for in advance. It does not pass. The bar does not move after the fact.

Next

Proposed for Brian and Claude to decide: a confirmation test of the same frozen method on fresh months (October 2026 on), with a new bar written before those months are seen. Separately, improvements may be explored on the discovery years only (fresher model runs as the live system has them, a smoother treatment of lead hours), never tuned on the test year.

L-001complete, not promoted

Does the NWS blend (NBM) or the wind direction sharpen our highs and lows?

opened October 16, 2026 · closed October 16, 2026
The question

Our temperature correction knows wind speed but not wind direction, and it has never been offered the National Blend of Models (NBM), the NWS's own calibrated forecast. Do either of them help?

Why it might work

Wind direction carries air from somewhere else: an overnight south wind brings warmer, moister air and holds the low up, a northwest wind does the opposite (advection). NBM is already bias-corrected by NWS against thousands of stations, so where it disagrees with the raw models it may know something.

The plan, written before the test

Add to the current temperature model, one variant at a time: (A) NBM minus the models' average, with a flag when NBM is missing (it starts Oct 2024); (B) the four models' average wind components during the period, east-west and north-south; (C) both. Same walk-forward test as everything else, test period Oct 1 2025 onward, days 1-3 ahead (archived wind direction only goes 3 days out).

What counts as a pass, written before the test: Keep a variant only if it cuts the average error for days 1-2 by at least 0.05 degrees for highs or for lows, and makes no day 1-3 worse by more than 0.03 degrees. This is the same bar the forecast discussions faced.
Results

Average error (degrees F) on days the model never saw, Oct 2025 to Oct 2026.

Highs
VariantDay 1Day 2Day 3Days 1-2
Today's model 1.574° 1.971° 2.400° 1.774°
A: + NBM 1.576° 1.979° 2.408° 1.778°
B: + wind direction 1.569° 1.974° 2.366° 1.773°
C: + both 1.570° 1.982° 2.378° 1.777°
Lows
VariantDay 1Day 2Day 3Days 1-2
Today's model 1.697° 2.034° 2.304° 1.867°
A: + NBM 1.699° 2.039° 2.300° 1.870°
B: + wind direction 1.652° 2.022° 2.288° 1.838°
C: + both 1.652° 2.024° 2.282° 1.839°
  • Wind direction for lows, days 1-2: 0.028 degrees better; resampling the test days 2,000 times puts it between 0.004 and 0.051, and better in 98% of resamples. Real, but small.
  • Where it helps: nights with a northerly wind 0.075 better (138 nights), light winds 0.035 better (310), southerly winds no change (264).
  • By month it is noisy (from 0.06 worse in June to 0.13 better in October), as expected for an effect this small.
What it means

Neither idea clears the bar, so nothing changes in the forecast. NBM adds nothing to highs or lows: our own station correction already does what NBM's calibration does, for this spot. Wind direction is a real but small help to lows, and it lives almost entirely on northerly-wind nights, when cold air flows in behind a front. That fits the physics: those are nights when the air mass itself is changing, which a correction built around clouds and calm can't see.

Next

Follow the signal where it lives: test northerly flow together with clear skies (a clear night behind a front is both cold-air flow and radiational cooling). But this test year has now been looked at, so a sharper version tested on it again would flatter itself. The honest move is to confirm it on months nobody has looked at yet: re-run in the January 2027 lab session on Oct-Dec 2026 alone, with the bar written down now.

Lab log

One line for every lab session, including the ones where nothing was worth testing.

Ideas waiting their turn