REX-10: perf ratchet: per-tag §5.1 job on primary cell, ratchet file, knob sweep per train #4365

Activity

1. noahgift

   added this to the 0.70.1 milestone on Sep 25, 2026
2. noahgift

   added a parent issue on Sep 25, 2026

   • PROMETHEUS (PRM-001 v3) epic: Qwen 3.5 4B local lane — fast enough and good enough to break quorum ties #4354
3. noahgift

   added a commit that references this issue on Sep 25, 2026

   REX-10: perf ratchet — paired Harrell–Davis p95 rise over the best ta…

   ...

   634916c
4. noahgift commented on Sep 25, 2026

   

   noahgift

   on Sep 25, 2026

   MemberAuthor

   More actions

   REX-10 landed on rex/001 @ 634916c (no PR: aprender is at the 10-PR cap, so it folds into B3).

   • review-lane-perf-ratchet-v1 (contract validates with pv): rex ratchet record records a tag from its warm, non-rerun receipts; rex ratchet check exits 10 on RED.
   • Rule: after 3 tags, each tag is compared with the best earlier tag using a paired bootstrap of the Harrell–Davis p95 ratio. It is RED if the CI lower bound is above 1. The only level used is the prereg's ALPHA (R-7).
   • Finding (FALSIFY-RXR-002): the §4 unpaired CI-overlap rule misses a +10% rise at n=130, and so does a paired nearest-rank p95. The HD paired rule catches the planted +10% in 40/40 runs with 0/40 false alarms, at a modelled 3% per-item run noise.
   • Planted RED: weakening the rule to lo>1.5 fails RXR-001. On the CLI, a +20% tag gives rc=10.
   • Acceptance is partly open: 3 real tags recorded = NotRun{NoDeclaredExecutor}, blocked on infra#1088 (no C4 harness executor). Until that lands, the per-tag job cannot run.
5. noahgift

   modified the milestones: 0.70.1, 0.71.0 on Sep 25, 2026
6. noahgift

   added

   P2Medium priorityMedium priority

   size:MK̂ 3-4 hK̂ 3-4 h

   on Oct 4, 2026
7. noahgift commented on Oct 4, 2026

   

   noahgift

   on Oct 4, 2026

   MemberAuthor

   More actions

   Folded into #4354 as a checklist line (ISSUE-TREE-001: a family of rows becomes one parent with checklist lines). Work and status are tracked there. aprender-wtix-even
8. noahgift

   closed this as not plannedon Oct 4, 2026