# Forecast protocol (Phase 4)

Two arms, run as separate subagents with no shared context. Each arm forecasts each study five times; every run is kept.

| | BASELINE | FRAMEWORK |
|---|---|---|
| Study protocol | `forecasts/{study_id}/protocol.md` | same file |
| Required reading | `forecasts/primer_baseline.md` | `forecasts/primer_framework.md` |
| Also available | nothing else | `data/claims.csv`, `corpus/` |
| Must cite | nothing | `claim_ids_used` from `data/claims.csv` |
| May see pistomechanics material | **no** | yes |

The two required-reading files are matched in length (within 10% by character count; see `forecasts/primer_lengths.txt`). The FRAMEWORK arm additionally has the claim registry and the corpus on disk. That extra material is the treatment being tested, not a leak.

Neither arm may browse the web. A forecaster who searched for the study could find its results, and the pool only admits studies whose results were not public when checked, so a search would either be useless or contaminate the run.

Each run is a fresh agent. Runs do not see one another's output.

## Output: one JSON file per run

`forecasts/{study_id}/{agent}_{run}.json`, where agent is `baseline` or `framework` and run is 1-5.

```json
{
  "study_id": "NCT00000000",
  "agent": "baseline",
  "run": 1,
  "p_primary_hypothesis_supported": 0.42,
  "effect_direction": "favours intervention | favours control | null | other: <text>",
  "effect_size_90pct_interval": {"low": 0.05, "high": 0.45, "metric": "Cohen's d on <outcome>"},
  "moderator_prediction": "<which subgroup responds more, and why> or \"none\"",
  "claim_ids_used": [],
  "rationale": "<= 120 words"
}
```

- `p_primary_hypothesis_supported` is the probability that the study's own primary hypothesis, as stated in `protocol.md`, will be reported as supported on its primary outcome at the stated alpha.
- The interval is in the study's own metric if `protocol.md` names one with a scale; otherwise it is a standardised mean difference (Cohen's d) or an odds ratio, and `metric` says which. The sign convention is positive = in the hypothesised direction.
- `claim_ids_used` is empty for BASELINE. For FRAMEWORK it lists every claim_id that the rationale relies on; at least one is required.