a11y-content-judgmentlisted

# Content Judgment Skill (draft-and-ratify for the judgment criteria) > **Status: CANDIDATE — promoted 2026-09-02 from one engagement run; eval lane added the same day, gate not met as the rubric stood (rubric revised, re-draw pending — `evals/results/content-judgment-2026-09/`).** Origin: the > zivtech/a11y-audits (private) EPA interactive-retest engagement, OpenACR lane phase P3 bucket B3, > where the owner ruled that "determining these things are actually perfect use cases for using AI > to help with a11y tests" and delegated the *drafting* of B3 judgments to the agent while keeping > ratification human. First run: 43 views across two products, 1,899 deduplicated rows, 27 judge > batches. The promotion bar this skill has not yet met: a fixture set with planted defective and > planted *clean* rows (the false-alarm half matters more), scored across two model tiers and two > draws. Until then, treat the drafts as detector output behind a mandatory human pass. Apply this skill when a WCAG-EM or ICT-Baseline audit reaches the rows the crosswalk marks `partial` — the tests where a scanner can enumerate the elements but cannot say whether the text a person receives does its job. It sits **after** a11y-test (which measures) and **before** acr-reporting (which serializes ratified outcomes). It never replaces either. --- ## Core Mandate **The agent drafts. A named human ratifies. Nothing in this skill's output is a criterion outcome until the `ratified_by` column is filled by a