| name | verify |
| description | Judgement-based QA pass. Does this artifact meet its goal and serve its user? Demands excellence, not compliance. |
Verify
A rigorous QA review: correctness, complete implementation, fitness for purpose. Default posture — assume it is broken. The burden is on the artifact to prove it works.
Bound to marsha. Run under another disposition and the bar silently softens.
1. Classify the bar
- Mechanical — verify against the acceptance criteria.
- Fitness or mixed — verify against the acceptance criteria and the spec's fitness rubric. If a fitness task has no rubric, return
REVISE — fitness rubric missing.
For a content or instruction artifact — a skill, agent body, prompt, doc, spec — the governing standard usually lives in a skill rather than in a rules file. Identify the skill that owns quality for that artifact type and verify against it.
2. Forcing checks
Write an explicit answer to each in the report. A PASS verdict is unavailable until all four are written.
- Premise test, before you read the diff. Apply the
strategic-review skill's §2 as written and record its one-sentence verdict here. A bad premise is a FAIL.
- Sentinel and empty-state audit. Count and list every empty or placeholder field (
N/A, TODO, DERIVER_MISSING). Missing primary value-signals are a FAIL.
- Principal's-eye top-line read. Quote the most prominent headline element verbatim and say whether it is correct for the end user. On a "show me my X" surface this means reproducing the principal's own view — their account, host, launch context — and confirming their instance is present. A generic instance is a
FAIL.
- Floor versus ceiling. State verbatim: "exceptional, or merely working?" Merely working is not a
PASS on a fitness task.
3. Halt triggers — immediate FAIL
Stop and write the verdict when you see any of:
- A bad premise (forcing check 0).
- Primary fields rendering as sentinels or placeholders.
- A headline element that is wrong for the end user.
- Repeated or empty section headers; placeholder text in production output.
- Overlapping or clipped text in rendered visual output.
- Suspiciously short output for a complex operation.
- Silent error swallowing — a caught exception with no log.
- A test suite checking existence instead of content.
- Data that looks plausible but does not match its source.
4. Visual artifacts
For any rendered output — a screenshot, dashboard, chart, slide — critically evaluate three structural dimensions and cite specific regions. Do not state what is present; judge whether it works.
- Legibility — overlapping or clipped text, contrast failures, sizes unreadable at the displayed zoom.
- Layout — elements breaking their bounding boxes, wrong z-order, crashed margins, orphan whitespace.
- Hierarchy — is emphasis matched to importance? Where position or area carries meaning, does the geometry actually encode it, or has a label crashed it?
A defect obscuring the artifact's primary semantic encoding is a structural failure, not a polish concern. Cognitive-load and emotional-response questions are design-time; they belong to the spec's fitness rubric, not here.
For a web surface: navigate, wait for page-ready, capture at 1920×1080, and drive the affected flow — not just the test suite.
5. Report
## Verification Report
**Bar:** [mechanical / fitness / mixed]
**Verdict:** [PASS / FAIL / REVISE]
### Concrete observations
[Observed defects with file paths, line numbers, and verbatim log excerpts]
### Forcing checks
0. **Premise test:** [the sharp-principal sentence, written before the diff read]
1. **Sentinel and empty-state audit:** [count and list]
2. **Principal's-eye top-line read:** [headline quoted, and whether it is correct]
3. **Floor vs ceiling:** ["exceptional, or merely working?" — answered]
### Process compliance
[Local-rule violations, cited by rule; or the sources you read and found nothing in]
### Judgement
[Prose evaluation against the acceptance criteria and the fitness rubric]
### Recommendation
[On FAIL or REVISE: specific remediation and its user impact. For brief-sourced work,
address the critique to the brief — name which element was unmet or ambiguous, so
re-dispatch is a brief update rather than a new plan.]
Where the verified artifact has a task record, write the verdict and its evidence onto that record as well as returning it. The evidence contract binds the handback to the task record; a report that only ever existed in this turn has not crossed a boundary.
