Content QA Rubric

1) The problem: a strong average can hide a severe failure

Imagine a service page with excellent style, writing and SEO, but one unsupported performance claim, a form label that is not programmatically associated with its control, and a CTA that contradicts the promised next step. A single average can still look healthy. Professional QA must do the opposite: evaluate important dimensions independently and prevent strength in one dimension from compensating for a critical failure in another.

2) COM-010 scope: content QA, not SEO design or AI lineage

This lesson owns Commercial Content QA Rubric v1. COM-009 owns SEO Page Schema and intent-driven content, while COM-011 owns Human Anchor + AI Variants. COM-010 does not decide how content is generated or published; it builds an auditable review contract for one exact artifact version.

3) Mental model: Artifact Version → Dimension Checks → Critical Blocker? → Repair/Retest → Final State

Every QA run binds to an exact artifact_id and artifact_version. Dimensions are evaluated independently. An open critical blocker cannot be overridden by an average or a favorable overall impression. After repair, rerun only affected dimensions unless the artifact version or governing evidence changed more broadly.

4) FACT: Google's people-first guidance is useful evidence, not MAKNOW release authority

Google Search Central continues to recommend content made primarily to help people and offers self-assessment questions about usefulness, expertise, trust and whether a reader can accomplish their goal. Those questions are useful QA inputs, but Google does not define MAKNOW release thresholds and the guidance is not a ranking guarantee.

5) FACT: accessibility is an independent, testable dimension

WCAG 2 provides testable success criteria for web accessibility, and W3C encourages use of the latest version, WCAG 2.2. An accessibility defect should therefore not disappear inside a generic writing-quality score. At the same time, WCAG conformance alone does not prove factual accuracy, usefulness or conversion quality.

6) The 22-field contract: no decision without trace

Each QA row carries enough context and evidence to reproduce the decision: qa_run_id, artifact_id, artifact_version, reviewer, dimension, criticality, criterion, observable_evidence, state/score, severity, blocker, claim references, accessibility/intent/helpfulness/duplication/conversion checks, repair, retest, final state and evidence_ref.

7) Criticality, severity and blocking flag are not the same thing

criticality states whether a dimension may block release. severity describes the observed defect. blocking_flag applies policy to that defect. If a critical accessibility failure is merely relabeled advisory, the rubric is no longer trustworthy; the correct state is FAIL_SEVERITY_POLICY.

8) Claim support and critical accessibility are non-compensating

An unsupported factual or commercial claim on a critical dimension becomes FAIL_CLAIM_SUPPORT. A critical accessibility defect becomes FAIL_ACCESSIBILITY. Style, SEO or conversion strength cannot turn either into PASS. An aggregate score may summarize evidence, but it is never the authority that overrides a blocker.

9) Do not collapse helpfulness and intent fit into SEO

A page can have excellent keywords and structure while still leaving the reader unable to complete the promised task. helpfulness_check asks whether the artifact helps the real audience; intent_fit_check asks whether promise, content and next step align with the defined intent. Neither is equivalent to ranking signals or keyword density.

10) Judge duplication with purpose and audience, not similarity alone

High similarity is not automatically a defect. But when a new variant is substantially duplicative and lacks a clear separate audience intent, the safe result is HOLD_INTENT_DUPLICATION. If intent is clear but duplication still needs reduction, use REVISE_DUPLICATION. A similarity tool never receives release authority by itself.

11) Missing evidence is not zero; it is HOLD

If observable_evidence or evidence_ref is missing for a critical dimension, do not guess and do not enter zero just to finish an average. Use HOLD_MISSING_EVIDENCE. This typed non-success state distinguishes “the artifact failed” from “we do not yet have enough evidence to decide.”

12) Version binding: an old PASS does not survive a new artifact version automatically

A QA run is bound to exact bytes or an exact artifact version. If the page changes after review, copying the previous PASS onto the new artifact_version is invalid. Use HOLD_STALE_ARTIFACT until affected dimensions are re-evaluated. This blocks approval-by-copy.

13) Repair loop: smallest valid repair, smallest valid retest

For each defect, repair_action states the bounded fix and retest_result records the new evidence state. Fixing a form label does not require rebuilding claim-support evidence from zero; changing a claim does require claim-support and dependent dimensions to rerun. A strong rubric preserves unaffected evidence while following dependency changes.

14) Major failure modes and typed states

• all critical gates closed, current version, required retests PASS → PASS
• open noncritical defect → REVISE_ADVISORY
• reducible duplication with clear intent → REVISE_DUPLICATION
• unsupported critical claim → FAIL_CLAIM_SUPPORT
• critical accessibility failure → FAIL_ACCESSIBILITY
• critical failure downgraded to advisory → FAIL_SEVERITY_POLICY
• missing decisive evidence → HOLD_MISSING_EVIDENCE
• prior PASS copied to a changed version → HOLD_STALE_ARTIFACT
• material duplication plus ambiguous intent → HOLD_INTENT_DUPLICATION

15) Lab and transfer: rerun only the affected dimensions

The offline lab contains eight fixtures and 48 deterministic assertions. In transfer, claim support and accessibility have already been repaired, but a new variant is substantially duplicative and its audience intent is ambiguous. Rerun only duplication and intent, preserve unrelated accepted evidence, and derive the final state from visible evidence. This lesson does not prove live publishing, legal approval or Search ranking.