A11Y Content Judgment

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. Ratification records the human judgment, not a criterion outcome: a yes remains sample-scoped; a no may proceed through the separate finding/receipt path.

Two reasons this line is where it is, and both are load-bearing:

1. These rows land in conformance-adjacent documents. A fluent model opinion on whether an alt text is "adequate," multiplied across thousands of cells, is exactly the dead output an ACR must not carry. The rationale column exists so the ratifier can see why and disagree.
2. The judgment is about a person's experience, not a rule. "Learn more" is fine inside a paragraph that already names the destination and a failure in a grid of five identical cards. The rubric forces the draft to name what the person loses; a no without that is rejected at spot-check.

What it covers, and what it does NOT

Not covered — route elsewhere and say so: 1.3.1 beyond the level-skip flag (structure needs the rendered page and often AT); 1.4.1 use of color (visual); 3.3.x error handling and 3.2.2 on-input (interaction); 2.2.x timing (temporal observation); 4.1.2/4.1.3 name-role-value and status messages (screen-reader receipts); media alternatives 1.2.x, flashing 2.3.1, sensory 1.3.3, images of text 1.4.5, CAPTCHA (human observation). A request to extend the CSV to any of those is a scoping error.

The skill also does not compute accessible names. name in every row is a DOM approximation (aria-labelledby > aria-label > text content > image alt > title), recorded as name_source. When the approximation is what makes a row no, the ratifier verifies in a browser.

Pipeline

All four steps are files under references/. Run them from a directory where playwright resolves (they import it as a peer dependency).

1. Inventory — references/content-inventory.mjs

node references/content-inventory.mjs --urls-file views.txt --out ./content-inventory
# views.txt lines: url | id<TAB>url | id,product,url[,settle_ms]   (# comments allowed)
# options: --settle 2600 --viewport 1440x900 --only ID,ID --product name --engagement label --no-screenshots

Per URL, at one viewport, read-only, nothing activated: title, lang, h1 list, meta description, skip links; every heading with level, name source, and a section preview; every link with name, text, resolved href, file extension, new-window/external/icon-only signals, and the nearest enclosing text block as context; every image with alt presence, alt emptiness, aria-hidden, the control it sits inside, figcaption, size, and adjacent text; every form field with its computed label and where the label came from; every navigation landmark with its ordered link list. Caps per view (headings 250, links 500, images 250, fields 120) are recorded when hit. A navigation error is an environment note, never product evidence.

2. Build rows — references/build-judgment-rows.mjs --build

node references/build-judgment-rows.mjs --build --inventory ./content-inventory   # writes judgment-units.json + batches/*.jsonl

• Dedupes shared chrome. A unit is keyed on product + type + name + destination (+ a context hash for links, + alt/control for images), so a footer link on 19 pages is judged once and fans out with view_count and the views list. On the origin run this took 43 views' raw elements down to 1,899 rows.
• Attaches deterministic flags so the ratifier sees the machine's reasons separately from the model's opinion: title_shared_by_N_views, title_shares_no_word_with_h1, no_h1, empty_h1, heading_empty, heading_generic, heading_numeric_only, heading_repeated_in_page, level_skip_hN_to_hM, link_empty_name, link_generic_text, link_name_is_url, file_<ext>_not_indicated, new_window_not_indicated, icon_only_link_weak_alt, same_name_different_hrefs_in_page, img_missing_alt_attr, alt_generic, alt_is_filename, functional_image_no_name, alt_duplicates_control_text, alt_repeats_adjacent_text, svg_unnamed_not_hidden, complex_image_short_alt, field_unlabeled_<source>, label_generic, required_not_in_label, same_href_multiple_names. Flags are evidence to weigh, not verdicts; the rubric says so to the judge.
• 3.2.3 is decided here, not by a model. Per product and host, the navigation items present on at least half the views (minimum two) form the shared set; each view is checked for the same relative order of the shared items it carries. Extra or absent items are informational (a sub-navigation on detail pages is allowed); an order change is the signal.
• Writes judge worklists of ≤ 90 rows as batches/<product>-<type>-NN.jsonl.

3. Judge — one subagent per batch, rubric-bound

Give each judge exactly four things: references/judgment-rubric.md, references/judge-prompt.md, one batch, and a one-paragraph product-and-audience note (what the product is, who uses it) — the rubric's audience-shorthand rule cannot be applied without it, and the batch line does not carry it (the origin run leaked product only through the batch filename, which is the mechanism behind its over-hedging bias). It writes <batch>.judged.jsonl with, per row: judgment (yes | no | unsure), confidence, rationale (≤ 25 words naming what the person experiences), fix (≤ 20 words), needs_human, drafted_by (model id). Constraints that matter: native Read/Write only, no nested agents, one output line per input line, never invent what a destination page contains.

Model routing: judgment on the hosted tier (Sonnet-class handled the origin run; see the calibration record below). A local model may run as a detector to pre-sort, never as the drafted_by of record — the local tier's "detector, not verdict authority" rule from this repo's benchmark lane applies without exception here.

4. Merge, spot-check, hand off — --merge

node references/build-judgment-rows.mjs --merge --inventory ./content-inventory
# writes draft-judgments.csv, draft-judgments-<product>.csv, nav-consistency.csv, draft-judgments.json

Before handing the CSV over, the orchestrating session spot-checks: every unsure, every row where a heuristic flag and the draft disagree (flagged but yes, clean but no), and a random 10 % of the rest, reading the row's context and, where the call hinges on it, the screenshot or the live page. Write the results to spot-checks.jsonl ({id, spot_check: agree|overturn, judgment?, note?}); the merge renders them in the spot_check column and carries the effective value in session_judgment (the overturned value where one exists, else the draft). --sample writes spot-check-sample.jsonl with exactly that selection (seeded, reproducible). This is the step that makes the drafts the session's draft rather than a delegated one — still a draft. Run it with a stronger tier than the first pass where one is available, and have the reader check each rationale against the row: on the origin run, first-pass rationales occasionally named chemicals, landmarks, or titles that were not in the captured row, with the verdict still correct on what the row did show.

CSV columns: status, id, product, type, sc, view_count, views, name, detail, href, context, landmark, selector, visible, flags, draft_judgment, confidence, rationale, fix, needs_human, drafted_by, spot_check, session_draft_judgment, ratified_by, ratified_judgment, ratifier_note. status is the first column on every row (DRAFT_NOT_RATIFIED until a person signs that row) so no consumer can read a draft as a verdict; session_draft_judgment is named for what it is. The last three are blank on delivery and are display fields populated from the human's durable return, not edits to the generated CSV. Rows sort no → unsure → yes, widest fan-out first, so the ratifier's first hour lands on the rows that matter.

For a portable human handoff, use the content-review return loop. review-content-judgments.mjs --export generates a draft worklist, a CSV for an ordinary tracker workspace, and a blank response template. --apply validates the returned observation against the exact inventory snapshot, unit pin, and prior decision before updating ratifications.jsonl; run --merge afterward to refresh the views. It does not write to a tracker or authenticate the person named. Assignments, disagreements, and next work actions remain in the receiving tracker. The wrapper covers WCAG content judgments only; direct JSONL retains the separate client scope.

Receipt discipline

• draft-judgments.json carries status (DRAFT_NOT_RATIFIED → PARTIALLY_RATIFIED → RATIFIED), the units file hash, the rubric filename, and a coverage_basis block (views inventoried, viewport, caps, activation = none). It is never an input to an outcome map.
• A yes row is evidence of no defect in the captured sample only; it never supports a criterion outcome. One viewport, nothing activated, capped counts: everything behind interaction, every other viewport, and everything past the caps was never inventoried, so an all-yes ratified CSV is not evidence for "Supports" on 2.4.4 or 1.1.1. Only a ratified no travels (as a finding).
• Ratification is a file, not a CSV edit: ratifications.jsonl lines of {id, ratified_by, ratified_judgment, ratifier_note, ratified_utc, ruling}; --merge fills the ratifier columns from it. A family ruling ("every logo that is a link is named by its destination") is one ruling id fanned out over its rows, so the CSV shows which sentence of the owner's decided each row. Rows a ratifier defers or skips carry a note and no ratified_by.
• A name alone does not complete a return. --merge requires a name, a yes | no | unsure judgment (or a nonblank client result), and a valid review date before exposing effective ratifier fields. New pinned or correction records require a real UTC timestamp ending in Z. Historical records without pins or correction fields also accept a calendar-valid YYYY-MM-DD, labeled date_precision: day and legacy_unpinned; no time is invented. Shape alone cannot authenticate a historical record's age. Incomplete records remain draft with diagnostics. Malformed JSON, invalid IDs/scopes/fields/values, stale supplied unit pins, or conflicting records exit 2 before generated outputs change; only a missing optional file is treated as absent. This is validation atomicity, not a transaction across sequential output writes on a failing filesystem.
• Corrections name what they replace. Identical records are retries, not additional decisions. A changed record requires supersedes equal to the prior computed decision ID and a nonblank supersession_reason. Orphan and forward references fail. The merged rows expose decision IDs, including incomplete drafts, along with ratification_state, unit pin, pin status, and diagnostics; client scope has independent client_ratification_* columns. Existing complete unpinned records remain compatible and explicitly legacy_unpinned; they are not retroactively pinned evidence.
• A client-standard ruling is a separate scope: {..., scope: "client", ratified_client_result, ruling: "<rule id>"} renders as client_ratified_* columns and leaves the WCAG judgment column alone. A page title that is fine under 2.4.2 can sit on a page with three h1s that fails the client's one-h1 rule; the two verdicts must not overwrite each other, and the receipt that reaches an outcome map cites only the WCAG-scope one.
• Hand the ratifier a worklist, not the CSV: family decisions first (one answer settles many rows), then the rows that need a look (unsure, or a no with no fact class behind it), then the fact classes to sign once (missing alt attribute, empty link name, unlabeled field, raw-URL text). On the origin run that turned 1,847 rows into 8 questions, 53 rows, and 10 signatures; the owner answered the 8 questions in one message.
• A client may publish its own web standards beside WCAG. The origin run applied one such standard as a separate client scope — see the worked example in references/client-standards-example-epa.md — and the merge accepts a standards.jsonl ({id, <prefix>_rules, <prefix>_result, <prefix>_note} — any single prefix, read by suffix; the worked example uses epa_) that it renders as client_rules / client_result / client_note. That is the whole of what this skill provides: the matcher is engagement-specific, this skill does not prescribe a client-standards pass, and generalizing the rules to a table is deliberately not done until a second client standard exists. Client-scope results never reach the WCAG-scope receipt.
• A ratified row feeds the evaluation report's per-SC outcome map only through a receipt that names ratified_by, the ratification date, and the row id. The report contract's outcomes entry cites the receipt, not the CSV. drafted_by travels with it so nobody later mistakes the draft for the judgment.
• Rows with view_count > 1 ratify once and apply to every listed view; the receipt lists the views.
• Keep new captures in new inventory directories and retain their predecessors. IDs are stable grouping keys, not universal evidence revisions: a title unit, for example, is keyed to its view even when its text changes. New returned decisions therefore carry unit_sha256, the canonical unit snapshot hash; a supplied pin must match the current unit. The portable loop also pins the whole inventory snapshot. Changed source requires a fresh capture/review bundle, not editing or discarding historical ratifications to pass validation.

Gotchas from the origin run (do not re-learn)

• URL fragments AND query strings are part of identity. Skip links (/#main) collapsed into the home page until the destination key kept the hash; then 25 city and region links collapsed into the bare home URL, two PubMed articles into one, and ten map tabs into one, until it kept the query string too. Each collapse produced a false "different names for one destination" row that the first-pass judge could only mark unsure. The second reader caught the pattern from the rationale wording ("likely a lost parameter"); the fix is in the builder, not the rubric.
• javascript:/void(0) hrefs are not destinations. Basemap and zoom controls marked up as links grouped as seven names for one destination. Excluded from the 3.2.4 rows.
• Paired table columns are not a 3.2.4 case. An ID column and a name column that both link to the same record, on the same page, in main, produced 60 rows the first pass judged yes and the second reader confirmed as the table pattern. The builder skips a destination whose every name variant sits on the identical view set inside main. On the origin run this took the 3.2.4 rows from 103 to 25, all of which are now real questions (a version-number link in the header and a "Release Notes" link in the footer for one destination, on nearly every view, is the clearest).
• Informational flags must not drive spot-check selection. no_h1, h1_count_N, level_skip_*, name_from_*, title_duplicates_text, new_window_not_indicated describe the page, not the row's own judgment; counting them as "flagged" pulled 25 structurally fine titles into the disagreement sample. The sampler ignores them.
• Navigation landmarks fragment and nest. One site wrapped each top-level menu button in its own <nav> (nineteen nav landmarks, most with one item); another nested three navs and reused an id on the detail-page tab nav. Comparing nav signatures across views flagged half the pages as inconsistent; comparing relative order of the shared items flagged none, which matched the pages. That is why 3.2.3 is relative-order, not equality.
• Two-view groups make every item "shared." Minimum shared count is two, so a pair of pages is compared only on items both carry.
• Detail-page sub-navigation is not a 3.2.3 failure. Absent shared items are reported as informational; only order changes are the signal.
• Application shells (maps, single-page tools) have no shared navigation. The note says so; it is not a finding.
• A bare site name as <title> on most pages is the single most common 2.4.2 miss and the title_shared_by_N_views flag catches it before any model runs. On the origin run most views of one product shared a single bare title.
• Grid-heavy pages produce hundreds of near-identical field rows (column filters). Dedupe on type + label + label source keeps them to one row each; the ratifier is not asked 300 times.
• Never treat alt="" as automatically right or wrong. The image row records the control it sits inside; an empty alt on the only content of a link is functional_image_no_name, an empty alt beside a text label is usually correct. The rubric routes on that distinction.

Calibration record

Filled per run. A run that skips this section has not finished. Read the split, not the average. The agreement column is second-reader agreement — one model class reading another's drafts — not a human base rate; the ratifier's own overturn rate is a separate column filled from ratifications.jsonl and was not captured on the origin run. The unsure and clean-but-no rates say where the first pass needs the second reader most, so a budget-constrained run should second-read those two groups first and sample the rest.

Boundaries with sibling skills

• a11y-test measures and enumerates; it never judges descriptiveness. This skill consumes the same kind of URL list and can run beside references/baseline-url-scan.mjs on the same views.
• a11y-critic reviews design decisions and plans; if a no row implies a systemic pattern (every card grid uses "Learn more"), that pattern goes to the critic as one finding, not 40 rows.
• bug-reporting turns a ratified no into a filable issue; the row's selector, href, views, and fix are the inputs it needs.
• acr-reporting serializes ratified outcomes; its own text now carries the refusal rule (a content-judgment CSV row whose status is not RATIFIED is never an outcome input), and even a ratified yes is sample-scoped evidence, not a supports term.