Evidence workspace

Corpus

What is recorded, comparable, scored, and available to train from.

evidence-20260919195500 · 2026-09-19 19:55 UTC

Recorded coverage

Recorded samples by task. Bar length compares availability, not skill. Select the full evidence record for admission rules.

Hourly apparent temperature

urban · degC

1 lane

47,920

temperature_2m (air)

air · °C

1 lane

21,926

Sea-surface temperature

sea · degC

1 lane

824

Weekly precipitation total

agriculture · mm/week

1 lane

252

PM2.5 concentration

air · ug/m3

1 lane

Recorded datasets

Figures are counts from the pinned releases (snapshot-20261020 to episodes-v3-20261020) and dated database totals, not live reads. All held data is cleared for Zeno sharing and training; only leakage and missing-outcome gates apply.

Global weather model forecasts

Marine, agricultural, urban and air-quality sites; lead times 0–72h and beyond.

3.08M forecasts issued Dec 2023 – Sep 2026; 84,179 revisions with evidenced availability clocks

Inspect coverage and access

Temperature, wind, pressure and precipitation at fixed sites, from nine global numerical weather models.

Available material: Point forecasts · Constructed 90% ranges · Revisions · Observed outcomes
Supported analysis: Historical comparison · Calibration · Co-failure analysis · Prospective evaluation
Evidence limitations: Ranges are constructed from the cross-model spread, not provider-declared. Short fully-shared history per site
Records: Prospectively captured, with archive backfill labelled separately
Access: Public sample; full data on request

Download a sample

CDC FluSight influenza forecasts

Reference dates Oct 2023 – Apr 2026; horizons 0–3 weeks.

616,711 forecasts; 13,572 events joined to first observed result vintages

Inspect coverage and access

Weekly influenza hospitalisations by US state and nation, quantile forecasts from many teams.

Available material: Quantile distributions · Outcome vintages (revised values) · Team metadata
Supported analysis: Historical comparison · Interval coverage · Team dependence
Evidence limitations: Publication times bounded by repository commits. Outcomes are revised; first vintage is used for scoring
Records: Historical reconstruction from the public hub
Access: Redistributable; full data on request

Prediction market prices and answers

Venues kept separate; never pooled.

About 3.6M questions and 11.2M venue-stored prices held; 75,200 cross-venue links, 99,449 price revisions and 253,569 Polymarket trader trades frozen

Inspect coverage and access

Probabilities for yes/no questions on Manifold, Polymarket, Kalshi and Metaculus.

Available material: Timestamped prices · Resolved answers · Resolution times
Supported analysis: Historical comparison · Calibration
Evidence limitations: Manifold is play money. Polymarket timing uses resolved_at, not closes_at. The full question and price export is still running in resumable batches
Records: Prospective collection plus provider history
Access: Request access

ECB Survey of Professional Forecasters

Quarterly rounds; several horizons.

156,912 individual probability distributions across 42 rounds; realised-value vintages

Inspect coverage and access

Euro-area inflation, growth and unemployment: individual panel probability distributions.

Available material: Probability distributions · Point forecasts · Outcome vintages
Supported analysis: Historical comparison · Calibration
Evidence limitations: 104 of 111 round release dates are bounded, not exact. Outcome is the latest vintage, not first release
Records: Historical reconstruction
Access: Public frozen supplement

Download a sample

ForecastBench AI and human forecasts

Jul 2024 – Sep 2026; 808 forecasters in 82 families.

13,523,509 forecast rows, 6,311,674 with real answers; 7,248 still-open questions frozen daily

Inspect coverage and access

Probabilities from named AI models, human groups and baselines on world-event and dataset questions.

Available material: Named forecaster identities · Resolved answers · Daily open-question freezes
Supported analysis: Historical comparison · Family dependence · Prospective evaluation
Evidence limitations: Answers often arrive months after the forecast. Reasoning text excluded
Records: Public ForecastBench releases (CC BY-SA 4.0)
Access: Frozen release on Hugging Face

Numinous miner forecasts

Aug – Oct 2026; 6,939 markets.

1,081,071 miner forecasts; 432,862 market price points

Inspect coverage and access

Probabilities from many Numinous miners on prediction-market questions.

Available material: Per-miner forecasts · Market prices · Resolved outcomes
Supported analysis: Crowd vs market comparison · Calibration
Evidence limitations: Miners count as one family until architecture metadata exists. PRICE scores are claims, never labels
Records: Prospective capture
Access: Frozen release on Hugging Face

Crypto price forecasts

Jul – Oct 2026; hourly.

255,570 forecasts, 49,587 independent-family quotes, 85,646 hourly observations

Inspect coverage and access

Kalshi price ladders, Deribit futures marks and reference methods for BTC, ETH and SOL.

Available material: Venue-timestamped quotes · Hourly closes · Settlement indices
Supported analysis: Family comparison · Calibration
Evidence limitations: Each family is scored on its own settlement index. Deribit rows wait for settlement
Records: Prospective capture with venue timestamps
Access: Frozen release on Hugging Face

Philadelphia Fed Survey of Professional Forecasters

1968 – 2026; 232 rounds, 64 indicators, 463 forecasters.

1,979,271 responses; 2,250 resolved growth and inflation episodes scored on first-release figures

Inspect coverage and access

US growth, inflation and other indicators from individual survey respondents.

Available material: Individual responses · First-release outcome figures
Supported analysis: Historical comparison · Calibration
Evidence limitations: Release dates are conservative, stated bounds. 62 indicators still need official-figure files
Records: Historical reconstruction from the public survey
Access: Public feed

GB electricity demand forecasts

Since 18 Oct 2026; collected hourly.

2,489 forecasts at release (1,449 admitted, 95 resolved half-hours); outcomes now cover every half-hour since 18 Oct

Inspect coverage and access

NESO national demand forecasts for each half-hour.

Available material: Issued forecasts with revisions · Elexon outturn vintages · Local weather context
Supported analysis: Calibration · Weather-conditioned analysis
Evidence limitations: One forecaster family, so no family comparison yet. Weather context is local, not national
Records: Prospective capture
Access: Public feed and frozen release

Download a sample

Last verified 2026-10-20.