# E124 rank feed pipeline

Seven deterministic agents run in order and each writes a SHA-256 hash-chained receipt:

1. INGEST — hash and read the official 500-task results, the public submission leaderboard, and the prediction file.
2. NORMALIZE — build one row per submission (family/scaffold tags, resolved, rate, horizon).
3. SCORE — CAIN resolved/rate, Wilson 95% interval, difficulty buckets, patch diagnostics.
4. RANK — rank = 1 + count(submissions with resolved > CAIN resolved); neighbours, gaps, thresholds, what-if curve.
5. TRAJECTORY — append a real snapshot; deltas only from previous real snapshots.
6. PUBLISH — atomically write /rank/live.json + /rank/history.jsonl + raw sources.
7. VERIFY — independently recompute rank/rate/checksum from the sources; refuse to publish on mismatch.

Source: `scripts/e124/rank_agents.py`. Refresh: cron, every 5 minutes. No rank is ever typed in by hand.
