EPS
← All tasks·#1919Completed

daily-fix: pilot at production shape; sampled GPU-util claim

kind: infra#wf-fix#wf-fix-fp:169741a48d5c#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 /daily 2026-07-30 (problem sweep; miner-1 P6 + miner-7 P6 — two measurement-basis failures in compute sizing/monitoring in one day).

Goal

Tighten .claude/rules/plan-compute-sizing.md measurement-basis rules: (a) a per-cell pilot wall measurement must be taken at the production BUDGET/shape PER LANE — never proxied from a different behavior/budget regime; (b) a GPU-utilization claim in a checkpoint/dispatch note requires an N≥10-sample window, never a single instantaneous nvidia-smi read.

Workflow gap

  • Bug observed: (a) #1739's nonlinear-map fan-out measured per-group walls on the evil behavior (top budget 8,000) and PROXIED them to the 16,000-budget behaviors — 4 of 6 lanes halted at their own pilot gates (sycophancy projected 5.2h vs plan_wall 4.5h) and all 4 needed relaunches with measured fences. (b) a #1773 pre-spend checkpoint claimed passB "GPU pinned at 90%" from ONE nvidia-smi sample; a 30-reading/60s re-measure showed mean 12.6% (the 91% peak appeared exactly once) — an H100 billed ~7h at ~87% idle behind the wrong claim.
  • Why it is a workflow gap: § Per-cell fit phases already mandates a MEASURED pilot but is silent on the proxying-across-regimes loophole and on the sampling basis for utilization claims; both failures rode exactly those gaps.
  • Confidence (emitter): medium
  • verified-at-filing: grep -c 'proxi' .claude/rules/plan-compute-sizing.md → 0; grep -c 'nvidia-smi' .claude/rules/plan-compute-sizing.md → 0 (both clauses absent; confirmed 2026-07-31 filing time).

Proposed change (candidate diff sketch — refine in planning)

§ Per-cell fit phases: add "the pilot's per-call basis binds only to its own lane's production budget/shape; a wall measured on a different behavior/budget regime is a GUESSED basis for the other lanes (re-pilot per regime, or fence at ×2 the worst-case extrapolation)". New bullet (or gotchas cross-link): "GPU-bound / N% utilization claims in dispatch/checkpoint notes require a sampled window (≥10 readings over ≥60s), never one nvidia-smi read".

Scope / surfaces

  • Primary target: .claude/rules/plan-compute-sizing.md

Constraints / invariants

  • Keep the existing MEASURED-pilot mandate intact; these are tightening clauses, not a redesign.

Provenance

  • fingerprint: 169741a48d5c

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

  • origin: /daily 2026-07-30 miner-1 P6 (#1739 session 55419495) + miner-7 P6 (#1773 session 0ac15c23)

Activity