# Offdiagonal data request for the next Zeno snapshot

## Priority order

1. **Longer consecutive weather and sea history** — ideally 12 months, minimum 6, with unchanged forecaster identities. The current short windows leave too little temporal diversity and too little later data for a reliable verdict.
2. **Deeper health forecaster histories** — COVID-19, RSV and FluSight panels with stable team/model IDs and enough earlier resolved forecasts per forecaster. RSV and FluSight remain too noisy to justify a correction.
3. **More genuinely distinct forecasts per crypto and market question** — third-party model or desk forecasts, not more price rows. Several agents sharing one base model must be marked as one model family, not counted as independent views.
4. **Larger multi-forecaster benchmark panels** — more resolved ForecastBench or research-authorised Metaculus-style questions with dated revisions. ForecastBench is currently too small to measure.
5. **Prospective AI-forecaster panels** — future questions recorded before resolution, with model-family identity. This tests whether Numinous-style behavioural learning transfers without pretending correlated agents are independent evidence.

## Required episode fields

- Stable question, source, forecaster, team and model-family IDs.
- Native forecast type and complete values: probability, point, quantiles or ordered bins.
- `decision_at`, each forecast's `available_at`, `target_at`, and outcome `label_available_at`.
- Full revision history, withdrawals and explicit missingness reasons; not latest-value-only rows.
- First-reported and final outcomes with finality or dispute state.
- Named reference forecast and stable outcome-event key.
- Forecaster history and matched co-miss counts computed only from outcomes known before the decision.
- Provenance, licence/training permission, frozen manifest totals and content hashes.

## Acceptance checklist

- At least two genuinely distinct forecast families per question.
- Every forecast was available by the decision time; every label became available later.
- No duplicate real-world outcome crosses a split.
- Missing values remain missing, never zero-filled or reconstructed.
- Learning, development, confirmation and sealed periods each contain enough resolved episodes.
- Every source total and file hash reconciles with the frozen manifest.

This is a request specification only. New data is not imported or used until a new immutable snapshot passes both Offdiagonal's provenance checks and Zeno's chronology/schema checks.