ultra11y 2.31.2 → 2.32.0

package/src/llm.ts

@@ -20,7 +20,7 @@
 // changes: the engine's "no keys, no install" promise holds everywhere else.
 //
 // Zero dependencies: global `fetch`, no SDK.
-import type { AdjudicationItem, AgentFinding, CriterionVerdict } from "./adjudicate.js";
+import type { AdjudicationItem, AgentFinding, CriterionVerdict, Evidence } from "./adjudicate.js";
 
 export const DEFAULT_MODEL = "claude-sonnet-5";
 const API_VERSION = "2023-06-01";
@@ -97,6 +97,21 @@ const VERDICT_TOOL = {
                 required: ["file", "line", "message", "normativeRef"],
               },
             },
+            citations: {
+              type: "array",
+              description:
+                "Required for C and NA when evidence was presented: the evidence items you cleared (or ruled out of scope). Copy `file`, `line` and `snippet` VERBATIM from the evidence list of this criterion — a citation that is not among them, or that no longer matches the source, is rejected.",
+              items: {
+                type: "object",
+                properties: {
+                  file: { type: "string" },
+                  line: { type: "number" },
+                  selector: { type: "string" },
+                  snippet: { type: "string" },
+                },
+                required: ["file", "line"],
+              },
+            },
             recommendations: {
               type: "array",
               description: "Non-normative good practices. They never change a status.",
@@ -126,7 +141,7 @@ const SYSTEM = `You are an accessibility auditor ruling on the criteria a static
 Rules, in order of importance:
 1. NEVER assert conformity you did not verify. "manual" with a reason is always available and is a correct answer.
 2. An NC must cite a real file:line taken from the evidence you were given, and the criterion's OWN numbered test as normativeRef. A citation that does not resolve against the real source is rejected downstream and wastes the whole batch.
-3. C and NA require a justification that says what you saw, not that you checked.
+3. C and NA require a justification that says what you saw, AND a "citations" array naming the evidence items you cleared — file and line copied verbatim from the evidence presented for that criterion. A criterion presented with NO evidence cannot be C: record "manual" (reason "undecidable"), or NA if nothing in scope is concerned.
 4. A criterion that needs a rendered page (computed contrast, visible focus, zoom, reflow) and was given only source evidence is "manual" with reason "needs-rendered-dom".
 5. Rule only on the criteria presented. Never introduce another.`;
 
@@ -146,6 +161,7 @@ export interface RawVerdict {
   justification?: string;
   reason?: string;
   findings?: AgentFinding[];
+  citations?: Evidence[];
   recommendations?: AgentFinding[];
 }
 
@@ -252,6 +268,11 @@ export function applyRawVerdicts(items: AdjudicationItem[], verdicts: RawVerdict
     item.justification = v.justification ?? "";
     item.reason = v.verdict === "manual" ? (v.reason ?? null) : null;
     item.findings = v.verdict === "NC" ? (v.findings ?? []) : [];
+    // Citations belong to the clearing verdicts, the way findings belong to NC. Dropping
+    // them here (as this function once dropped everything a non-NC verdict carried) would
+    // make every model-produced C fail the gate that now requires them.
+    if (v.verdict === "C" || v.verdict === "NA") item.citations = v.citations ?? [];
+    else delete item.citations;
     if (v.recommendations?.length) item.recommendations = v.recommendations;
     filled++;
   }

package/src/pages-report.ts

@@ -19,12 +19,14 @@
 //      `verify` (AUDITOR_OCCURRENCE, src/verify.ts).
 //
 // The screenshot is the one thing the page report has that no other output does. It is
-// referenced, never copied: `.ultra11y/pages/<id>/screen.png` is already on disk, so the
-// caller resolves a relative path and passes it in — this module stays pure and testable.
+// passed in as an href, never resolved here: `.ultra11y/pages/<id>/screen.png` is already
+// on disk, and the caller decides whether to reference it relatively or copy it beside the
+// report (it does copy, so an uploaded `audits/` artefact keeps its images), which keeps
+// this module pure and testable.
 import { prdUnits } from "./prd.js";
 import { renderAuditorUnit } from "./auditor.js";
 import { pageView } from "./pages.js";
-import { CORE, type StandardId, derivePackResults, isCore, loadPack, themeName, titlePlain } from "./standards/index.js";
+import { CORE, type StandardId, derivePackResults, isCore, loadPack, packTestIds, themeName, titlePlain } from "./standards/index.js";
 import type { AuditResult, Lang, PageResult, Status } from "./types.js";
 import { compareSC, scTitle } from "./wcag.js";
 
@@ -61,9 +63,13 @@ const L = {
     viewport: "Fenêtre",
     producer: "Producteur",
     auth: "Authentification requise",
-    rate: "Taux de conformité",
+    rate: "Taux de réussite automatique (vérifications statiques)",
     rateNote: "sous-ensemble décidable : C ÷ (C + NC)",
     tally: (c: number, nc: number, na: number, m: number) => `${c} conforme(s) · ${nc} non conforme(s) · ${na} non applicable(s) · ${m} à évaluer`,
+    coverage: (decided: number, total: number) =>
+      `Couverture : ${decided}/${total} critère(s) évalué(s) — le taux ci-dessus ne porte que sur eux, et ne dit rien des ${total - decided} autres.`,
+    tests: "Tests",
+    agentMark: "`C*` : conformité tranchée par l'agent IA à partir des évidences citées (gaté), et non prouvée par le moteur déterministe.",
     screenshotAlt: (n: string) => `Capture d'écran de la page ${n}`,
     noScreenshot: "Aucune capture d'écran pour cette page (le producteur n'en a pas fourni) — le tier pixel est donc inactif ici.",
     gridTitle: "Grille des critères",
@@ -101,9 +107,13 @@ const L = {
     viewport: "Viewport",
     producer: "Producer",
     auth: "Authentication required",
-    rate: "Conformance rate",
+    rate: "Automatic static-check pass rate",
     rateNote: "decidable subset: C ÷ (C + NC)",
     tally: (c: number, nc: number, na: number, m: number) => `${c} conforming · ${nc} non-conforming · ${na} not applicable · ${m} to assess`,
+    coverage: (decided: number, total: number) =>
+      `Coverage: ${decided}/${total} criteria assessed — the rate above covers only those, and says nothing about the other ${total - decided}.`,
+    tests: "Tests",
+    agentMark: "`C*`: conformity ruled by the AI agent from the evidence it cited (gated), not proven by the deterministic engine.",
     screenshotAlt: (n: string) => `Screenshot of the ${n} page`,
     noScreenshot: "No screenshot for this page (the producer supplied none) — the pixel tier is therefore inactive here.",
     gridTitle: "Criteria grid",
@@ -144,6 +154,12 @@ interface Row {
   label: string;
   group: string;
   status: Status;
+  // The criterion's own numbered tests (pack standards only — the WCAG core's analogue is
+  // its techniques, which are advisory rather than the thing you rule on). Rendered in the
+  // grid so a sheet says WHAT has to be checked for every criterion, not only for the ones
+  // that happened to trigger a non-conformity.
+  tests: string[];
+  decidedBy?: "engine" | "agent" | "scan";
 }
 
 /** The criterion rows for ONE page, in the active standard's own vocabulary and order.
@@ -152,15 +168,24 @@ function rowsFor(result: AuditResult, page: PageResult, standard: StandardId, la
   if (isCore(standard)) {
     return [...page.criteria]
       .sort((a, b) => compareSC(a.id, b.id))
-      .map((c) => ({ id: c.id, label: `${c.id} ${scTitle(c.id, lang) ?? ""}`.trim(), group: c.guideline, status: c.status }));
+      .map((c) => ({
+        id: c.id,
+        label: `${c.id} ${scTitle(c.id, lang) ?? ""}`.trim(),
+        group: c.guideline,
+        status: c.status,
+        tests: [],
+        decidedBy: c.decidedBy,
+      }));
   }
   const pack = loadPack(standard);
-  const byId = new Map(derivePackResults(pageView(result, page), standard).map((r) => [r.id, r.status]));
+  const byId = new Map(derivePackResults(pageView(result, page), standard).map((r) => [r.id, r]));
   return pack.criteria.map((pc) => ({
     id: pc.id,
     label: `${pc.id} — ${titlePlain(pack, pc, lang)}`,
     group: `${pc.theme}. ${themeName(pack, pc.theme, lang) ?? ""}`.trim(),
-    status: byId.get(pc.id) ?? "manual",
+    status: byId.get(pc.id)?.status ?? "manual",
+    tests: packTestIds(pack, pc.id),
+    decidedBy: byId.get(pc.id)?.decidedBy,
   }));
 }
 
@@ -202,7 +227,10 @@ export function renderPageReport(result: AuditResult, page: PageResult, opts: Pa
   const rows = rowsFor(result, page, standard, lang);
   const t = tally(rows);
   out.push(`- **${s.rate}** : **${ratePct(rows)} %** _(${s.rateNote})_`);
-  out.push(`- ${s.tally(t.c, t.nc, t.na, t.manual)}`, "");
+  out.push(`- ${s.tally(t.c, t.nc, t.na, t.manual)}`);
+  // The rate alone reads as a verdict on the page. Naming the denominator next to it is
+  // what stops "100 %" over six decided criteria from being quoted as a conformant page.
+  out.push(`- ${s.coverage(t.c + t.nc, rows.length)}`, "");
 
   // A source-only page cannot earn conformity by silence. Say so ON THE PAGE, not only in a
   // legend the reader may never reach — this is the sentence that stops a clean-looking
@@ -213,17 +241,28 @@ export function renderPageReport(result: AuditResult, page: PageResult, opts: Pa
   if (shot) out.push(`![${s.screenshotAlt(page.name)}](${shot})`, "");
   else out.push(`_${s.noScreenshot}_`, "");
 
+  // The grid carries the criterion's own numbered TESTS, not just its title and status:
+  // those tests are what has to be checked, and listing them only inside the
+  // non-conformity blocks meant a sheet said nothing about the work still to do on the
+  // ~99 criteria that never triggered. Test ids go in backticks and are never followed by
+  // an em dash — `check`'s criterion scanner matches `(\d+\.\d+)\s*—` and would capture
+  // "2.1" out of "6.2.1 — …".
+  const withTests = rows.some((r) => r.tests.length);
   out.push(`${h}# ${s.gridTitle}`, "", `> ${s.gridNote}`, "");
-  out.push(`| ${s.criterion} | ${s.status} |`, "| --- | --- |");
+  out.push(withTests ? `| ${s.criterion} | ${s.tests} | ${s.status} |` : `| ${s.criterion} | ${s.status} |`);
+  out.push(withTests ? "| --- | --- | --- |" : "| --- | --- |");
   let group = "";
   for (const row of rows) {
     if (row.group !== group) {
       group = row.group;
-      out.push(`| **${group}** | |`);
+      out.push(withTests ? `| **${group}** | | |` : `| **${group}** | |`);
     }
-    out.push(`| ${row.label} | ${MARK[row.status]} |`);
+    const mark = row.decidedBy === "agent" && row.status === "C" ? `${MARK[row.status]}*` : MARK[row.status];
+    if (withTests) out.push(`| ${row.label} | ${row.tests.map((t) => `\`${t}\``).join(" ")} | ${mark} |`);
+    else out.push(`| ${row.label} | ${mark} |`);
   }
   out.push("", `> ${s.manualWarn}`, "");
+  if (rows.some((r) => r.decidedBy === "agent" && r.status === "C")) out.push(`> ${s.agentMark}`, "");
 
   // The non-conformities, rendered by the ONE auditor block (invariant 2). Feeding
   // `prdUnits` the page view is what scopes them to this page without a second grouping.

package/src/prd.ts

@@ -14,6 +14,7 @@ import {
   loadPack,
   derivePackResults,
   packConformancePct,
+  packTestIds,
   standardLabel,
   themeName,
   titlePlain as packTitlePlain,
@@ -363,8 +364,8 @@ export function toRuleOnSection(r: AuditResult, standard: StandardId, lang: Lang
   for (const pc of manual) {
     const crit = pack.criteria.find((c) => c.id === pc.id);
     const title = crit ? packTitlePlain(pack, crit, lang) : pc.id;
-    const tests = Object.keys(crit?.tests ?? {});
-    out.push(`- [ ] **${pack.name} ${pc.id}** — ${title}${tests.length ? `  ·  ${s.tests}: ${tests.map((k) => `\`${pc.id}.${k}\``).join(" ")}` : ""}`);
+    const tests = packTestIds(pack, pc.id);
+    out.push(`- [ ] **${pack.name} ${pc.id}** — ${title}${tests.length ? `  ·  ${s.tests}: ${tests.map((t) => `\`${t}\``).join(" ")}` : ""}`);
   }
   out.push("");
   return out;

package/src/report.ts

@@ -4,7 +4,7 @@
 // WCAG-keyed result. Both keep the honest structure: per-guideline/theme synthesis,
 // non-conformities, conforming + not-applicable lists, and the manual worklist
 // (never silently C).
-import { existsSync, mkdirSync, writeFileSync } from "node:fs";
+import { copyFileSync, existsSync, mkdirSync, writeFileSync } from "node:fs";
 import { join, relative } from "node:path";
 import type { AuditResult, Finding, Lang, Severity, Status } from "./types.js";
 import { guidelineTitle, scTitle } from "./wcag.js";
@@ -18,10 +18,10 @@ import {
   CORE,
   isCore,
   loadPack,
-  getCriterion as getPackCriterion,
   derivePackResults,
   packCriteriaForFinding,
   packConformancePct,
+  packTestIds,
   title as packTitle,
   themeName,
   type StandardPack,
@@ -56,6 +56,9 @@ const L = {
     sev: { bloquant: "Bloquant", majeur: "Majeur", mineur: "Mineur" } as Record<Severity, string>,
     none: "Aucune non-conformité détectée par le moteur statique.",
     cTitle: "3. Critères conformes (C)",
+    cAgentTitle: "Conformes par adjudication de l'agent (jugement, non prouvé par le moteur)",
+    cAgentNote:
+      "Ces critères ont été tranchés par l'agent IA à partir des évidences citées, non décidés par le moteur déterministe. Ils sont gatés (chaque verdict cite une évidence résolvable) mais restent un jugement : ils ne sont pas comptés dans le taux de réussite automatique ci-dessus.",
     naTitle: "4. Critères non applicables (NA)",
     manualTitle: "5. Critères à adjuger (jugement / rendu) — non décidés par le moteur statique",
     manualWarn:
@@ -65,6 +68,7 @@ const L = {
       "Générez la worklist : `verify --manual --in <audit.json> --standard <pack> --out <dir>`. Chaque item y porte l'énoncé complet de ses tests, sa note technique, ses cas particuliers, sa guidance et les termes que le référentiel définit.",
     outOfScope: "Hors périmètre moteur — mappé sur des SC hors WCAG 2.2 AA ; vérification manuelle.",
     scopedOut: "Les non-conformités WCAG relevées concernent des éléments hors du périmètre de ce critère — à évaluer séparément.",
+    judgment: "L’énoncé du critère demande davantage que les CS WCAG auxquels il est rattaché — le moteur n’y a pas répondu, à trancher.",
     nothing: "Aucun.",
     dedup: "Dédup",
     canonical: "fichier(s) canonique(s) audité(s)",
@@ -117,6 +121,9 @@ const L = {
     sev: { bloquant: "Blocking", majeur: "Major", mineur: "Minor" } as Record<Severity, string>,
     none: "No non-conformity detected by the static engine.",
     cTitle: "3. Conforming criteria (C)",
+    cAgentTitle: "Conforming by agent adjudication (judgement, not proven by the engine)",
+    cAgentNote:
+      "These criteria were ruled on by the AI agent from the evidence it cited, not decided by the deterministic engine. They are gated (every verdict cites resolvable evidence) but remain a judgement: they are not counted in the automatic pass rate above.",
     naTitle: "4. Not-applicable criteria (NA)",
     manualTitle: "5. Criteria to adjudicate (judgment / rendering) — not decided by the static engine",
     manualWarn:
@@ -126,6 +133,7 @@ const L = {
       "Generate the worklist: `verify --manual --in <audit.json> --standard <pack> --out <dir>`. Each item carries the full wording of its tests, its technical note, its particular cases, its guidance and the terms the standard defines.",
     outOfScope: "Out of engine scope — mapped to SCs outside WCAG 2.2 AA; manual verification.",
     scopedOut: "The WCAG failures found concern elements outside this criterion's scope — assess separately.",
+    judgment: "The criterion asks more than the WCAG SCs it maps to — the engine did not answer it; rule on it.",
     nothing: "None.",
     dedup: "Dedup",
     canonical: "canonical file(s) audited",
@@ -194,6 +202,11 @@ interface Row {
   status: Status;
   findings: Finding[];
   justification?: string;
+  // Who decided this criterion. Absent/"engine" = the deterministic engine; "agent" = an
+  // adjudication (gated, but a judgement call); "scan" = the rendered tier. Rendered so a
+  // conformity the engine PROVED is never presented as the same thing as one an agent
+  // RULED — see the split in section 3 and the header rate.
+  decidedBy?: "engine" | "agent" | "scan";
 }
 
 interface Group {
@@ -329,16 +342,27 @@ function render(
       out.push(`- ${nc.length} ${s.ncCount}${adv.length ? ` · ${adv.length} ${s.advCount}` : ""}`);
       const notes = pageScope.find((x) => x.id === pg.id)?.notes;
       if (notes) out.push(`- _${notes}_`);
-      // The screenshot the snapshot already holds. Referenced relative to the report's own
-      // directory — a copy would double the artefact's weight for no gain.
+      // The screenshot the snapshot already holds. Copied next to the report when there is
+      // an output directory, so the directory travels intact — CI uploads `audits/` alone,
+      // and a `../.ultra11y/…` reference is a broken image in the artifact the reviewer
+      // opens. Without an --out (stdout), stay relative: there is nothing to be
+      // self-contained about.
       const shot = join(PAGES_DIR, pg.id, "screen.png");
-      if (existsSync(shot))
-        out.push(
-          "",
-          `![${s.screenshotAlt(pg.name)}](${relative(opts.outDir ?? ".", shot)
-            .split("\\")
-            .join("/")})`,
-        );
+      if (existsSync(shot)) {
+        let href = relative(opts.outDir ?? ".", shot)
+          .split("\\")
+          .join("/");
+        if (opts.outDir) {
+          try {
+            mkdirSync(join(opts.outDir, "assets"), { recursive: true });
+            copyFileSync(shot, join(opts.outDir, "assets", `${pg.id}.png`));
+            href = `./assets/${pg.id}.png`;
+          } catch {
+            // Unwritable output dir: keep the relative reference rather than lose the image.
+          }
+        }
+        out.push("", `![${s.screenshotAlt(pg.name)}](${href})`);
+      }
       for (const f of nc.slice(0, 30)) {
         const crits = pack ? packCriteriaForFinding(pack, f) : [];
         const label = crits.length ? crits.join(", ") : f.criteriaId; // graceful fallback to the WCAG SC
@@ -348,10 +372,23 @@ function render(
     }
   }
 
-  // 3. conforming
+  // 3. conforming — split by PROVENANCE. A criterion the deterministic engine decided and
+  // one an agent ruled on are both "C", but they are not the same claim, and merging them
+  // into one list (and one headline rate) is how an adjudication pass could publish dozens
+  // of conformities nobody had verified. The gate makes an agent C evidence-bound; this
+  // keeps it legible as an agent's judgement all the way to the reader.
   out.push(`## ${s.cTitle}`, "");
   const conform = rows.filter((x) => x.status === "C");
-  out.push(conform.length ? conform.map((x) => `- ${x.label}`).join("\n") : s.nothing, "");
+  const byEngine = conform.filter((x) => x.decidedBy !== "agent");
+  const byAgent = conform.filter((x) => x.decidedBy === "agent");
+  if (!conform.length) out.push(s.nothing, "");
+  else {
+    if (byEngine.length) out.push(...byEngine.map((x) => `- ${x.label}`), "");
+    if (byAgent.length) {
+      out.push(`### ${s.cAgentTitle}`, "", `> ${s.cAgentNote}`, "");
+      out.push(...byAgent.map((x) => `- ${x.label}${x.justification ? ` — _${x.justification}_` : ""}`), "");
+    }
+  }
 
   // 4. not applicable
   out.push(`## ${s.naTitle}`, "");
@@ -368,8 +405,8 @@ function render(
   else {
     const pack5 = isCore(opts.standard) ? undefined : loadPack(opts.standard);
     for (const x of manual) {
-      const tests = pack5 ? Object.keys(getPackCriterion(pack5, x.id)?.tests ?? {}) : [];
-      const testRef = tests.length ? ` — ${s.testsToRule}: ${tests.map((k) => `\`${x.id}.${k}\``).join(" · ")}` : "";
+      const tests = pack5 ? packTestIds(pack5, x.id) : [];
+      const testRef = tests.length ? ` — ${s.testsToRule}: ${tests.map((t) => `\`${t}\``).join(" · ")}` : "";
       out.push(`- ${x.label}${x.justification ? ` — _${x.justification}_` : ""}${testRef}`);
     }
     out.push("", `> ${s.manualHowTo}`, "");
@@ -384,7 +421,14 @@ export function renderReport(r: AuditResult, lang: Lang = "en", outDir?: string)
   const byGuideline = new Map<string, Row[]>();
   for (const c of r.criteria) {
     const title = scTitle(c.id, lang);
-    const row: Row = { id: c.id, label: title ? `${c.id} — ${title}` : c.id, status: c.status, findings: c.findings, justification: c.justification };
+    const row: Row = {
+      id: c.id,
+      label: title ? `${c.id} — ${title}` : c.id,
+      status: c.status,
+      findings: c.findings,
+      justification: c.justification,
+      decidedBy: c.decidedBy,
+    };
     (byGuideline.get(c.guideline) ?? byGuideline.set(c.guideline, []).get(c.guideline)!).push(row);
   }
   // `g.title` on the AuditResult's GuidelineTally is the baked-in English title (kept for
@@ -409,15 +453,18 @@ export function renderPackReport(r: AuditResult, pack: StandardPack, lang: Lang
       label: `${pack.name} ${pr.id} — ${packTitle(pack, pc, lang)}`,
       status: pr.status,
       findings: pr.findings,
+      ...(pr.decidedBy ? { decidedBy: pr.decidedBy } : {}),
       // outOfScope / scopedOut criteria are "manual" (not NA) with their own dedicated
       // justification — never mixed with the ordinary NA reason (see the manual section above).
       ...(pr.outOfScope
         ? { justification: s.outOfScope }
         : pr.scopedOut
           ? { justification: s.scopedOut }
-          : pr.status === "NA"
-            ? { justification: naReason }
-            : {}),
+          : pr.judgment
+            ? { justification: s.judgment }
+            : pr.status === "NA"
+              ? { justification: naReason }
+              : {}),
     };
     (byTheme.get(pr.theme) ?? byTheme.set(pr.theme, []).get(pr.theme)!).push(row);
   }
@@ -434,6 +481,10 @@ export function renderPackReport(r: AuditResult, pack: StandardPack, lang: Lang
     standard: pack.key,
     partialAudit: untestedNeedsRendering(r),
     headerRatePct: packConformancePct(derived),
+    // Forwarded, unlike before: without it the per-page screenshots resolved against the
+    // CWD instead of the report's own directory, so a pack report written to `audits/`
+    // carried links that only worked when read from the repo root.
+    outDir,
   });
 }
 

package/src/standards/derive.ts

@@ -23,6 +23,16 @@ export interface PackCriterionResult {
   // sibling's failure. It derives as `manual` (assess separately) with a dedicated
   // scoped-out justification (see src/report.ts renderPackReport).
   scopedOut?: boolean;
+  // Set when the pack marks this criterion `judgment` and the projection WOULD have
+  // inherited a `C` from a mapped SC that never answered the criterion's own question
+  // (RGAA 8.6 "is the title pertinent?" off WCAG 2.4.2 "a title exists"). It derives
+  // `manual` instead — see `judgmentGuard` and src/report.ts for the justification.
+  judgment?: boolean;
+  // Recorded on the adjudicated branch below (an agent verdict at the pack's own
+  // granularity wins over the derivation). Declared here because renderings distinguish a
+  // conformity the engine PROVED from one an agent RULED — they are not the same claim.
+  justification?: string;
+  decidedBy?: "engine" | "agent" | "scan";
 }
 
 // NC dominates (a real failure anywhere fails the criterion); then a decided C; then
@@ -68,7 +78,11 @@ export function packCriteriaForFinding(pack: StandardPack, finding: Finding): st
  *  same as core. Exists so a pack report's header rate can't drift from its own table once
  *  pack overrides (advisory/severity flips) make the two diverge from core `conformancePct`. */
 export function packConformancePct(derived: PackCriterionResult[]): number {
-  const c = derived.filter((d) => d.status === "C").length;
+  // Agent-adjudicated conformities are deliberately excluded, mirroring the core
+  // (`recomputeTallies` in src/adjudicate.ts): this number is the AUTOMATIC pass rate, and
+  // a judgement — however well gated — is not an automatic verification. An agent NC still
+  // counts; lowering the rate off evidenced findings is the safe direction.
+  const c = derived.filter((d) => d.status === "C" && d.decidedBy !== "agent").length;
   const nc = derived.filter((d) => d.status === "NC").length;
   return c + nc === 0 ? 100 : Math.round((c / (c + nc)) * 100);
 }
@@ -123,6 +137,23 @@ function applySecondaryMappings(
   return { ...base, findings };
 }
 
+/** A criterion the pack marks `judgment` never INHERITS a `C` from its mapped WCAG SCs.
+ *
+ *  The projection folds the SCs DINUM's own crosswalk cites, and `aggregate` returns `C`
+ *  as soon as one of them is `C`. That is the right answer when the RGAA question and the
+ *  SC ask the same thing, and a fabricated conformity when the RGAA wording asks more:
+ *  RGAA 8.6 wants a *pertinent* page title, WCAG 2.4.2 only a present one; RGAA 13.3 wants
+ *  an accessible version of a downloaded document, and no mapped SC ever opened it. Those
+ *  derived `C`s were the one error this tool must not make, so they become `manual` and go
+ *  to the agent instead.
+ *
+ *  Only `C` is intercepted: an `NC` was evidenced by a rule that really fired on this
+ *  criterion, and an `NA` means nothing in scope is concerned — both stay. */
+function judgmentGuard(r: PackCriterionResult, pc: PackCriterion): PackCriterionResult {
+  if (!pc.judgment || r.status !== "C") return r;
+  return { ...r, status: "manual" as Status, judgment: true };
+}
+
 export function derivePackResults(audit: AuditResult, packKey: string): PackCriterionResult[] {
   const pack = loadPack(packKey);
   const byScId = new Map(audit.criteria.map((c) => [c.id, c]));
@@ -207,7 +238,7 @@ export function derivePackResults(audit: AuditResult, packKey: string): PackCrit
         decidedBy: "agent" as const,
       };
     }
-    const base = deriveBase(pc);
+    const base = judgmentGuard(deriveBase(pc), pc);
     return enabledSecondary.length ? applySecondaryMappings(base, pc, enabledSecondary, secondarySources, pack.defaultLocale) : base;
   });
 }

package/src/standards/pack.ts

@@ -41,6 +41,19 @@ export function titlePlain(pack: StandardPack, c: PackCriterion, lang: Lang): st
   return localize(pack, c.titlePlain, lang);
 }
 
+/** The criterion's own numbered test ids, in pack order: RGAA 11.1 → ["11.1.1", "11.1.2",
+ *  "11.1.3"]. These are what an auditor actually rules on, and what a `normativeRef` must
+ *  cite, so several surfaces render them (the auditor block, the report's to-rule-on list,
+ *  the PRD, the adjudication worklist, the per-page grid). Centralised here so the id shape
+ *  is defined once — it used to be re-derived inline at each call site.
+ *
+ *  Callers render these inside backticks, never followed by an em dash: `check`'s criterion
+ *  scanner matches `(\d+\.\d+)\s*—` and would capture "2.1" out of "6.2.1 — …". */
+export function packTestIds(pack: StandardPack, id: string): string[] {
+  const tests = getCriterion(pack, id)?.tests;
+  return tests ? Object.keys(tests).map((k) => `${id}.${k}`) : [];
+}
+
 export function resolveGlossary(packKey: string, anchor: string): GlossaryEntry | undefined {
   return packGlossary(packKey)?.[anchor];
 }

package/src/standards/types.ts

@@ -37,6 +37,19 @@ export interface PackCriterion {
   // failure). Optional/additive: a pack WITHOUT `appliesTo` keeps the legacy fan-out
   // (every mapped SC's findings attach) so third-party packs are unaffected.
   appliesTo?: { ruleIds: string[] };
+  // The criterion's own wording asks MORE than the WCAG SCs it maps to, so a `C` on those
+  // SCs is not an answer to it. RGAA 8.6 asks whether the page title is *pertinent*; WCAG
+  // 2.4.2 only that a title exists. RGAA 13.3 asks whether a downloadable document has an
+  // accessible version; no mapped SC ever opened the document. Without this flag the
+  // projection returns `C` as soon as one mapped SC is `C` (see `aggregate` in derive.ts)
+  // and publishes a conformity nobody assessed.
+  //
+  // A flagged criterion can still derive `NC` (a rule actually fired on it) and `NA`
+  // (nothing applicable in scope); it simply never inherits a `C` — it derives `manual`
+  // and goes to the agent, who rules on it against the criterion's own numbered tests.
+  // `appliesTo` already carries the mirror-image promise for NC ("never NC from a
+  // sibling's failure"); this is the same guarantee on the conforming side.
+  judgment?: boolean;
 }
 
 // The localized DISPLAY vocabulary a standard uses when its audit is rendered for an

package/src/types.ts

@@ -1,6 +1,6 @@
 // Single source of truth for shared types + the embedded version. sync-version.mjs
 // rewrites VERSION at release time (kept in lockstep with package.json + SKILL.md).
-export const VERSION = "2.31.2";
+export const VERSION = "2.32.0";
 export const SCHEMA_VERSION = 2;
 
 // Lang = the UI FRAME's language — the `L` localization tables each module (report,