INTEL_LAB_SWEEP.md

Intelligence-Lab campaign sweep at scale (Wave-4)

dummy/autoresearch/campaign_sweep.py runs the Wave-3 campaign across many cohorts at once and controls the multiple-comparisons risk that scale creates.

Why scale needs FDR control

Wave-3 runs one campaign over one cohort. Searching every stream × subject × market_type × phase the ledger carries means hundreds of (cohort × candidate) tests — at q=0.05, dozens of "edges" survive by chance alone. Reporting the naive survivor count at scale is the multiple-comparisons trap the Intelligence-Lab is built to avoid.

What the sweep does

1. Group a stream's records into cohorts by cohort_scope (the same key the partition plan enforces).
2. Skip and disclose cohorts too small to partition (≥3 decision dates required) or with too few visible rows — every skip carries a reason; nothing is dropped silently.
3. Score each complexity-passing candidate on the cohort's VISIBLE_DEVELOPMENT partition only (no lookahead), with a paired-Brier statistic vs the market prior → a one-sided p-value per (cohort, candidate).
4. FDR-control the pooled p-values with Benjamini-Hochberg at level q, so the disclosure reports the family size searched, the naive significant count, and the FDR-survivor count.

The paired-Brier gain is (prior − y)² − (candidate − y)² per row; positive means the candidate beat the market prior. Rows are independent event-cluster snapshots (the partition plan purges cross-partition clusters), so paired-by-row is a defensible cluster-independent test.

SweepResult.to_dict() is the reviewable artifact. reaches_execution is always false — this is research evidence, never an execution path.