The corpus
Big volume. Little independent evidence.
v0.80.2 trains on 2,840,152 panels with known outcomes in 21 fields (12,134,209 forecasts). Counted apart: the served model's frozen snapshot (280,956 groups) and 22,169 airport-checked weather cases.
01 · forecast rows
225,235
Six systems, four variables, hourly, 59 places.
02 · outcomes
44,801
÷5 from the step before
Several systems forecast the same hour and place; one answer scores them all.
03 · place-days
996
÷45 from the step before
Hours at one place on one day share weather, so they count once.
04 · weeks
6
÷166 from the step before
Splits run by week; a source needs enough later weeks to be judged.
ensemble weather · corpus-v1-20261020 · range-history-v1 · manifest 624a4aa2 · 12 Sep – 17 Oct 2026 · bar length on a log scale
00 · v0.80.2 data
What the next model trains on.
Counted from the build itself, after the time audit.
- forecasts
- 12,134,209
- panels
- 4,599,358
- with outcome
- 2,840,152
- distinct events
- 1,265,459
Time audit passed: 0 forecasts used after their decision time, 0 outcomes known before it. Most of the growth over v0.80 comes from the new air pressure and wind fields. Every field counts equally in judging, however large it is.
00a · every field
24 fields, one model of models.
Size, forecasters and Zeno's latest result in each field.
24 fields · 7 areas · 6 already beat every rival
- Warnings · v0.82 warning test
- Brier score of confident-miss warnings vs each forecaster's own track record (tree). historical: development and later confirmation slices of past data. exploratory · release gate not yet passed.
- Corrections · v0.80.2 public beta
- forecast loss vs calibrated reference vs calibrated reference. historical: two sealed test periods. passed fields are served in the opt-in beta.
Atmosphere
- Sea-level pressure1.13M outcomes · 2 forecastersWarnings: no edge yetCorrections: calibrated reference
- Wind1.13M outcomes · 2 forecastersWarnings: no edge yetCorrections: calibrated reference
- Urban heat154k outcomes · 2 forecastersWarnings: no edge yetCorrections: insufficient evidence
- Weather vs analysis49k outcomes · 6 forecastersWarnings: beats every rival (+7.9% vs track record, +10.5% later)Corrections: calibrated reference
- Single-run weather23k outcomes · 7 forecastersWarnings: too few misses to judgeCorrections: calibrated reference
- Weather vs airports22k outcomes · 5 forecastersWarnings: beats every rival (+5.6% vs track record, +6.3% later)Corrections: validated improvement
- Air quality644 outcomes · 1 forecastersWarnings: too few misses to judgeCorrections: insufficient evidence
- Farming rain31 outcomes · 2 forecastersWarnings: too small to judgeCorrections: insufficient evidence
- Ensemble weather0 outcomes · 6 forecastersWarnings: outcomes arrivingCorrections: not trained
Space and sea
- Space weather26k outcomes · 4 forecastersWarnings: too few misses to judgeCorrections: calibrated reference
- Sea12k outcomes · 24 forecastersWarnings: beats every rival (+28% vs track record, +23% later)Corrections: calibrated reference
Markets and crypto
- Crypto84k outcomes · 3 forecastersWarnings: beats every rival (+14.2% vs track record, +13.4% later)Corrections: calibrated reference
- Prediction markets66k outcomes · 1 forecastersWarnings: no edge yetCorrections: calibrated reference
- Market traders4k outcomes · 8248 forecastersWarnings: too few misses to judgeCorrections: calibrated reference
- Cross-venue markets325 outcomes · 617 forecastersWarnings: too few misses to judgeCorrections: insufficient evidence
Health
- COVID-196k outcomes · 16 forecastersWarnings: ahead of most rivals (+0.1% vs track record, +1.1% later)Corrections: calibrated reference
- RSV3k outcomes · 6 forecastersWarnings: ahead of most rivals (+2.3% vs track record, +24% later)Corrections: calibrated reference
- Flu2k outcomes · 66 forecastersWarnings: no edge yet (-15.1% vs track record, +9% later)Corrections: calibrated reference
Economy
- Philadelphia Fed (points)4k outcomes · 197 forecastersWarnings: ahead of most rivals (+37.8% vs track record, +23.8% later)Corrections: calibrated reference
- Philadelphia Fed (ranges)2k outcomes · 195 forecastersWarnings: too small to judgeCorrections: calibrated reference
- ECB1k outcomes · 116 forecastersWarnings: ahead of most rivals (+7.8% vs track record, +16.8% later)Corrections: calibrated reference
Energy
- Energy demand3k outcomes · 1 forecastersWarnings: too few misses to judgeCorrections: validated improvement
General forecasting
- ForecastBench109k outcomes · 806 forecastersWarnings: beats every rival (+45.3% vs track record, +23.1% later)Corrections: not trained
- Numinous154 outcomes · 2 forecastersWarnings: too small to judgeCorrections: not trained
00a · evidence accounting
What the model actually learned from.
Collected, with outcome, actually trained on and distinct outcome events, per field.
| Field | Forecasts | Panels | With outcome | Trained on | Outcome events | |
|---|---|---|---|---|---|---|
| Air pressure | 1,898,196 | 1,898,196 | 1,134,793 | 1,134,793 | 233,762 | |
| Wind 10 m | 2,111,869 | 2,111,869 | 1,134,793 | 1,134,793 | 233,762 | |
| Urban heat | 196,004 | 196,004 | 154,115 | 154,115 | 4,731 | |
| Crypto | 250,512 | 83,504 | 83,504 | 83,504 | 83,504 | |
| Market prices | 65,849 | 65,849 | 65,849 | 65,849 | 65,849 | |
| Weather vs analysis | 243,267 | 48,708 | 48,708 | 48,708 | 48,708 | |
| Space weather | 103,623 | 27,952 | 25,855 | 25,855 | 3,158 | |
| Single-run weather | 350,346 | 54,458 | 22,960 | 22,960 | 22,960 | |
| Weather vs airports | 110,845 | 22,187 | 22,187 | 22,187 | 22,187 | |
| Sea | 93,161 | 12,096 | 12,096 | 12,096 | 12,096 | |
| COVID-19 | 38,610 | 6,365 | 6,365 | 6,365 | 6,365 | |
| Polymarket traders | 54,949 | 4,469 | 4,469 | 4,469 | 4,469 | |
| Phil. Fed (numbers) | 154,524 | 4,190 | 4,190 | 4,190 | 4,190 | |
| RSV | 7,887 | 2,891 | 2,891 | 2,891 | 2,891 | |
| Energy demand | 2,667 | 2,667 | 2,667 | 2,667 | 129 | |
| Flu | 9,582 | 2,042 | 2,042 | 2,042 | 2,042 | |
| Phil. Fed (prob.) | 56,700 | 699 | 1,717 | 1,717 | 243 | |
| ECB survey | 58,362 | 1,098 | 1,098 | 1,098 | 1,098 | |
| Air quality PM2.5 | 50,197 | 50,197 | 644 | 644 | 164 | |
| Cross-venue markets | 650 | 325 | 325 | 325 | 325 | |
| Agriculture rain | 3,592 | 3,592 | 31 | 31 | 31 | too few outcomes to judge |
| ForecastBench | 6,200,321 | 0 | 108,699 | 0 | 0 | outcomes not linked to events · excluded |
| Numinous | 308 | 0 | 154 | 0 | 25 | excluded |
| Ensemble weather | 72,188 | 0 | 0 | 0 | 0 | outcomes not linked to events · excluded |
Volume is not evidence volume: a panel counts once per decision time, hours and nearby places are related, and one-family fields cannot show shared failures on their own.
00b · v0.80 census · dated 20 Oct
The earlier v0.80 inventory.
Historical: the 397,002-panel census of the v0.80 run, before the v6.5 rebuild.
- collected
- 7,853,462
- admitted
- 7,846,807
- labelled
- 3,762,020
- labelled cases
- 397,002
- trained
- after the run
- evaluated
- after the run
Not included yet: wildfire (no forecast episodes delivered yet) · air quality (forecast list empty) · urban (forecast list empty) · agriculture (forecast list empty). Offdiagonal judgments are empty in every pinned file, so they are tested as an ablation and never claimed.
00c · v0.79e admission · dated
From raw data to judged cases.
What Offdiagonal audited, what we admitted, and what the tests can judge. Removed data is crossed out.
Weather ensembles
trains + judged · 1 flag carried
audited · 1,072,357 · 5 families
admitted · 44,801
judged · 15,698 · 4,482 missed
Weather single runs
trains + judged · 1 flag carried
audited · 1,072,357 · 6 families
admitted · 30,280
judged · 5,408 · 306 missed
ForecastBench
trains · answers dated from the pinned pack · 2 flags carried
audited · 106,760 resolutions
admitted · 108,757
judged · not testable
Crypto
trains as replays · one family · 2 flags carried
audited · 84,900 · 1 family
admitted · 84,597
judged · not testable
Prediction markets
trains · one venue per question · 2 flags carried
audited · Manifold 97,788 · Polymarket 71,195
admitted · 65,896
judged · not testable
Sea
trains + judged
audited · not audited
admitted · 12,096
judged · 2,540 · 56 missed
COVID-19
trains + judged · 1 flag carried
audited · 536,748 archive
admitted · 6,365
judged · 1,281 · 155 missed
Space weather
trains + judged · reconstructed clock
audited · not audited
admitted · 3,158
judged · 709 · 14 missed
RSV
trains + judged · 1 flag carried
audited · 18,008 archive
admitted · 2,891
judged · 323 · 42 missed
US economy · point
trains + judged
audited · within 2,250 SPF
admitted · 423
judged · 441 · 74 missed
US economy · probability
trains + judged
audited · 2,250 SPF
admitted · 102
judged · 392 · 84 missed
Numinous AI crowd
trains · no valid test period · 1 flag carried
audited · 1,113 resolved
admitted · 94
judged · not testable
Euro economy
trains · 27 test outcomes, too few
audited · 124
admitted · 90
judged · not testable
removed before training · 7 defects
- Scoring mixes forecast revisions · health
- FluSight WIS normalised by weight sum · health
- Metaculus multiple-choice first option stored as YES · markets
- Replay forecast picked using the final answer · markets
- Flood run time invented · hydrology
- Seasonal run time invented · weather
- Zoo forecasts with targets at or before issuance · own-engine
not sources in this run
- ECB point survey · 81 yearly outcomes, too few for a miss threshold
- Polymarket wallet trades · 12,251 trades, none resolved in the pinned release
- Energy · 95 outcomes, one family · exploratory only
- Metaculus · 0 stored resolutions
- Jev · no dated judgments
17 more defects stay in the data, flagged on every row they touch.
Bars use a log scale. Judged counts are group checks in the test periods; a source with one family or too few outcomes trains but cannot be judged on group misses.
01 · failures
Counting failures honestly.
The same misses, counted by independent unit and after correcting the ranges.
raw ranges · counted three ways
- single outcomes2,646
- place-days624
- broader events72
Thousands of misses shrink to 72 separate weather events.
later weeks · before → after range correction
- hours1,639 → 129
- place-days374 → 59
Most "everyone missed" moments were ranges drawn too narrow. What remains is rarer, and still far above chance.
group failure = every family missed its own 90% range · corpus-v1-20261020 · range-history-v1 · Zeno v0.77 audit (later weeks, ranges widened from earlier weeks only)
02 · what survives
Volume shrinks. Real failures shrink more.
COVID-19
2,199 outcomes · 23 wk
28 → 16 · 57% stay
RSV
529 outcomes · 10 wk
24 → 10 · 42% stay
Ensemble weather
15,314 outcomes · 2 wk
857 → 30 · 4% stay
Weather's shared misses were mostly ranges that were too narrow. In health, about half survive: real confidently-wrong consensus, still too rare to train on.
Evidence +−
corpus
The complete frozen collection of forecast panels, histories, timestamps and outcomes.
domain
A broad subject area, such as weather or public health.
source
One named forecast stream inside a domain, such as Ensemble weather or FluSight.
evidence role
How a source contributes: learned from, measured on later data, learn-only, or not yet measurable.
Recorded groups
Unique forecast-outcome groups present in the frozen pack.
Admitted episodes
Groups that pass the source admission rules; this is not an evaluation count.
Distinct forecasters
Shown only where the frozen receipt reports the count.
Earlier history
Not yet counted in the frozen receipt; unavailable is not zero.
Matched-pair history
Not yet counted in the frozen receipt; unavailable is not zero.
Eligible evaluation outcomes
Known only after split reconciliation. ECB is designated learn-only; no evaluation count is claimed here.
snapshot mirror-v044-20261004T1300Z · 4 October 2026 · 12:39 UTC · frozen · complete
Served answers come from v057-main-l40sx1-20261007T1138Z: Crypto price ranges (Zeno correction) · Ensemble weather (Zeno correction) · Deterministic weather (Zeno correction without history) · COVID-19 hospitalisations (Zeno correction without history) · Sea conditions (Zeno correction without history) · Prediction markets (reference kept) · RSV (reference kept) · FluSight (reference kept) · ForecastBench (reference kept) · Numinous AI forecasters (reference kept). The finished v0.79e research run used its own frozen cache; v0.80 reads every source again from the original Offdiagonal files.