loop-verifier

Independent verifier for loop outputs. Checks artifacts and rubric only — not maker reasoning. Enforces cache citations, L1 no-auto-fix, and loop-budget compliance. Use after loop-triage.

SKILL.md

GitHub

Loop Verifier (maker/checker split)

You are the checker. You did not write the triage report. Judge the artifact only.

Inputs (read only these)

  1. The report under reports/loops/ (path from argument or latest by date)
  2. LOOP.md — level must match (L1 = report-only)
  3. loop-budget.md — L1 allows zero source file reads
  4. STATE.md — must be updated consistently with report
  5. loop-run-log.md — new row present

Do not re-read the codebase to "validate" findings. If the report claims a cache miss, accept or reject based on whether the miss is documented.

Rubric (pass/fail)

Check Pass criteria
Cache cited Report lists manifest + ≥2 cache paths/sections
L1 compliance No auto-fix, no commit/PR language, no source exploration claims
STATE sync STATE.md Last session + Cache used updated
Run log Append row with timestamp and artifact path
Brevity Executive summary ≤120 words
Token discipline No large code fences; section refs only

Output

## Loop Verification — [loop name]

**Artifact:** path
**Result:** PASS | FAIL

**Checks:**
1. Cache cited: pass/fail — note
2. L1 compliance: pass/fail
3. STATE sync: pass/fail
4. Run log: pass/fail
5. Brevity: pass/fail

**Required fixes:** (if FAIL, numbered; maker re-runs triage)

Fail closed: if any critical check fails, do not mark loop complete in STATE.