EPS Dashboard

Overview / Motivation

Auto-filed by /daily 2026-07-16 from two user corrections in transcript mining (09f28ede ~08:35Z; ea3a7991 ~06:08Z).

Goal

Close the two disclosure gaps today's user corrections exposed in the ad-hoc results-summary rule.

Workflow gap

Bug observed: two gaps in the 'Ad-hoc results summaries state per-arm provenance' bullet surfaced today: (a) the #1345 examples dashboard silently display-substituted 'ARIA'->'Assistant' in text presented as model generations (user had to ask whether the story really said ARIA); (b) a #1092 chat table compared prefix-R2 vs context-R2 arms scored against DIFFERENT targets and the user had to flag 'these results seem wrong' before a matched-target comparison was built
Why it is a workflow gap: Both incidents required the user to detect a silent presentation choice; the rule exists to make those disclosures automatic.
Confidence (emitter): high (two user corrections in one day)
verified-at-filing: grep -n 'Ad-hoc results summaries' CLAUDE.md -> L185 (bullet present); grep -c 'display-substitution\|matched-target' CLAUDE.md -> 0 (absence of both clauses)

Proposed change (candidate diff sketch — refine in planning)

extend the CLAUDE.md ad-hoc provenance bullet with: any display-substitution of raw generation text needs inline per-passage disclosure (dashboards included), and any cross-arm metric table shown in chat states matched-target/matched-corpus (or names the mismatch)

Scope / surfaces

Primary target: CLAUDE.md
Grep the workflow surface for the pattern before editing and update every hit; list them in the plan.

Constraints / invariants

scripts/workflow_lint.py --check-asks passes; ruff on touched files passes.
This session runs under a workflow_fix_target: Provenance line — it MUST NOT auto-route its own subagents' workflow-fix candidates (recursion guard).

Provenance

workflow_fix_target: CLAUDE.md
fingerprint: 07b219d40592