EPS Dashboard

Overview / Motivation

Auto-filed by the /daily 2026-07-28 problem sweep (transcript miners over the day's 55 session transcripts). Sources: group-D P2/P3/P4/P6, group-I P9 (2 miners, one interactive session cluster).

Goal

Extend the compose-time verification / presentation-register discipline with five user-correction classes from 2026-07-28's interactive sessions.

Workflow gap

Bug observed: (1) A 'largest n ever tested = 3,600' figure was committed and pushed; Thomas corrected it — #779's scaling fits reach n=963,444 (the wrong ceiling also drove several turns of sizing advice). (2) The four-arm theory plan claimed kernel/MLP fits were 'new compute'; #779 had them banked at 50k AND 963k — two corrections before the doc was fixed. (3) '12 vs 36 vs 56 arms' took 3 asks to define; a GPU-hours question was answered in milestones and had to be re-asked. (4) When Thomas ordered removing the 200-draw bootstrap, the ack led with mechanics rather than the ~44 A100-h already-paid split — a 2-turn round trip ending in a reversed directive. (5) A prior session's null-space analysis existed only in chat; Thomas had to ask whether it ever reached the doc (it hadn't).
Why it is a workflow gap: the compose-time re-grep clause covers numeric headline re-reads but not superlative/coverage sweeps, chat-authored plan docs are outside the /issue reuse checklist, and no clause covers count-definition, sunk-compute-led acks, or same-turn doc folds for chat analysis.
Confidence (emitter): medium (interaction patterns read verbatim from transcripts; fixes are prose-rule additions)
verified-at-filing: grep -c 'Compose-time re-grep' CLAUDE.md → 1 (clause exists; extensions are additive); artifact-reuse.md governs /issue plans only (no chat-doc clause) (2026-07-29 UTC).

Proposed change (candidate diff sketch — refine in planning)

Five short clauses at the § 'Ad-hoc results summaries' / § 'Interim/chat writeup presentation register' family + one artifact-reuse.md line for chat-authored plan docs.

Scope / surfaces

Primary targets: CLAUDE.md, .claude/rules/artifact-reuse.md

Constraints / invariants

Workflow-surface only — never experiment code, configs/, or tasks/.
scripts/workflow_lint.py --check-asks passes; ruff on touched files passes.
The spawned session runs under a workflow_fix_target: Provenance line — recursion guard applies (it parks, never auto-routes, its own subagents' workflow-fix candidates).

Provenance

fingerprint: 9e3adbaf45e0
workflow_fix_target: CLAUDE.md