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.

1. v0.86.2Trained · being judged
   Checks passed · full 27-model run training on H200

   v0.86.1 plus the 25 Oct review fixes: error floors set per quantity and unit, a stability check that works for negative scores, dropped-forecaster counts kept with the saved data, and saves that wait out Hugging Face limits. Same 13-field, 7.24M-forecast snapshot; 11 fields trained (ECB and energy admitted, not trained). At 07:10 UTC 25 Oct, 7 of 9 check models had finished 4,000 steps with finite losses (first large model: held-back error 5.01 → best 1.405); all 9 checks then passed the scale check on every trained field, and the full 27-model run (job 6addaf0a) started at 07:26 UTC. An exact measurement found the per-quantity floors changed no training-split scale. No win verdict yet: the matched comparison against simple methods and the served model decides.

2. v0.86.1Practice stopped
   Full run stopped at 13 of 27 models · save limit

   Same 7.24M-forecast snapshot with a train-fitted resolution floor, a bounded loss and a confidence floor. All 9 full-length checks passed. The full run restarted at 04:18 UTC on 25 Oct after a storage stop; 13 of 27 models finished stably (held-back error about 1.3–1.4) before it stopped at about 05:47 UTC on the Hugging Face limit of 320 saves per hour. Its models are kept but not used. ECB and energy are admitted but not trained (no separate development history). Stable loss is not yet evidence of better forecasts: the matched comparison against simple methods and the served model decides.

3. v0.86.0Practice stopped
   Full run stopped · held-back error diverged

   On Offdiagonal v6.13 (7.24M forecasts) the checks passed, and the full run started at 23:47 UTC on 24 Oct. Training error stayed near 1, but error on held-back cases jumped to about 1.6 million in most models after step 600. The cause was sea panels where every forecaster said 0: those got a near-zero scale, so a 0.1 outcome counted as about 15,000 scales off. The run was stopped at 01:27 UTC on 25 Oct. No model from it is used.

4. v0.85.11Practice stopped
   Trial failed its planted-pattern check · full run not started

   Trained 45 practice models (3 sizes × 5 versions × 3 seeds) on about 858,000 verified forecasts in 10 fields. Small and medium models cut their error on unseen cases by about 85%. But 9 of the 15 large models got stuck on one constant score, and the planted-pattern check failed (0.57 against a 0.5 bar) because it trained and scored in different frames. The full run was not started.

5. v0.82.1Practice stopped
   Practice stopped · full training not started

   Checked on Hugging Face on 22 Oct at 02:03 UTC. Nine practice checkpoints were saved before the transfer test requested ensemble weather, which has zero labelled panels in this build. The job stopped at 01:22 UTC; the full 30-round, three-seed run never started. The frozen attempt and saved checkpoints remain unchanged. No final verdict or warning-skill pass is available.

6. v0.83Too few outcomes
   Inflation rematch: keep the mean

   On 366 eligible HICP inflation cases, the tested ridge and tree corrections underperformed the panel mean: +0.058 and +0.056 MAE on 190 test cases across 12 held-out outcomes. Their 90% outcome-bootstrap intervals are above zero. This tests two summary-feature recipes, not the shared model of models or other ECB targets. Offdiagonal's archive recovery found no qualifying identical-target matches for the 744 excluded cases; first-print status remains unestablished. Valid inflation examples can still contribute to joint training. Serving is unchanged.

7. v0.82Candidate
   Finished · energy, sea and crypto lead

   The full v0.82 neural run finished on the H200 on 22 Oct with no failures and was judged against its locked rules; the result is public on Hugging Face. On the confirmation period it corrects forecasts best in energy (17.5% better than track records and the tree learner, 15.6% better than the calibrated reference), sea (25% better than the calibrated reference) and crypto (23%), and helps weather vs airports (10% better than track records). The two warning skills were not saved by this run, and two rivals (the market's own price and a scrambled-relations control) were not measured, so no skill passes on this run alone. v0.82.1 stopped during practice before full training. Serving does not change.

8. v0.80.2Candidate
   Public on Hugging Face · 21 fields · 2 beat strong references

   Rebuilt on Offdiagonal v6.5: 2,840,152 forecast panels with known outcomes in 21 trained fields, including air quality, urban heat, agriculture, world conditions and shared answer sources. The time audit found no forecast used after its decision time. No real-world event is scored in two periods. The approved full run trained all 21 starts on 21 Oct (7 variants × 3 seeds). The corrected judge (averaged forecasts, one test across all fields) passed two fields on both test periods: energy demand (one forecasting family) and weather vs airports (airport timing still assumed). Flu improved on development but failed confirmation. Every other field keeps the calibrated reference or has too few outcomes. It now answers as an opt-in beta (/api/public/v1/beta/check, same answers as the frozen files on 721 real cases) and powers the public Forecast Failure Watch; its weights are public on Hugging Face (ZenoDivergent/zeno-divergent-v0802) as a research beta; the default answer stays v0.57.

9. v0.80Candidate
   Full H200 run done: two research candidates, now being repaired

   Trained on 20 Oct on 454,490 cases from 16 sources (5 versions × 3 seeds, 13 minutes on an H200). Flu (+32% against the reference on confirmation) and energy (+4% against a tree model) passed both test periods; every other field kept the calibrated reference, and crypto, RSV, economy and Polymarket traders scored worse. An outside audit then found defects: forecasts dated after the run, energy missing its outcome clock, events grouped by question instead of real-world event, and two empty inputs. So both wins stay research candidates until v0.80.1 reproduces them on repaired data and a fresh period.

10. v0.79eExploratory
    13-source panel reader: scored, research only

    Owner approved on 20 Oct. Data rebuilt with corrected answer timing and defect flags; the H200 measured its own speed, then trained all 134 research models (0.5M and 2M) that learn forecaster miss risk and confidently wrong consensus; all were saved and re-checked. Official verdict (v0.79f, 20 Oct): it beats four time-safe references in 7 of 10 testable sources, ties on COVID-19 and Phil. Fed ranges, and loses to a tree on crypto; pair history and the 2M size add nothing. Training is not a result: comparisons against simple references decide anything, and nothing switches the served model.

11. v0.79Failed
    Quick rerun on older data: stopped

    A 16-model rerun of the v0.78 recipe on the older data. It was stopped on 20 Oct because it skipped the new sources and the agreed simpler comparisons. Its partial numbers are diagnostic only.

12. v0.78tToo few outcomes
    Neural trust campaign: all 50 models trained, saved and scored

    Reading the whole panel together beat each forecaster's track record alone, crypto gained a little, and ensemble weather looked better on few test groups. ForecastBench did worse than the tree, pair history gave no reliable edge, and 2M was no better than 0.5M. Not good enough everywhere to replace the served model.

13. v0.77Exploratory
    CPU tree experiment across seven sources (provisional)

    Not a neural training run: a small tree model on a CPU in about 4 minutes. An outside review found two average rates computed with future answers included, so we re-ran it with every input time-valid and every model calibrated the same way (v0.77r, a diagnostic re-check of weeks already seen). Against each forecaster's own track record it cut error by about 11% on ensemble weather, 21% on single-run weather and 8% on crypto. The earlier ~25% ForecastBench gain disappeared: it now only matches the track record. The economy and Numinous sets are too small to judge.

14. v0.76bToo few outcomes
    Failure Atlas: who misses together

    The earlier 35% → 45% readout used ranges drawn too narrow. With ranges corrected on earlier weeks, the test window holds zero group failures, so warnings cannot be judged yet.

15. v0.75Failed
    One shared model for all sources

    Did not beat separate per-source models, so the larger run never started.

16. v0.74Candidate
    Ensemble-wind correction

    +7.8% against the team forecast on its sealed wind weeks — but simply removing each system's usual error at each place scored +9.1% on the same cases. The gain is local habit, not learned failure structure; it will not be promoted.

17. v1.06Served
    Served model

    Answers the public Check, API and MCP. Corrects only the sources it passed on, keeps the reference everywhere else.