EPS Dashboard

Overview / Motivation

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

Goal

The code-review throughput check REQUIRES the diff (or review verdict) to name the batched helper implementing any per-layer/per-cell/per-fold fit loop, or an explicit not-batchable justification — absence is a FAIL, not a note.

Workflow gap

Bug observed: serial inner loops shipped past plan AND review twice in one day: #1332's serial per-layer loop had to be killed mid-run and vectorized (caught only by the mid-run ≥2× deviation check; 6d38a307 14:28-14:30Z), and #825's reused issue_779 MLP helper ran 120 serial CPU SGD fits after passing plan review (09f28ede 04:19-04:20Z).
Why it is a workflow gap: the reviewer's existing throughput checks are conditional — Step 0.68 fires only when the plan/body NAMES a helper, and the compute-throughput anti-pattern list flags specific shapes (batch-1 forwards, generate loops) — so a per-cell fit loop with no plan-named twin and none of the listed shapes passes with at most a note.
Severity: medium
verified-at-filing: grep -n 'batched' .claude/agents/code-reviewer.md → Step 0.68 "Named-helper adherence check" (L804-825: fires only for helpers the body/plan names by module::fn; records N/A when none named) and "Compute-throughput anti-patterns" (L942-961: flags batch-1 forwards / GPU→CPU transfers / HF generate / per-row serialization — no per-cell FIT-loop batched-helper requirement) — the proposed unconditional requirement for fit loops is absent from both (2026-07-16 UTC).

Proposed change (refine in planning)

Extend .claude/agents/code-reviewer.md's compute-throughput check (the anti-patterns block at L942, coordinating with Step 0.68 at L804): for ANY diff introducing a per-layer/per-cell/per-fold/per-draw fit or dense-factorization loop, the review verdict must name the batched helper implementing the inner loop (e.g. vectorized_mlp_skill.py, batched perm_null_draws, Gram/dual-space ridge) or record an explicit not-batchable justification — absence is a Major/FAIL finding, not a note, regardless of whether the plan named a helper. Mirror into the v2 efficiency-critic implementation-mode lens if trivially co-editable.

Scope / surfaces

Primary target: .claude/agents/code-reviewer.md (anchors: Step 0.68 L804; compute-throughput anti-patterns L942)
Secondary: .claude/agents/efficiency-critic.md (v2 implementation mode)

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: 49321aa6268a
workflow_fix_target: .claude/agents/code-reviewer.md

Mined from 2026-07-15 session transcripts by the /daily problem sweep. Evidence: 6d38a307 (#1332) 14:28-14:30Z (batch 04 P13); 09f28ede (#825) 04:19-04:20Z (batch 08 P8 leg a).