EPS Dashboard

Overview / Motivation

Auto-filed by the /daily 2026-07-15 problem sweep (route 2 — behavior/logic change, independent review required).

Goal

Enforcement leg: the critic (and/or verify_plan) REVISEs any §9 wall-time estimate for a permutation/bootstrap/null battery or many-cell fit that does not cite a measured pilot basis — escalating the existing WARN-only mechanical check to a binding review outcome.

Workflow gap

Bug observed: the #1092 refit plan undersized compute 2.6× (5 h/box planned vs 12.8 h/box measured) because the dominant 200-draw permutation null bands were priced by assumption, not the measured 1-cell pilot .claude/rules/plan-compute-sizing.md already mandates; Thomas asked "why did it take so long" (ea3a7991 22:30-22:31Z).
Why it is a workflow gap: the sizing rule mandates a measured pilot basis but the only mechanical check (verify_plan c32) is WARN-only and the critic's Methodology lens has no explicit REVISE trigger for an unpiloted battery estimate — so an asserted per-call cost survives plan review.
Severity: high
verified-at-filing: grep -n 'pilot' scripts/verify_plan.py → hits L5118-5213 (check c32: "an ASSERTED per-call cost is never a basis... Ground the row on a measured 1-cell pilot at production shape"; landed 2026-07-09 as WARN-only per commit 06a0dfced7 "verify_plan.py c32 fit-family §9 basis grounding (WARN-only)") — the mechanical check EXISTS but is non-binding; grep -n 'pilot\|measured' .claude/agents/critic.md → 1 hit (L142, a lens-roster mention of "measured wall-time" basis) with no REVISE trigger text for unpiloted perm/null-battery estimates — the binding enforcement leg is absent (2026-07-16 UTC).

Proposed change (refine in planning)

Two coordinated legs: (a) in .claude/agents/critic.md's Methodology lens, add an explicit REVISE trigger — a §9 wall-time estimate for a permutation/bootstrap/null-draw battery or many-cell fit whose basis is asserted (no measured 1-cell pilot figure, no cited prior-issue measured figure, no pilot-gated flag) is a REVISE, not a note; (b) evaluate escalating verify_plan.py's c32 from WARN to FAIL for the fit-family/battery rows specifically (keeping WARN for the broader class if false-positive risk is high). The #1092 2.6× undersize post-dates c32's landing, showing WARN-only does not bind.

Scope / surfaces

Primary target: .claude/agents/critic.md (Methodology lens; anchor L142)
Secondary: scripts/verify_plan.py (check c32, L5118-5213)

Constraints / invariants

Workflow-surface only — never experiment code, configs/, or tasks/.
scripts/workflow_lint.py --check-asks passes; ruff on touched files passes.
This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route its own subagents' workflow-fix candidates (recursion guard).

Provenance

fingerprint: 85adc172452b
workflow_fix_target: .claude/agents/critic.md

Mined from 2026-07-15 session transcripts by the /daily problem sweep. Evidence: ea3a7991 (#1092) 22:30-22:31Z (batch 06 P20; P21-P22 costs downstream).