Removes legacy result layouts and summaries after leaderboard-v2 was uploaded and checksum-verified. The removed files remain recoverable from repository history.
Add audited leaderboard-v2 rollouts and fractional SetHit metrics
Adds full agent_hidden rollouts for ten model/reasoning arms. SetHit is recomputed as |G intersection R| / |G| and metrics include report-validity and coverage audits.
Standardize benchmark queries to four essential fields
Keep candidate_id, partition, query and positive_items only. Preserve all examples and their ordering; update schema manifests, hashes and documentation. Original records retained at before-minimal-query-schema-2026-10-13 (a712ed01d3d01905d98251bfcb97af5a1320c58a). Saved evaluation results and construction audits are unchanged.