EPS Dashboard

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #1333 (emitting agent: clean-result-critic, round 1).

Goal

Add a WARN-level v4 check to verify_task_body.py counting sentences per interpretation paragraph in ### <result> prose (cap 1-3 per the v4 spec register), so the #385 incident class stops burning LM critic rounds.

Workflow gap

Bug observed: four of six results in #1333's body carried single >=4-sentence interpretation paragraphs; the v4 conciseness check (check 20) counts words only, so the spec's 1-3-sentence-per-analytical-paragraph rule was caught only by the clean-result-critic (a full REVISE round for paragraph splits).
Why it is a workflow gap: the sentence-per-paragraph cap is spec text (clean-result-critic-lens-reference.md Lens 12) with no mechanical backstop in verify_task_body.py, so it recurs on dense bodies.
Confidence (emitter): medium
verified-at-filing: grep -n "sentence" scripts/verify_task_body.py | grep -iE "per.paragraph|paragraph|cap" -> 0 hits (23 'sentence' mentions exist, none implement a per-paragraph sentence cap) (2026-07-15)

Proposed change (candidate diff sketch — refine in planning)

def check_v4_paragraph_sentence_cap(body): # WARN-only
```
for result in iter_v4_results(body):
```

    for para in interpretation_paragraphs(result):  # prose after the blockquote caption

        n = count_sentences(para)  # split on [.!?] guarding decimals, parens, "e.g."/"vs."

        if n >= 4: warn(f"result '{result.head[:40]}' has a {n}-sentence paragraph (v4 cap 1-3)")

Scope / surfaces

Primary target: scripts/verify_task_body.py (WARN-only; forward-only for v4 bodies — grandfathered v3/v2 never newly flagged)
Keep .claude/skills/clean-results/SPEC.md + clean-result-critic-lens-reference.md Lens 12 consistent (the spec text is the source; the check is the backstop).

Constraints / invariants

WARN-level only, never a FAIL (register judgment stays with the critic).
tests/test_verify_task_body.py pins the new check (a 4-sentence paragraph fixture WARNs; a 3-sentence one does not; decimals/e.g./vs. do not split).
This session runs under EPM_WORKFLOW_FIX_SESSION=1 (recursion guard applies).

Provenance

workflow_fix_target: scripts/verify_task_body.py
fingerprint: 3cd16468deda

(Verbatim candidate block preserved in origin_prompt.)