EPS
← All tasks·#1967Completed

daily-fix: per-tier pilot walls + c12 screen vocabulary

kind: infra#wf-fix#wf-fix-fp:8bc22471216c#daily-auto-filed
track:

Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.

No comments yet.

Overview / Motivation

Auto-filed by the /daily 2026-07-31 problem sweep (CONSOLIDATED H5; miner-3:P8, miner-1:P4/P3, miner-7:P1). Source sessions: 55419495 (#1739 — per-group fence walls measured on the evil behavior at its 8k token budget and proxied to 16k-budget behaviors: sycophancy pilot 302 s vs the assumed 140 s, so 4 of 6 lanes rc=7 self-halted and were relaunched with measured fences; round ~43-45 GPU-h vs ~30 named), 74476b0d (#1901 — a near-dupe screen shipped serial at 12,339 µs/row = ~3.3 h vs the 1.0 h sub-budget; died fail-loud and needed a mid-round 175× vectorize fix round that plan critics had passed), 3318f0b2 + 3a60e6ee (#1902 — pilot timed at n=354 vs production n=13,674, ~12× §9 deviation; adherence/vintage context below).

Goal

Add a heterogeneous-tier wall-time re-pilot clause to plan-compute-sizing and extend verify_plan's c12 trigger vocabulary to dedupe/near-dupe screens over a large fixed pool.

Workflow gap

  • Bug observed: (1) A measured per-cell/per-group WALL transferred across cells whose token-budget tier differed 2× (#1739) — the fence sized off the wrong tier self-halted 4/6 healthy lanes; no clause forbids the transfer or requires a re-pilot per distinct tier. (2) A near-dupe SCREEN phase escaped verify_plan's mechanical c12 battery check entirely (#1901) — the trigger regex covers permutation/bootstrap/null-draw/resample vocabulary only, while the plan-compute-sizing RULE's pool-scale clause explicitly names "pairwise similarity / near-dupe screens ~ pool²" as the covered kernel class: the rule and the verifier vocabulary have diverged.
  • Why it is a workflow gap: The 2026-07-29 #1798 landing (commit 6c75fcce7b) added production-shape/pool-scale pilot clauses and RAM largest-cell keying (the latter already citing the #1739 RAM incident) — so the CONSOLIDATED entry's pilot-at-production-n and RAM-row halves are ALREADY LANDED and are SCOPED OUT of this filing (the #1902 wrong-n pilot is plan-vintage/adherence against the landed clause). The two residuals above are verified absent: no tier-transfer clause anywhere in the rule, and no screen vocabulary in the c12 trigger.
  • Confidence (emitter): medium-high (residuals grep-verified; the #1739 tier-transfer attribution is the session's own compute-deviation marker diagnosis)
  • verified-at-filing: grep -n -iE 'does not transfer|re-pilot|per (distinct )?tier|budget tier|group size|pilot-shape' .claude/rules/plan-compute-sizing.md → 0 hits (absence = evidence for residual 1); grep -n -iE 'screen|dedup|near-dup|pairwise|per-row' scripts/verify_plan.py → hits only in unrelated contexts (L3034 precedence-phrase screen, L3040 quantifier screen, L1719 display-dedupe); _BATTERY_TRIGGER_RE + commitment regex read at L1305-1318: permutation/battery/null-draw/resample/bootstrap/B=\d{3,} tokens only, no screen/dedupe tokens (residual 2 confirmed). Landed-fix check: git log --oneline --since='7 days ago' -- .claude/rules/plan-compute-sizing.md scripts/verify_plan.py → 6c75fcce7b (2026-07-29, #1798) context read — implements pool-scale pilots + RAM keying (citing #1739's 163-GiB OOM), NOT the tier-transfer clause or the c12 vocabulary (2026-08-01).

Proposed change (candidate diff sketch — refine in planning)

.claude/rules/plan-compute-sizing.md § Per-cell fit phases:
+ TIER TRANSFER — a measured per-cell/per-group wall does NOT transfer
+ across cells whose group/budget/context size differs (a 2x token
+ budget is a different shape): heterogeneous lanes re-pilot per
+ distinct tier, or state the scaling and extrapolate — the wall-time
+ sibling of the RAM largest-cell keying clause (#1739: fence walls
+ measured at the 8k-budget behavior proxied to 16k lanes; 4/6 lanes
+ rc=7 self-halted).
scripts/verify_plan.py (check 12):
+ extend _BATTERY_TRIGGER_RE (and the commitment vocabulary) with
+ dedupe/near-dupe/similarity-screen tokens over a large fixed pool
+ (e.g. r"near[- ]dup(e|licate)? screen", r"dedup(e|lication)? (pass|
+ screen)", r"pairwise similarity"), so a pool-quadratic screen phase
+ triggers the same batched-commitment check as a draw battery (#1901).
+ pin test in tests/test_verify_plan.py.

Scope / surfaces

  • Primary target: .claude/rules/plan-compute-sizing.md; secondary: scripts/verify_plan.py (+ tests/test_verify_plan.py pin).
  • Grep the workflow surface for the pattern before editing (grep -rn 'pool-scale\|largest-cell' .claude/rules/ scripts/verify_plan.py) — keep the new clause adjacent + consistent with the landed #1798 blocks, and update the rule's frontmatter description line if the trigger summary changes.

Constraints / invariants

  • Workflow-surface only — never experiment code, configs/, or tasks/.
  • scripts/workflow_lint.py no-flags run passes; ruff on verify_plan.py passes; c12's documented non-trigger carve-outs (bare "bootstrap CI") must not regress — calibrate the new tokens against the corpus like #1796 did for c39.
  • Do NOT re-add the already-landed pilot-at-production-shape / pool-scale / RAM-keying clauses.
  • This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates.

Provenance

  • fingerprint: 8bc22471216c

  • workflow_fix_target: .claude/rules/plan-compute-sizing.md

  • fingerprint: (driver-computed; tag authoritative)

Activity