ultra11y 2.31.2 → 2.31.3

package/src/adjudicate.ts

@@ -77,6 +77,18 @@ export interface AdjudicationItem {
   justification: string; // REQUIRED for C and NA
   reason: string | null; // REQUIRED for a still-`manual` verdict ("needs-rendered-dom" | "undecidable")
   findings: AgentFinding[]; // REQUIRED (≥1, groundable, each with a normativeRef) for NC
+  // What the agent CLEARED, for a C (or ruled out of scope, for an NA). The mirror image of
+  // `findings`, and required for the same reason: without a slot to cite in, a conforming
+  // verdict was pure prose, and the gate could only ever check that the prose was non-empty
+  // — so a model answering "C" to everything with the justification "x" passed, and the
+  // report published 91 conformant criteria nobody had assessed.
+  //
+  // Each citation must ground (the file/line/snippet must resolve, exactly like an NC
+  // finding) AND match one of this item's own `evidence` anchors by file+line — that pairing
+  // is what turns "cite something real" into "cite the evidence you were shown". When the
+  // harvester found no evidence at all there is nothing to clear, and the honest verdicts
+  // are `manual` or `NA`, never `C`.
+  citations?: Evidence[];
   // Non-normative good practices the agent noted on this criterion — folded back into the
   // audit as ADVISORY findings (grounded exactly like an NC finding, but never affecting
   // status: they cannot flip the criterion to NC nor enter conformancePct). Optional.
@@ -427,12 +439,29 @@ export function applyAdjudication(audit: AuditResult, adj: AdjudicationFile, opt
   // means every pack criterion that derives `manual` — the pack's own granularity, which is
   // what the worklist was built at.
   const packMode = !isCore(adj.standard);
+  const open = new Set<string>();
   if (packMode) {
     for (const pc of derivePackResults(audit, adj.standard)) {
-      if (pc.status === "manual" && !byId.has(pc.id)) issues.push(`criterion ${pc.id}: missing from the adjudication (coverage gap)`);
+      if (pc.status !== "manual") continue;
+      open.add(pc.id);
+      if (!byId.has(pc.id)) issues.push(`criterion ${pc.id}: missing from the adjudication (coverage gap)`);
     }
   } else {
-    for (const r of audit.residualRisks) if (!byId.has(r.criteriaId)) issues.push(`criterion ${r.criteriaId}: missing from the adjudication (coverage gap)`);
+    for (const r of audit.residualRisks) {
+      open.add(r.criteriaId);
+      if (!byId.has(r.criteriaId)) issues.push(`criterion ${r.criteriaId}: missing from the adjudication (coverage gap)`);
+    }
+  }
+
+  // …and the other direction. The coverage check above only proves every open criterion was
+  // ruled on; it says nothing about a SURPLUS item. That mattered: the fold resolved a
+  // criterion by id against the whole audit, so an extra `{criteriaId: "3.1.1", verdict:
+  // "C"}` overwrote a non-conformity the deterministic engine had decided — no finding, no
+  // citation, no normativeRef. Adjudication may only ever decide what the engine left open.
+  for (const it of adj.items) {
+    if (!open.has(it.criteriaId)) {
+      issues.push(`criterion ${it.criteriaId}: not open for adjudication — the engine already decided it, or it is not part of ${adj.standard}`);
+    }
   }
 
   // Per-item fail-closed validation.
@@ -443,6 +472,36 @@ export function applyAdjudication(audit: AuditResult, adj: AdjudicationFile, opt
       issues.push(`criterion ${it.criteriaId}: unadjudicated (verdict is null)`);
     } else if (v === "C" || v === "NA") {
       if (!it.justification || !it.justification.trim()) issues.push(`criterion ${it.criteriaId}: a ${v} verdict requires a justification`);
+      // A clearing verdict is gated exactly like an accusing one. Before this, the only
+      // check was "the justification is a non-empty string", so `"x"` cleared a criterion —
+      // and a model answering C to everything published a conformance nobody had assessed.
+      const cites = it.citations ?? [];
+      if (it.evidence.length === 0) {
+        // Nothing was harvested for this criterion, so there is nothing the agent could have
+        // read to clear it. `NA` is still legitimate (the honest "no element in scope is
+        // concerned"); `C` is not — it must stay `manual` and be ruled on against the
+        // criterion's own tests, from a rendered capture or by hand.
+        if (v === "C") {
+          issues.push(
+            `criterion ${it.criteriaId}: a C verdict needs evidence to cite, and none was harvested for this criterion — record "manual" (reason "undecidable"), or "NA" if nothing in scope is concerned`,
+          );
+        }
+      } else if (cites.length === 0) {
+        issues.push(
+          `criterion ${it.criteriaId}: a ${v} verdict must cite at least one of the ${it.evidence.length} evidence item(s) it was shown (citations: [{file, line, …}])`,
+        );
+      } else {
+        // Each citation must name evidence THIS item carried. Grounding alone would only
+        // prove the anchor exists somewhere in the tree; the pairing proves the agent ruled
+        // on what it was actually given.
+        const anchors = new Set(it.evidence.map((e) => `${e.file}:${e.line}`));
+        for (const c of cites) {
+          if (!anchors.has(`${c.file}:${c.line}`)) {
+            issues.push(`criterion ${it.criteriaId}: citation ${c.file}:${c.line} is not among this criterion's harvested evidence (fabricated?)`);
+          }
+          groundInputs.push({ file: c.file, line: c.line, selector: c.selector, snippet: c.snippet });
+        }
+      }
     } else if (v === "NC") {
       if (!it.findings || it.findings.length === 0) issues.push(`criterion ${it.criteriaId}: an NC verdict requires at least one groundable finding`);
       for (const f of it.findings ?? []) {
@@ -471,7 +530,8 @@ export function applyAdjudication(audit: AuditResult, adj: AdjudicationFile, opt
     for (const rec of it.recommendations ?? []) groundInputs.push({ file: rec.file, line: rec.line, selector: rec.selector, snippet: rec.snippet });
   }
 
-  // Content-level grounding of every agent NC finding + every recommendation.
+  // Content-level grounding of every agent NC finding, every C/NA citation, and every
+  // recommendation — the same check, whichever direction the verdict points.
   const grounding = groundItems(groundInputs, { cwd: opts.cwd });
   for (const gi of grounding.issues) issues.push(gi);
 
@@ -596,7 +656,12 @@ function recomputeTallies(a: AuditResult): void {
     g.na = inG.filter((c) => c.status === "NA").length;
     g.manual = inG.filter((c) => c.status === "manual").length;
   }
-  const decided = a.criteria.filter((c) => c.status === "C" || c.status === "NC");
+  // `conformancePct` is labelled everywhere as the AUTOMATIC static-check pass rate, so an
+  // agent's C must not enter it — otherwise one adjudication pass turns a judgement into a
+  // machine-verified number, which is exactly the claim the label denies. An agent NC does
+  // count: a non-conformity is evidenced (grounded finding + normativeRef) and reporting
+  // it lowers the rate, which is the safe direction.
+  const decided = a.criteria.filter((c) => c.status === "NC" || (c.status === "C" && c.decidedBy !== "agent"));
   const conform = decided.filter((c) => c.status === "C").length;
   a.conformancePct = decided.length === 0 ? 100 : Math.round((conform / decided.length) * 100);
 }
@@ -611,13 +676,13 @@ const T = {
     intro:
       "Pour CHAQUE critère, lisez les évidences ci-dessous (extraites de la source auditée) et attribuez un verdict dans `ADJUDICATE.todo.json` (champ `verdict`) :",
     verdicts: [
-      "- `C` — conforme (renseignez `justification`) ;",
+      "- `C` — conforme (renseignez `justification` ET `citations[]` : les évidences que vous avez levées, `file`/`line` recopiés depuis la liste du critère) ;",
       "- `NC` — non conforme (ajoutez au moins un `findings[]` : file/line/message, avec un `snippet` groundable ET un `normativeRef` citant le test précis échoué) ;",
-      "- `NA` — non applicable (renseignez `justification`) ;",
+      "- `NA` — non applicable (renseignez `justification` ; si des évidences sont présentées, citez-les aussi dans `citations[]` pour dire lesquelles sortent du périmètre) ;",
       "- `manual` — indécidable statiquement (renseignez `reason` : `needs-rendered-dom` → `scan`, ou `undecidable`).",
     ],
-    rule: "> Ne signalez une NC que si un test précis du référentiel actif échoue — citez-le (`normativeRef`). Une bonne pratique sans test normatif est une recommandation (`recommendations[]`, non normative). Une simple préoccupation UX n'est ni l'un ni l'autre.",
-    then: "Puis : `ultra11y verify --apply ADJUDICATE.todo.json --in <audit.json> --out <dir>` (échoue si un verdict manque une justification/finding/reason).",
+    rule: "> Ne signalez une NC que si un test précis du référentiel actif échoue — citez-le (`normativeRef`). Une bonne pratique sans test normatif est une recommandation (`recommendations[]`, non normative). Une simple préoccupation UX n'est ni l'un ni l'autre.\n>\n> Symétriquement, un `C` se cite comme une NC : il faut nommer dans `citations[]` les évidences levées. **Un critère présenté sans aucune évidence ne peut pas être `C`** — c'est `manual` (`undecidable`), ou `NA` si rien n'est concerné.",
+    then: "Puis : `ultra11y verify --apply ADJUDICATE.todo.json --in <audit.json> --out <dir>` (échoue si un verdict manque sa justification, ses citations, son finding ou sa raison).",
     evidence: "Évidences",
     none: "(aucune évidence automatique — décidez depuis la source, ou laissez `manual` avec une raison)",
     questions: "À vérifier manuellement",
@@ -635,13 +700,13 @@ const T = {
     title: "# Criteria adjudication (ultra11y)",
     intro: "For EACH criterion, read the evidence below (harvested from the audited source) and set a verdict in `ADJUDICATE.todo.json` (field `verdict`):",
     verdicts: [
-      "- `C` — conformant (fill `justification`);",
+      "- `C` — conformant (fill `justification` AND `citations[]`: the evidence you cleared, `file`/`line` copied from the criterion's own list);",
       "- `NC` — non-conformant (add at least one `findings[]`: file/line/message, with a groundable `snippet` AND a `normativeRef` citing the precise failed test);",
-      "- `NA` — not applicable (fill `justification`);",
+      "- `NA` — not applicable (fill `justification`; when evidence is presented, cite it too in `citations[]` to say which items are out of scope);",
       "- `manual` — not statically decidable (fill `reason`: `needs-rendered-dom` → `scan`, or `undecidable`).",
     ],
-    rule: "> Report NC only if a precise test of the active standard fails — cite it (`normativeRef`). A good practice without a normative test is a recommendation (`recommendations[]`, non-normative). A purely UX concern is neither.",
-    then: "Then: `ultra11y verify --apply ADJUDICATE.todo.json --in <audit.json> --out <dir>` (fails if any verdict lacks its justification/finding/reason).",
+    rule: "> Report NC only if a precise test of the active standard fails — cite it (`normativeRef`). A good practice without a normative test is a recommendation (`recommendations[]`, non-normative). A purely UX concern is neither.\n>\n> A `C` is cited the same way an NC is: name the evidence you cleared in `citations[]`. **A criterion presented with no evidence at all cannot be `C`** — record `manual` (`undecidable`), or `NA` if nothing in scope is concerned.",
+    then: "Then: `ultra11y verify --apply ADJUDICATE.todo.json --in <audit.json> --out <dir>` (fails if any verdict lacks its justification, citations, finding or reason).",
     evidence: "Evidence",
     none: "(no automatic evidence — decide from source, or leave `manual` with a reason)",
     questions: "To verify manually",

package/src/auditor.ts

@@ -11,7 +11,7 @@ import type { AuditResult, Finding, Lang, Severity } from "./types.js";
 import { prdUnits, partitionUnits, effortOf, guidanceFor, guidanceExampleBlock, acceptanceCriteria, type PrdUnit, type PrdFile } from "./prd.js";
 import { getSC, guidelineTitle, principleTitle, techniques as scTechniques } from "./wcag.js";
 import { resolveMessage, resolveRemediation, resolveNote } from "./messages.js";
-import { type StandardId, isCore, loadPack, standardLabel, themeName, vocabularyFor } from "./standards/index.js";
+import { type StandardId, isCore, loadPack, packTestIds, standardLabel, themeName, vocabularyFor } from "./standards/index.js";
 
 const SEV_ORDER: Severity[] = ["bloquant", "majeur", "mineur"];
 const ICON: Record<Severity, string> = { bloquant: "🔴", majeur: "🟠", mineur: "🟡" };
@@ -186,7 +186,7 @@ export function renderAuditorUnit(unit: PrdUnit, standard: StandardId, lang: Lan
     const pc = pack.criteria.find((c) => c.id === unit.criteriaId);
     if (pc) out.push(`**${v.theme}** : ${pc.theme}. ${themeName(pack, pc.theme, lang) ?? ""}`.trimEnd());
     out.push(`**${v.criterion}** : ${unit.criteriaId} — ${unit.title}`);
-    const testNums = pc?.tests ? Object.keys(pc.tests).map((k) => `${unit.criteriaId}.${k}`) : [];
+    const testNums = packTestIds(pack, unit.criteriaId);
     if (testNums.length) out.push(`**${v.test}(s)** : ${testNums.join(" · ")}`);
     if (unit.refs.length) out.push(`**WCAG** : ${unit.refs.map((sc) => `${sc}${scLevel(sc)}`).join(" · ")}`);
   }

package/src/cli.ts

@@ -1,4 +1,4 @@
-import { realpathSync, writeFileSync, mkdirSync, existsSync, readFileSync, appendFileSync } from "node:fs";
+import { realpathSync, writeFileSync, mkdirSync, existsSync, readFileSync, appendFileSync, copyFileSync } from "node:fs";
 import { join, relative, sep, dirname } from "node:path";
 import { fileURLToPath, pathToFileURL } from "node:url";
 import { VERSION, type Lang, type AuditResult, type DynamicResult, type SampleConfig, type Severity } from "./types.js";
@@ -1157,14 +1157,30 @@ async function cmdPages(p: ParsedArgs): Promise<number> {
   const split = p.flags.split === "page";
   const outDir = typeof p.flags.out === "string" && p.flags.out ? (p.flags.out as string) : undefined;
 
-  // The screenshot lives beside the snapshot, on disk. Reference it relatively from wherever
-  // the file lands — a copy would double the repo's weight for no gain, and an absolute path
-  // would break the moment the report is read anywhere else.
-  const shotsFor = (fileDir: string): Map<string, string> => {
+  // The screenshot lives beside the snapshot, in `.ultra11y/pages/`. Referencing it
+  // relatively is right when the report is read in place, and wrong the moment the report
+  // DIRECTORY travels on its own: CI uploads `audits/` as the artifact, and a
+  // `../../.ultra11y/…` link then points outside it — every image broken in the artifact
+  // the reviewer actually opens. So when writing to a directory we copy each screenshot
+  // into it (`assets/<page-id>.png`) and link that; stdout mode keeps the relative
+  // reference, since there is no output directory to be self-contained.
+  const shotOf = (id: string): string => join(PAGES_DIR, id, "screen.png");
+  const shotsRelative = (fileDir: string): Map<string, string> => {
     const m = new Map<string, string>();
     for (const pg of derived) {
-      const shot = join(PAGES_DIR, pg.id, "screen.png");
-      if (existsSync(shot)) m.set(pg.id, relative(fileDir, shot).split("\\").join("/"));
+      if (existsSync(shotOf(pg.id))) m.set(pg.id, relative(fileDir, shotOf(pg.id)).split("\\").join("/"));
+    }
+    return m;
+  };
+  const shotsCopiedInto = (dir: string): Map<string, string> => {
+    const m = new Map<string, string>();
+    const assets = join(dir, "assets");
+    for (const pg of derived) {
+      const src = shotOf(pg.id);
+      if (!existsSync(src)) continue;
+      mkdirSync(assets, { recursive: true });
+      copyFileSync(src, join(assets, `${pg.id}.png`));
+      m.set(pg.id, `./assets/${pg.id}.png`);
     }
     return m;
   };
@@ -1172,19 +1188,19 @@ async function cmdPages(p: ParsedArgs): Promise<number> {
   if (!outDir) {
     // No --out: stream the whole document to stdout. Screenshots are resolved against the
     // CWD, which is where a reader piping this would sit.
-    console.log(renderPagesDocument(result, derived, { standard, lang, screenshots: shotsFor(".") }));
+    console.log(renderPagesDocument(result, derived, { standard, lang, screenshots: shotsRelative(".") }));
     return 0;
   }
 
   mkdirSync(outDir, { recursive: true });
   if (!split) {
     const file = join(outDir, `pages-${result.date}.md`);
-    writeFileSync(file, `${renderPagesDocument(result, derived, { standard, lang, screenshots: shotsFor(outDir) })}\n`);
+    writeFileSync(file, `${renderPagesDocument(result, derived, { standard, lang, screenshots: shotsCopiedInto(outDir) })}\n`);
     console.log(file);
     return 0;
   }
 
-  const shots = shotsFor(outDir);
+  const shots = shotsCopiedInto(outDir);
   // Sheets are `page-<id>.md`, never `<id>.md`: a page whose id is `index` — the ordinary id
   // for an `index.html` target — would otherwise be written to the same path as the index and
   // one of the two would silently vanish. The prefix makes the collision impossible by

package/src/data/standards/rgaa.json

@@ -1258,7 +1258,8 @@
       "wcag": ["1.4.2"],
       "appliesTo": {
         "ruleIds": ["autoplay-media", "axe:no-autoplay-audio"]
-      }
+      },
+      "judgment": true
     },
     {
       "id": "4.11",
@@ -1897,7 +1898,8 @@
       "wcag": ["3.1.1"],
       "appliesTo": {
         "ruleIds": ["axe:html-lang-valid", "lang-invalid"]
-      }
+      },
+      "judgment": true
     },
     {
       "id": "8.5",
@@ -1933,7 +1935,8 @@
       "wcag": ["2.4.2"],
       "appliesTo": {
         "ruleIds": []
-      }
+      },
+      "judgment": true
     },
     {
       "id": "8.7",
@@ -3327,7 +3330,8 @@
       "wcag": ["1.1.1", "1.3.1", "1.3.2", "2.4.1", "2.4.3", "3.1.1", "4.1.2"],
       "appliesTo": {
         "ruleIds": []
-      }
+      },
+      "judgment": true
     },
     {
       "id": "13.4",
@@ -3349,7 +3353,8 @@
       "wcag": ["1.1.1", "1.3.1", "1.3.2", "2.4.1", "2.4.3", "3.1.1", "4.1.2"],
       "appliesTo": {
         "ruleIds": []
-      }
+      },
+      "judgment": true
     },
     {
       "id": "13.5",

package/src/data/wcag-universe.json

@@ -2,7 +2,7 @@
   "wcagVersion": "2.2",
   "source": "https://www.w3.org/TR/WCAG22/",
   "criteriaSource": "https://github.com/w3c/wcag",
-  "provenance": "Full WCAG 2.x SC universe (all levels incl. AAA, and the removed 4.1.1 Parsing) fetched from raw.githubusercontent.com/w3c/wcag@main on 2026-08-11 via `node scripts/build-standards.mjs --refresh-universe`. Classification: core-AA = ships in src/data/wcag.json (the shipped WCAG 2.2 AA core); out-of-core = WCAG AAA; removed = obsolete (4.1.1).",
+  "provenance": "Full WCAG 2.x SC universe (all levels incl. AAA, and the removed 4.1.1 Parsing) fetched from raw.githubusercontent.com/w3c/wcag@main on 2026-08-12 via `node scripts/build-standards.mjs --refresh-universe`. Classification: core-AA = ships in src/data/wcag.json (the shipped WCAG 2.2 AA core); out-of-core = WCAG AAA; removed = obsolete (4.1.1).",
   "criteria": [
     {
       "id": "1.1.1",

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;