What the model learned from, and what it revealed
Every domain the current release trained and evaluated on, what was computed from each source, and the gate verdict — promoted or refused, with the reason. Failures are listed, not hidden.
Domains: considered, computed, revealed
Release v0.9.0. Chronological train / calibration / test splits; nothing published unless the pre-declared gate clears.
loading release summary…
Upstream providers — who actually produced the numbers
Behind the 7 real corpora and the partner feeds. Satellite operators appear here only with the exact, narrow thing they contribute.
Promoted domain. Event magnitudes and inter-event gaps.
row fields: event_time · available_at · magnitude · region entity id · catalog revision
Promoted domain — the one decisive baseline win (cold extreme).
row fields: station id · observation date · available_at · element code · QC flag
Trained and evaluated; refused at the gate.
row fields: gauge id · observation date · available_at · discharge · provisional flag
Trained and evaluated; refused at the gate.
row fields: index name · observation time · available_at · estimated vs final flag
Wildfire corpus; refused — coverage still too thin.
row fields: detection time · available_at · satellite/instrument · confidence · FRP
Macro corpus; refused at the gate.
row fields: series id · observation date · available_at (vintage) · revision
Markets corpus; refused at the gate.
row fields: instrument / chain id · bar close time · available_at · venue
Primary direct input to AntiFogo's early-ignition system, relayed to Zeno inside wildfire rows. Zeno claims it only where the row's own lineage names it; FIRMS corroborates when the feed is dark.
row fields: provider/product · native resolution · processing level · channel class · available_at
row-level contract →Optional derived features only. Raw imagery is not in the public corpus. A scene is claimed only with a persisted scene ID and capture time — a sentinel zero is a no-scene marker, not evidence.
row fields: scene id · capture time · available_at · processing level · channel class = derived
row-level contract →Derived products, licence-checked per channel. Marine is an evaluation lane, not a trained domain.
row fields: product id · native resolution · channel class · licence · available_at
row-level contract →Labelled derived, never treated as station observations. Partner model bands enter as reference predictions.
row fields: model name + version · horizon · channel class = model output · available_at
row-level contract →No pixel ingestion path exists. Zeno consumes tabular channels with row-level lineage; imagery would need its own contract and licence review before any claim.
how to readA provider being registered proves nothing. Only a row's own lineage decides what may be claimed, which is why EUMETSAT and Open Cosmos are marked narrow: they contribute specific derived channels inside partner rows, not the corpus at large. Nothing here is a promoted domain unless the gate said so. Open any chart segment in the workspace to see which of these contracts produced that exact view.
Incoming sources — schemas pinned, data pending
The apps adapt to the model's contract, never the reverse. Each source enters only after a live probe passes: contract version, availability timestamps, licences and leakage checks — then a forecast-gate rerun decides if it can contribute.
Live at antifire.live. History starts 2026-08-29 — real but thin, so it widens entity coverage without moving the wildfire gate yet. Publication-time available_at on every row.
Live. Joins the seismic corpus as a namespaced second source alongside USGS; quiet days stay quiet — no synthetic events.
Live as the `airquality` pack. Stations carry observations; wards carry the AirWatch PM2.5 model (name, version, horizon attached) and join station days as reference predictions. The forecast era began 2026-08-24, so the reference overlap window is thin and grows daily.
Live as the new `agriculture` pack. All channels are derived products (satellite + reanalysis lineage), labelled as such — never treated as direct observations.
Live as the new `marine` pack. Copernicus Marine lineage; licences verified per channel before ingestion.
Live as the `urban` pack. Real model peak-heat rows (name, version, horizon) are emitted and land as reference predictions, never as observations.
Binance/CoinGecko fallback for CELO price (Coinbase carries no CELO pair), as a context channel on markets.
how to readObserved is a measurement. Derived is computed from measurements. Model output is another model's forecast — never blended into observations; its disagreement with reality is itself evidence the surprise latent can use. Model-output rows must declare when they were issued, so nothing enters a backtest before it was knowable.
Training run v0.9.0 — final status
v090-run1b · 1x NVIDIA L4 (Hugging Face Jobs) · read from the job log at 13:00Z
- Preflight, schema and leakage checkscomplete06:04Z
- Availability-aware feature construction, 8 domainscomplete06:32Z
- 12 epochs, shared trunk + per-domain headscomplete08:40Z
- Held-out evaluation per domaincomplete09:20Z
- Leave-one-domain-out retrain, 8 foldscomplete10:35Z
- Frozen-trunk transfer (GiftEval, LOTSA, 4 partner lanes)complete11:40Z
- Calibration fit on the calibration split onlycomplete12:05Z
- Per-head promotion gatescomplete12:20Z
- Checkpoint snapshot + gated ONNX exportcomplete12:36Z
per-domain checkpoints
checkpoint snapshot
complete
Run complete. Checkpoint multi-v0.9.0-candidate.pt (7.6 MB), full report and promoted-only ONNX exports were snapshotted at 12:36 UTC.
calibration & gates
complete
Calibration is fitted on the calibration split only, then each head faces its pre-declared gate. Only heads that pass are exported.
how to readGreen means finished, amber means running right now, grey means queued. A finished run is not a promoted model: the gates at the end decide which domains ship and which are published as refusals. The run finished and is frozen. Climate and seismic cleared every gate criterion, including the paired lift over a logistic regression on the same features, so v0.9.0 is published as an immutable promoted revision. Hydrology, markets, macro and on-chain score above chance but fail the paired lift gate; wildfire has too few test positives; space weather's compound head sits at chance. Air quality, agriculture, marine and urban ran as frozen-trunk transfer lanes only.
What was computed
per-domain adapters → shared trunk → shared surprise latent z^S → per-domain heads
- Seismic. event magnitudes and inter-event gaps, multi-scale windows (5/20/100), robust median/IQR scaling
- Streamflow. daily discharge per gauge, multi-scale windows, calendar day-of-year, robust scaling
- Markets. daily OHLCV bars, log returns and range, multi-scale windows, robust scaling
- Space weather. NOAA indices (Kp, Dst, solar wind), multi-scale windows, calendar features
- Climate. station daily temperatures and precipitation, calendar features, robust scaling
- Macro. 204 FRED series, month-over-month transforms, multi-scale windows
- Wildfire. FIRMS detection counts and radiative power, calendar features, robust scaling
- Chaos control. synthetic chaotic systems — control corpus, never evidence about the world
Training run v0.8.0
1x NVIDIA L4 (Hugging Face Jobs) · started Fri, 28 Aug 2026 11:30 · status: completed
Run finished and is frozen. One real-corpus domain (climate) promoted; two are required. The refusal report is published in full rather than the run being re-tuned after the fact.
- — focal loss + per-entity resampling on rare targets — heads left chance in seismic, macro and wildfire, but none of them beat the logistic baseline with a CI excluding zero
- — per-target calibration refit on a held-out calibration split — worst-target ECE cleared the bound everywhere except wildfire (0.151)
- — macro corpus widened to 198 FRED series — entity minimum cleared; only one target beat the baseline
- — declared abstention operating point — climate gains 4.7 points of selective accuracy at 80% coverage; hydrology's gain was too small to be useful and blocked its promotion
- — ONNX export gated on the same run report — only climate qualified, so in-browser scoring keeps the v0.7.0 heads elsewhere
A running job is never edited: corrections become a new immutable versioned run. Domains in scope: seismic, hydrology, markets, spaceweather, climate, macro, wildfire, chaos. Held out for transfer: gifteval, lotsa.