EPS Dashboard

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a prose follow-up surfaced by the codex-critic (alternatives lens) reviewing #810 plan v13, echoed by the Claude alternatives + methodology critics.

Goal

Add a verify_plan.py check that a plan's registered causal-branch wording does not exceed what its reported diagnostics can distinguish.

Workflow gap

Bug observed: #810 plan v13's falsification branch was worded as demonstrated "answer-content integration" ("they really were carrying answer content, the echo story dies") although the design cannot distinguish integration from OOD empty-turn degradation; three reviewers independently required a wording scope-down before approval.
Why it is a workflow gap: no mechanical check catches registered branch text whose causal claim outstrips the diagnostics; the pattern recurs whenever an ablation/intervention has a confounded falsification branch.
Confidence (emitter): medium

Proposed change (candidate diff sketch — refine in planning)

In scripts/verify_plan.py, add a check (WARN-level initially): for each registered hypothesis branch,
flag unqualified causal verbs/claims (e.g. integration =>, carries/carried <X> content, the <account> dies,
rewrites the <round> interpretation) unless the same paragraph names the undistinguished alternative
(OOD, endpoint, not uniquely diagnostic, degradation remains live, or equivalent).

Scope / surfaces

Primary target: scripts/verify_plan.py
Sibling: tests/test_verify_plan.py (pin the new check)

Constraints / invariants

Workflow-surface only; workflow_lint + ruff pass; WARN-level first (no retro hard-FAILs).

Provenance

workflow_fix_target: scripts/verify_plan.py
fingerprint: 4f6f8a6a7655

Surfaced prose (codex-critic alternatives, #810 v13, verbatim): "This is a recurring workflow-surface verifier candidate: causal branch wording should not exceed what the reported diagnostics can distinguish."