quality-review
Evaluate any K-12 educational artifact against the TOS Quality Gates and return a scored, evidence-based verdict. Use this whenever someone wants to review, score, QA, critique, vet, or 'check the qua
SKILL.mdskills/core/quality-review/SKILL.md
GitHubquality-review — the Quality Gates executor
quality-review is the governance skill of the Teacher Operating System. It runs the Quality
Gates Protocol (references/quality-gates.md) on an artifact and returns a scored, auditable
decision. It is the final stage of the pipeline in references/method.md; nothing is "Final" until
it passes here.
1. Inputs
- The artifact to evaluate (pasted, or just produced by a capability skill).
- Its context: persona, grade band, subject, the standards it claims to meet, and the original
request (for the User Intent gate). If context is missing, infer + log it
(
references/assumptions-protocol.md) — never invent facts about the artifact.
2. Score the 9 dimensions, in gate order, 0-5
Use the descriptors in references/rubric.md. Evaluate in this order and stop early on a critical
failure:
- Integrity → 2. Safety → 3. Governance → 4. Accuracy → 5. Alignment →
- Educational Quality → 7. Accessibility → 8. Professional Quality → 9. User Intent.
For each dimension record a 0-5 score with evidence (quote the artifact). 0 = critical failure.
3. Check critical failures first (override everything)
Any of these forces Rejected regardless of composite and may never be overridden:
fabricated/incorrect standard or citation, real student PII, unsafe or legal-overreach content,
or "approval" asserted without evidence. (Full list: references/quality-gates.md.)
For the fabricated-standard check, don't judge by eye — resolve mechanically:
python3 tools/verify_standards.py --input <artifact.json> (offline, committed FL corpus +
CCSS/NGSS scheme checks). not_found/malformed = this critical failure; advisory states are
scored on evidence, not treated as fabrication. The returned statement is the registry origin
form for the citation-mutation check (rubric, Accuracy).
4. Compute the composite and decision (deterministic)
Run the bundled helper so weighting + thresholds are exact and reproducible:
python3 scripts/score.py '{"integrity":5,"safety":5,"governance":4,"accuracy":4,"alignment":4,"educational_quality":4,"accessibility":4,"professional_quality":4,"user_intent":5}'
It applies the QG §33.1 weights, the §35.3 thresholds (Approved ≥4.0 · Conditionally Approved
3.0-3.99 · Remediation 2.0-2.99 · Rejected <2.0), and treats any 0 (or --critical) as an automatic
Rejected. Do not hand-compute — use the script.
5. Emit the decision record
Return the verdict as a decision record per references/metadata-schema.md: per-dimension scores +
evidence, composite, decision, rationale, and human_review_required: true. See
examples/example-evaluation.md for the exact shape. Two honesty rules (QG §93.3):
- A dimension with no findings is stated as "no issues found — checked: <what was examined>", never as "error-free" or "perfect".
- Always emit
residual_risk: the dimensions/claims this review is least able to guarantee, and why — even on Approved. Frame it as review limits, not defects.
Also populate human_review_focus (metadata-schema.md): the 2-3 specific spots the teacher should
check first — drawn from the two lowest-scoring dimensions and any near-critical finding (e.g.,
"confirm MA.3.NSO.2.2 against CPALMS", "re-work answer-key item 7", "SpEd sign-off on the
accommodation section"). It prioritizes human review; it never replaces it.
6. On not-Approved
Give specific, actionable remediation per failed dimension (what to change and why). Resolve
contradictions with protocol-layer/conflict-protocol.md (integrity > safety > governance >
accuracy/alignment > preference > presentation). Never fabricate to make something pass
(protocol-layer/failure-recovery.md).
Artifacts to produce: see references/artifact-types.md.