@mutagent/evaluator 0.1.0-alpha.2

package/.claude/skills/mutagent-evaluator/scripts/render-eval-report.ts

@@ -0,0 +1,3823 @@
+/**
+ * scripts/render-eval-report.ts — P4 (D-1/D-2/D-3): the v2 `*evaluate` reporting.
+ * ---------------------------------------------------------------------------
+ * The `*evaluate` path (P5) produces a Scorecard (GATE + variance + per-criterion
+ * verdicts); this module is the v2 HTML reporting. It is a faithful TS PORT of the
+ * operator-APPROVED 5-tab component spec (the sample-eval-1 reference renderer —
+ * `.mutagent/evaluator/sample-eval-1/scripts/render-report.py`), NOT a flat table:
+ *
+ *   - D-1  renderEvalCards   : terminal per-criterion cards from the Scorecard.
+ *   - D-2/D-3 renderEvalReport: the 5-tab HTML eval-report —
+ *       §1 Overview          — KPIs + coverage-contract note + gating-criteria
+ *                              table (pass-rate over applicable) + top-findings teaser.
+ *       §2 Trajectory·Judge  — per-trace ledger (filter/search) + click-row drill:
+ *                              Target-Agent‖Judge SIDE-BY-SIDE (gather-context band,
+ *                              two lanes, judge micro-steps anchored to agent steps,
+ *                              localize band, judge-health) when judge_steps exist,
+ *                              else the per-trajectory scorecard.
+ *       §3 Eval Scorecard    — cohort heatmap (criteria × route cohort) + nested
+ *                              subcards (grounding / verdict / why-abstained) +
+ *                              inline calibration buttons.
+ *       §4 Findings          — verbatim evidence bullets + judge reasoning chain +
+ *                              agree/revise/refute alignment review.
+ *       §5 Self-Eval [INTERNAL] — judge-health + methodology + EV-051 routed
+ *                              decisions; carries a `strip-for-client` marker.
+ *   - autoOpenCommand        : cross-platform open helper (built; the orchestrator
+ *           fires it post-render — never fired here).
+ *
+ * GRACEFUL DEGRADE: the rich tabs CONSUME the §9.4 judge contract (a dense
+ * na-explicit per-criterion map + `judge_steps[]` anchored to agent steps) WHEN
+ * the judge emits it — and fall back to the scorecard-derived rendering when it is
+ * absent. The judge-AGENT emission of dense-map + judge_steps is a SEPARATE change
+ * (see TODO[§9.4-judge-emit] in buildEvalReportInput + the additive optional
+ * fields the renderer reads off `MatrixVerdictFile`).
+ *
+ * Brand: rendered from the evaluator's OWN bundled `assets/brand/theme.css` (the
+ * unified design-system tokens, aligned to @mutagent/templates/tokens.css) +
+ * report-component CSS referencing those same tokens — no diagnostics-package
+ * runtime ref (sealed-sibling). SHARP corners · SUBTLE non-black cards · TONED
+ * status · RESTRAINED glow. Tabs are <button> (NOT <s>) — no strikethrough.
+ *
+ * DETERMINISTIC + null-guarded: the ONLY non-deterministic input is the injected
+ * `generatedAt`, so two renders of the same report input are BYTE-IDENTICAL after
+ * the generatedAt mask (`mask.ts`). v1 render-report.ts / evaluate.ts / mask.ts are
+ * UNTOUCHED — this is a separate v2 renderer.
+ */
+import { readFileSync } from "node:fs";
+import { dirname, join } from "node:path";
+import { fileURLToPath } from "node:url";
+import {
+  AssumptionKind,
+  CriterionFlag,
+  Grounding,
+  OutcomeVerdict,
+  type CriterionVerdict,
+  type DiscoveryRationale,
+  type EvalTrace,
+  type MinedCriterion,
+} from "./contracts/eval-types.ts";
+import { localizeText, PROFILE_UNKNOWN } from "./contracts/eval-matrix.ts";
+import type {
+  ExpectedStep,
+  MatrixCriterion,
+  MatrixVerdictFile,
+  SubjectProfile,
+  Understanding,
+} from "./contracts/eval-matrix.ts";
+import { deriveWalkHealth } from "./matrix-judge.ts";
+import { assessEmitCompleteness, type EmitCompleteness } from "./emit-completeness.ts";
+import type { Scorecard } from "./evaluate.ts";
+import type { SourceMap } from "./source-map.ts";
+import { routeFailures, type FailureRef, type HandoverBundle } from "./route-failures.ts";
+
+const HERE = dirname(fileURLToPath(import.meta.url));
+const BRAND_DIR = join(HERE, "..", "assets", "brand");
+
+/** HTML-escape (null-guarded — no throw on undefined). Mirrors render-report.ts. */
+export function esc(s: unknown): string {
+  return String(s ?? "")
+    .replace(/&/g, "&amp;")
+    .replace(/</g, "&lt;")
+    .replace(/>/g, "&gt;")
+    .replace(/"/g, "&quot;");
+}
+
+// ── Render input (rich; every rich field OPTIONAL → graceful degrade) ─────────
+
+/** §9.4.4 R4 — where a criterion CAME FROM (defined-in-suite vs source-mined). */
+export interface CriterionProvenance {
+  /** `defined` = an authored matrix row · `source` = mined from the trace batch. */
+  kind: "defined" | "source";
+  /** a human-readable origin label (e.g. "eval-matrix.yaml" / "mined: ✓/✗ split"). */
+  label: string;
+  /** the source detail (e.g. how many traces it was seen in), when known. */
+  detail?: string;
+}
+
+/** One criterion as the report needs it (mined OR matrix-derived). */
+export interface ReportCriterion {
+  id: string;
+  statement: string;
+  severity: string; // CRIT | HIGH | MED | LOW
+  gating?: boolean;
+  dimension?: string;
+  level?: string; // context | output | cross-stage
+  appliesTo?: string[];
+  /** §9.4.4 R4 — the concrete binary pass condition (the full definition for the hover). */
+  passCondition?: string;
+  /** UI-7 — what the judge READS to evaluate this criterion (the "how it's judged"
+   *  inputs, eval-matrix `judgeInputs`). Render-layer surfacing of an existing field. */
+  judgeInputs?: string[];
+  /** §9.4.4 R4 — where the criterion came from (defined vs source). */
+  provenance?: CriterionProvenance;
+  /** UI-10 — the code-vs-judge ROUTER for this criterion (eval-matrix `checkMethod`
+   *  / mined `metadata.check_method`): `deterministic` (CODE — a code-eval, no judge
+   *  tokens) · `hybrid` (code pre-filter + judge) · `llm-judge` (an LLM-judge
+   *  criterion). Defaults to `llm-judge` when the source row omits it (the legacy
+   *  all-judge behaviour). Render-layer surfacing — drives the CODE vs JUDGE chip;
+   *  NO judge-protocol change. */
+  checkMethod?: string;
+  /** the §5c DR-2 discovery-rationale block (present iff mined). */
+  discovery?: DiscoveryRationale;
+}
+
+/** One ordered judge micro-step (§9.4 `judge_steps[]`), anchored to an agent step. */
+export interface JudgeStep {
+  /** context | examine | detect | bind | ground | critique | decide | verify */
+  kind: string;
+  text?: string;
+  /** a grounding citation — a string OR a structured {obs,path,value} ref (coerced in the renderer). */
+  ref?: string | { obs: string; path: string; value: string };
+  /** the agent-step index this judge step anchors to (0 / "context" = gather band).
+   *  TOLERANT: real judges may emit a stringified index / label / range (coerced below). */
+  anchor?: number | string;
+}
+
+/** One target-agent step (the left lane of the §2 side-by-side). */
+export interface AgentStep {
+  n: number;
+  tool?: string;
+  /** ok | error | warn | false-success */
+  status?: string;
+  detail?: string;
+  /** The trace OBSERVATION id this step corresponds to. Carried best-effort by
+   *  `enrichAgentStepsObs` (verdict files don't emit it natively): recovered from
+   *  the per-criterion verdict `refs` (`path:"name"` → `{obs, value:<tool>}`) and
+   *  the `tool@<obs>` tokens the judge embeds in its critiques. Lets the §2 judge
+   *  lane map a step to the criterion verdicts that EXAMINED it via `ref.obs ===
+   *  step.obs` (the precise key) on top of the tool-name fallback. Undefined when
+   *  no obs id could be recovered for the step (then tool-name match still applies). */
+  obs?: string;
+}
+
+/** WS-1 — one per-criterion JUDGE verdict for a SINGLE trajectory, read verbatim
+ *  from the verdict file's `verdicts[]` (critique-before-verdict + grounding refs).
+ *  Distinct from the FOLDED `CriterionVerdict` (one-per-criterion across the batch):
+ *  this is the per-trace judgement bound to the §2 drill so EVERY evaluated trace
+ *  surfaces its judge reasoning (result · critique · refs), whether or not a
+ *  `judge_steps[]` walk was emitted for it. */
+export interface CriterionTrajectoryVerdict {
+  criterionId: string;
+  /** pass | fail | uncertain | na */
+  result: string;
+  confidence?: number;
+  confidenceBand?: string;
+  critique?: string;
+  /** grounding citations — a string OR a structured {obs,path,value} (coerced in the UI). */
+  refs?: Array<string | { obs?: string; path?: string; value?: string }>;
+}
+
+/** Per-trajectory judge-health micro-summary (the §2 footer). */
+export interface JudgeHealthRow {
+  contextGathered?: boolean;
+  grounded?: number;
+  assumed?: number;
+  stoppedAtSymptom?: boolean;
+}
+
+/** WS-5 — the resolution class for an INDETERMINATE (uncertain) per-criterion verdict.
+ *  `na` = the criterion's TRIGGER/precondition was ABSENT in this trace (not applicable —
+ *  DROPPED from the applicable denominator). `needs-evidence` = the precondition WAS
+ *  present but the judge could not decide (stays in the denominator, carries a concrete
+ *  next-action). NEVER a bare "indeterminate". */
+export type IndeterminateResolutionClass = "na" | "needs-evidence";
+/** WS-5 — the chain next-action for a `needs-evidence` item (the deterministic decision
+ *  chain: re-check the code, get a 2nd judge, revise the criterion, or HITL-spot-check). */
+export type IndeterminateNextAction = "code-recheck" | "2nd-judge" | "revise-criterion" | "hitl-spot-check";
+/** WS-5 — one resolved indeterminate per-criterion verdict (§3 resolution chain). */
+export interface IndeterminateResolution {
+  criterionId: string;
+  resolutionClass: IndeterminateResolutionClass;
+  /** present iff `needs-evidence` — the concrete chain next-action. */
+  nextAction?: IndeterminateNextAction;
+  /** one-line WHY: the missing precondition (na) or the missing signal (needs-evidence). */
+  reason: string;
+}
+
+/** One per-trace ledger row (§2). The side-by-side fields are present iff the
+ *  judge emitted `judge_steps[]` for this trajectory. */
+export interface LedgerRow {
+  trajectoryId: string;
+  route?: string;
+  /** PASS | FAIL | INDETERMINATE | INCOMPLETE */
+  verdict: string;
+  /** dense, na-explicit per-criterion map: cid -> pass|fail|uncertain|na. */
+  perCriterion: Record<string, string>;
+  failingCriteria: string[];
+  root?: string;
+  grounding?: string;
+  /** WS-1 — the per-criterion JUDGE verdicts for THIS trajectory (result · critique-
+   *  before-verdict · grounding refs), read verbatim from the verdict file's
+   *  `verdicts[]`. Bound to the §2 drill ("How the Judge Reasoned") so EVERY trace —
+   *  PASS or FAIL, walk or no walk — surfaces its judge reasoning. Present whenever
+   *  the verdict file carried ≥1 per-criterion verdict (i.e. all judged trajectories). */
+  criterionVerdicts?: CriterionTrajectoryVerdict[];
+  // ── §2 side-by-side (present iff judge_steps[] emitted) ──
+  context?: { harness?: string; scenario?: string; exitStates?: string };
+  /** Gap A — the RAW triggering INPUT for this trajectory (the thing that fired
+   *  the subject agent). Sourced from the TRACE (`EvalTrace.input.prompt`, fallback
+   *  packet `transcript[0]`) keyed by traceId in `buildEvalReportInput` — NOT a
+   *  judge-emitted field (the input is ground-truth in the trace, no judge echo
+   *  needed). ABSENT when the trace carried no readable input ⇒ the §2 drill cell
+   *  renders "—" (never faked). The judge's `context.scenario` LABEL still shows
+   *  beneath it — the input is the WHAT, the scenario is the route summary. */
+  input?: string;
+  agentSteps?: AgentStep[];
+  judgeSteps?: JudgeStep[];
+  localize?: string;
+  health?: JudgeHealthRow;
+  // ── §9.4.4 v2.2 (present iff the judge emitted them) ──
+  /** M2 node-0 train-of-thought (rephrase + given-vs-inferred). */
+  understanding?: Understanding;
+  /** M3 node-0.5 expected-trajectory (how the target SHOULD have acted). */
+  expectedTrajectory?: ExpectedStep[];
+  /** M1 the subject profile the judge reasoned under (echoed / reconstructed). */
+  subjectProfile?: SubjectProfile;
+  /** §9.4.5 E3 — the trace's wall-clock timestamp (ISO), when the source carried it.
+   *  Powers the eval-HEALTH temporal heatmap (correctness over time). ABSENT ⇒ that
+   *  trajectory falls into the data-pending bucket (structure renders, never faked). */
+  timestamp?: string;
+  /** WS-5 — per-criterion INDETERMINATE resolution for THIS trace (keyed by criterionId):
+   *  the deterministic precondition gate's verdict (`na` precondition-absent vs
+   *  `needs-evidence` + next-action). Present only for criteria whose raw verdict was
+   *  uncertain on this trace. */
+  indeterminateResolution?: Record<string, IndeterminateResolution>;
+  /** WS-2 — how this trajectory's judgement RESOLVED, made honest + visible (a trace
+   *  with no §2 walk is NOT "unjudged"). One of:
+   *   • `judge-walk`            — the judge emitted the §9.4 walk (agentSteps/judgeSteps)
+   *   • `judged-walk-not-captured` — the trace WAS judged (per-criterion verdicts present)
+   *                                  but the emit-contract dropped the structured walk
+   *   • `truncated`             — INCOMPLETE fidelity (node-1 gate); never walked by design
+   *  Drives the §2 ledger resolution badge + the per-trace drill routing line. */
+  resolution?: TrajectoryResolution;
+}
+
+/** WS-2 — the resolution class of one ledger row (see `LedgerRow.resolution`). */
+export type TrajectoryResolution = "judge-walk" | "judged-walk-not-captured" | "truncated";
+
+/** WS-2 — derive the resolution class for one verdict file. PURE.
+ *  truncated > judge-walk (walk emitted) > judged-walk-not-captured (judged, walk dropped). */
+export function deriveTrajectoryResolution(f: {
+  fidelity?: { complete?: boolean };
+  agentSteps?: unknown[];
+  judgeSteps?: unknown[];
+  verdicts?: unknown[];
+}): TrajectoryResolution {
+  if (f.fidelity?.complete === false) return "truncated";
+  const hasWalk = (f.judgeSteps?.length ?? 0) > 0 || (f.agentSteps?.length ?? 0) > 0;
+  if (hasWalk) return "judge-walk";
+  return "judged-walk-not-captured";
+}
+
+/** WS-2 — the human badge label + one-line routing explanation per resolution class. */
+export function resolutionMeta(r: TrajectoryResolution): { badge: string; cls: string; routing: string } {
+  switch (r) {
+    case "judge-walk":
+      return {
+        badge: "JUDGE·WALK",
+        cls: "res-walk",
+        routing:
+          "Routed through the full §9.4 judge walk — the target step lane + the anchored judge reasoning were captured, so the side-by-side below is complete.",
+      };
+    case "judged-walk-not-captured":
+      return {
+        badge: "judged · walk not captured",
+        cls: "res-nowalk",
+        routing:
+          "This trace WAS judged (per-criterion verdicts + critiques are present below) — the judge simply did not persist the structured §9.4 walk (agentSteps/judgeSteps/M2/M3). The verdict stands; only the step-by-step lane is missing. The emit-completeness gate (§5) tracks this gap; a fresh run with the hardened emit-contract captures the walk natively.",
+      };
+    case "truncated":
+      return {
+        badge: "TRUNCATED",
+        cls: "res-trunc",
+        routing:
+          "The node-1 fidelity gate fired — the trace was truncated/unreadable, so it was NEVER walked per-criterion (no fabricated verdicts from a partial trace). Resolved INCOMPLETE by design, not a judge omission.",
+      };
+  }
+}
+
+/** §9.4.4 R3/M5 — one DETECTED-but-unmatched flag (a node-2.5 candidate). NOT a
+ *  verdict: a real behaviour the judge flagged with NO matching criterion. Surfaced
+ *  in §4 CLEARLY SEPARATED from fails-on-existing-criteria, and routed to discover. */
+export interface DetectedFlag {
+  /** `eval` = a candidate criterion to mine · `dataset` = a candidate test case. */
+  kind: string;
+  detection: string;
+  trajectoryId: string;
+  anchor?: number | string;
+  ref?: string | { obs: string; path: string; value: string };
+}
+
+/** One §4 finding (one failing/at-risk criterion, aggregated over the ledger). */
+export interface TopFinding {
+  criterion: string;
+  severity?: string;
+  applicable: number;
+  failCount: number;
+  /** % over the APPLICABLE denominator (GA-D2b — na never counted). */
+  prevalencePctOverApplicable: number;
+  root?: string;
+  /** the trace whose judge-walk powers the verbatim evidence. */
+  exampleTraceId?: string;
+}
+
+/** One gating-criterion roll (§1 table + §3 subcards). */
+export interface GatingRoll {
+  criterion: string;
+  severity?: string;
+  applicable: number;
+  fail: number;
+  /** WS-5 — GENUINE under-evidence (precondition present, undecided). NOT N/A. */
+  indeterminate?: number;
+  /** WS-5 — N/A (precondition/trigger absent) — DROPPED from the applicable denominator. */
+  na?: number;
+  passRateOverApplicable: number;
+  denominatorNote?: string;
+  root?: string;
+}
+
+/** WS-5 — one per-criterion INDETERMINATE resolution roll (§3 resolution chain). Covers
+ *  ALL criteria (not just gating). The applicable denominator EXCLUDES `na`. */
+export interface CriterionResolution {
+  criterion: string;
+  severity?: string;
+  /** applicable = pass + fail + needsEvidence (N/A excluded — denominator-honest). */
+  applicable: number;
+  pass: number;
+  fail: number;
+  /** N/A (precondition absent) — shown SEPARATELY, never in the applicable denominator. */
+  na: number;
+  /** genuine under-evidence (precondition present, undecided) — carries next-actions. */
+  needsEvidence: number;
+  passRateOverApplicable: number;
+  /** the dominant N/A reason (the missing precondition), when na > 0. */
+  naReason?: string;
+  /** aggregated next-actions for the needsEvidence items (action → count + a reason). */
+  nextActions: Array<{ action: IndeterminateNextAction; count: number; reason: string }>;
+}
+
+export interface EvalReportInput {
+  subject: { name: string; org?: string; source?: string; models?: string[] };
+  /** the v2 *evaluate Scorecard (GATE + variance). */
+  scorecard: Scorecard;
+  /** the folded per-criterion verdicts. */
+  verdicts: CriterionVerdict[];
+  /** the criteria (mined → DR-2 cards; matrix → statement/severity only). */
+  criteria: ReportCriterion[];
+  // ── rich (all OPTIONAL — derived by `buildEvalReportInput` from real pipeline
+  //     data; the renderer degrades gracefully when absent) ──
+  /** per-trace ledger (§2). */
+  ledger?: LedgerRow[];
+  /** coverage contract (§1). */
+  coverage?: { judged?: number; triaged?: number; byVerdict?: Record<string, number>; gapNote?: string };
+  /** route cohorts (§3 heatmap): counts per cohort + cid×cohort cells. */
+  cohorts?: {
+    counts: Record<string, number>;
+    matrix: Record<string, Record<string, { pass: number; fail: number; indeterminate: number }>>;
+  };
+  /** the gating-criteria rolls (§1 table + §3 subcards). */
+  gatingCriteria?: GatingRoll[];
+  /** WS-5 — per-criterion INDETERMINATE resolution chain (ALL criteria): the
+   *  applicable·pass·fail·N/A·needs-evidence split that resolves the single
+   *  indeterminate bucket into precondition-absent (N/A, denominator-excluded) vs
+   *  genuine under-evidence (carries a next-action). §3 Scorecard. */
+  criterionResolutions?: CriterionResolution[];
+  /** the §4 top findings (verbatim-evidence cards). */
+  topFindings?: TopFinding[];
+  /** §9.4.4 R3/M5 — the DETECTED-but-unmatched flags (node-2.5 candidates), shown
+   *  in §4 SEPARATED from criterion fails. A detection routed to discover, never minted. */
+  detectedFlags?: DetectedFlag[];
+  /** §9.4.4 M1/R2 — the subject profile (identity·purpose·tools·skill·scope) for the
+   *  INTERNAL calibration tab + the §2 "who is the agent" band. */
+  subjectProfile?: SubjectProfile;
+  /** §9.4.5 E2 — per-tool OBSERVED call-count (from the trace observations / trajectory
+   *  steps). Drives the entity-hero tool CHIPS' "obs N" badge — honest in the trace-only
+   *  (no-code-access) case where the only tool evidence is what the traces show. */
+  toolObservations?: Record<string, number>;
+  /** judge-health (§5 self-eval). */
+  judgeHealth?: { groundedPct?: number; abstentionRatePct?: number; stoppedAtSymptomPct?: number };
+  /** P2 source-map topology (coverage note). */
+  sourceMap?: SourceMap;
+  /** the EV-051 diagnostics handover (§5 decisions); null when nothing failed. */
+  handover?: HandoverBundle | null;
+  /** WS-3 — the §5 Self-Eval AGGREGATE judge-calibration derivation (behavior map +
+   *  the 9 components). Deterministic over the verdict files; ABSENT ⇒ §5 degrades to
+   *  the legacy calibration view. */
+  selfEval?: SelfEvalAggregate;
+  /** Overview provenance meta-strip — the RUN config (run-id · date · source · JUDGE
+   *  substrate · pinned judge model · temp-0 · C-PIN). Distinct from the TARGET model
+   *  (`subject.models`). All OPTIONAL: a genuinely-unavailable field is MARKED, never
+   *  faked; substrate/temp/C-PIN carry sensible engine defaults. */
+  runConfig?: RunConfig;
+  /** the ONLY non-deterministic input (masked for byte-identity). */
+  generatedAt: string;
+}
+
+/** §Overview-redesign — the run-configuration the provenance meta-strip renders. The
+ *  JUDGE substrate + pinned judge model are DISTINCT from the TARGET model under eval. */
+export interface RunConfig {
+  /** the *evaluate run id (the artifact dot-root key). */
+  runId?: string;
+  /** the run date/timestamp (defaults to `generatedAt` when absent). */
+  date?: string;
+  /** the trace SOURCE platform (langfuse · otel · local-ndjson · …). */
+  source?: string;
+  /** the JUDGE substrate: agent-dispatch (host runtime) · ai-sdk · code. */
+  judgeSubstrate?: string;
+  /** the PINNED judge model (C-PIN) — NOT the target model. */
+  judgeModel?: string;
+  /** the pinned judge temperature (C-PIN ⇒ 0). */
+  temperature?: number;
+  /** C-PIN engaged (pinned model + temp ⇒ byte-identical reruns). */
+  cPin?: boolean;
+}
+
+// ── helpers ─────────────────────────────────────────────────────────────────
+
+function sevCls(sev: string | undefined): string {
+  const s = String(sev ?? "med").toLowerCase();
+  if (s === "crit") return "crit";
+  if (s === "high") return "high";
+  if (s === "med" || s === "medium") return "med";
+  return "low";
+}
+/** Heatmap / pass-rate colour by % (fail<80 · warn<95 · pass). Deterministic. */
+function rateCol(r: number): string {
+  return r < 80 ? "var(--fail)" : r < 95 ? "var(--warn)" : "var(--pass)";
+}
+/** Map a per-criterion result string to its display token (uncertain→indeterminate). */
+function resultToken(r: string): string {
+  return r === OutcomeVerdict.Uncertain ? "indeterminate" : r;
+}
+function verdictCls(v: string): string {
+  const u = v.toUpperCase();
+  return u === "PASS" ? "pass" : u === "FAIL" ? "fail" : "inc";
+}
+function criterionById(input: EvalReportInput, id: string): ReportCriterion | undefined {
+  return input.criteria.find((c) => c.id === id);
+}
+/** Render structured grounding refs `{obs,path,value}` as compact citations. */
+function refsText(refs: { obs: string; path: string; value: string }[]): string {
+  return refs.map((r) => `${r.obs}${r.path ? "/" + r.path : ""}: "${r.value}"`).join(" · ");
+}
+
+/** UI-6 — pass-RATE → an accent class (driven by SUCCESS rate, NOT severity): green-ish
+ *  high · amber mid · red-ish low. Mirrors the heatmap thresholds (≥95 ok · ≥80 mid ·
+ *  <80 low) so a criterion card's colour reflects how it actually scored. Deterministic. */
+export function rateClass(r: number): "rate-ok" | "rate-mid" | "rate-low" {
+  return r >= 95 ? "rate-ok" : r >= 80 ? "rate-mid" : "rate-low";
+}
+
+// ── UI-2 — verdict-state legend (Overview + Scorecard) ───────────────────────
+/**
+ * UI-2 — the verdict-state LEGEND: plain-words for every run/trajectory verdict state
+ * so a reader never has to guess what INCOMPLETE means. Rendered in the Overview AND the
+ * Scorecard. INCOMPLETE is called out as DISTINCT from FAIL (a CRIT/HIGH criterion is
+ * uncertain → the run can't be certified, but it is not a fail). PURE.
+ */
+function verdictLegend(): string {
+  const items: [string, string, string][] = [
+    ["pass", "PASS", "every applicable criterion was met — the run is certified as pass."],
+    ["fail", "FAIL", "at least one gating (CRIT/HIGH) criterion failed."],
+    [
+      "inc",
+      "INCOMPLETE",
+      "a CRIT/HIGH criterion is UNCERTAIN (or the trace was too truncated to judge) — the run can't be certified as pass, but it is NOT a fail.",
+    ],
+    ["indet", "INDETERMINATE", "judged, but no confident pass/fail on this criterion (not certain either way)."],
+    ["na", "na", "not applicable — out of scope for this trajectory; never counted as a fail."],
+  ];
+  const cells = items
+    .map(
+      ([cls, k, v]) =>
+        `<div class="vl-item"><span class="vl-k ${cls}">${esc(k)}</span><span class="vl-v">${esc(v)}</span></div>`,
+    )
+    .join("");
+  return `<div class="vlegend"><div class="vl-h">▾ what the verdicts mean</div>${cells}</div>`;
+}
+
+// ── UI-7 — the criterion DEFINITION block (explainability) ────────────────────
+/**
+ * UI-7 — the criterion DEFINITION (what it is + HOW it's judged), shown BESIDE the
+ * verdict so a card reads as explainable, not robotic. Sourced purely from the
+ * criterion's EXISTING fields (statement · passCondition · dimension/severity/level ·
+ * judgeInputs · provenance · mined discovery). Every field OPTIONAL → graceful. PURE.
+ */
+function criterionDefn(c: ReportCriterion | undefined, fallbackId: string): string {
+  if (c === undefined) return `<div class="row">${esc(fallbackId)}</div>`;
+  const meta = [
+    c.dimension ? `dimension · ${c.dimension}` : "",
+    c.severity ? `severity · ${String(c.severity).toUpperCase()}` : "",
+    c.level ? `level · ${c.level}` : "",
+    c.gating ? "gating" : "",
+  ]
+    .filter(Boolean)
+    .join("    ·    ");
+  const judged =
+    c.judgeInputs && c.judgeInputs.length > 0
+      ? `<div class="row"><span class="ref">judged from:</span> ${esc(c.judgeInputs.join(", "))}</div>`
+      : "";
+  const disc =
+    c.discovery !== undefined
+      ? `<div class="row"><span class="ref">mined:</span> ${esc(c.discovery.targets)} — ${esc(c.discovery.why_problem)}</div>`
+      : "";
+  return (
+    `<div class="row">${esc(c.statement)}</div>` +
+    (c.passCondition && c.passCondition !== c.statement
+      ? `<div class="row"><span class="ref">pass when:</span> ${esc(c.passCondition)}</div>`
+      : "") +
+    (meta ? `<div class="row defn-meta">${esc(meta)}</div>` : "") +
+    judged +
+    `<div class="row"><span class="ref">provenance:</span> ${esc(c.provenance?.label ?? "—")}${c.provenance?.detail ? " · " + esc(c.provenance.detail) : ""}</div>` +
+    disc
+  );
+}
+
+/**
+ * WS-3 — a ONE-LINE plain-language gloss for a criterion: de-jargons WHAT it measures
+ * and WHY it matters, so a non-author can read the scorecard/findings without parsing
+ * the "Pass = …; Fail = …" statement scaffolding or the dimension/refs.
+ *
+ * SOURCED honestly, in priority order — NEVER fabricates beyond what the criterion
+ * already states:
+ *   1) the mined discovery rationale — `discovery.targets` (the behaviour it guards)
+ *      + `discovery.why_problem` (the user/correctness consequence). These are the
+ *      purpose-built readable fields, used verbatim.
+ *   2) DERIVED deterministically from the `statement` by stripping the
+ *      `Pass = …` / `Fail = …` scaffolding into "Measures: <pass clause>. Why it
+ *      matters: failing means <fail clause>." (a pure restructure — no new content).
+ *   3) the bare statement, when it carries no Pass/Fail scaffolding.
+ * Always returns a non-empty string for any criterion with a statement. PURE.
+ */
+export function plainExplainer(c: ReportCriterion | undefined): string {
+  if (c === undefined) return "";
+  const clean = (s: string): string => s.replace(/\s+/g, " ").trim();
+  // (1) mined criteria carry the most readable, purpose-built fields.
+  if (c.discovery !== undefined) {
+    const what = clean(c.discovery.targets);
+    const why = clean(c.discovery.why_problem);
+    if (what !== "") return `Measures: ${what}${why !== "" ? `. Why it matters: ${why}` : ""}.`;
+  }
+  // (2) derive from the statement — strip the Pass/Fail scaffolding (tolerant of
+  // `=`/`:` and casing; the pass clause runs up to a `;` or the `Fail` marker).
+  const stmt = clean(c.statement);
+  if (stmt === "") return "";
+  const passM = stmt.match(/pass\s*[:=]\s*([^;]+?)(?:;\s*|\s+fail\s*[:=]|$)/i);
+  const failM = stmt.match(/fail\s*[:=]\s*(.+)$/i);
+  if (passM) {
+    const what = clean(passM[1]);
+    const why = failM ? clean(failM[1]).replace(/\.$/, "") : "";
+    return `Measures: ${what}${why !== "" ? `. Why it matters: failing means ${why}` : ""}.`;
+  }
+  // (3) no scaffolding — present the statement as the measure.
+  return `Measures: ${stmt}`;
+}
+
+/** WS-3 — the muted plain-language banner element (above the technical block). Empty
+ *  source ⇒ empty string (no stray banner). Shared by the §3 subcards + §4 findings. */
+function plainBanner(c: ReportCriterion | undefined): string {
+  const txt = plainExplainer(c);
+  return txt === "" ? "" : `<div class="plain">${esc(txt)}</div>`;
+}
+
+/**
+ * WS-5 — the deterministic INDETERMINATE-resolution gate. Resolves ONE uncertain
+ * per-criterion verdict into the decision chain, NEVER leaving a bare "indeterminate":
+ *
+ *  ① PRECONDITION GATE (deterministic) — is the criterion's TRIGGER/precondition even
+ *     present in this trace? Signals, in priority:
+ *       (a) the judge's DENSE na-map marks this criterion `na` on this trace — its
+ *           explicit, per-trace applicability call (the strongest deterministic signal);
+ *       (b) the abstain is `blockedBy.kind === "scope"` — the criterion's referent is
+ *           out of scope for this route;
+ *       (c) a derivable precondition TRIGGER (tool/event) did NOT fire in the observed
+ *           agent steps (used when the criterion carries one + the trace has steps).
+ *     Any ⇒ `na` (precondition absent) — DROPPED from the applicable denominator.
+ *  ② Otherwise the precondition WAS present but the judge could not decide ⇒
+ *     `needs-evidence`, with a concrete NEXT-ACTION: code-recheck (a deterministic/hybrid
+ *     criterion) · 2nd-judge (re-ground a `factual-intent` abstain) · revise-criterion
+ *     (a `normative` value-call — operator-owned) · hitl-spot-check (default).
+ * PURE. `reason` is the judge's own `blockedBy.text` when present (the unbound term),
+ * never invented.
+ */
+export function classifyIndeterminate(
+  criterion: ReportCriterion | undefined,
+  verdict: { result?: string; blockedBy?: { kind?: string; text?: string } } | undefined,
+  denseValue: string | undefined,
+  agentSteps?: AgentStep[],
+): IndeterminateResolution {
+  const cid = criterion?.id ?? "";
+  const bb = verdict?.blockedBy;
+  const trigger = preconditionTrigger(criterion);
+  const norm = (s: string): string => s.trim().toLowerCase();
+  const triggerAbsent =
+    trigger !== undefined && agentSteps !== undefined && agentSteps.length > 0
+      ? !agentSteps.some((s) => norm(s.tool ?? "") === norm(trigger))
+      : false;
+  // ① precondition-absent ⇒ N/A (excluded from the applicable denominator)
+  if (denseValue === "na" || bb?.kind === AssumptionKind.Scope || triggerAbsent) {
+    const reason =
+      bb?.text && bb.text.trim().length > 0
+        ? bb.text.trim()
+        : trigger
+          ? `trigger '${trigger}' not invoked in this trace`
+          : "criterion not applicable to this trace — precondition/trigger absent";
+    return { criterionId: cid, resolutionClass: "na", reason };
+  }
+  // ② precondition present but undecided ⇒ needs-evidence + a chain next-action
+  const cm = criterion?.checkMethod;
+  const nextAction: IndeterminateNextAction =
+    cm === "deterministic" || cm === "hybrid"
+      ? "code-recheck"
+      : bb?.kind === AssumptionKind.FactualIntent
+        ? "2nd-judge"
+        : bb?.kind === AssumptionKind.Normative
+          ? "revise-criterion"
+          : "hitl-spot-check";
+  const reason =
+    bb?.text && bb.text.trim().length > 0
+      ? bb.text.trim()
+      : "judge abstained — evidence present but inconclusive; needs a confirming pass";
+  return { criterionId: cid, resolutionClass: "needs-evidence", nextAction, reason };
+}
+
+/** WS-5 — the precondition TRIGGER (tool/event) a criterion requires before it applies,
+ *  when derivable. RESERVED hook: `ReportCriterion` carries no codeEval today, so this
+ *  returns undefined and the deterministic gate runs off the judge's dense na-map + the
+ *  typed `blockedBy`. A codeEval-bearing criterion (discover-side) can override later. */
+function preconditionTrigger(_criterion: ReportCriterion | undefined): string | undefined {
+  return undefined;
+}
+
+// ── §9.4.4 R3 — per-tab "what this tab shows" description ─────────────────────
+/** A standardized tab-description banner (R3): every tab states what it shows. */
+function tabDesc(text: string): string {
+  return `<div class="tabdesc"><span class="td-i">ⓘ</span> ${esc(text)}</div>`;
+}
+
+// ── §9.4.4 R4 — criterion provenance + human-readable statement + hover ───────
+/**
+ * The provenance chip for a criterion (R4): `defined` (authored matrix row) vs
+ * `source` (mined from the trace ✓/✗ split), with a hover title carrying the origin.
+ */
+function provChip(c: ReportCriterion | undefined): string {
+  const p = c?.provenance;
+  if (p === undefined) return "";
+  const cls = p.kind === "defined" ? "defined" : "source";
+  const title = `${p.label}${p.detail ? " — " + p.detail : ""}`;
+  return `<span class="prov ${cls}" title="${esc(title)}">${esc(p.kind)}</span>`;
+}
+
+/**
+ * UI-10 — the CODE vs JUDGE chip for a criterion, from its `checkMethod` router:
+ * `deterministic` → CODE (a deterministic code-eval — zero judge tokens) · `hybrid`
+ * → CODE+JUDGE (code pre-filter, then judge on the residual) · else JUDGE (scored by
+ * the critique-before-verdict LLM judge). Lets a reader tell at a glance HOW each
+ * criterion is evaluated — code-checked vs judged. Absent checkMethod ⇒ no chip.
+ */
+function methodKind(c: ReportCriterion | undefined): "code" | "hybrid" | "judge" | undefined {
+  const m = c?.checkMethod;
+  if (m === undefined) return undefined;
+  return m === "deterministic" ? "code" : m === "hybrid" ? "hybrid" : "judge";
+}
+function methodChip(c: ReportCriterion | undefined): string {
+  const kind = methodKind(c);
+  if (kind === undefined) return "";
+  const label = kind === "code" ? "CODE" : kind === "hybrid" ? "CODE+JUDGE" : "JUDGE";
+  const title =
+    kind === "code"
+      ? "code-eval — deterministic check, no LLM judge (zero judge tokens)"
+      : kind === "hybrid"
+        ? "hybrid — deterministic code pre-filter, then LLM judge on the residual"
+        : "llm-judge — scored by the critique-before-verdict LLM judge";
+  return `<span class="method ${kind}" title="${esc(title)}">${label}</span>`;
+}
+
+/**
+ * An info-hover badge carrying the criterion's FULL definition (R4): the human-
+ * readable statement + pass condition + provenance. Rendered as a `title=`-tooltip
+ * on a small ⓘ glyph so every criterion id is self-describing on hover.
+ */
+function critHover(c: ReportCriterion | undefined): string {
+  if (c === undefined) return "";
+  const kind = methodKind(c);
+  const parts = [
+    `STATEMENT: ${c.statement}`,
+    c.passCondition ? `PASS WHEN: ${c.passCondition}` : "",
+    // UI-10 — the code-vs-judge method, spelled out in the hover.
+    kind !== undefined ? `EVALUATED BY: ${kind === "code" ? "CODE (deterministic check)" : kind === "hybrid" ? "HYBRID (code pre-filter + judge)" : "LLM JUDGE"}` : "",
+    c.provenance ? `PROVENANCE: ${c.provenance.label}${c.provenance.detail ? " (" + c.provenance.detail + ")" : ""}` : "",
+  ].filter(Boolean);
+  return `<span class="ihover" title="${esc(parts.join("\n"))}">ⓘ</span>`;
+}
+
+/**
+ * Gap A — the RAW triggering INPUT for a trace: the thing that fired the subject
+ * agent. PRIMARY source is `EvalTrace.input.prompt` (the canonical contract field,
+ * contracts/eval-types.ts). FALLBACK is the packet-style `transcript[0]` raw user
+ * turn carried on some exports (`input.transcript[0].content|text|message`). An
+ * absent / non-string / empty value ⇒ `undefined` (the row gets no input; the
+ * drill cell renders "—" — never a fabricated input). PURE, read-only.
+ */
+function traceInputText(t: EvalTrace): string | undefined {
+  const input = t.input;
+  if (input === undefined || input === null) return undefined;
+  // PRIMARY: the canonical `input.prompt`.
+  const prompt = (input as { prompt?: unknown }).prompt;
+  if (typeof prompt === "string" && prompt.trim().length > 0) return prompt;
+  // FALLBACK: a packet/transcript-style first user turn (`transcript[0]`).
+  const transcript = (input as { transcript?: unknown }).transcript;
+  if (Array.isArray(transcript) && transcript.length > 0) {
+    const first = transcript[0] as Record<string, unknown> | string | undefined;
+    if (typeof first === "string" && first.trim().length > 0) return first;
+    if (first !== null && typeof first === "object") {
+      for (const key of ["content", "text", "message", "prompt"] as const) {
+        const v = first[key];
+        if (typeof v === "string" && v.trim().length > 0) return v;
+      }
+    }
+  }
+  return undefined;
+}
+
+/**
+ * WS-6 — reconstruct the subject's SYSTEM PROMPT from a raw trace. The eval
+ * `subjectProfile` carries no `systemPrompt` (0/N verdicts), but the agent's static
+ * system message rides on every LLM-call observation as the first `role:"system"`
+ * message — so it is RECOVERABLE from the trace batch, not confabulated.
+ *
+ * Scans, in priority order: (1) each observation's `input` — a bare messages ARRAY
+ * (e.g. Langfuse `ai.generateText.doGenerate.input = [{role:"system",content:…}, …]`)
+ * OR a `{messages:[…]}` wrapper (the `ai.generateText` SPAN shape); (2) the trace-level
+ * `input` in those same two shapes. Returns the FIRST `role:"system"` message's text
+ * (string content, or the joined `text` parts of an array content). Returns `undefined`
+ * when no system message exists — the caller then leaves the prompt UNAVAILABLE
+ * (NEVER fabricated). PURE, read-only over the trace. REUSABLE: the discover-side
+ * entity hero can import this to mirror the same reconstruction.
+ */
+export function reconstructSystemPrompt(t: EvalTrace): string | undefined {
+  // content → text: a message content is a string OR an array of {type,text} parts.
+  const contentText = (content: unknown): string => {
+    if (typeof content === "string") return content;
+    if (Array.isArray(content)) {
+      return content
+        .map((part) =>
+          part !== null && typeof part === "object"
+            ? String((part as { text?: unknown }).text ?? "")
+            : typeof part === "string"
+              ? part
+              : "",
+        )
+        .join("");
+    }
+    return "";
+  };
+  // scan a message ARRAY for the first role:"system" → its (non-empty) content text.
+  const fromMessages = (msgs: unknown): string | undefined => {
+    if (!Array.isArray(msgs)) return undefined;
+    for (const m of msgs) {
+      if (m !== null && typeof m === "object" && (m as { role?: unknown }).role === "system") {
+        const txt = contentText((m as { content?: unknown }).content).trim();
+        if (txt.length > 0) return txt;
+      }
+    }
+    return undefined;
+  };
+  // a container is either a bare messages array OR a {messages:[…]} wrapper.
+  const fromContainer = (c: unknown): string | undefined => {
+    const direct = fromMessages(c);
+    if (direct !== undefined) return direct;
+    if (c !== null && typeof c === "object" && !Array.isArray(c)) {
+      return fromMessages((c as { messages?: unknown }).messages);
+    }
+    return undefined;
+  };
+  for (const o of t.observations ?? []) {
+    const hit = fromContainer(o.input);
+    if (hit !== undefined) return hit;
+  }
+  return fromContainer(t.input);
+}
+
+/**
+ * UI-9 (§9.4.5 E3) — the trace WALL-CLOCK timestamp, sourced from the RAW input
+ * prompt's `<current_time>…ISO…</current_time>` tag. Real agent harnesses stamp the
+ * wall-clock into the prompt they fire the subject with (e.g. a sample subject:
+ * `<current_time>2026-05-05T18:39:19.934Z</current_time>`), so the time the run
+ * happened is GROUND-TRUTH in the trace — NOT a judge-emit field. The eval-health
+ * heatmap buckets pass-rate by this timestamp. TOLERANT: returns the inner ISO when
+ * it parses as a real date (never a fabricated stamp); absent/garbage ⇒ undefined
+ * (the row lands in the heatmap data-pending bucket). PURE, read-only over traces.
+ */
+export function traceTimestamp(t: EvalTrace): string | undefined {
+  const raw = traceInputText(t);
+  if (raw === undefined) return undefined;
+  const m = raw.match(/<current_time>\s*([^<]+?)\s*<\/current_time>/i);
+  if (m === null) return undefined;
+  const iso = m[1].trim();
+  // validate: a real, date-parseable stamp of ISO-ish length — never fake.
+  if (iso.length < 10 || Number.isNaN(Date.parse(iso))) return undefined;
+  return iso;
+}
+
+/** A criterion's human-readable one-line statement (R4), truncated for table rows. */
+function critStatement(c: ReportCriterion | undefined, max = 88): string {
+  const s = c?.statement ?? "";
+  return s.length > max ? s.slice(0, max - 1) + "…" : s;
+}
+
+/** Coerce a tolerant context band (string | string[]) to readable strings — real
+ *  judges emit exit-states / scenarios as a LIST; join them for the §2 band. */
+function coerceContext(
+  ctx: { harness?: unknown; scenario?: unknown; exitStates?: unknown },
+): { harness?: string; scenario?: string; exitStates?: string } {
+  // TOLERANT — real judges emit exitStates/scenario as a string, a string[], OR a
+  // structured object (e.g. `{steps: 7}`). Coerce all three to a readable string so
+  // the §2 drill never renders a raw "[object Object]" (WS-1: the rich verdict files
+  // that carry an object exitStates are exactly the ones with the judge_steps walk).
+  const one = (v: unknown): string | undefined =>
+    v === undefined || v === null
+      ? undefined
+      : Array.isArray(v)
+        ? v.join(" · ")
+        : typeof v === "object"
+          ? Object.entries(v as Record<string, unknown>)
+              .map(([k, val]) => `${k}: ${val}`)
+              .join(", ")
+          : String(v);
+  return {
+    ...(ctx.harness !== undefined ? { harness: one(ctx.harness) } : {}),
+    ...(ctx.scenario !== undefined ? { scenario: one(ctx.scenario) } : {}),
+    ...(ctx.exitStates !== undefined ? { exitStates: one(ctx.exitStates) } : {}),
+  };
+}
+
+// ── criterion adapters (mined ⇄ matrix → ReportCriterion) ────────────────────
+
+const GATING_SEVERITIES = new Set(["CRIT", "HIGH"]);
+
+/** A mined criterion carries the full §5b metadata + §5c DR-2 rationale. Its
+ *  provenance is `source` (R4): mined from the ✓/✗ trace split. */
+export function fromMinedCriteria(criteria: MinedCriterion[]): ReportCriterion[] {
+  return criteria.map((c) => ({
+    id: c.id,
+    statement: c.statement,
+    severity: c.metadata.severity,
+    gating: GATING_SEVERITIES.has(c.metadata.severity),
+    dimension: c.metadata.dimension,
+    level: c.metadata.level,
+    appliesTo: c.metadata.applies_to,
+    // UI-10 — carry the code-vs-judge router (default llm-judge when the metadata omits it).
+    checkMethod: c.metadata.check_method ?? "llm-judge",
+    ...(c.metadata.judge_inputs !== undefined ? { judgeInputs: c.metadata.judge_inputs } : {}),
+    provenance: {
+      kind: "source" as const,
+      label: "mined from trace ✓/✗ split",
+      ...(c.discovery !== undefined
+        ? { detail: `grounding ${c.discovery.evidence.grounding} · seen in ${c.discovery.evidence.seen_in_traces} trace(s)` }
+        : {}),
+    },
+    discovery: c.discovery,
+  }));
+}
+
+/** A matrix criterion carries statement + pass condition + severity. Its provenance
+ *  is `defined` (R4): an authored eval-matrix row. */
+export function fromMatrixCriteria(criteria: MatrixCriterion[]): ReportCriterion[] {
+  return criteria.map((c) => ({
+    id: c.criterionId,
+    statement: c.statement,
+    severity: c.severity,
+    gating: GATING_SEVERITIES.has(c.severity),
+    passCondition: c.passCondition,
+    // UI-10 — carry the code-vs-judge router (default llm-judge when the matrix row omits it).
+    checkMethod: c.checkMethod ?? "llm-judge",
+    provenance: { kind: "defined" as const, label: "defined eval-matrix criterion" },
+    ...(c.dimension !== undefined ? { dimension: c.dimension } : {}),
+    ...(c.judgeInputs !== undefined ? { judgeInputs: c.judgeInputs } : {}),
+  }));
+}
+
+// ── D-1 terminal eval-cards ─────────────────────────────────────────────────
+
+/**
+ * Render per-criterion terminal cards from the Scorecard: criterion · verdict ·
+ * severity · critique snippet · the §5c DR-2 grounding tag (when mined). Plain
+ * text, deterministic (criteria in `verdicts` order). The D-1 dogfood fix.
+ */
+export function renderEvalCards(input: EvalReportInput): string {
+  const lines: string[] = [];
+  const g = input.scorecard.gate;
+  lines.push(`╔═ EVAL CARDS — ${input.subject.name} ═══`);
+  lines.push(`║ GATE: ${g.passed ? "PASS" : (g.runVerdict ?? "fail").toUpperCase()}  (${g.passCount}/${g.total} pass, ${g.gatedBy.length} gating)`);
+  lines.push("╠══════════════════════════════════════");
+  for (const v of input.verdicts) {
+    const crit = criterionById(input, v.criterionId);
+    const sev = crit?.severity ?? "MED";
+    // UI-10 — the code-vs-judge tag, so the terminal cards distinguish code-checked
+    // criteria from judged ones (CODE · HYBRID · JUDGE). Default JUDGE when absent.
+    const mkind = methodKind(crit);
+    const method = mkind === "code" ? "CODE" : mkind === "hybrid" ? "HYBRID" : "JUDGE";
+    const grounding = crit?.discovery?.evidence.grounding ?? Grounding.Inferred;
+    const prevalence = crit?.discovery?.evidence.prevalence ?? "n/a";
+    const snippet = v.critique.length > 80 ? v.critique.slice(0, 77) + "…" : v.critique;
+    lines.push(`║ ▸ ${v.criterionId}  [${sev}·${method}]  ${resultToken(v.result).toUpperCase()}  (conf ${v.confidence})`);
+    lines.push(`║     grounding: ${grounding} (${prevalence})`);
+    lines.push(`║     ${snippet}`);
+  }
+  lines.push("╚══════════════════════════════════════");
+  return lines.join("\n");
+}
+
+// ── wiring: build the rich report input from real pipeline data ──────────────
+
+function ledgerVerdict(perCriterion: Record<string, string>): string {
+  const vals = Object.values(perCriterion).filter((v) => v !== "na");
+  if (vals.includes(OutcomeVerdict.Fail)) return "FAIL";
+  if (vals.includes(OutcomeVerdict.Uncertain)) return "INDETERMINATE";
+  if (vals.length === 0) return "INCOMPLETE";
+  return "PASS";
+}
+
+/** Optional §9.4 trajectory-level fields the judge MAY emit (additive). */
+interface MatrixVerdictFileRich extends MatrixVerdictFile {
+  route?: string;
+  context?: { harness?: string | string[]; scenario?: string | string[]; exitStates?: string | string[] };
+  agentSteps?: AgentStep[];
+  judgeSteps?: JudgeStep[];
+  localize?: string;
+  health?: JudgeHealthRow;
+  // §9.4.4 v2.2
+  understanding?: Understanding;
+  expectedTrajectory?: ExpectedStep[];
+  subjectProfile?: SubjectProfile;
+  /** §9.4.5 E3 — the trace wall-clock timestamp (ISO), when the source carried it.
+   *  Common spellings tolerated below (`timestamp` · `traceTimestamp` · `startTime`). */
+  timestamp?: string;
+  traceTimestamp?: string;
+  startTime?: string;
+}
+
+/**
+ * Best-effort carry of an observation id onto each `AgentStep` so the §2 judge
+ * lane can map a step to the criterion verdicts that EXAMINED it by the precise
+ * `ref.obs === step.obs` key (on top of the tool-name fallback). The verdict
+ * files DON'T emit a per-step obs id — but the per-criterion verdict `refs` and
+ * critiques DO carry obs ids bound to a tool:
+ *   (a) `refs[]` with `path:"name"` → `{obs:<id>, value:<toolName>}`.
+ *   (b) `tool@<obsid>` tokens the judge embeds in `critique` text
+ *       (e.g. "sendMessage@04ff44e20dbacb37") — the ONLY source for tools that
+ *       were cited only by an `output.*` ref (which carries no tool name).
+ * We gather tool → ordered-unique obs ids from both, then assign positionally per
+ * step (nth call of a tool → nth obs id). A tool with no recovered obs id leaves
+ * the step's obs UNDEFINED (honest — the tool-name match still covers it). PURE.
+ */
+export function enrichAgentStepsObs(f: {
+  agentSteps?: AgentStep[];
+  verdicts?: Array<{ critique?: string; refs?: Array<{ obs?: string; path?: string; value?: string }> }>;
+}): AgentStep[] | undefined {
+  const steps = f.agentSteps;
+  if (steps === undefined) return undefined;
+  const toolObs = new Map<string, string[]>();
+  const push = (tool: string, obs: string): void => {
+    if (!tool || !obs) return;
+    const list = toolObs.get(tool) ?? [];
+    if (!list.includes(obs)) {
+      list.push(obs);
+      toolObs.set(tool, list);
+    }
+  };
+  for (const v of f.verdicts ?? []) {
+    for (const rf of v.refs ?? []) {
+      if (rf && typeof rf === "object" && rf.path === "name" && rf.value && rf.obs) {
+        push(String(rf.value), String(rf.obs));
+      }
+    }
+    // `tool@<obs>` tokens in the critique (obs ids are 6+ hex chars).
+    const crit = typeof v.critique === "string" ? v.critique : "";
+    const re = /([A-Za-z_]\w*)@([0-9a-f]{6,})/g;
+    let m: RegExpExecArray | null;
+    while ((m = re.exec(crit)) !== null) push(m[1], m[2]);
+  }
+  if (toolObs.size === 0) return steps;
+  const counter: Record<string, number> = {};
+  return steps.map((s) => {
+    const tool = s.tool ?? "";
+    const list = toolObs.get(tool);
+    if (list === undefined || list.length === 0) return s;
+    const i = counter[tool] ?? 0;
+    counter[tool] = i + 1;
+    const obs = list[i];
+    return obs !== undefined ? { ...s, obs } : s;
+  });
+}
+
+/**
+ * Derive the rich `EvalReportInput` fields (ledger · coverage · cohorts · gating
+ * rolls · top findings · judge-health) from the REAL pipeline outputs — the
+ * per-trajectory Judge-Agent verdict files + the folded scorecard + the criteria.
+ * Invents NO data: every cell is read from a verdict file or the scorecard.
+ *
+ * §9.4-judge-emit CLOSED (T2): the JUDGE AGENT now EMITS the side-by-side fields
+ * (`judgeSteps`, `agentSteps`, `context`, `localize`, `health`, `route`), a dense
+ * na-explicit `denseMap`, a `fidelity` gate, and per-verdict `confidenceBand` —
+ * mandated in `evaluator.md #mode-judge-trajectory`. The renderer uses an emitted
+ * `denseMap` verbatim when present and otherwise synthesizes one from `verdicts[]`
+ * (na for any unjudged row), so a judge that omits the walk still renders. PURE.
+ */
+export function buildEvalReportInput(params: {
+  subject: EvalReportInput["subject"];
+  scorecard: Scorecard;
+  verdicts: CriterionVerdict[];
+  criteria: ReportCriterion[];
+  matrixVerdictFiles?: MatrixVerdictFile[];
+  sourceMap?: SourceMap;
+  handover?: HandoverBundle | null;
+  generatedAt: string;
+  triaged?: number;
+  /** §9.4.4 M1/R2 — the subject profile (given or reconstructed) for the report. */
+  subjectProfile?: SubjectProfile;
+  /** Overview provenance meta-strip — the run config (judge substrate · pinned model · …). */
+  runConfig?: RunConfig;
+  /**
+   * Gap A — the INGESTED trace batch (the EvalTraces under evaluation). Used ONLY
+   * to source each ledger row's RAW triggering INPUT (`EvalTrace.input.prompt`,
+   * keyed by trace id == `MatrixVerdictFile.trajectoryId`). The input is
+   * ground-truth in the trace — NOT a judge-emit field, so this stays non-protocol
+   * (the judge schema is untouched). ABSENT ⇒ rows carry no `input` and the §2
+   * drill cell renders "—" (graceful, never faked). PURE — read-only over traces.
+   */
+  traces?: EvalTrace[];
+}): EvalReportInput {
+  const files = (params.matrixVerdictFiles ?? []) as MatrixVerdictFileRich[];
+  const critIds = params.criteria.map((c) => c.id);
+  // Gap A — index the raw triggering INPUT by trace id (== trajectoryId). Read
+  // `EvalTrace.input.prompt`; an absent/empty input leaves the row without `input`
+  // (the drill renders "—"). NO judge echo — the input is read straight from the
+  // trace the verdict was produced over.
+  const inputByTrace = new Map<string, string>();
+  // UI-9 (§9.4.5 E3) — index the wall-clock timestamp by trace id (== trajectoryId),
+  // extracted from the trace's `<current_time>` prompt stamp. This is the REAL data
+  // path the eval-health heatmap needs; verdict files carry no timestamp, but the
+  // ingested traces do. Mirror inputByTrace; NO contract change.
+  const timestampByTrace = new Map<string, string>();
+  for (const t of params.traces ?? []) {
+    const raw = traceInputText(t);
+    if (raw !== undefined) inputByTrace.set(t.id, raw);
+    const ts = traceTimestamp(t);
+    if (ts !== undefined) timestampByTrace.set(t.id, ts);
+  }
+  const sevById = new Map(params.criteria.map((c) => [c.id, c.severity]));
+  const critById = new Map(params.criteria.map((c) => [c.id, c]));
+  const gatingIds = new Set(params.criteria.filter((c) => c.gating).map((c) => c.id));
+
+  // ── ledger: one dense row per trajectory ──
+  const ledger: LedgerRow[] = files.map((f) => {
+    const perCriterion: Record<string, string> = {};
+    for (const cid of critIds) perCriterion[cid] = "na"; // dense, na-explicit
+    // §9.4.2 node 9: prefer the judge's EMITTED dense map (na-explicit) verbatim;
+    // else synthesize from verdicts[] (any unjudged row stays na).
+    const emittedDense = (f as { denseMap?: Record<string, string> }).denseMap;
+    if (emittedDense !== undefined) {
+      for (const cid of critIds) if (emittedDense[cid] !== undefined) perCriterion[cid] = emittedDense[cid];
+    }
+    for (const v of f.verdicts) perCriterion[v.criterionId] = v.result;
+    // WS-5 — the INDETERMINATE resolution chain: deterministically resolve every
+    // uncertain per-criterion verdict into `na` (precondition absent — DROPPED from the
+    // applicable denominator) or `needs-evidence` (+ a next-action). A precondition-
+    // absent uncertain is RECLASSIFIED to `na` in `perCriterion` so EVERY downstream
+    // count (rolls · cohorts · heatmap · trace verdict) becomes denominator-honest —
+    // no bare "indeterminate" that conflates not-applicable with under-evidenced.
+    const verdictByCid = new Map(
+      f.verdicts.map((v) => [v.criterionId, v as { result?: string; blockedBy?: { kind?: string; text?: string } }]),
+    );
+    const indeterminateResolution: Record<string, IndeterminateResolution> = {};
+    for (const cid of critIds) {
+      if (perCriterion[cid] === OutcomeVerdict.Uncertain) {
+        const res = classifyIndeterminate(critById.get(cid), verdictByCid.get(cid), emittedDense?.[cid], f.agentSteps);
+        indeterminateResolution[cid] = res;
+        if (res.resolutionClass === "na") perCriterion[cid] = "na";
+      }
+    }
+    const failingCriteria = f.verdicts.filter((v) => v.result === OutcomeVerdict.Fail).map((v) => v.criterionId);
+    const grounding = f.verdicts
+      .flatMap((v) => v.refs ?? [])
+      .map((r) => `${r.obs}/${r.path}`)
+      .slice(0, 3)
+      .join(" · ");
+    // #3 EXAMINE fidelity gate: a trajectory whose deterministic fidelity gate
+    // fired (truncated / unterminated) is INCOMPLETE regardless of its dense map —
+    // it was never walked per-criterion.
+    const isIncomplete = (f as { fidelity?: { complete?: boolean } }).fidelity?.complete === false;
+    const row: LedgerRow = {
+      trajectoryId: f.trajectoryId,
+      verdict: isIncomplete ? "INCOMPLETE" : ledgerVerdict(perCriterion),
+      perCriterion,
+      failingCriteria,
+      ...(f.route !== undefined ? { route: f.route } : {}),
+      ...(grounding ? { grounding } : {}),
+      // WS-1 — bind the per-criterion JUDGE verdicts (result · critique · refs) for
+      // THIS trajectory, read verbatim off the verdict file's `verdicts[]`. This is
+      // the rich "How the Judge Reasoned" payload that EVERY judged trace carries
+      // (PASS or FAIL), independent of whether a judge_steps[] walk was emitted —
+      // so the §2 drill never renders an empty reasoning panel when the verdict file
+      // has data. NO invention: a row with no verdicts simply omits the field.
+      ...(f.verdicts.length > 0
+        ? { criterionVerdicts: f.verdicts as unknown as CriterionTrajectoryVerdict[] }
+        : {}),
+      // Gap A — the raw triggering INPUT for this trajectory, sourced from the
+      // trace by id (== trajectoryId). ABSENT ⇒ the drill cell renders "—".
+      ...(inputByTrace.has(f.trajectoryId) ? { input: inputByTrace.get(f.trajectoryId) } : {}),
+      ...(f.context !== undefined ? { context: coerceContext(f.context) } : {}),
+      // Carry recovered obs ids onto each agent step so the §2 judge lane can map
+      // a step → the criterion verdicts that examined it by `ref.obs === step.obs`.
+      ...(f.agentSteps !== undefined ? { agentSteps: enrichAgentStepsObs(f) } : {}),
+      ...(f.judgeSteps !== undefined ? { judgeSteps: f.judgeSteps } : {}),
+      ...(f.localize !== undefined ? { localize: localizeText(f.localize) } : {}),
+      // §9.4.4 v2.2 — node-0 understanding · node-0.5 expected-trajectory · M1 profile.
+      ...(f.understanding !== undefined ? { understanding: f.understanding } : {}),
+      ...(f.expectedTrajectory !== undefined ? { expectedTrajectory: f.expectedTrajectory } : {}),
+      ...(f.subjectProfile !== undefined ? { subjectProfile: f.subjectProfile } : {}),
+      // T4: DERIVE health from the walk when judge_steps were emitted (don't trust
+      // the self-reported `health`); fall back to the self-reported field otherwise.
+      ...((f.judgeSteps?.length ?? 0) > 0
+        ? { health: deriveWalkHealth(f) }
+        : f.health !== undefined
+          ? { health: f.health }
+          : {}),
+      // §9.4.5 E3 / UI-9 — carry the trace timestamp for the eval-HEALTH temporal
+      // heatmap. PRIMARY = the trace's `<current_time>` wall-clock (real ground-truth,
+      // sourced from params.traces); FALLBACK = a verdict-file timestamp (tolerant
+      // spellings) if a judge ever emits one. ABSENT ⇒ this row lands in the
+      // data-pending bucket (never faked).
+      ...(((timestampByTrace.get(f.trajectoryId) ?? f.timestamp ?? f.traceTimestamp ?? f.startTime) !== undefined)
+        ? { timestamp: timestampByTrace.get(f.trajectoryId) ?? f.timestamp ?? f.traceTimestamp ?? f.startTime }
+        : {}),
+      // WS-5 — the per-criterion indeterminate resolution for this trace (na vs needs-evidence).
+      ...(Object.keys(indeterminateResolution).length > 0 ? { indeterminateResolution } : {}),
+      // WS-2 — how this trajectory resolved (judge-walk · judged-walk-not-captured ·
+      // truncated). Makes the "no walk" traces HONEST: judged, not unjudged.
+      resolution: deriveTrajectoryResolution(f),
+    };
+    return row;
+  });
+
+  // ── §9.4.5 E2 — per-tool OBSERVED call-count (from the trajectory agent-steps). In
+  //    the trace-only (no-code-access) case this is the ONLY honest tool evidence: how
+  //    often each tool was actually seen called across the judged trajectories. PURE. ──
+  const toolObservations: Record<string, number> = {};
+  for (const f of files) {
+    for (const step of f.agentSteps ?? []) {
+      const name = step.tool;
+      if (typeof name === "string" && name.length > 0) {
+        toolObservations[name] = (toolObservations[name] ?? 0) + 1;
+      }
+    }
+  }
+
+  // ── §9.4.4 R3/M5 — collect the DETECTED-but-unmatched flags (node-2.5 candidates),
+  //    surfaced in §4 SEPARATED from criterion fails (a detection, never a verdict). ──
+  const detectedFlags: DetectedFlag[] = [];
+  for (const f of files) {
+    for (const cand of f.candidates ?? []) {
+      detectedFlags.push({
+        kind: cand.kind,
+        detection: cand.detection,
+        trajectoryId: f.trajectoryId,
+        ...(cand.anchor !== undefined ? { anchor: cand.anchor } : {}),
+        ...(cand.ref !== undefined ? { ref: cand.ref } : {}),
+      });
+    }
+  }
+  // §9.4.4 M1/R2 — the subject profile: prefer the run-supplied one, else the first
+  // one a judge echoed/reconstructed on its verdict file.
+  const subjectProfileBase = params.subjectProfile ?? files.find((f) => f.subjectProfile !== undefined)?.subjectProfile;
+  // WS-6 — reconstruct the SYSTEM PROMPT from the trace batch when the profile lacks
+  // one. The eval subjectProfile never carries `systemPrompt`, but the agent's static
+  // system message rides on each LLM-call observation (role:"system"). Extract it ONCE
+  // from the first trace that yields one, attach it, and TAG it reconstructed (added to
+  // `inferredFields` ⇒ the entity hero shows the "system prompt · reconstructed"
+  // collapsible). NEVER confabulate: if no trace carries a system message, leave it
+  // UNAVAILABLE. If no profile exists at all but a system prompt is recoverable, mint a
+  // MINIMAL reconstructed profile so the recovered prompt still surfaces.
+  let reconstructedSystemPrompt: string | undefined;
+  if (subjectProfileBase?.systemPrompt === undefined || subjectProfileBase.systemPrompt.trim() === "") {
+    for (const t of params.traces ?? []) {
+      const sys = reconstructSystemPrompt(t);
+      if (sys !== undefined) {
+        reconstructedSystemPrompt = sys;
+        break;
+      }
+    }
+  }
+  const subjectProfile: SubjectProfile | undefined =
+    reconstructedSystemPrompt !== undefined
+      ? ({
+          ...(subjectProfileBase ?? { provenance: "reconstructed" }),
+          systemPrompt: reconstructedSystemPrompt,
+          inferredFields: Array.from(
+            new Set([...(subjectProfileBase?.inferredFields ?? []), "systemPrompt"]),
+          ),
+        } as SubjectProfile)
+      : subjectProfileBase;
+
+  // ── coverage: byVerdict tally over the ledger ──
+  const byVerdict: Record<string, number> = { PASS: 0, FAIL: 0, INDETERMINATE: 0, INCOMPLETE: 0 };
+  for (const r of ledger) byVerdict[r.verdict] = (byVerdict[r.verdict] ?? 0) + 1;
+  const coverage = {
+    judged: ledger.length,
+    triaged: params.triaged ?? ledger.length,
+    byVerdict,
+    gapNote:
+      ledger.length > 0
+        ? `${ledger.length} trajectory verdict(s) folded; every in-scope trajectory carries a per-criterion judgement (0 silently dropped).`
+        : "no per-trajectory verdict files supplied for this render — scorecard-only.",
+  };
+
+  // ── cohorts: criteria × route cohort (route defaults to "all") ──
+  const counts: Record<string, number> = {};
+  const matrix: Record<string, Record<string, { pass: number; fail: number; indeterminate: number }>> = {};
+  for (const r of ledger) {
+    const cohort = r.route ?? "all";
+    counts[cohort] = (counts[cohort] ?? 0) + 1;
+    for (const [cid, res] of Object.entries(r.perCriterion)) {
+      if (res === "na") continue;
+      matrix[cid] ??= {};
+      matrix[cid][cohort] ??= { pass: 0, fail: 0, indeterminate: 0 };
+      if (res === OutcomeVerdict.Pass) matrix[cid][cohort].pass += 1;
+      else if (res === OutcomeVerdict.Fail) matrix[cid][cohort].fail += 1;
+      else matrix[cid][cohort].indeterminate += 1;
+    }
+  }
+
+  // ── per-criterion roll (over the ledger). WS-5 — the applicable denominator EXCLUDES
+  //    `na` (precondition absent); the remaining `uncertain` are GENUINE needs-evidence
+  //    (the precondition-absent uncertains were already reclassified to `na` above). ──
+  function rollOf(cid: string): {
+    applicable: number;
+    pass: number;
+    fail: number;
+    needsEvidence: number;
+    na: number;
+    example?: string;
+  } {
+    let applicable = 0,
+      pass = 0,
+      fail = 0,
+      needsEvidence = 0,
+      na = 0;
+    let example: string | undefined;
+    for (const r of ledger) {
+      const res = r.perCriterion[cid];
+      if (res === undefined) continue;
+      if (res === "na") {
+        na += 1;
+        continue;
+      }
+      applicable += 1;
+      if (res === OutcomeVerdict.Fail) {
+        fail += 1;
+        example ??= r.trajectoryId;
+      } else if (res === OutcomeVerdict.Uncertain) {
+        needsEvidence += 1;
+        example ??= r.trajectoryId;
+      } else pass += 1;
+    }
+    return { applicable, pass, fail, needsEvidence, na, ...(example !== undefined ? { example } : {}) };
+  }
+
+  // #6 — the AUDITABLE second-judge note: surfaces that an INDEPENDENT verifier ran
+  // over a gating fail and what it concluded (upheld vs refuted→downgraded), so the
+  // report shows the refutation result (not just "eligible for"). PURE.
+  function ivNote(folded?: CriterionVerdict): string {
+    const iv = folded?.independentVerify;
+    if (iv === undefined) return "";
+    const who = iv.reviewerId !== undefined ? ` (${iv.reviewerId})` : "";
+    return ` · [INDEPENDENT VERIFY — 2nd judge${who}: ${iv.upheld ? "UPHELD" : "REFUTED → downgraded"}; ${iv.reason}]`;
+  }
+
+  // ── gating-criteria rolls (the gating subset; GA-D2b applicable denominator) ──
+  const gatingCriteria: GatingRoll[] = [...gatingIds]
+    .map((cid) => {
+      const roll = rollOf(cid);
+      const passRate = roll.applicable > 0 ? Math.round((100 * roll.pass) / roll.applicable) : 100;
+      const folded = params.verdicts.find((v) => v.criterionId === cid);
+      return {
+        criterion: cid,
+        severity: sevById.get(cid) ?? "MED",
+        applicable: roll.applicable,
+        fail: roll.fail,
+        indeterminate: roll.needsEvidence,
+        na: roll.na,
+        passRateOverApplicable: passRate,
+        denominatorNote: `N/A excluded — denominator = ${roll.applicable} applicable (na ${roll.na} dropped, GA-D2b)`,
+        ...(folded?.critique ? { root: folded.critique + ivNote(folded) } : {}),
+      };
+    })
+    .sort((a, b) => a.passRateOverApplicable - b.passRateOverApplicable);
+
+  // ── WS-5 — per-criterion INDETERMINATE resolution chain (ALL criteria, not just
+  //    gating). Splits each criterion's indeterminate into N/A (precondition absent,
+  //    denominator-excluded) vs needs-evidence (carries aggregated chain next-actions).
+  //    The reasons/next-actions are aggregated from each trace's deterministic
+  //    `indeterminateResolution`. ──
+  const criterionResolutions: CriterionResolution[] = params.criteria.map((c) => {
+    const roll = rollOf(c.id);
+    const naReasons: Record<string, number> = {};
+    const actionAgg: Record<string, { count: number; reason: string }> = {};
+    for (const r of ledger) {
+      const res = r.indeterminateResolution?.[c.id];
+      if (res === undefined) continue;
+      if (res.resolutionClass === "na") naReasons[res.reason] = (naReasons[res.reason] ?? 0) + 1;
+      else if (res.nextAction !== undefined) {
+        const a = (actionAgg[res.nextAction] ??= { count: 0, reason: res.reason });
+        a.count += 1;
+      }
+    }
+    const naReason = Object.entries(naReasons).sort((a, b) => b[1] - a[1])[0]?.[0];
+    return {
+      criterion: c.id,
+      severity: c.severity,
+      applicable: roll.applicable,
+      pass: roll.pass,
+      fail: roll.fail,
+      na: roll.na,
+      needsEvidence: roll.needsEvidence,
+      passRateOverApplicable: roll.applicable > 0 ? Math.round((100 * roll.pass) / roll.applicable) : 100,
+      ...(naReason !== undefined ? { naReason } : {}),
+      nextActions: Object.entries(actionAgg)
+        .map(([action, v]) => ({ action: action as IndeterminateNextAction, count: v.count, reason: v.reason }))
+        .sort((a, b) => b.count - a.count),
+    };
+  });
+
+  // ── top findings: every criterion with ≥1 applicable fail, worst-first ──
+  const topFindings: TopFinding[] = params.criteria
+    .map((c) => {
+      const roll = rollOf(c.id);
+      const prev = roll.applicable > 0 ? Math.round((100 * roll.fail) / roll.applicable) : 0;
+      const folded = params.verdicts.find((v) => v.criterionId === c.id);
+      return {
+        criterion: c.id,
+        severity: c.severity,
+        applicable: roll.applicable,
+        failCount: roll.fail,
+        prevalencePctOverApplicable: prev,
+        ...(folded?.critique ? { root: folded.critique + ivNote(folded) } : {}),
+        ...(roll.example !== undefined ? { exampleTraceId: roll.example } : {}),
+      };
+    })
+    .filter((f) => f.failCount > 0)
+    .sort((a, b) => b.prevalencePctOverApplicable - a.prevalencePctOverApplicable || b.failCount - a.failCount);
+
+  // ── judge-health: grounded% (verdicts citing refs) · abstention% · symptom% ──
+  const allV = files.flatMap((f) => f.verdicts);
+  // UI-12-B — groundedPct is computed over DECIDED verdicts ONLY. An `uncertain`
+  // abstain carries empty refs BY DESIGN (it's blockedBy / na for grounding, NOT
+  // "ungrounded"), so counting it in the denominator dilutes the honest grounded
+  // rate (a clean 100%-of-decided run reads e.g. ~92% just because some rows abstained).
+  // Abstains are surfaced separately as the abstention rate. When there are NO decided
+  // verdicts to ground (all abstained / none judged), groundedPct is undefined →
+  // the cell renders "capture-unavailable" (never a fabricated 100%/0%).
+  const decidedV = allV.filter((v) => v.result !== OutcomeVerdict.Uncertain);
+  const grounded = decidedV.filter((v) => (v.refs?.length ?? 0) > 0).length;
+  const abstained = allV.filter((v) => v.result === OutcomeVerdict.Uncertain).length;
+  const symptom = ledger.filter((r) => r.health?.stoppedAtSymptom).length;
+  const judgeHealth = {
+    ...(decidedV.length > 0 ? { groundedPct: Math.round((100 * grounded) / decidedV.length) } : {}),
+    abstentionRatePct: allV.length > 0 ? Math.round((100 * abstained) / allV.length) : 0,
+    stoppedAtSymptomPct: ledger.length > 0 ? Math.round((100 * symptom) / ledger.length) : 0,
+  };
+
+  // WS-3 — derive the §5 Self-Eval AGGREGATE (behavior map + 9 components) from the
+  // SAME parsed files + ledger. Deterministic; the emit-completeness meter (5.7) reads
+  // the WS-1 gate. PURE — no extra dispatch (lean ~0 latency).
+  const selfEval = deriveSelfEval({ ledger, criteria: params.criteria, emit: assessEmitCompleteness(files) });
+
+  return {
+    subject: params.subject,
+    scorecard: params.scorecard,
+    verdicts: params.verdicts,
+    criteria: params.criteria,
+    ledger,
+    selfEval,
+    coverage,
+    cohorts: { counts, matrix },
+    gatingCriteria,
+    criterionResolutions,
+    topFindings,
+    detectedFlags,
+    judgeHealth,
+    ...(Object.keys(toolObservations).length > 0 ? { toolObservations } : {}),
+    ...(subjectProfile !== undefined ? { subjectProfile } : {}),
+    ...(params.sourceMap !== undefined ? { sourceMap: params.sourceMap } : {}),
+    ...(params.runConfig !== undefined ? { runConfig: params.runConfig } : {}),
+    handover: params.handover ?? null,
+    generatedAt: params.generatedAt,
+  };
+}
+
+// ── WS-3 §5 Self-Eval — the AGGREGATE judge-calibration derivation ────────────
+
+/** WS-3 — one behavior cluster in the §5.0 Judge Behavior Map. */
+export interface BehaviorCluster {
+  label: string;
+  count: number;
+  /** good | warn | bad | neutral — drives the chip tint. */
+  tone: string;
+  note: string;
+}
+
+/** WS-3 — the §5 Self-Eval AGGREGATE: a deterministic judge-calibration roll-up over
+ *  the WHOLE population of verdicts (NOT per-trace). Each field maps to a wireframe
+ *  component; panels that need ground-truth the run does not have are MARKED proxy. */
+export interface SelfEvalAggregate {
+  /** the headline trust band (0..100 + label + the basis sentence). */
+  trust: { score: number; label: string; basis: string };
+  /** 5.0 Judge Behavior Map — how the judge handled the population. */
+  behaviorMap: {
+    total: number;
+    verdict: { pass: number; fail: number; indeterminate: number; incomplete: number };
+    resolution: { walk: number; noWalk: number; truncated: number };
+    clusters: BehaviorCluster[];
+    /** where fails CONCENTRATE (top criteria by fail count). */
+    concentration: { id: string; statement: string; fails: number; pct: number }[];
+  };
+  /** 5.1 Reference Integrity — grounded-ref presence over decided verdicts. */
+  refIntegrity: { decided: number; grounded: number; ungrounded: number; groundedPct: number; note: string };
+  /** 5.2 Confidence Calibration — verdicts bucketed by confidence band. */
+  confidence: { band: string; count: number; pct: number }[];
+  /** 5.3 Per-Criterion Reliability. */
+  perCriterion: {
+    id: string;
+    statement: string;
+    severity: string;
+    pass: number;
+    fail: number;
+    unc: number;
+    na: number;
+    groundedPct: number;
+    flag: string;
+  }[];
+  /** 5.4 Criterion Value / Applicability Audit. */
+  criterionValue: { id: string; statement: string; severity: string; applicablePct: number; fails: number; verdict: string }[];
+  /** 5.5 Diligence / Trace Coverage. */
+  diligence: { examined: number; total: number; avgGroundingRefs: number; note: string };
+  /** 5.6 Reasoning Transparency (blocked-aware). */
+  transparency: { m2Pct: number; m3Pct: number; walkPct: number; blocked: boolean; note: string };
+  /** 5.7 Emit-Completeness — the WS-1 self-honesty meter. */
+  emit: EmitCompleteness;
+  /** 5.8 Spot-Check / Disagreement queue — the brittle verdicts worth an operator look. */
+  spotCheck: { trace: string; criterion: string; verdict: string; conf: string; reason: string }[];
+}
+
+/** WS-3 — derive the §5 Self-Eval aggregate. PURE + deterministic (sorted outputs). */
+export function deriveSelfEval(args: {
+  ledger: LedgerRow[];
+  criteria: ReportCriterion[];
+  emit: EmitCompleteness;
+}): SelfEvalAggregate {
+  const { ledger, criteria, emit } = args;
+  const total = ledger.length;
+  const critById = new Map(criteria.map((c) => [c.id, c]));
+
+  // ── verdict + resolution tallies ──
+  const verdict = { pass: 0, fail: 0, indeterminate: 0, incomplete: 0 };
+  const resolution = { walk: 0, noWalk: 0, truncated: 0 };
+  for (const r of ledger) {
+    const v = (r.verdict || "").toUpperCase();
+    if (v === "PASS") verdict.pass++;
+    else if (v === "FAIL") verdict.fail++;
+    else if (v === "INCOMPLETE") verdict.incomplete++;
+    else verdict.indeterminate++;
+    if (r.resolution === "judge-walk") resolution.walk++;
+    else if (r.resolution === "truncated") resolution.truncated++;
+    else resolution.noWalk++;
+  }
+
+  // ── per-criterion roll-up (pass/fail/unc/na + grounding) ──
+  type Agg = { pass: number; fail: number; unc: number; na: number; refd: number; decided: number };
+  const agg = new Map<string, Agg>();
+  for (const c of criteria) agg.set(c.id, { pass: 0, fail: 0, unc: 0, na: 0, refd: 0, decided: 0 });
+  for (const r of ledger) {
+    for (const [cid, res] of Object.entries(r.perCriterion)) {
+      const a = agg.get(cid);
+      if (!a) continue;
+      if (res === "pass") a.pass++;
+      else if (res === "fail") a.fail++;
+      else if (res === "uncertain" || res === "indeterminate") a.unc++;
+      else a.na++;
+    }
+    for (const cv of r.criterionVerdicts ?? []) {
+      const a = agg.get(cv.criterionId);
+      if (!a) continue;
+      const decided = cv.result === "pass" || cv.result === "fail";
+      if (decided) {
+        a.decided++;
+        if ((cv.refs?.length ?? 0) > 0) a.refd++;
+      }
+    }
+  }
+
+  const perCriterion = criteria.map((c) => {
+    const a = agg.get(c.id)!;
+    const groundedPct = a.decided > 0 ? Math.round((100 * a.refd) / a.decided) : 100;
+    const flag = a.fail > 0 ? "fails-present" : a.pass + a.fail + a.unc === 0 ? "never-applies" : "solid";
+    return { id: c.id, statement: c.statement, severity: c.severity, pass: a.pass, fail: a.fail, unc: a.unc, na: a.na, groundedPct, flag };
+  });
+
+  // 5.4 criterion value: applicable% + ever-fails → keep/low-value verdict.
+  const criterionValue = criteria.map((c) => {
+    const a = agg.get(c.id)!;
+    const applicable = a.pass + a.fail + a.unc;
+    const denom = applicable + a.na;
+    const applicablePct = denom > 0 ? Math.round((100 * applicable) / denom) : 0;
+    const everFails = a.fail > 0;
+    const verdictTag = applicable === 0 ? "never-applies → review" : everFails ? "keep — discriminates" : "keep — guards";
+    return { id: c.id, statement: c.statement, severity: c.severity, applicablePct, fails: a.fail, verdict: verdictTag };
+  });
+
+  // 5.0 concentration — where fails concentrate (top criteria by fail count).
+  const concentration = perCriterion
+    .filter((p) => p.fail > 0)
+    .sort((a, b) => b.fail - a.fail)
+    .slice(0, 6)
+    .map((p) => ({ id: p.id, statement: p.statement, fails: p.fail, pct: total > 0 ? Math.round((100 * p.fail) / total) : 0 }));
+
+  // 5.1 ref integrity — grounded-ref presence over decided verdicts (population).
+  let decided = 0;
+  let groundedRefs = 0;
+  let totalRefs = 0;
+  for (const r of ledger) {
+    for (const cv of r.criterionVerdicts ?? []) {
+      if (cv.result === "pass" || cv.result === "fail") {
+        decided++;
+        const n = cv.refs?.length ?? 0;
+        totalRefs += n;
+        if (n > 0) groundedRefs++;
+      }
+    }
+  }
+  const refIntegrity = {
+    decided,
+    grounded: groundedRefs,
+    ungrounded: decided - groundedRefs,
+    groundedPct: decided > 0 ? Math.round((100 * groundedRefs) / decided) : 100,
+    note:
+      `${totalRefs} structured {obs,path,value} refs cited across ${decided} decided verdicts. ` +
+      "Presence is checked here; full resolveRef-against-trace (GA-1) is the engine readiness gate.",
+  };
+
+  // 5.2 confidence calibration — bucket decided verdicts by confidence band.
+  const bands: { band: string; lo: number; hi: number }[] = [
+    { band: "0.9–1.0", lo: 0.9, hi: 1.01 },
+    { band: "0.7–0.9", lo: 0.7, hi: 0.9 },
+    { band: "0.5–0.7", lo: 0.5, hi: 0.7 },
+    { band: "<0.5", lo: -1, hi: 0.5 },
+  ];
+  const bandCount = bands.map(() => 0);
+  let confTotal = 0;
+  for (const r of ledger) {
+    for (const cv of r.criterionVerdicts ?? []) {
+      const conf = typeof cv.confidence === "number" ? cv.confidence : undefined;
+      if (conf === undefined) continue;
+      confTotal++;
+      const i = bands.findIndex((b) => conf >= b.lo && conf < b.hi);
+      if (i >= 0) bandCount[i]++;
+    }
+  }
+  const confidence = bands.map((b, i) => ({
+    band: b.band,
+    count: bandCount[i],
+    pct: confTotal > 0 ? Math.round((100 * bandCount[i]) / confTotal) : 0,
+  }));
+
+  // 5.5 diligence — traces examined + avg grounding refs per trace.
+  const examined = ledger.filter((r) => (r.criterionVerdicts?.length ?? 0) > 0).length;
+  const avgGroundingRefs = examined > 0 ? Math.round((10 * totalRefs) / examined) / 10 : 0;
+  const diligence = {
+    examined,
+    total,
+    avgGroundingRefs,
+    note: `${examined}/${total} trajectories carry ≥1 per-criterion judge verdict; avg ${avgGroundingRefs} grounding refs cited per examined trace.`,
+  };
+
+  // 5.6 transparency — read off the emit-completeness (blocked when M2 is 0%).
+  const fieldPct = (f: string): number => emit.fields.find((x) => x.field === f)?.pct ?? 0;
+  const transparency = {
+    m2Pct: fieldPct("understanding"),
+    m3Pct: fieldPct("expectedTrajectory"),
+    walkPct: fieldPct("judgeSteps"),
+    blocked: fieldPct("understanding") === 0,
+    note:
+      fieldPct("understanding") === 0
+        ? "BLOCKED — the judge did not persist M2 understanding / M3 expected-trajectory (emit-contract gap, §5.7). A fresh run with the hardened contract captures them."
+        : "the judge's train-of-thought (M2 understanding · M3 expected-trajectory) is captured and auditable.",
+  };
+
+  // 5.8 spot-check queue — the brittle verdicts (fails + low-confidence) worth a look.
+  const spotCheck: SelfEvalAggregate["spotCheck"] = [];
+  for (const r of ledger) {
+    for (const cv of r.criterionVerdicts ?? []) {
+      const conf = typeof cv.confidence === "number" ? cv.confidence : undefined;
+      const lowConf = conf !== undefined && conf < 0.6;
+      const isFail = cv.result === "fail";
+      if (!lowConf && !isFail) continue;
+      spotCheck.push({
+        trace: r.trajectoryId,
+        criterion: critById.get(cv.criterionId)?.statement ?? cv.criterionId,
+        verdict: String(cv.result || "").toUpperCase(),
+        conf: conf !== undefined ? conf.toFixed(2) : "—",
+        reason: isFail ? (lowConf ? "FAIL · low-confidence" : "FAIL") : "low-confidence call",
+      });
+    }
+  }
+  spotCheck.sort((a, b) => (a.conf === "—" ? 1 : b.conf === "—" ? -1 : Number(a.conf) - Number(b.conf)));
+  const spotCheckTop = spotCheck.slice(0, 20);
+
+  // headline trust band — a conservative composite (grounding · emit-completeness · fail-concentration).
+  const emitPct = emit.completePct;
+  const trustScore = Math.round(0.5 * refIntegrity.groundedPct + 0.5 * emitPct);
+  const trustLabel =
+    transparency.blocked || emitPct < 50
+      ? "PARTIAL — reasoning transparency blocked by the emit-contract gap"
+      : trustScore >= 85
+        ? "STRONG — grounded + transparent"
+        : "MODERATE — grounded, transparency improving";
+  const trust = {
+    score: trustScore,
+    label: trustLabel,
+    basis: `grounding ${refIntegrity.groundedPct}% of decided verdicts · emit-completeness ${emitPct}% · ${verdict.fail} failing trajectories across ${total}.`,
+  };
+
+  return {
+    trust,
+    behaviorMap: { total, verdict, resolution, clusters: behaviorClusters(verdict, resolution, total), concentration },
+    refIntegrity,
+    confidence,
+    perCriterion,
+    criterionValue,
+    diligence,
+    transparency,
+    emit,
+    spotCheck: spotCheckTop,
+  };
+}
+
+/** WS-3 — derive the §5.0 behavior clusters from the population tallies. PURE. */
+function behaviorClusters(
+  verdict: { pass: number; fail: number; indeterminate: number; incomplete: number },
+  resolution: { walk: number; noWalk: number; truncated: number },
+  total: number,
+): BehaviorCluster[] {
+  const pct = (n: number): number => (total > 0 ? Math.round((100 * n) / total) : 0);
+  const out: BehaviorCluster[] = [
+    { label: "Passed (criteria satisfied)", count: verdict.pass, tone: "good", note: `${pct(verdict.pass)}% of the population judged PASS on all applicable criteria.` },
+    { label: "Failed (≥1 criterion)", count: verdict.fail, tone: "bad", note: `${pct(verdict.fail)}% carry ≥1 failing criterion — the discriminating cases.` },
+    { label: "Abstained (indeterminate)", count: verdict.indeterminate, tone: "warn", note: `${pct(verdict.indeterminate)}% abstained — inputs could not decide (blockedBy), routed to calibration not gate.` },
+    { label: "Truncated (INCOMPLETE)", count: verdict.incomplete, tone: "neutral", note: `${pct(verdict.incomplete)}% gated INCOMPLETE by the node-1 fidelity short-circuit (never walked).` },
+    { label: "Walk captured", count: resolution.walk, tone: "good", note: `${pct(resolution.walk)}% persisted the full §9.4 judge walk (target lane + anchored reasoning).` },
+    { label: "Walk dropped (judged)", count: resolution.noWalk, tone: "warn", note: `${pct(resolution.noWalk)}% were judged but did NOT persist the structured walk — the emit-contract gap (§5.7).` },
+  ];
+  return out.filter((c) => c.count > 0).sort((a, b) => b.count - a.count);
+}
+
+// ── §1 Overview ──────────────────────────────────────────────────────────────
+
+/**
+ * Overview-redesign — the full-width ENTITY HERO card: rich subject identity from the
+ * §9.4.4 M1 subject profile (name · entityType · purpose · tools · skill · scope ·
+ * version · code-access · system-prompt). HONEST: a field the profile cannot establish
+ * renders the `unknown` sentinel (never confabulated); an INFERRED (reconstructed)
+ * field carries an `inferred` honesty chip; the system prompt is COLLAPSED. PURE.
+ */
+function entityHero(input: EvalReportInput): string {
+  const p = input.subjectProfile;
+  const inferred = new Set(p?.inferredFields ?? []);
+  const name = p?.identity ?? input.subject.name;
+  const codeAccess = p?.provenance ?? "reconstructed";
+  // §9.4.5 E1 — TRACE-ONLY when there is no code access (reconstructed profile, or no
+  // profile at all). Drives the honest "CODE-ACCESS: NONE · trace-only" header + the
+  // UNAVAILABLE (never-confabulated) marking of harness + system-prompt.
+  const isTraceOnly = codeAccess !== "given";
+  const nTraces =
+    input.sourceMap?.traceCount ?? input.coverage?.judged ?? input.ledger?.length ?? 0;
+
+  // a value cell: shows the value, or the `unknown` sentinel when absent; an `inferred`
+  // chip flags a reconstructed (not-given) field so given-vs-inferred is honest.
+  const cell = (value: string | undefined, key?: string): string => {
+    const has = typeof value === "string" && value.trim().length > 0 && value !== PROFILE_UNKNOWN;
+    const body = has ? esc(value as string) : `<span class="unk">${esc(PROFILE_UNKNOWN)}</span>`;
+    const inf = key !== undefined && inferred.has(key) ? ` <span class="inf">inferred</span>` : "";
+    return `${body}${inf}`;
+  };
+
+  // §9.4.5 E2 — tools as diagnostics-style CHIPS with a per-tool OBSERVED call-count
+  // ("obs N", from the trajectory agent-steps) instead of a flat comma list. In the
+  // trace-only case this is the only honest tool evidence (what was actually called).
+  const obs = input.toolObservations ?? {};
+  const tools = p?.tools ?? [];
+  const toolsInferred = inferred.has("tools");
+  const toolChips =
+    tools.length > 0
+      ? tools
+          .map((t) => {
+            const n = obs[t] ?? 0;
+            const count =
+              n > 0
+                ? ` <b class="obs" title="observed ${n} call(s) across the judged trajectories">obs ${n}</b>`
+                : ` <b class="obs none" title="present in the inventory; 0 calls observed in the judged trajectories">obs 0</b>`;
+            return `<span class="tchip">${esc(t)}${count}</span>`;
+          })
+          .join("")
+      : `<span class="unk">${esc(PROFILE_UNKNOWN)}</span>`;
+  const toolsCell =
+    `<div class="tchips">${toolChips}${toolsInferred ? ` <span class="inf">inferred</span>` : ""}</div>`;
+
+  const targetModels = (input.subject.models ?? []).join(", ");
+
+  // §9.4.5 E1 — harness is UNAVAILABLE (never confabulated) in the trace-only case.
+  const harnessHas = typeof p?.harness === "string" && p.harness.trim().length > 0 && p.harness !== PROFILE_UNKNOWN;
+  const harnessCell = harnessHas
+    ? cell(p?.harness, "harness")
+    : `<span class="unk">${esc(PROFILE_UNKNOWN)}</span> <span class="recon-note">UNAVAILABLE — ${isTraceOnly ? "reconstructed, never confabulated" : "not supplied"}</span>`;
+
+  const rows: [string, string][] = [
+    ["purpose", cell(p?.purpose, "purpose")],
+    ["tools", toolsCell],
+    ["skill", cell(p?.skill, "skill")],
+    ["scope", cell(p?.scope, "scope")],
+    ["harness", harnessCell],
+    ["source", cell(input.subject.source)],
+    ["org", cell(input.subject.org)],
+    ["target model(s)", cell(targetModels)],
+  ];
+  const grid = rows.map(([k, v]) => `<div class="hk">${esc(k)}</div><div class="hv">${v}</div>`).join("");
+
+  // version + entityType + code-access provenance ride in the hero header.
+  const verChip = p?.version ? `<span class="hchip">v ${esc(p.version)}</span>` : "";
+  const typeChip = `<span class="hchip type">${cell(p?.entityType, "entityType")}</span>`;
+  // §9.4.5 E1 — code-access chip: GIVEN reads green; trace-only reads the explicit
+  // "NONE · trace-only" so a reader never mistakes a reconstructed profile for a known one.
+  const accessChip = isTraceOnly
+    ? `<span class="hchip access recon">code-access: NONE · trace-only</span>`
+    : `<span class="hchip access given">code-access: given</span>`;
+
+  // system prompt — present when GIVEN (code access) OR RECONSTRUCTED from the traces
+  // (UI-14). `inferred.has("systemPrompt")` distinguishes the two for the provenance tag.
+  const sysHas = typeof p?.systemPrompt === "string" && p.systemPrompt.trim().length > 0;
+  const sysReconstructed = inferred.has("systemPrompt");
+
+  // §9.4.5 E1 — the trace-reconstructed banner: only when trace-only. States the N traces
+  // the identity/tools were inferred from. The system prompt is now RECONSTRUCTED from the
+  // batch when an LLM call carried it (UI-14); only harness stays UNAVAILABLE in that case.
+  const reconLine = isTraceOnly
+    ? `<div class="hrecon">⟲ TRACE-RECONSTRUCTED · inferred from ${esc(nTraces)} trace(s) — identity + tools${sysHas ? " + system-prompt" : ""} observed from the batch; harness${sysHas ? "" : " + system-prompt"} UNAVAILABLE (never confabulated).</div>`
+    : "";
+
+  // system prompt — COLLAPSED (it can be tens of KB); carries a provenance tag —
+  // "RECONSTRUCTED · from traces" when derived from a GENERATION message list, or
+  // "given" under code access. Marked UNAVAILABLE only when genuinely not found
+  // (never confabulated).
+  const sysProvTag = sysReconstructed
+    ? `<span class="hsys-prov recon">RECONSTRUCTED · from traces</span>`
+    : `<span class="hsys-prov given">given · code access</span>`;
+  const sysPrompt = sysHas
+    ? `<details class="hsys"><summary>system prompt ${sysProvTag}</summary><pre>${esc(p?.systemPrompt)}</pre></details>`
+    : `<div class="hsys-na">system prompt — <span class="unk">${esc(PROFILE_UNKNOWN)}</span> · UNAVAILABLE (${isTraceOnly ? "reconstructed, no code access — never confabulated" : "not supplied"})</div>`;
+
+  // honesty footer: which fields are inferred (reconstructed), not given.
+  const infList = [...inferred];
+  const honesty =
+    infList.length > 0
+      ? `<div class="hhonesty"><span class="inf">inferred</span> ${esc(infList.join(" · "))} — reconstructed from the trace batch, not given. The rest is GIVEN (code/metadata access).</div>`
+      : `<div class="hhonesty"><span class="inf given">all given</span> every field supplied with code/metadata access — nothing reconstructed.</div>`;
+
+  return (
+    `<div class="hero">` +
+    `<div class="hero-top"><div class="hname">${esc(name)}</div>${typeChip}${verChip}${accessChip}</div>` +
+    reconLine +
+    `<div class="hero-grid">${grid}</div>` +
+    sysPrompt +
+    honesty +
+    `</div>`
+  );
+}
+
+/**
+ * Overview-redesign — the thin PROVENANCE meta-strip under the header: the RUN config
+ * (run-id · date · source · JUDGE substrate · pinned judge model · temp-0 · C-PIN).
+ * The JUDGE substrate + model are DISTINCT from the TARGET model under eval. Defaults
+ * that are ALWAYS true for this engine (agent-dispatch host runtime · temp 0 · C-PIN)
+ * fill in when no run-config is threaded; a genuinely-unavailable field is MARKED. PURE.
+ */
+function provenanceStrip(input: EvalReportInput): string {
+  const rc = input.runConfig ?? {};
+  const substrate = rc.judgeSubstrate ?? "agent-dispatch";
+  const temp = rc.temperature ?? 0;
+  const cPin = rc.cPin !== false; // C-PIN is the engine default
+  const date = rc.date ?? input.generatedAt;
+  const source = rc.source ?? input.subject.source;
+  const unk = `<span class="unk">${esc(PROFILE_UNKNOWN)}</span>`;
+  const v = (val: string | undefined): string =>
+    typeof val === "string" && val.trim().length > 0 ? esc(val) : unk;
+  const item = (label: string, body: string): string =>
+    `<span class="ps-item"><span class="ps-k">${esc(label)}</span><span class="ps-v">${body}</span></span>`;
+  return (
+    `<div class="provstrip">` +
+    item("run-id", v(rc.runId)) +
+    item("date", v(date)) +
+    item("source", v(source)) +
+    item("judge substrate", esc(substrate)) +
+    item("judge model", v(rc.judgeModel)) +
+    item("temp", esc(`temp ${temp}`)) +
+    `<span class="ps-item"><span class="ps-badge ${cPin ? "on" : "off"}">${cPin ? "C-PIN" : "no C-PIN"}</span></span>` +
+    `</div>`
+  );
+}
+
+/**
+ * Overview-redesign — the segmented COVERAGE FUNNEL: ingested → triaged → judged →
+ * outcomes. INCOMPLETE is a FIRST-CLASS outcome (a distinct segment from indeterminate:
+ * one is "trace too truncated to judge", the other is "judged but not certain"). Counts
+ * read from `coverage.{triaged,judged,byVerdict}` + the source-map trace count. PURE.
+ */
+function coverageFunnel(input: EvalReportInput): string {
+  const cov = input.coverage ?? {};
+  const bv = cov.byVerdict ?? {};
+  const judged = cov.judged ?? input.ledger?.length ?? 0;
+  const triaged = cov.triaged ?? judged;
+  const ingested = input.sourceMap?.traceCount ?? triaged;
+  const stage = (label: string, n: number, note: string): string =>
+    `<div class="fstage"><div class="fn">${esc(n)}</div><div class="fl">${esc(label)}</div><div class="fnote">${esc(note)}</div></div>`;
+  const arrow = `<div class="farrow">›</div>`;
+  // the OUTCOME segment — four FIRST-CLASS buckets, INCOMPLETE separate from indeterminate.
+  const outcomes: [string, number, string][] = [
+    ["pass", bv.PASS ?? 0, "pass"],
+    ["fail", bv.FAIL ?? 0, "fail"],
+    ["indeterminate", bv.INDETERMINATE ?? 0, "indet"],
+    ["incomplete", bv.INCOMPLETE ?? 0, "inc"],
+  ];
+  const outcomePills = outcomes
+    .map(([l, n, cls]) => `<div class="fpill ${cls}"><b>${esc(n)}</b> ${esc(l)}</div>`)
+    .join("");
+  return (
+    `<div class="funnel">` +
+    stage("Ingested", ingested, "traces pulled from source") +
+    arrow +
+    stage("Triaged", triaged, "tier-0 + fidelity gate") +
+    arrow +
+    stage("Judged", judged, "per-criterion verdicts") +
+    arrow +
+    `<div class="fstage outc"><div class="fl">Outcomes</div><div class="foutc">${outcomePills}</div></div>` +
+    `</div>`
+  );
+}
+
+/**
+ * §9.4.5 E3 — the eval-HEALTH temporal heatmap: CORRECTNESS over time (NOT latency —
+ * latency belongs to diagnostics). Each bucket is an hour window; its COLOUR is the
+ * pass-rate over the trajectories judged in that window (fail-clustering pops as a red
+ * window → a deploy/incident regression signal), and its NUMBER is the count of
+ * trajectories judged in the bucket. Computed from the ledger verdicts + the per-trace
+ * timestamps (E3 data path). When NO trajectory carries a timestamp, the heatmap renders
+ * its STRUCTURE with a single data-pending cell + an explicit note (never faked). PURE.
+ */
+function evalHealthHeatmap(input: EvalReportInput): string {
+  const ledger = input.ledger ?? [];
+  const timed = ledger.filter((r) => typeof r.timestamp === "string" && (r.timestamp as string).length >= 13);
+  const head = `<h3>eval-health over time — correctness, not latency</h3>`;
+
+  if (timed.length === 0) {
+    // structure + data-pending: the wiring is live, the batch just lacks timestamps.
+    return (
+      head +
+      `<div class="ehm" data-pending="1">` +
+      `<div class="ehm-cell skip" title="awaiting timestamped traces">·</div>` +
+      `</div>` +
+      `<div class="ehm-pending">⏲ data-pending — no per-trace timestamps in this batch. Correctness-over-time wiring is active; it renders real pass-rate windows as soon as the source carries trace timestamps.</div>`
+    );
+  }
+
+  // bucket by hour window (ISO "YYYY-MM-DDTHH"), deterministic chronological order.
+  const buckets = new Map<string, { judged: number; pass: number; fail: number }>();
+  for (const r of timed) {
+    const key = (r.timestamp as string).slice(0, 13); // YYYY-MM-DDTHH
+    const b = buckets.get(key) ?? { judged: 0, pass: 0, fail: 0 };
+    b.judged += 1;
+    const v = r.verdict.toUpperCase();
+    if (v === "PASS") b.pass += 1;
+    else if (v === "FAIL") b.fail += 1; // INCOMPLETE/INDETERMINATE count as judged, not in the rate
+    buckets.set(key, b);
+  }
+  const keys = [...buckets.keys()].sort();
+  const cells = keys
+    .map((k) => {
+      const b = buckets.get(k)!;
+      const denom = b.pass + b.fail;
+      const rate = denom > 0 ? Math.round((100 * b.pass) / denom) : -1;
+      const cls = rate < 0 ? "skip" : rate < 80 ? "fail" : rate < 95 ? "indet" : "pass";
+      const hour = k.slice(11, 13);
+      const day = k.slice(5, 10);
+      const title = `${esc(k)}:00 — ${b.judged} judged · ${b.pass} pass / ${b.fail} fail${rate >= 0 ? ` · ${rate}% pass-rate` : " · rate n/a"}`;
+      return (
+        `<div class="ehm-col">` +
+        `<div class="ehm-cell ${cls}" title="${title}">${esc(b.judged)}</div>` +
+        `<div class="ehm-lab">${esc(day)}<br>${esc(hour)}h</div>` +
+        `</div>`
+      );
+    })
+    .join("");
+  return (
+    head +
+    `<div class="ehm-legend">colour = pass-rate per window (<span class="sw fail"></span>&lt;80% &nbsp;<span class="sw indet"></span>&lt;95% &nbsp;<span class="sw pass"></span>≥95%) · number = trajectories judged</div>` +
+    `<div class="ehm">${cells}</div>`
+  );
+}
+
+function overviewTab(input: EvalReportInput): string {
+  const g = input.scorecard.gate;
+  const sm = input.sourceMap;
+  const cov = input.coverage ?? {};
+  const runVerdict = g.runVerdict ?? (g.passed ? "pass" : "fail");
+  const indetCount = g.indeterminateBy?.length ?? 0;
+  const verdictClass = runVerdict === "pass" ? "pass" : runVerdict === "incomplete" ? "skip" : "fail";
+  const verdictText =
+    runVerdict === "pass"
+      ? "GATE PASS — no CRIT/HIGH criterion failed or was indeterminate."
+      : runVerdict === "incomplete"
+        ? `GATE INCOMPLETE — ${indetCount} CRIT/HIGH criterion/criteria indeterminate (${(g.indeterminateBy ?? []).map((x) => esc(x.criterionId)).join(", ")}). No CRIT/HIGH fail, but the run cannot be certified — NOT a pass.`
+        : `GATE FAIL — ${g.gatedBy.length} gating criterion/criteria failed (${g.gatedBy.map((x) => esc(x.criterionId)).join(", ")}).`;
+
+  // 6-tile big-stat row (theme.css .big-stat .s/.v/.l) — status-acuity: the OUTCOME
+  // tiles (pass/fail/indeterminate) carry a status tint + left-accent so the eye
+  // separates outcomes from the neutral COUNT tiles (criteria/trajectories/traces).
+  const tileRows: [string, string, string][] = [
+    ["Criteria", String(g.total), "count"],
+    ["Pass", String(g.passCount), "pass"],
+    ["Fail", String(g.gatedBy.length), "fail"],
+    ["Indeterminate", String(indetCount), "indet"],
+    ["Trajectories", String(input.ledger?.length ?? 0), "count"],
+    ["Traces (coverage)", String(sm?.traceCount ?? cov.judged ?? 0), "count"],
+  ];
+  const tiles = tileRows
+    .map(([l, v, kind]) => `<div class="s ${kind}"><div class="v">${esc(v)}</div><div class="l">${esc(l)}</div></div>`)
+    .join("");
+
+  const coverageNote = cov.gapNote
+    ? `<div class="note"><span class="tag">★ COVERAGE CONTRACT</span>&nbsp;${esc(cov.gapNote)}</div>`
+    : sm
+      ? `<div class="note"><span class="tag">★ COVERAGE CONTRACT</span>&nbsp;${esc(sm.traceCount)} trace(s) · ${esc(sm.nullOutputTraces)} with null trace-output (read from GENERATION, SV-1) · platform <code>${esc(sm.platform)}</code>.</div>`
+      : "";
+
+  // gating-criteria table — pass-rate over applicable (R4: statement + provenance + hover)
+  const gating = input.gatingCriteria ?? [];
+  const gatingRows =
+    gating.length > 0
+      ? gating
+          .map((d) => {
+            const c = criterionById(input, d.criterion);
+            return (
+              `<tr><td class="cn"><b class="mono">${esc(d.criterion)}</b> ${critHover(c)} ${provChip(c)} ${methodChip(c)}` +
+              `<div class="cstmt">${esc(critStatement(c))}</div></td>` +
+              `<td>${esc(d.applicable)}</td><td>${esc(d.fail)}</td>` +
+              `<td style="color:${rateCol(d.passRateOverApplicable)};font-weight:600">${esc(d.passRateOverApplicable)}%</td></tr>`
+            );
+          })
+          .join("")
+      : `<tr><td colspan="4" style="color:var(--dim)">no gating criteria</td></tr>`;
+  const gatingTable =
+    `<h3>gating criteria — pass-rate over applicable</h3>` +
+    `<table><thead><tr><th>criterion (statement · provenance)</th><th>applicable</th><th>fail</th><th>pass-rate</th></tr></thead><tbody>${gatingRows}</tbody></table>`;
+
+  // top-findings teaser
+  const teaser = (input.topFindings ?? [])
+    .slice(0, 5)
+    .map(
+      (fd) =>
+        `<div class="teaser-row"><span class="sev ${sevCls(fd.severity)}">${esc((fd.severity ?? "med").toUpperCase())}</span>` +
+        `<b class="mono">${esc(fd.criterion)}</b>` +
+        `<span class="teaser-num">${esc(fd.failCount)}/${esc(fd.applicable)} fail · ${esc(fd.prevalencePctOverApplicable)}%</span></div>`,
+    )
+    .join("");
+  const teaserBlock = (input.topFindings ?? []).length > 0 ? `<h3>top findings</h3>${teaser}` : "";
+
+  return (
+    `<h2>① Overview</h2><div class="sub">subject identity · provenance · coverage · gate · top findings</div>` +
+    tabDesc("What this tab shows: WHO the subject is (entity hero), HOW the run was judged (provenance strip — judge substrate + pinned model, distinct from the target), the coverage funnel (ingested → triaged → judged → outcomes, INCOMPLETE first-class), the GATE verdict, the gating criteria with their pass-rate over the applicable denominator, and a teaser of the top findings.") +
+    entityHero(input) +
+    provenanceStrip(input) +
+    coverageFunnel(input) +
+    `<div class="verdict ${verdictClass}"><strong>${esc(verdictText)}</strong></div>` +
+    verdictLegend() +
+    `<div class="big-stat">${tiles}</div>` +
+    coverageNote +
+    evalHealthHeatmap(input) +
+    gatingTable +
+    teaserBlock
+  );
+}
+
+// ── §2 Trajectory · Judge Behaviour (ledger + drill) ─────────────────────────
+
+function trajectoryTab(input: EvalReportInput): string {
+  const ledger = input.ledger ?? [];
+  if (ledger.length === 0) {
+    return (
+      `<h2>② Trajectory · Judge Behaviour</h2>` +
+      `<div class="sub">per-trace judge ledger — no per-trajectory verdict files supplied for this render.</div>`
+    );
+  }
+  const walks = ledger.filter((r) => r.judgeSteps && r.judgeSteps.length > 0).length;
+  return (
+    `<h2>② Trajectory · Judge Behaviour</h2>` +
+    tabDesc("What this tab shows: the Target-Agent trajectory ‖ Judge trajectory SIDE-BY-SIDE. Click a row to drill in — the agent's steps render on the LEFT lane, the judge's reasoning walk (anchored to each step) on the RIGHT, with the judge's gather-context, expected-trajectory and root localization.") +
+    `<div class="sub">all ${esc(ledger.length)} — click a row; ✦ = full agent‖judge walk (${esc(walks)} with judge_steps)</div>` +
+    `<div class="lfilter">` +
+    `<b class="on" data-flt="all">all</b><b data-flt="FAIL">fail</b><b data-flt="INDETERMINATE">indeterminate</b><b data-flt="PASS">pass</b>` +
+    `<input id="q" placeholder="search route / trace id…"><span class="sp" id="cnt"></span></div>` +
+    `<table class="ledger"><thead><tr><th>trace</th><th>route</th><th>verdict</th><th>resolution</th><th>pass</th><th>fail</th><th>indet</th><th>failing criteria</th></tr></thead><tbody id="lrows"></tbody></table>` +
+    `<div id="virt" class="mono virt"></div>` +
+    `<div id="drill"></div>`
+  );
+}
+
+// ── §3 Eval Scorecard (heatmap + subcards + calibration) ─────────────────────
+
+function heatmapHtml(input: EvalReportInput): string {
+  const cohorts = input.cohorts;
+  if (!cohorts || Object.keys(cohorts.counts).length === 0) {
+    return `<div class="sub">criteria × route cohort — no per-trajectory cohort data for this render.</div>`;
+  }
+  const cohortKeys = Object.entries(cohorts.counts)
+    .sort((a, b) => b[1] - a[1])
+    .slice(0, 7)
+    .map(([c]) => c);
+  const head =
+    `<span class="lab"></span>` +
+    cohortKeys.map((c) => `<span class="colh">${esc(c)}<br>${esc(cohorts.counts[c] ?? 0)}</span>`).join("");
+  const cids = Object.keys(cohorts.matrix).slice(0, 14);
+  const rows = cids.map((cid) => {
+    let cells = `<span class="lab" title="${esc(cid)}">${esc(cid.slice(0, 26))}</span>`;
+    for (const co of cohortKeys) {
+      const cell = cohorts.matrix[cid]?.[co];
+      if (!cell) {
+        cells += `<span class="cell skip"></span>`;
+        continue;
+      }
+      const tot = cell.pass + cell.fail + cell.indeterminate;
+      if (tot === 0) {
+        cells += `<span class="cell skip"></span>`;
+        continue;
+      }
+      const r = Math.round((100 * cell.pass) / tot);
+      const cls = r >= 95 ? "pass" : r >= 80 ? "indet" : "fail";
+      cells += `<span class="cell ${cls}" title="${esc(cid)} × ${esc(co)}: ${cell.pass}/${tot} pass">${r}</span>`;
+    }
+    return cells;
+  });
+  // UI-3 — wrap in a horizontal-scroll container so long cohort names / many cohorts
+  // never overflow the page; the grid columns use minmax(0,…) so a long unbreakable
+  // header word can't force the track wider than its share (it ellipsis-clips instead).
+  return (
+    `<div class="hm-scroll">` +
+    `<div class="hm" style="grid-template-columns:150px repeat(${cohortKeys.length},minmax(58px,1fr))">` +
+    head +
+    rows.join("") +
+    `</div>` +
+    `</div>`
+  );
+}
+
+/**
+ * UI-5 — the success-rate + key-metrics SUMMARY row under the §3 heatmap. The heatmap
+ * alone "dangled" into the subcards with no headline numbers; this anchors it with a
+ * compact metrics strip (criteria pass-rate · trajectory pass-rate · indeterminate ·
+ * incomplete · grounded%). All numbers derive from the scorecard gate + the folded
+ * coverage + judge-health — nothing fabricated; an empty source renders 0/—. PURE.
+ */
+function scorecardMetrics(input: EvalReportInput): string {
+  const g = input.scorecard.gate;
+  const bv = input.coverage?.byVerdict ?? {};
+  const judged = input.coverage?.judged ?? input.ledger?.length ?? 0;
+  const critPassRate = g.total > 0 ? Math.round((100 * g.passCount) / g.total) : 100;
+  const trajPass = bv.PASS ?? 0;
+  const trajRate = judged > 0 ? Math.round((100 * trajPass) / judged) : 0;
+  const grounded = input.judgeHealth?.groundedPct;
+  const tile = (label: string, value: string, cls: string, note: string): string =>
+    `<div class="scm ${cls}"><div class="scm-v">${esc(value)}</div><div class="scm-l">${esc(label)}</div><div class="scm-n">${esc(note)}</div></div>`;
+  return (
+    `<div class="sc-metrics">` +
+    tile("criteria pass-rate", `${critPassRate}%`, rateClass(critPassRate), `${g.passCount}/${g.total} criteria pass`) +
+    tile("trajectory pass-rate", `${trajRate}%`, rateClass(trajRate), `${trajPass}/${judged} trajectories pass`) +
+    tile("needs-evidence", String(bv.INDETERMINATE ?? 0), (bv.INDETERMINATE ?? 0) > 0 ? "rate-mid" : "rate-ok", "precondition present, undecided (N/A excluded)") +
+    tile("incomplete", String(bv.INCOMPLETE ?? 0), (bv.INCOMPLETE ?? 0) > 0 ? "rate-mid" : "rate-ok", "too truncated to judge") +
+    // UI-12-B — grounded is HONEST about capture: 0 / undefined (judge emitted 0
+    // structured refs) renders "capture-unavailable" (data-pending), NEVER a silent
+    // green 0%; a real >0 value is colour-coded by rate (low = critical).
+    (grounded === undefined || grounded === 0
+      ? tile("grounded", "capture-unavailable", "rate-na pending", "judge emitted 0 structured refs — capture-unavailable")
+      : tile("grounded", `${grounded}%`, rateClass(grounded), "verdicts citing evidence")) +
+    `</div>`
+  );
+}
+
+/**
+ * WS-3 follow-up — a compact per-criterion KEY rendered DIRECTLY UNDER the §3
+ * heatmap. The heatmap rows are raw technical criterion ids (e.g.
+ * `send-delivery-success`); this legend maps EVERY criterion on the scorecard axis —
+ * gating AND non-gating — to its one-line plain-language gloss (reusing the same
+ * `plainExplainer` the Findings cards use), so no scorecard criterion is left without
+ * "what it measures · why it matters". Covers ALL `input.criteria` (not the gating
+ * subset). PURE — empty criteria ⇒ empty string.
+ */
+function criterionLegend(input: EvalReportInput): string {
+  const crits = input.criteria ?? [];
+  if (crits.length === 0) return "";
+  const rows = crits
+    .map((c) => {
+      const sev = sevCls(c.severity);
+      return (
+        `<div class="plain plain-leg">` +
+        `<span class="pl-id mono">${esc(c.id)}</span>` +
+        `<span class="sev ${sev}">${esc(String(c.severity ?? "med").toUpperCase())}</span>` +
+        `<span class="pl-tx">${esc(plainExplainer(c))}</span>` +
+        `</div>`
+      );
+    })
+    .join("");
+  return (
+    `<div class="plain-legend">` +
+    `<div class="pl-h">what each criterion measures — plain-language key (all ${crits.length})</div>` +
+    rows +
+    `</div>`
+  );
+}
+
+function subcardsHtml(input: EvalReportInput): string {
+  const gating = input.gatingCriteria ?? [];
+  if (gating.length === 0) return `<div class="sub">no gating criteria to calibrate.</div>`;
+  return gating
+    .map((d) => {
+      const sev = sevCls(d.severity);
+      const c = criterionById(input, d.criterion);
+      // UI-6 — the card colour is driven by SUCCESS RATE (green-ish high · amber mid ·
+      // red-ish low), on a purplish base — NOT a uniform severity red. Severity stays
+      // legible as its own pill in the header.
+      const rate = rateClass(d.passRateOverApplicable);
+      // WS-5 — the indeterminate RESOLUTION split: N/A (precondition absent, excluded
+      // from the denominator) vs needs-evidence (precondition present, undecided). No
+      // bare "indeterminate".
+      const whyAb =
+        (d.indeterminate ?? 0) > 0 || (d.na ?? 0) > 0
+          ? `<div class="nest"><div class="nest-h">▾ indeterminate resolution</div><div class="row">` +
+            `<span class="rz na">N/A ${esc(d.na ?? 0)}</span> precondition absent — excluded from denominator · ` +
+            `<span class="rz ne">needs-evidence ${esc(d.indeterminate ?? 0)}</span> precondition present, undecided</div></div>`
+          : "";
+      return (
+        `<div class="subc ${rate}"><div class="subc-h"><span class="sev ${sev}">${esc((d.severity ?? "med").toUpperCase())}</span>` +
+        `<b>${esc(d.criterion)}</b> ${critHover(c)} ${provChip(c)} ${methodChip(c)}<span class="chip">gate</span>` +
+        `<span class="vp" style="margin-left:auto;color:${rateCol(d.passRateOverApplicable)};border-color:${rateCol(d.passRateOverApplicable)}">${esc(d.passRateOverApplicable)}%</span></div>` +
+        // OPERATOR-REQUESTED: the one-line plain-language "Measures / Why it matters" gloss
+        // INLINE in every gating subcard (not only in the §3 legend) — so reading a gating
+        // criterion's detail explains, in plain language, what it measures right there.
+        plainBanner(c) +
+        // UI-7 — the full criterion DEFINITION in-card (statement · pass condition ·
+        // dimension/severity/level · judged-from inputs · provenance · mined details).
+        `<div class="nest"><div class="nest-h">▾ criterion definition — what it is &amp; how it's judged</div>` +
+        criterionDefn(c, d.criterion) +
+        `</div>` +
+        `<div class="nest"><div class="nest-h">▾ grounding</div><div class="row"><span class="ref">obs:</span> ${esc(d.fail)}/${esc(d.applicable)} applicable fail · ${esc((d.denominatorNote ?? "").slice(0, 90))}</div></div>` +
+        `<div class="nest"><div class="nest-h">▾ verdict reasoning</div><div class="row">${esc((d.root ?? "all applicable pass").slice(0, 160))}</div></div>` +
+        whyAb +
+        // UI-8 — calibrate tags grouped for MUTUAL EXCLUSION (radio semantics): the
+        // keep/revise/retire group and the verify/eliminate group each allow ONE active
+        // selection (the client JS clears siblings within a `data-calgroup`).
+        `<div class="cal"><span class="l">calibrate</span>` +
+        `<span class="calgroup" data-calgroup="keep-revise-retire"><b>keep</b><b>revise</b><b>retire</b></span>` +
+        `<span style="color:var(--dim)">·</span>` +
+        `<span class="calgroup" data-calgroup="verify-eliminate"><b>verify</b><b>eliminate</b></span>` +
+        `</div></div>`
+      );
+    })
+    .join("");
+}
+
+/** WS-5 — a readable label for a chain next-action. */
+function nextActionLabel(a: string): string {
+  return a === "code-recheck"
+    ? "code-recheck"
+    : a === "2nd-judge"
+      ? "2nd-judge"
+      : a === "revise-criterion"
+        ? "revise-criterion"
+        : "HITL-spot-check";
+}
+
+/**
+ * WS-5 — the per-criterion INDETERMINATE resolution chain (§3). For EVERY criterion
+ * (gating + non-gating) it shows the denominator-honest split:
+ *   applicable · pass · fail · N/A (precondition absent) · needs-evidence (+ next-action)
+ * N/A is shown SEPARATELY and is NEVER in the applicable/fail denominator. A
+ * needs-evidence count carries its aggregated chain next-actions + the missing-signal
+ * reason — so no criterion ever shows a bare, actionless "indeterminate". PURE.
+ */
+function criterionResolutionTable(input: EvalReportInput): string {
+  const rows = input.criterionResolutions ?? [];
+  if (rows.length === 0) return "";
+  const body = rows
+    .map((r) => {
+      const sev = sevCls(r.severity);
+      const naCell =
+        r.na > 0
+          ? `<span class="rz na">N/A ${esc(r.na)}</span>` +
+            (r.naReason ? `<span class="rz-reason">precondition absent — ${esc(r.naReason.slice(0, 120))}</span>` : "")
+          : `<span class="rz na zero">N/A 0</span>`;
+      const neCell =
+        r.needsEvidence > 0
+          ? `<span class="rz ne">needs-evidence ${esc(r.needsEvidence)}</span>` +
+            r.nextActions
+              .map(
+                (a) =>
+                  `<span class="rz-act">→ ${esc(nextActionLabel(a.action))} ×${esc(a.count)}` +
+                  `<span class="rz-reason">${esc(a.reason.slice(0, 120))}</span></span>`,
+              )
+              .join("")
+          : `<span class="rz ne zero">needs-evidence 0</span>`;
+      return (
+        `<div class="rzrow ${sev}">` +
+        `<div class="rz-h"><span class="sev ${sev}">${esc(String(r.severity ?? "med").toUpperCase())}</span>` +
+        `<b class="mono">${esc(r.criterion)}</b>` +
+        `<span class="rz-rate" style="color:${rateCol(r.passRateOverApplicable)}">${esc(r.passRateOverApplicable)}% pass / applicable</span></div>` +
+        `<div class="rz-stats">` +
+        `<span class="rz applicable">applicable ${esc(r.applicable)}</span>` +
+        `<span class="rz pass">pass ${esc(r.pass)}</span>` +
+        `<span class="rz fail">fail ${esc(r.fail)}</span>` +
+        naCell +
+        neCell +
+        `</div></div>`
+      );
+    })
+    .join("");
+  return (
+    `<h3>indeterminate resolution chain — applicable · pass · fail · N/A (precondition absent) · needs-evidence</h3>` +
+    `<div class="sub">every indeterminate is RESOLVED by a deterministic precondition gate: <b>N/A</b> (the criterion's trigger/precondition was absent in the trace — DROPPED from the applicable denominator) vs <b>needs-evidence</b> (precondition present but undecided — carries a concrete next-action: code-recheck · 2nd-judge · revise-criterion · HITL-spot-check). No bare indeterminate.</div>` +
+    `<div class="rztable">${body}</div>`
+  );
+}
+
+function scorecardTab(input: EvalReportInput): string {
+  return (
+    `<h2>③ Eval Scorecard</h2>` +
+    tabDesc("What this tab shows: the DEFINED criteria × results. Each criterion carries its STATEMENT, its provenance (defined vs source-mined), and a hover with the full definition; the heatmap is criteria × route cohort, with nested subcards + inline calibration on the applicable-denominator-honest pass-rate.") +
+    `<div class="sub">cohort heatmap · nested subcards · inline calibration · GA-D2b applicable-denominator honest</div>` +
+    verdictLegend() +
+    `<h3>criteria × route cohort — cell = % pass over (pass+fail+indet)</h3>` +
+    heatmapHtml(input) +
+    // WS-3 follow-up — the per-criterion plain-language KEY directly under the heatmap,
+    // covering ALL criteria (gating + non-gating) so every technical heatmap id has a
+    // "what it measures · why it matters" gloss.
+    criterionLegend(input) +
+    // UI-5 — anchor the heatmap with a success-rate + key-metrics summary row (no dangle).
+    scorecardMetrics(input) +
+    `<h3>gating criteria — nested subcards + HITL calibration</h3>` +
+    subcardsHtml(input) +
+    // WS-5 — the INDETERMINATE resolution chain for ALL criteria (gating + non-gating):
+    // splits the single indeterminate bucket into N/A (precondition absent, excluded)
+    // vs needs-evidence (carries a next-action). interest-logging-integrity's 283 land here.
+    criterionResolutionTable(input)
+  );
+}
+
+// ── §4 Findings (verbatim evidence + judge chain + agree/revise/refute) ───────
+
+const KIND_ORDER: Record<string, number> = { context: 0, examine: 1, detect: 2, bind: 3, ground: 4, critique: 5, decide: 6, verify: 7 };
+
+/** One §4 finding from the top-findings roll; uses DR-2 / judge-walk for evidence. */
+function findingFromTop(input: EvalReportInput, fd: TopFinding): string {
+  const crit = criterionById(input, fd.criterion);
+  const sev = sevCls(fd.severity ?? crit?.severity);
+  const ex = fd.exampleTraceId ?? "";
+  const row = input.ledger?.find((r) => r.trajectoryId === ex);
+  const grounds = (row?.judgeSteps ?? []).filter((s) => s.ref);
+  // a judge-step ref may be a string OR a structured {obs,path,value} — coerce.
+  const refToStr = (rf: JudgeStep["ref"]): string =>
+    rf === undefined ? "" : typeof rf === "string" ? rf : [rf.obs, rf.path, rf.value].filter(Boolean).join(":");
+  const evItems =
+    grounds.length > 0
+      ? grounds
+          .slice(0, 4)
+          .map((s) => { const r = refToStr(s.ref); return `<li><span class="ref">${esc(r.split(":")[0])}</span> ${esc(r.split(":").slice(1).join(":").slice(0, 120))}</li>`; })
+          .join("")
+      : crit?.discovery && crit.discovery.evidence.refs.length > 0
+        ? crit.discovery.evidence.refs
+            .slice(0, 4)
+            .map((r) => `<li><span class="ref">${esc(r.obs)}</span> ${esc(((r.path ? r.path + ": " : "") + r.value).slice(0, 120))}</li>`)
+            .join("")
+        : `<li><span class="ref">trace</span> <code>${esc(ex || "(suite)")}</code> — search §2 for the judge walk</li>`;
+  const chain =
+    (row?.judgeSteps ?? []).length > 0
+      ? [...(row?.judgeSteps ?? [])]
+          .filter((s) => s.kind !== "context")
+          .sort((a, b) => (KIND_ORDER[a.kind] ?? 9) - (KIND_ORDER[b.kind] ?? 9))
+          .map((s) => `<span class="step">${esc(s.kind)}</span>`)
+          .join(" → ")
+      : "detect → ground → critique → decide → verify";
+  const link =
+    row?.judgeSteps && row.judgeSteps.length > 0 ? ` <span class="lnk" data-drill="${esc(ex)}">[open §2 side-by-side ▸]</span>` : "";
+  const gate = sev === "crit" || sev === "high" ? `<span class="gatetag">GATING</span>` : "";
+  const prev = fd.prevalencePctOverApplicable;
+  const assumptionsLine =
+    crit?.discovery && crit.discovery.assumptions.length > 0
+      ? crit.discovery.assumptions.map((a) => `${esc(a.text)} <span class="badge b-skip">${esc(a.status)}</span>`).join(" · ")
+      : `<span class="none">grounded</span> — judge-health surfaced (see §5)`;
+  const whyProblem = crit?.discovery ? `<div class="k">why a<br>problem</div><div class="v">${esc(crit.discovery.why_problem)}</div>` : "";
+  return (
+    `<div class="find ${sev === "crit" || sev === "high" ? "gate" : ""}">` +
+    `<div class="find-h"><span class="sev ${sev}">${esc((fd.severity ?? "med").toUpperCase())}</span><b>${esc(fd.criterion)}</b>${gate}` +
+    `<span class="verd fail" style="margin-left:auto">${esc(fd.failCount)}/${esc(fd.applicable)} fail · ${esc(prev)}%</span></div>` +
+    // WS-3 — plain-language gloss (what · why) above the technical claim/evidence grid.
+    plainBanner(crit) +
+    `<div class="fg">` +
+    `<div class="k">claim</div><div class="v">Criterion <code>${esc(fd.criterion)}</code> fails on ${esc(fd.failCount)} of ${esc(fd.applicable)} applicable trajectories.</div>` +
+    whyProblem +
+    `<div class="k">evidence<br>(verbatim)</div><div class="v" data-pii="evidence"><ul>${evItems}</ul></div>` +
+    `<div class="k">judge<br>reasoning</div><div class="v reason">${chain}${link}</div>` +
+    `<div class="k">root</div><div class="v">${esc(fd.root ?? "")}</div>` +
+    `<div class="k">prevalence</div><div class="v"><b>${esc(prev)}%</b> over applicable (${esc(fd.applicable)}) <span class="prev"><i style="width:${Math.min(prev, 100)}%"></i></span> <span class="mono na-note">na excluded (GA-D2b)</span></div>` +
+    `<div class="k">assumptions</div><div class="v">${assumptionsLine}</div>` +
+    `</div>` +
+    `<div class="areview"><span class="l">⊕ alignment review:</span><b class="agree">✓ agree</b><b>~ revise</b><b class="refute">✗ refute</b><input placeholder="reviewer note…"></div>` +
+    `</div>`
+  );
+}
+
+/** Fallback DR-2 finding card (no ledger / no top-findings — mined criteria only). */
+function findingFromCriterion(c: ReportCriterion, verdict: CriterionVerdict | undefined): string {
+  const sev = sevCls(c.severity);
+  const d = c.discovery;
+  if (!d) {
+    return (
+      `<div class="find"><div class="find-h"><span class="sev ${sev}">${esc(c.severity)}</span><b>${esc(c.id)}</b>` +
+      `${verdict ? `<span class="verd ${verdictCls(resultToken(verdict.result))}" style="margin-left:auto">${esc(resultToken(verdict.result))}</span>` : ""}</div>` +
+      plainBanner(c) +
+      `<div class="fg"><div class="k">statement</div><div class="v">${esc(c.statement)}</div>` +
+      `${verdict ? `<div class="k">critique</div><div class="v">${esc(verdict.critique)}</div>` : ""}</div></div>`
+    );
+  }
+  const ev = d.evidence;
+  const refs = ev.refs.length > 0 ? `<li><span class="ref">refs</span> <code>${esc(refsText(ev.refs))}</code></li>` : "";
+  const assumptions = d.assumptions.map((a) => `${esc(a.text)} <span class="badge b-skip">${esc(a.status)}</span>`).join(" · ");
+  return (
+    `<div class="find ${sev === "crit" || sev === "high" ? "gate" : ""}">` +
+    `<div class="find-h"><span class="sev ${sev}">${esc(c.severity)}</span><b>${esc(c.id)}</b>` +
+    `${verdict ? `<span class="verd ${verdictCls(resultToken(verdict.result))}" style="margin-left:auto">${esc(resultToken(verdict.result))}</span>` : ""}</div>` +
+    plainBanner(c) +
+    `<div class="fg">` +
+    `<div class="k">targets</div><div class="v">${esc(d.targets)}</div>` +
+    `<div class="k">why a<br>problem</div><div class="v">${esc(d.why_problem)}</div>` +
+    `<div class="k">evidence<br>(verbatim)</div><div class="v" data-pii="evidence">grounding <strong>${esc(ev.grounding)}</strong> · seen-in-traces ${esc(ev.seen_in_traces)} · prevalence <strong>${esc(ev.prevalence)}</strong><ul>${refs}</ul></div>` +
+    `<div class="k">judge<br>reasoning</div><div class="v reason">${esc(d.reasoning)}</div>` +
+    `<div class="k">assumptions</div><div class="v">${assumptions}</div>` +
+    `</div>` +
+    `<div class="areview"><span class="l">⊕ alignment review:</span><b class="agree">✓ agree</b><b>~ revise</b><b class="refute">✗ refute</b><input placeholder="reviewer note…"></div>` +
+    `</div>`
+  );
+}
+
+function findingsTab(input: EvalReportInput): string {
+  const top = input.topFindings ?? [];
+  let body: string;
+  if (top.length > 0) {
+    body = top.map((fd) => findingFromTop(input, fd)).join("");
+  } else {
+    // no ledger-derived findings — fall back to DR-2 cards for failing/at-risk criteria.
+    const verdictById = new Map(input.verdicts.map((v) => [v.criterionId, v]));
+    const failing = input.criteria.filter((c) => {
+      const v = verdictById.get(c.id);
+      return v ? v.result !== OutcomeVerdict.Pass : c.discovery !== undefined;
+    });
+    const pool = failing.length > 0 ? failing : input.criteria;
+    body = pool.map((c) => findingFromCriterion(c, verdictById.get(c.id))).join("");
+  }
+  // §9.4.4 R3/M5 — DETECTED-but-unmatched flags, CLEARLY SEPARATED from the
+  // fails-on-existing-criteria above. These are detections the judge flagged with NO
+  // matching criterion — routed to *discover-evals, never minted into evals mid-judging.
+  const flags = input.detectedFlags ?? [];
+  const refToStr = (rf: DetectedFlag["ref"]): string =>
+    rf === undefined ? "" : typeof rf === "string" ? rf : [rf.obs, rf.path, rf.value].filter(Boolean).join(":");
+  const detectedSection =
+    flags.length > 0
+      ? `<h3 class="detected-h">⚠ Detected — unmatched (no defined criterion) <span class="badge b-skip">NOT SCORED</span></h3>` +
+        `<div class="sub">M5 JUDGE-WHAT-IS: these behaviours were DETECTED + FLAGGED but have no matching criterion — routed to <code>*discover-evals</code> / <code>*build-dataset</code>, never scored or minted into evals mid-judging.</div>` +
+        flags
+          .map((fl) => {
+            const r = refToStr(fl.ref);
+            return (
+              `<div class="detected"><div class="detected-top"><span class="dkind">${esc(fl.kind)}</span>` +
+              `<b>${esc(fl.detection)}</b><span class="dtrace mono">${esc(fl.trajectoryId)}${fl.anchor !== undefined ? " · step " + esc(fl.anchor) : ""}</span></div>` +
+              (r ? `<div class="detected-ev mono"><span class="ref">${esc(r.split(":")[0])}</span> ${esc(r.split(":").slice(1).join(":"))}</div>` : "") +
+              `</div>`
+            );
+          })
+          .join("")
+      : `<h3 class="detected-h">⚠ Detected — unmatched (no defined criterion)</h3><div class="sub">none — every detected behaviour had a matching defined criterion this run.</div>`;
+
+  return (
+    `<h2>④ Findings</h2>` +
+    tabDesc("What this tab shows: TWO clearly separated groups — (1) FAILURES on the existing DEFINED criteria (with verbatim evidence + judge reasoning chain), and (2) DETECTED-but-unmatched flags the judge raised with no matching criterion (routed to discover, never scored).") +
+    `<div class="sub">alignment-auditable — verbatim evidence + judge reasoning chain</div>` +
+    `<h3 class="fails-h">✗ Failures on defined criteria</h3>` +
+    body +
+    detectedSection +
+    `<div class="note"><span class="tag">★ AUDITABLE</span>&nbsp;Each finding: verbatim evidence (obs/diff refs from the example trace's judge walk), the judge reasoning chain, prevalence over the applicable denominator, agree/revise/refute to overturn. Open §2 to see the full side-by-side.</div>`
+  );
+}
+
+// ── §5 Self-Eval · Calibration [INTERNAL — stripped on publish] ───────────────
+
+const PHASE_LENS_ORDER = ["gather", "expect", "context", "examine", "detect", "bind", "ground", "critique", "decide", "verify", "localize"];
+
+/** §9.4.4 R2/M1 — the subject-profile card for the INTERNAL calibration lens. */
+function subjectProfileCard(p: SubjectProfile | undefined): string {
+  if (p === undefined) {
+    return `<div class="sub">no subject profile supplied for this run — the judge reconstructed identity at reason-time.</div>`;
+  }
+  const inferred = new Set(p.inferredFields ?? []);
+  const tag = (field: string): string => (inferred.has(field) ? ` <span class="badge b-skip">inferred</span>` : "");
+  const rows: [string, string, string][] = [
+    ["identity", p.identity, "identity"],
+    ["purpose", p.purpose, "purpose"],
+    ["scope", p.scope, "scope"],
+    ["skill", p.skill ?? "—", "skill"],
+    ["tools", (p.tools ?? []).join(", ") || "—", "tools"],
+    ["harness", p.harness, "harness"],
+    ["version", p.version ?? "—", "version"],
+  ];
+  const prov = `<span class="prov ${p.provenance === "given" ? "defined" : "source"}">${esc(p.provenance)}</span>`;
+  const body = rows
+    .map(([k, v, field]) => `<div class="pk">${esc(k)}${tag(field)}</div><div class="pv">${esc(v)}</div>`)
+    .join("");
+  return (
+    `<div class="sub">M1 — who the judge understood the agent to be (${prov}; <span class="badge b-skip">inferred</span> = reconstructed, not given · harness <code>unknown</code> is MARKED, never confabulated).</div>` +
+    `<div class="profile-grid">${body}</div>`
+  );
+}
+
+/** §9.4.4 R2/M4 — the PHASE-BLOCK lens: the judge's reasoning grouped by DAG node
+ *  across the whole run — the calibration surface the builder reads like an AI eng. */
+function phaseLens(input: EvalReportInput): string {
+  const ledger = input.ledger ?? [];
+  const byPhase = new Map<string, { count: number; samples: string[] }>();
+  for (const r of ledger) {
+    for (const s of r.judgeSteps ?? []) {
+      const k = s.kind;
+      const entry = byPhase.get(k) ?? { count: 0, samples: [] };
+      entry.count += 1;
+      if (s.text && entry.samples.length < 3) entry.samples.push(s.text);
+      byPhase.set(k, entry);
+    }
+  }
+  if (byPhase.size === 0) {
+    return `<div class="sub">no judge_steps emitted this run — phase-block lens unavailable (scorecard-only render).</div>`;
+  }
+  const orderedKinds = [
+    ...PHASE_LENS_ORDER.filter((k) => byPhase.has(k)),
+    ...[...byPhase.keys()].filter((k) => !PHASE_LENS_ORDER.includes(k)),
+  ];
+  return (
+    `<div class="phase-lens">` +
+    orderedKinds
+      .map((k) => {
+        const e = byPhase.get(k)!;
+        const samples = e.samples.map((t) => `<div class="ps-row">${esc(t.slice(0, 140))}</div>`).join("");
+        return (
+          `<div class="phase-blk"><div class="phase-h"><span class="k ${esc(k)}">${esc(k)}</span><span class="phase-n">${esc(e.count)} step(s)</span></div>${samples}</div>`
+        );
+      })
+      .join("") +
+    `</div>`
+  );
+}
+
+/** §9.4.4 R2/M2+M3 — an example node-0 understanding (train-of-thought) + node-0.5
+ *  expected-trajectory, from the first trajectory that carries them. */
+function reasoningExample(input: EvalReportInput): string {
+  const ledger = input.ledger ?? [];
+  const withU = ledger.find((r) => r.understanding !== undefined);
+  const withE = ledger.find((r) => r.expectedTrajectory && r.expectedTrajectory.length > 0);
+  if (withU === undefined && withE === undefined) {
+    return `<div class="sub">no node-0 understanding / node-0.5 expected-trajectory emitted this run.</div>`;
+  }
+  let out = "";
+  if (withU?.understanding !== undefined) {
+    const u = withU.understanding;
+    const given = (u.given ?? []).map((g) => `<li>${esc(g)}</li>`).join("");
+    const inferred = (u.inferred ?? []).map((g) => `<li>${esc(g)}</li>`).join("");
+    out +=
+      `<div class="re-blk"><div class="re-h">node-0 GATHER — understanding (M2) · <span class="mono">${esc(withU.trajectoryId)}</span></div>` +
+      `<div class="re-rephrase">“${esc(u.rephrase)}”</div>` +
+      (given ? `<div class="re-sub">given</div><ul class="re-ul">${given}</ul>` : "") +
+      (inferred ? `<div class="re-sub">inferred</div><ul class="re-ul">${inferred}</ul>` : "") +
+      `</div>`;
+  }
+  if (withE?.expectedTrajectory !== undefined) {
+    const steps = withE.expectedTrajectory
+      .map((s: ExpectedStep, i: number) => `<li><b>${esc(s.step ?? i + 1)}.</b> ${esc(s.expected)}${s.rationale ? ` <span class="re-rat">— ${esc(s.rationale)}</span>` : ""}</li>`)
+      .join("");
+    out +=
+      `<div class="re-blk"><div class="re-h">node-0.5 EXPECTED-TRAJECTORY (M3) · <span class="mono">${esc(withE.trajectoryId)}</span></div>` +
+      `<div class="re-sub">how the target SHOULD have acted — built BEFORE examine</div><ol class="re-ul">${steps}</ol></div>`;
+  }
+  return out;
+}
+
+// (UI-12-B groundedHealthCell removed in the §5 self-eval rebuild (WS-3): the per-cell
+//  judge-health grounding readout was superseded by the aggregate Reference Integrity
+//  panel (§5.1), which keeps the same no-false-green honesty. `rateClass` stays — used
+//  by the §3 pass-rate tiles + the per-criterion rows.)
+
+/**
+ * WS-4 — one parsed routed-failure. The EV-051 handover encodes each routed item as a
+ * deterministic acceptance string (`routeFailures`):
+ *   `<criterionId> [<severity>/<flag>] FAILED on trace <traceId>: <critique>`
+ * so the §5 handover is reconstructed STRUCTURALLY (WHAT · WHY · WHERE · TARGET) from
+ * it — no new data, just a parse. A string that doesn't match the format degrades to
+ * `{criterionId: <raw>}` (rendered as WHAT, never dropped).
+ */
+interface RoutedItem {
+  criterionId: string;
+  severity: string;
+  flag: string;
+  traceId: string;
+  critique: string;
+  raw: string;
+}
+function parseRoutedCriterion(c: string): RoutedItem {
+  const m = c.match(/^(.+?)\s+\[(.+?)\/(.+?)\]\s+FAILED on trace\s+(.+?):\s+([\s\S]*)$/);
+  if (m) return { criterionId: m[1], severity: m[2], flag: m[3], traceId: m[4], critique: m[5].trim(), raw: c };
+  return { criterionId: c, severity: "", flag: "", traceId: "", critique: "", raw: c };
+}
+
+/**
+ * WS-4 — the refined EV-051 "Routed to diagnostics" handover. Restructures the raw
+ * acceptance-criteria dump into a clean, COPY-PASTEABLE handover block: a meta strip
+ * (subject · target stage · escalation · produced-by) + one card per routed item with
+ * WHAT (the criterion/finding) · WHY (the judge critique) · WHERE (the locus — trace +
+ * dimension) · TARGET (diagnostics). Each card carries a per-item "copy" button (this
+ * item as markdown) and the block ends with a "copy full handover as markdown" button —
+ * the operator hands the bundle straight to diagnostics. PURE; deterministic markdown.
+ */
+function routedHandover(input: EvalReportInput): string {
+  const h = input.handover;
+  if (h === null || h === undefined) {
+    return `<div class="sub">No failures routed — nothing handed to diagnostics this run (judge-only: the evaluator flags + routes, it never fixes).</div>`;
+  }
+  const subjName = h.subject?.name ?? input.subject.name;
+  const target = h.intent?.command ?? "*diagnose";
+  const items = h.acceptance.criteria.map(parseRoutedCriterion);
+
+  // per-item markdown (one routed finding, copy-pasteable into diagnostics).
+  const itemMd = (p: RoutedItem, crit: ReportCriterion | undefined): string =>
+    [
+      `## ${p.criterionId}${p.severity ? ` [${p.severity}/${p.flag}]` : ""}`,
+      `- **WHAT:** ${crit?.statement ?? p.criterionId}`,
+      `- **WHY:** ${p.critique || "(no critique recorded)"}`,
+      `- **WHERE:** ${p.traceId ? `trace ${p.traceId}` : "—"}${crit?.dimension ? ` · ${crit.dimension}` : ""}`,
+      `- **TARGET:** diagnostics (${target})`,
+    ].join("\n");
+
+  const cards = items
+    .map((p, i) => {
+      const crit = criterionById(input, p.criterionId);
+      const sev = sevCls(p.severity || crit?.severity);
+      const sevTxt = String(p.severity || crit?.severity || "MED").toUpperCase();
+      const where = p.traceId
+        ? `trace <code>${esc(p.traceId)}</code>${crit?.dimension ? ` · <span class="hov-dim">${esc(crit.dimension)}</span>` : ""}`
+        : "—";
+      return (
+        `<div class="hov">` +
+        `<div class="hov-h"><span class="hov-n">#${i + 1}</span><span class="sev ${sev}">${esc(sevTxt)}</span>` +
+        `<b>${esc(p.criterionId)}</b>${p.flag ? `<span class="chip">${esc(p.flag)}</span>` : ""}` +
+        `<button class="copy-md hov-copy" data-md="${esc(itemMd(p, crit))}">copy item ▸</button></div>` +
+        `<div class="hov-g">` +
+        `<div class="hk">WHAT</div><div class="hv">${esc(crit?.statement ?? p.criterionId)}</div>` +
+        `<div class="hk">WHY</div><div class="hv">${esc(p.critique || "—")}</div>` +
+        `<div class="hk">WHERE</div><div class="hv">${where}</div>` +
+        `<div class="hk">TARGET</div><div class="hv">→ diagnostics (<code>${esc(target)}</code>)</div>` +
+        `</div></div>`
+      );
+    })
+    .join("");
+
+  // full-bundle markdown (the whole routed set, ready to hand to diagnostics).
+  const fullMd = [
+    `# EV-051 — Routed to diagnostics`,
+    ``,
+    `**Subject:** ${subjName}  ·  **Stage:** ${h.adl_stage}  ·  **Escalation:** ${h.escalation_policy}  ·  **By:** ${h.provenance.produced_by ?? "evaluator"}`,
+    ``,
+    `**Goal:** ${h.acceptance.goal}`,
+    ``,
+    ...items.flatMap((p) => [itemMd(p, criterionById(input, p.criterionId)), ``]),
+  ].join("\n");
+
+  return (
+    `<div class="sub">EV-051 — failing criteria ROUTED to diagnostics (judge-only; the evaluator flags + routes, it never fixes). Hand the block below straight to <code>*diagnose</code>.</div>` +
+    `<div class="hov-meta">` +
+    `<span class="hm-c"><span class="l">subject</span><span class="v">${esc(subjName)}</span></span>` +
+    `<span class="hm-c"><span class="l">target stage</span><span class="v"><code>${esc(target)}</code></span></span>` +
+    `<span class="hm-c"><span class="l">routed</span><span class="v">${items.length} finding(s)</span></span>` +
+    `<span class="hm-c"><span class="l">escalation</span><span class="v">${esc(h.escalation_policy)}</span></span>` +
+    `<span class="hm-c"><span class="l">produced by</span><span class="v">${esc(h.provenance.produced_by)}</span></span>` +
+    `</div>` +
+    `<div class="hov-list">${cards}</div>` +
+    `<button class="copy-md" data-md="${esc(fullMd)}">⧉ Copy full handover as markdown</button>`
+  );
+}
+
+function selfEvalTab(input: EvalReportInput): string {
+  const se = input.selfEval;
+  const decisions = routedHandover(input);
+  const header =
+    `<h2>⑤ Self-Eval · Judge Calibration <span class="badge b-skip">INTERNAL</span></h2>` +
+    tabDesc("What this tab shows [INTERNAL — stripped on publish]: the AGGREGATE judge-calibration surface — the human-in-the-loop for the JUDGE. It audits the METHODOLOGY the judge followed across the WHOLE population (not one trace): ① is the judge applying the method · ② do the criteria make sense / add value · ③ did it read enough traces (diligence) · ④ are results grounded · ⑤ are the refs real or fabricated. Accuracy-first ordering.") +
+    `<div class="sub" data-strip="strip-for-client">internal — calibrate the judge · stripped on publish</div>`;
+
+  if (se === undefined) {
+    // graceful fallback — no aggregate derived (older run): keep the EV-051 decisions.
+    return header + `<h3>Routed to diagnostics (EV-051)</h3>` + decisions;
+  }
+
+  const checksLegend =
+    `<div class="se-checks"><div class="se-checks-h">the 5 operator checks this tab answers</div><div class="se-checks-g">` +
+    [
+      ["①", "methodology adherence", "is the judge applying the judging method (bind · ground · critique-before-verdict · abstain-on-silence)?"],
+      ["②", "criterion value", "do the criteria make sense, apply, and discriminate — or are they dead weight?"],
+      ["③", "diligence", "did the judge read enough of each trace, or pick one sample and stop?"],
+      ["④", "grounded", "are the results anchored in real trace evidence (refs), not vibes?"],
+      ["⑤", "refs real", "do the cited {obs,path,value} refs actually resolve in the trace — real or fabricated?"],
+    ]
+      .map(([n, k, v]) => `<div class="se-check"><span class="se-cn">${n}</span><span class="se-ck">${esc(k)}</span><span class="se-cv">${esc(v)}</span></div>`)
+      .join("") +
+    `</div></div>`;
+
+  return (
+    header +
+    checksLegend +
+    selfEvalTrustBand(se) +
+    selfEvalBehaviorMap(se) +
+    selfEvalRefIntegrity(se) +
+    selfEvalConfidence(se) +
+    selfEvalPerCriterion(se) +
+    selfEvalCriterionValue(se) +
+    selfEvalDiligence(se) +
+    selfEvalTransparency(se) +
+    selfEvalEmit(se) +
+    selfEvalSpotCheck(se) +
+    selfEvalGroundTruthFuture() +
+    // ── judge-reasoning DETAIL (M1 · M2 · M3 · M4) — the per-trace calibration lens
+    //    that backs the aggregate above (who the judge thinks the agent is, its
+    //    train-of-thought, the expected-trajectory it built). Retained below the
+    //    aggregate so a builder can drill from the population view into an example.
+    `<h3 class="se-h"><span class="se-num">5.D</span> Judge reasoning detail (M1 · M2 · M3 · M4)</h3>` +
+    `<div class="sub">the per-trace calibration lens behind the aggregate — who the judge understood the agent to be (M1), its phase-by-phase reasoning (M4), and an example understanding + expected-trajectory (M2 · M3).</div>` +
+    subjectProfileCard(input.subjectProfile) +
+    phaseLens(input) +
+    reasoningExample(input) +
+    `<h3 class="se-h">Routed to diagnostics (EV-051)</h3>` +
+    decisions
+  );
+}
+
+// ── WS-3 §5 Self-Eval — section renderers (PURE string assembly) ──────────────
+
+function seBar(pct: number, tone = "p"): string {
+  const w = Math.max(0, Math.min(100, pct));
+  return `<span class="sebar"><span class="sebar-f ${tone}" style="width:${w}%"></span></span>`;
+}
+
+function selfEvalTrustBand(se: SelfEvalAggregate): string {
+  const t = se.trust;
+  const tone = t.score >= 85 ? "good" : t.score >= 60 ? "warn" : "bad";
+  return (
+    `<div class="se-trust ${tone}"><div class="se-trust-l"><div class="se-trust-score">${esc(t.score)}<span class="se-trust-d">/100</span></div><div class="se-trust-cap">judge trust</div></div>` +
+    `<div class="se-trust-r"><div class="se-trust-label">${esc(t.label)}</div><div class="se-trust-basis">${esc(t.basis)}</div></div></div>`
+  );
+}
+
+function selfEvalBehaviorMap(se: SelfEvalAggregate): string {
+  const b = se.behaviorMap;
+  const clusters = b.clusters
+    .map(
+      (c) =>
+        `<div class="se-cl ${esc(c.tone)}"><div class="se-cl-top"><span class="se-cl-n">${esc(c.count)}</span><span class="se-cl-l">${esc(c.label)}</span></div><div class="se-cl-note">${esc(c.note)}</div></div>`,
+    )
+    .join("");
+  const conc = b.concentration.length
+    ? `<div class="se-conc"><div class="se-conc-h">where fails CONCENTRATE</div>` +
+      b.concentration
+        .map(
+          (c) =>
+            `<div class="se-conc-r"><span class="se-conc-id" title="${esc(c.statement)}">${esc(c.id)}</span>${seBar(c.pct, "bad")}<span class="se-conc-v">${esc(c.fails)} fails · ${esc(c.pct)}%</span></div>`,
+        )
+        .join("") +
+      `</div>`
+    : `<div class="se-empty">no failing criteria — nothing concentrates.</div>`;
+  return (
+    `<h3 class="se-h"><span class="se-num">5.0</span> Judge Behavior Map — clusters &amp; where it falls short <span class="se-tag agg">AGGREGATE · MULTI-TRAJECTORY</span></h3>` +
+    `<div class="sub">how the judge handled the POPULATION of ${esc(b.total)} trajectories — the cluster sizes + the fail concentration tell you where the judge (and the criteria) are working vs thin.</div>` +
+    `<div class="se-clusters">${clusters}</div>` +
+    conc
+  );
+}
+
+function selfEvalRefIntegrity(se: SelfEvalAggregate): string {
+  const r = se.refIntegrity;
+  // honesty (UI-12-B): with ZERO decided verdicts grounding is UNMEASURABLE — render
+  // capture-unavailable (muted), NEVER a false-green default.
+  if (r.decided === 0) {
+    return (
+      `<h3 class="se-h"><span class="se-num">5.1</span> Reference Integrity — real or fabricated? <span class="se-tag gt">check ⑤ · GROUND-TRUTH</span></h3>` +
+      `<div class="se-row"><div class="se-metric"><div class="se-metric-v capture-na">capture-unavailable</div><div class="se-metric-l">no decided verdicts to ground</div></div></div>` +
+      `<div class="se-note">${esc(r.note)}</div>`
+    );
+  }
+  const tone = r.groundedPct >= 90 ? "good" : r.groundedPct >= 70 ? "warn" : "bad";
+  return (
+    `<h3 class="se-h"><span class="se-num">5.1</span> Reference Integrity — real or fabricated? <span class="se-tag gt">check ⑤ · GROUND-TRUTH</span></h3>` +
+    `<div class="se-row"><div class="se-metric ${tone}"><div class="se-metric-v">${esc(r.groundedPct)}%</div><div class="se-metric-l">decided verdicts carry ≥1 structured ref</div></div>` +
+    `<div class="se-kv"><div><b>${esc(r.grounded)}</b> grounded</div><div><b>${esc(r.ungrounded)}</b> ungrounded</div><div><b>${esc(r.decided)}</b> decided</div></div></div>` +
+    `<div class="se-note">${esc(r.note)}</div>`
+  );
+}
+
+function selfEvalConfidence(se: SelfEvalAggregate): string {
+  const rows = se.confidence
+    .map(
+      (c) =>
+        `<div class="se-conf-r"><span class="se-conf-b">${esc(c.band)}</span>${seBar(c.pct, "p")}<span class="se-conf-v">${esc(c.count)} · ${esc(c.pct)}%</span></div>`,
+    )
+    .join("");
+  return (
+    `<h3 class="se-h"><span class="se-num">5.2</span> Confidence Calibration <span class="se-tag proxy">PROXY</span></h3>` +
+    `<div class="sub">how the judge's stated confidence is distributed across decided verdicts. Tracking this against actual correctness (vs a 2nd judge / ground-truth) is the §5.X future extension.</div>` +
+    `<div class="se-conf">${rows || '<div class="se-empty">no per-verdict confidence emitted.</div>'}</div>`
+  );
+}
+
+function selfEvalPerCriterion(se: SelfEvalAggregate): string {
+  const rows = se.perCriterion
+    .map((p) => {
+      const flagCls = p.flag === "fails-present" ? "bad" : p.flag === "never-applies" ? "warn" : "good";
+      return (
+        `<tr><td class="se-cid" title="${esc(p.statement)}">${esc(p.id)}</td><td><span class="sev ${esc(p.severity)}">${esc(p.severity)}</span></td>` +
+        `<td class="good">${esc(p.pass)}</td><td class="bad">${esc(p.fail)}</td><td class="warn">${esc(p.unc)}</td><td class="dim">${esc(p.na)}</td>` +
+        `<td>${esc(p.groundedPct)}%</td><td><span class="se-flag ${flagCls}">${esc(p.flag)}</span></td></tr>`
+      );
+    })
+    .join("");
+  return (
+    `<h3 class="se-h"><span class="se-num">5.3</span> Per-Criterion Reliability <span class="se-tag proxy">check ①④ · PROXY</span></h3>` +
+    `<table class="se-table"><thead><tr><th>criterion</th><th>sev</th><th>pass</th><th>fail</th><th>indet</th><th>na</th><th>grnd</th><th>flag</th></tr></thead><tbody>${rows}</tbody></table>`
+  );
+}
+
+function selfEvalCriterionValue(se: SelfEvalAggregate): string {
+  const rows = se.criterionValue
+    .map((c, i) => {
+      const cls = c.applicablePct === 0 ? "warn" : "good";
+      return (
+        `<tr><td class="se-cid" title="${esc(c.statement)}">${esc(c.id)}</td><td><span class="sev ${esc(c.severity)}">${esc(c.severity)}</span></td>` +
+        `<td class="${cls}">${esc(c.applicablePct)}%</td><td>${esc(c.fails)} fails</td><td class="se-vverdict">${esc(c.verdict)}</td>` +
+        `<td class="se-hitl"><span class="se-pick" data-vgroup="cval-${i}" data-v="keep">keep</span><span class="se-pick" data-vgroup="cval-${i}" data-v="revise">revise</span><span class="se-pick" data-vgroup="cval-${i}" data-v="retire">retire</span></td></tr>`
+      );
+    })
+    .join("");
+  return (
+    `<h3 class="se-h"><span class="se-num">5.4</span> Criterion Value / Applicability Audit <span class="se-tag hitl">check ② · HITL</span></h3>` +
+    `<div class="sub">does each criterion apply + discriminate, or is it dead weight? A criterion that NEVER applies or NEVER fails is a candidate to retire/revise. Your pick is captured below (export-ready).</div>` +
+    `<table class="se-table"><thead><tr><th>criterion</th><th>sev</th><th>applicable</th><th>fails</th><th>value</th><th>operator</th></tr></thead><tbody>${rows}</tbody></table>`
+  );
+}
+
+function selfEvalDiligence(se: SelfEvalAggregate): string {
+  const d = se.diligence;
+  const pct = d.total > 0 ? Math.round((100 * d.examined) / d.total) : 0;
+  const tone = pct >= 95 ? "good" : pct >= 80 ? "warn" : "bad";
+  return (
+    `<h3 class="se-h"><span class="se-num">5.5</span> Diligence / Trace Coverage <span class="se-tag proxy">check ③ · PROXY</span></h3>` +
+    `<div class="se-row"><div class="se-metric ${tone}"><div class="se-metric-v">${esc(d.examined)}/${esc(d.total)}</div><div class="se-metric-l">trajectories examined</div></div>` +
+    `<div class="se-metric"><div class="se-metric-v">${esc(d.avgGroundingRefs)}</div><div class="se-metric-l">avg grounding refs / examined trace</div></div></div>` +
+    `<div class="se-note">${esc(d.note)}</div>`
+  );
+}
+
+function selfEvalTransparency(se: SelfEvalAggregate): string {
+  const t = se.transparency;
+  const block = t.blocked
+    ? `<div class="se-blocked"><span class="se-blocked-tag">BLOCKED</span> ${esc(t.note)}</div>`
+    : `<div class="se-note">${esc(t.note)}</div>`;
+  return (
+    `<h3 class="se-h"><span class="se-num">5.6</span> Judge Reasoning Transparency (walk · M2 · M3) <span class="se-tag">TRANSPARENCY · check ①</span></h3>` +
+    `<div class="se-row"><div class="se-metric ${t.m2Pct > 0 ? "good" : "bad"}"><div class="se-metric-v">${esc(t.m2Pct)}%</div><div class="se-metric-l">M2 understanding</div></div>` +
+    `<div class="se-metric ${t.m3Pct > 0 ? "good" : "bad"}"><div class="se-metric-v">${esc(t.m3Pct)}%</div><div class="se-metric-l">M3 expected-trajectory</div></div>` +
+    `<div class="se-metric ${t.walkPct > 0 ? "good" : "bad"}"><div class="se-metric-v">${esc(t.walkPct)}%</div><div class="se-metric-l">judge walk emitted</div></div></div>` +
+    block
+  );
+}
+
+function selfEvalEmit(se: SelfEvalAggregate): string {
+  const e = se.emit;
+  const meter = (label: string, present: number, total: number): string => {
+    const pct = total > 0 ? Math.round((100 * present) / total) : 0;
+    const tone = pct >= 90 ? "good" : pct >= 40 ? "warn" : "bad";
+    return `<div class="se-emit-r"><span class="se-emit-l">${esc(label)}</span>${seBar(pct, tone === "good" ? "p" : tone)}<span class="se-emit-v">${esc(present)}/${esc(total)}</span></div>`;
+  };
+  const fieldRow = (f: { field: string; present: number }): string =>
+    meter(f.field, f.present, e.eligible);
+  return (
+    `<h3 class="se-h"><span class="se-num">5.7</span> Emit-Completeness — self-honesty meter <span class="se-tag">HONESTY</span></h3>` +
+    `<div class="sub">does the judge persist what it reasoned? A dropped walk does NOT change a verdict — but it STARVES this tab + §2. The engine gate (assessEmitCompleteness) tracks this every run.</div>` +
+    `<div class="se-emit">` +
+    meter("verdict + critique", e.eligible, e.eligible) +
+    e.fields.map(fieldRow).join("") +
+    `</div>` +
+    `<div class="se-note">complete emits (all of M2+M3+agentSteps+judgeSteps): <b>${esc(e.completeEmits)}/${esc(e.eligible)}</b> (${esc(e.completePct)}%).` +
+    (e.exemptIncomplete > 0 ? ` ${esc(e.exemptIncomplete)} INCOMPLETE traces exempt (node-1 short-circuit).` : "") +
+    `</div>`
+  );
+}
+
+function selfEvalSpotCheck(se: SelfEvalAggregate): string {
+  if (se.spotCheck.length === 0) {
+    return (
+      `<h3 class="se-h"><span class="se-num">5.8</span> Spot-Check / Disagreement Queue <span class="se-tag hitl">check ④ · HITL</span></h3>` +
+      `<div class="se-empty">no failing or low-confidence verdicts queued — nothing brittle to spot-check.</div>`
+    );
+  }
+  const rows = se.spotCheck
+    .map(
+      (s, i) =>
+        `<div class="se-spot"><span class="se-spot-t mono" title="${esc(s.criterion)}">${esc(s.trace.slice(0, 14))}</span>` +
+        `<span class="se-spot-c">${esc(s.criterion.slice(0, 54))}</span>` +
+        `<span class="verd ${s.verdict === "FAIL" ? "fail" : "inc"}">${esc(s.verdict)}</span>` +
+        `<span class="se-spot-conf">conf ${esc(s.conf)}</span><span class="se-spot-why">${esc(s.reason)}</span>` +
+        `<span class="se-spot-act"><span class="se-pick" data-vgroup="spot-${i}" data-v="agree">agree</span><span class="se-pick" data-vgroup="spot-${i}" data-v="disagree">disagree</span></span></div>`,
+    )
+    .join("");
+  return (
+    `<h3 class="se-h"><span class="se-num">5.8</span> Spot-Check / Disagreement Queue <span class="se-tag hitl">check ④ · HITL</span></h3>` +
+    `<div class="sub">the brittle calls — fails + low-confidence verdicts — sorted weakest-confidence-first. Your agree/disagree is captured (and becomes the ground-truth seed for §5.X).</div>` +
+    `<div class="se-spots">${rows}</div>`
+  );
+}
+
+function selfEvalGroundTruthFuture(): string {
+  return (
+    `<h3 class="se-h"><span class="se-num">5.X</span> Judge Accuracy vs Ground-Truth <span class="se-tag future">FUTURE — not built</span></h3>` +
+    `<div class="se-future"><div class="se-future-h">documented extension</div>` +
+    `<div class="se-future-b">The strongest calibration is judge-verdict vs a TRUSTED label. We do not have ground-truth labels in this run, so this is DOCUMENTED, not computed. The seed already exists: §5.8 Spot-Check collects operator agree/disagree per verdict — accumulate those into a labelled set and this panel becomes a real accuracy/precision/recall readout against the judge. Until then, §5.1 (refs real) + §5.2 (confidence) + §5.3 (per-criterion) are the PROXY signals.</div></div>`
+  );
+}
+
+// ── report-component CSS (references theme.css tokens) ────────────────────────
+
+const REPORT_CSS = `
+.mono{font-family:var(--mono)}
+/* status-acuity — color-code the OUTCOME big-stat tiles; COUNT tiles stay neutral (faint cyan) */
+/* ALIGNMENT FIX (parity w/ discover): theme.css ships .big-stat as content-sized flex
+ * (min-width:96px) → ragged tile widths. Override to an EQUAL-column grid so every tile
+ * is the same width AND height (grid row-stretch). */
+.big-stat{display:grid;grid-template-columns:repeat(auto-fit,minmax(120px,1fr));gap:12px;align-items:start}
+.big-stat .s{min-width:0;min-height:84px;display:flex;flex-direction:column;justify-content:center}
+.big-stat .s .v{line-height:1.15}
+.big-stat .s{border-left:4px solid var(--border-strong)}
+.big-stat .s.count{border-left-color:var(--cyan);opacity:.95}
+.big-stat .s.pass{background:var(--pass-bg);border-color:var(--pass);border-left-color:var(--pass)}
+.big-stat .s.pass .v{color:var(--pass)}
+.big-stat .s.fail{background:var(--fail-bg);border-color:var(--fail);border-left-color:var(--fail)}
+.big-stat .s.fail .v{color:var(--fail)}
+.big-stat .s.indet{background:var(--warn-bg);border-color:var(--warn);border-left-color:var(--warn)}
+.big-stat .s.indet .v{color:var(--warn)}
+/* Overview-redesign — entity HERO (full-width subject identity) */
+/* §9.4.5 E4 — entity-AS-HERO: a primary left-accent + tighter rhythm so the subject
+ * card reads as the page anchor (diagnostics-grade density, dark/branded preserved). */
+.hero{border:1px solid var(--border-strong);border-left:3px solid var(--primary);background:var(--surf);margin:10px 0 14px;padding:0}
+.hero-top{display:flex;gap:10px;align-items:center;flex-wrap:wrap;padding:11px 14px;border-bottom:1px solid var(--border);background:var(--surf-2)}
+.hero .hname{font-size:var(--fs-xl);font-weight:700;color:var(--fg-strong);font-family:var(--mono)}
+.hero .hchip{font-family:var(--mono);font-size:var(--fs-2xs);font-weight:700;padding:2px 8px;border:1px solid var(--border);color:var(--muted);text-transform:uppercase;letter-spacing:.04em}
+.hero .hchip.type{color:var(--cyan);border-color:var(--cyan);background:rgba(69,184,204,.10)}
+.hero .hchip.access{margin-left:auto}
+.hero .hchip.access.given{color:var(--pass);border-color:var(--pass);background:var(--pass-bg)}
+.hero .hchip.access.recon{color:var(--warn);border-color:var(--warn);background:var(--warn-bg)}
+.hero-grid{display:grid;grid-template-columns:max-content 1fr;gap:1px;background:var(--border)}
+.hero-grid .hk{background:var(--surf);font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted);text-transform:uppercase;letter-spacing:.04em;padding:7px 12px}
+.hero-grid .hv{background:var(--surf-2);font-size:var(--fs-sm);color:var(--fg);padding:7px 13px;line-height:1.5}
+.hero .unk{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--dim);text-transform:uppercase;border:1px dashed var(--border-strong);padding:0 5px}
+.hero .inf{font-family:var(--mono);font-size:var(--fs-2xs);font-weight:700;color:var(--warn);background:var(--warn-bg);padding:1px 5px;text-transform:uppercase}
+.hero .inf.given{color:var(--pass);background:var(--pass-bg)}
+.hero .hsys{margin:9px 12px;border:1px solid var(--border)}
+.hero .hsys summary{font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;color:var(--cyan);cursor:pointer;padding:6px 10px;background:var(--surf);display:flex;gap:8px;align-items:center}
+.hero .hsys-prov{font-family:var(--mono);font-size:var(--fs-2xs);font-weight:700;padding:1px 6px;text-transform:uppercase;letter-spacing:.03em}
+.hero .hsys-prov.recon{color:var(--warn);background:var(--warn-bg);border:1px solid var(--warn)}
+.hero .hsys-prov.given{color:var(--pass);background:var(--pass-bg);border:1px solid var(--pass)}
+.hero .hsys pre{margin:0;padding:9px 11px;font-family:var(--mono);font-size:var(--fs-xs);color:var(--fg);white-space:pre-wrap;word-break:break-word;line-height:1.5;background:var(--bg);max-height:420px;overflow:auto}
+.hero .hsys-na{margin:9px 12px;font-family:var(--mono);font-size:var(--fs-2xs);color:var(--dim)}
+.hero .hhonesty{padding:8px 13px;border-top:1px solid var(--border);font-size:var(--fs-2xs);color:var(--muted);line-height:1.5}
+/* §9.4.5 E1 — trace-only honesty: the reconstructed banner + the harness/prompt UNAVAILABLE note */
+.hero .hrecon{padding:7px 14px;border-bottom:1px solid var(--border);background:var(--warn-bg);font-family:var(--mono);font-size:var(--fs-2xs);font-weight:600;color:var(--warn);letter-spacing:.03em;line-height:1.5}
+.hero .recon-note{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--warn);text-transform:uppercase;letter-spacing:.03em}
+/* §9.4.5 E2 — tools as diagnostics-style chips with a per-tool observed call-count */
+.hero .tchips{display:flex;gap:5px;flex-wrap:wrap;align-items:center}
+.hero .tchip{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--fg);background:var(--surf-3);border:1px solid var(--border-strong);padding:2px 7px;display:inline-flex;gap:5px;align-items:center;white-space:nowrap}
+.hero .tchip .obs{font-weight:700;color:var(--cyan);font-size:var(--fs-2xs)}
+.hero .tchip .obs.none{color:var(--dim)}
+/* Overview-redesign — provenance META-STRIP (run-config + judge substrate) */
+.provstrip{display:flex;gap:7px 14px;flex-wrap:wrap;align-items:center;border:1px solid var(--border);border-left:3px solid var(--primary);background:var(--surf-2);padding:7px 12px;margin:10px 0}
+.provstrip .ps-item{display:inline-flex;gap:5px;align-items:baseline}
+.provstrip .ps-k{font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;letter-spacing:.04em;color:var(--muted)}
+.provstrip .ps-v{font-family:var(--mono);font-size:var(--fs-xs);color:var(--fg-strong)}
+.provstrip .unk{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--dim);text-transform:uppercase}
+.provstrip .ps-badge{font-family:var(--mono);font-size:var(--fs-2xs);font-weight:700;padding:2px 8px;border:1px solid;text-transform:uppercase}
+.provstrip .ps-badge.on{color:var(--primary-soft);border-color:var(--primary);background:rgba(126,71,215,.12)}
+.provstrip .ps-badge.off{color:var(--dim);border-color:var(--border)}
+/* Overview-redesign — segmented coverage FUNNEL (INCOMPLETE first-class) */
+.funnel{display:flex;align-items:stretch;gap:6px;flex-wrap:wrap;margin:14px 0 6px}
+.funnel .fstage{flex:1;min-width:118px;border:1px solid var(--border);background:var(--surf);padding:9px 11px}
+.funnel .fstage .fn{font-family:var(--mono);font-size:var(--fs-2xl);font-weight:700;color:var(--fg-strong);line-height:1.05}
+.funnel .fstage .fl{font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;letter-spacing:.05em;color:var(--cyan);margin-top:3px}
+.funnel .fstage .fnote{font-size:var(--fs-2xs);color:var(--dim);margin-top:3px;line-height:1.4}
+.funnel .farrow{display:flex;align-items:center;color:var(--dim);font-size:var(--fs-lg);font-family:var(--mono)}
+.funnel .fstage.outc{flex:1.6;min-width:200px}
+.funnel .foutc{display:flex;gap:5px;flex-wrap:wrap;margin-top:6px}
+.funnel .fpill{font-family:var(--mono);font-size:var(--fs-2xs);padding:2px 8px;border:1px solid var(--border);color:var(--muted)}
+.funnel .fpill b{color:var(--fg-strong)}
+.funnel .fpill.pass{border-color:var(--pass);color:var(--pass)}.funnel .fpill.pass b{color:var(--pass)}
+.funnel .fpill.fail{border-color:var(--fail);color:var(--fail)}.funnel .fpill.fail b{color:var(--fail)}
+.funnel .fpill.indet{border-color:var(--warn);color:var(--warn)}.funnel .fpill.indet b{color:var(--warn)}
+.funnel .fpill.inc{border-color:var(--border-strong);color:var(--muted);border-style:dashed}
+/* §9.4.5 E3 — eval-HEALTH temporal heatmap (correctness over time, NOT latency) */
+.ehm{display:flex;gap:4px;flex-wrap:wrap;align-items:flex-start;margin:6px 0 4px}
+.ehm-col{display:flex;flex-direction:column;align-items:center;gap:3px}
+.ehm-cell{min-width:30px;height:30px;padding:0 6px;display:flex;align-items:center;justify-content:center;border:1px solid rgba(0,0,0,.35);font-family:var(--mono);font-size:var(--fs-2xs);font-weight:700;color:#0a0a12}
+.ehm-cell.pass{background:var(--hm-pass,rgba(67,195,154,.50))}.ehm-cell.fail{background:var(--hm-fail,rgba(224,102,102,.54))}.ehm-cell.indet{background:var(--hm-indet,rgba(232,166,77,.52))}.ehm-cell.skip{background:var(--surf-3);color:var(--dim)}
+.ehm-lab{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted);text-align:center;line-height:1.25}
+.ehm-legend{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted);margin:4px 0;display:flex;gap:4px;align-items:center;flex-wrap:wrap}
+.ehm-legend .sw{display:inline-block;width:11px;height:11px;border:1px solid rgba(0,0,0,.35);vertical-align:middle;margin:0 2px}
+.ehm-legend .sw.pass{background:var(--hm-pass,rgba(67,195,154,.50))}.ehm-legend .sw.fail{background:var(--hm-fail,rgba(224,102,102,.54))}.ehm-legend .sw.indet{background:var(--hm-indet,rgba(232,166,77,.52))}
+.ehm-pending{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--dim);border:1px dashed var(--border-strong);background:var(--surf);padding:7px 11px;margin:2px 0 6px;line-height:1.5}
+/* note callout */
+.note{border-left:3px solid var(--recommend);background:var(--recommend-bg);padding:7px 11px;font-size:var(--fs-xs);margin:14px 0}
+.note .tag{font-family:var(--mono);font-size:var(--fs-2xs);font-weight:700;color:var(--recommend)}
+/* severity pills + verd */
+.sev{font-family:var(--mono);font-size:var(--fs-2xs);font-weight:700;padding:2px 7px;margin-right:4px}
+.sev.crit{color:var(--fail);background:var(--fail-bg)}.sev.high{color:var(--warn);background:var(--warn-bg)}.sev.med{color:var(--cyan);background:rgba(69,184,204,.10)}.sev.low{color:var(--muted);background:var(--surf-3)}
+.verd{font-family:var(--mono);font-size:var(--fs-xs);font-weight:700;padding:2px 8px;border:1px solid}
+.verd.fail{color:var(--fail);border-color:var(--fail)}.verd.pass{color:var(--pass);border-color:var(--pass)}.verd.inc{color:var(--warn);border-color:var(--warn)}
+.chip{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted);border:1px solid var(--border);padding:1px 6px;white-space:nowrap}
+/* teaser */
+.teaser-row{display:flex;gap:8px;align-items:baseline;margin:4px 0;font-size:var(--fs-sm)}
+.teaser-num{color:var(--fail);font-family:var(--mono);font-size:var(--fs-xs)}
+/* §2 ledger */
+.lfilter{display:flex;gap:6px;align-items:center;flex-wrap:wrap;margin:9px 0}
+.lfilter b{font-family:var(--mono);font-size:var(--fs-2xs);padding:3px 8px;border:1px solid var(--border);color:var(--fg);opacity:.8;cursor:pointer}
+.lfilter b.on{border-color:var(--primary);color:var(--primary-soft);background:rgba(126,71,215,.12)}
+.lfilter input{background:var(--surf-2);border:1px solid var(--border);color:var(--fg);font-family:var(--mono);font-size:var(--fs-xs);padding:3px 8px;flex:1;max-width:240px}
+.lfilter .sp{margin-left:auto;font-family:var(--mono);font-size:var(--fs-2xs);color:var(--dim)}
+.virt{font-size:var(--fs-2xs);color:var(--dim);text-align:center;padding:8px}
+table.ledger td{cursor:pointer;font-family:var(--mono);font-size:var(--fs-xs)}
+table.ledger tr:hover td{background:var(--surf-2)}table.ledger tr.sel td{background:rgba(126,71,215,.10)}
+/* status-acuity — verdict-tinted ledger rows + a status left-accent (eye-catch at a glance) */
+table.ledger tr.vrow.FAIL td{background:var(--fail-bg)}table.ledger tr.vrow.FAIL td.tid{box-shadow:inset 3px 0 0 var(--fail)}
+table.ledger tr.vrow.INDETERMINATE td{background:var(--warn-bg)}table.ledger tr.vrow.INDETERMINATE td.tid{box-shadow:inset 3px 0 0 var(--warn)}
+table.ledger tr.vrow.PASS td.tid{box-shadow:inset 3px 0 0 var(--pass)}
+table.ledger tr.vrow.INCOMPLETE td.tid{box-shadow:inset 3px 0 0 var(--dim)}
+table.ledger tr.vrow:hover td{background:var(--surf-2)}table.ledger tr.vrow.sel td{background:rgba(126,71,215,.10)}
+.tid{color:var(--cyan)}.haswalk{color:var(--primary-soft);font-size:var(--fs-2xs)}
+.mini{font-family:var(--mono);font-size:var(--fs-2xs);padding:1px 5px}
+.mini.PASS{color:var(--pass);background:var(--pass-bg)}.mini.FAIL{color:var(--fail);background:var(--fail-bg)}.mini.INDETERMINATE{color:var(--warn);background:var(--warn-bg)}.mini.INCOMPLETE{color:var(--muted);background:var(--surf-3)}
+/* WS-2 — resolution badge (ledger column) + drill routing line. Makes a no-walk trace
+   HONEST: judged, not unjudged. walk=cyan · walk-dropped=amber · truncated=dim. */
+.resbadge{font-family:var(--mono);font-size:var(--fs-2xs);padding:1px 6px;border:1px solid;white-space:nowrap;cursor:help}
+.resbadge.res-walk{color:var(--cyan);border-color:var(--cyan);background:rgba(69,184,204,.10)}
+.resbadge.res-nowalk{color:var(--warn);border-color:var(--warn);background:var(--warn-bg)}
+.resbadge.res-trunc{color:var(--muted);border-color:var(--border-strong);background:var(--surf-3)}
+.routing{display:flex;gap:9px;align-items:baseline;flex-wrap:wrap;margin:0 0 9px;padding:8px 11px;border-left:3px solid var(--border-strong);background:var(--surf-2)}
+.routing.res-walk{border-left-color:var(--cyan)}
+.routing.res-nowalk{border-left-color:var(--warn)}
+.routing.res-trunc{border-left-color:var(--dim)}
+.routing-k{flex:0 0 auto;font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;letter-spacing:.05em;color:var(--dim)}
+.routing-b{flex:0 0 auto;font-family:var(--mono);font-size:var(--fs-2xs);padding:1px 6px;border:1px solid}
+.routing-b.res-walk{color:var(--cyan);border-color:var(--cyan)}
+.routing-b.res-nowalk{color:var(--warn);border-color:var(--warn)}
+.routing-b.res-trunc{color:var(--muted);border-color:var(--border-strong)}
+.routing-v{flex:1 1 200px;font-size:var(--fs-sm);color:var(--fg);line-height:1.5}
+/* §2 side-by-side */
+.ctx{border:1px solid var(--border);background:var(--surf-2);margin:8px 0 4px}
+.ctx-h{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--cyan);text-transform:uppercase;letter-spacing:.06em;padding:6px 11px;border-bottom:1px solid var(--border)}
+.ctx-g{display:grid;grid-template-columns:repeat(3,1fr);gap:1px;background:var(--border)}.ctx-c{background:var(--surf);padding:8px 11px}
+.ctx-c .l{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted);text-transform:uppercase}.ctx-c .v{font-size:var(--fs-xs);margin-top:2px}
+/* Gap A — raw triggering INPUT in the §2 drill (input + scenario cell). font floor 11px. */
+.ctx-c .iraw{margin:0;font-family:var(--mono);font-size:var(--fs-2xs);line-height:1.5;color:var(--fg);white-space:pre-wrap;word-break:break-word;background:var(--bg);border:1px solid var(--border);padding:6px 8px}
+.ctx-c .iraw.clamp{max-height:120px;overflow:auto}
+.ctx-c .iexp summary{cursor:pointer;font-family:var(--mono);font-size:var(--fs-2xs);color:var(--cyan);text-transform:uppercase;letter-spacing:.04em;padding:2px 0}
+.ctx-c .iexp[open] summary{margin-bottom:4px}
+.ctx-c .iscn{margin-top:4px;font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted)}
+.ctx-c .ival{font-size:var(--fs-xs);color:var(--fg)}
+.lanehdr{display:grid;grid-template-columns:1fr 42px 1fr}.lanehdr div{font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;letter-spacing:.05em;padding:5px 8px;border-bottom:1px solid var(--border-strong)}
+.lanehdr .a{color:var(--fg-strong)}.lanehdr .x{text-align:center;color:var(--dim)}.lanehdr .j{color:var(--primary-soft)}
+.grid2{display:grid;grid-template-columns:1fr 42px 1fr;align-items:stretch}
+.band{grid-column:1/-1;border:1px solid var(--border);margin:8px 0;padding:7px 11px;background:var(--surf)}.band.loc{border-left:3px solid var(--fail);background:var(--fail-bg)}.band .bh{font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;color:var(--fail)}
+.node{display:flex;flex-direction:column;align-items:center}.node .n{width:22px;height:22px;display:flex;align-items:center;justify-content:center;font-family:var(--mono);font-size:var(--fs-xs);font-weight:600;border:1px solid var(--border-strong);background:var(--surf-2);margin-top:12px}.node .ln{flex:1;width:1px;background:var(--border)}
+.node.error .n,.node.false-success .n{border-color:var(--fail);color:var(--fail)}.node.warn .n{border-color:var(--warn);color:var(--warn)}.node.ok .n{border-color:var(--pass);color:var(--pass)}
+.evb{border:1px solid var(--border);background:var(--surf-2);padding:6px 9px;margin:8px 8px 8px 0}.evb.r{margin:8px 0 8px 8px}
+.evb .top{display:flex;align-items:center;gap:6px}.evb .tool{font-family:var(--mono);font-size:var(--fs-xs);color:var(--fg-strong)}.evb .st{font-family:var(--mono);font-size:var(--fs-2xs);padding:1px 6px;margin-left:auto}
+.evb .st.ok{color:var(--pass);background:var(--pass-bg)}.evb .st.error,.evb .st.false-success{color:var(--fail);background:var(--fail-bg)}.evb .st.warn{color:var(--warn);background:var(--warn-bg)}
+.evb .det{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted);margin-top:3px;line-height:1.4}
+.jstep{display:flex;gap:6px;align-items:baseline;font-family:var(--mono);font-size:var(--fs-2xs);margin-top:4px}
+.jstep .k{font-size:var(--fs-2xs);font-weight:600;padding:1px 5px;min-width:54px;text-align:center}
+.k.context{color:var(--cyan);background:rgba(69,184,204,.10)}.k.examine,.k.detect{color:var(--warn);background:var(--warn-bg)}.k.bind{color:var(--cyan);background:rgba(69,184,204,.10)}.k.ground{color:var(--pass);background:var(--pass-bg)}.k.critique{color:var(--primary-soft);background:rgba(126,71,215,.12)}.k.decide{color:var(--fail);background:var(--fail-bg)}.k.verify{color:var(--fg-strong);background:var(--surf-3)}
+.jstep .t{color:var(--fg);opacity:.92;line-height:1.4}.jstep .t .ref{color:var(--cyan)}
+.jstep.noexam .t{color:var(--dim);font-style:italic}
+/* §2 judge lane — per-step eval-coverage entry (which criterion examined this step). */
+.jcov{border-left:2px solid var(--border-strong);background:var(--surf);padding:5px 8px;margin-top:6px}
+.jcov.pass{border-left-color:var(--pass)}.jcov.fail{border-left-color:var(--fail)}
+.jcov.uncertain,.jcov.indeterminate{border-left-color:var(--warn)}.jcov.na{border-left-color:var(--border)}
+.jcov-h{display:flex;gap:6px;align-items:center;flex-wrap:wrap;margin-bottom:3px}
+.jcov .jm{font-family:var(--mono);font-size:var(--fs-2xs);font-weight:600;letter-spacing:.04em;padding:0 5px;border:1px solid var(--border-strong)}
+.jcov .jm.code{color:var(--cyan);border-color:var(--cyan)}.jcov .jm.judge{color:var(--primary-soft);border-color:var(--primary-soft)}
+.jcov .jcid{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--fg-strong);word-break:break-all}
+.jcov .jcrit{font-size:var(--fs-xs);color:var(--fg);opacity:.92;line-height:1.45}.jcov .jcrit .dim{color:var(--dim)}
+.jcov .cvrefs{margin-top:4px}
+.health{display:grid;grid-template-columns:repeat(4,1fr);gap:1px;background:var(--border);margin-top:8px}.health .hc{background:var(--surf);padding:8px 10px}.health .hc .l{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted);text-transform:uppercase}.health .hc .v{font-size:var(--fs-lg);font-weight:600;margin-top:2px}.health .hc .v.good{color:var(--pass)}.health .hc .v.warn{color:var(--warn)}
+.drillbox{border:1px solid var(--border-strong);background:var(--surf-2);margin-top:10px;padding:11px}
+/* §3 heatmap + subcards + calibration */
+.hm{display:grid;gap:3px;font-family:var(--mono);font-size:var(--fs-2xs);margin-top:6px}
+.hm .lab{color:var(--fg);opacity:.82;text-align:right;padding-right:5px;line-height:18px;white-space:nowrap;overflow:hidden;text-overflow:ellipsis}
+.hm .colh{color:var(--muted);text-align:center;font-size:var(--fs-2xs);line-height:13px}
+.hm .cell{height:18px;border:1px solid rgba(0,0,0,.35);display:flex;align-items:center;justify-content:center;color:#0a0a12;font-size:var(--fs-2xs);font-weight:700}
+/* status-acuity — heatmap fills nudged up for contrast (still toned, not neon) */
+.hm .cell.pass{background:var(--hm-pass,rgba(67,195,154,.50))}.hm .cell.fail{background:var(--hm-fail,rgba(224,102,102,.54))}.hm .cell.indet{background:var(--hm-indet,rgba(232,166,77,.52))}.hm .cell.skip{background:var(--surf-3)}
+/* UI-6 — criteria cards: a PURPLISH base (brand primary), accent driven by SUCCESS RATE
+ * (green-ish high · amber mid · red-ish low) — NOT a uniform severity red. */
+.subc{border:1px solid var(--border);border-left:4px solid var(--primary);background:linear-gradient(180deg,rgba(126,71,215,.12),rgba(126,71,215,.045));margin-top:10px}
+.subc.rate-ok{border-left-color:var(--pass)}
+.subc.rate-mid{border-left-color:var(--warn)}
+.subc.rate-low{border-left-color:var(--fail)}
+.subc-h{display:flex;gap:8px;align-items:center;padding:8px 11px;border-bottom:1px solid var(--border)}.subc-h b{font-size:var(--fs-sm);color:var(--fg-strong)}.subc-h .vp{font-family:var(--mono);font-size:var(--fs-xs);font-weight:700;padding:1px 7px;border:1px solid}
+.nest{border:1px solid var(--border);background:var(--surf);margin:7px 11px}.nest-h{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted);text-transform:uppercase;padding:4px 9px;border-bottom:1px solid var(--border)}
+.row{font-family:var(--mono);font-size:var(--fs-xs);color:var(--fg);opacity:.92;margin:4px 9px;line-height:1.5}.row .ref{color:var(--cyan)}.row .w{color:var(--warn)}
+.cal{display:flex;gap:5px;align-items:center;padding:8px 11px;border-top:1px solid var(--border)}.cal .l{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted);margin-right:3px}.cal b{font-family:var(--mono);font-size:var(--fs-2xs);padding:2px 8px;border:1px solid var(--border);color:var(--fg);opacity:.85;cursor:pointer}.cal b.on{border-color:var(--primary);color:var(--primary-soft);background:rgba(126,71,215,.12)}
+/* §4 findings */
+.find{border:1px solid var(--border);background:var(--surf-2);margin-top:11px}.find.gate{border-top:2px solid var(--fail);border-left:4px solid var(--fail);background:var(--fail-bg)}
+.find-h{display:flex;gap:9px;align-items:center;padding:9px 12px;border-bottom:1px solid var(--border);flex-wrap:wrap}.find-h b{font-size:var(--fs-sm);color:var(--fg-strong)}.find-h .gatetag{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--fail);border:1px solid var(--fail);padding:1px 6px}
+.fg{display:grid;grid-template-columns:92px 1fr;gap:1px;background:var(--border)}.fg .k{background:var(--surf);font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted);text-transform:uppercase;padding:8px 10px}
+.fg .v{background:var(--surf-2);padding:8px 11px;font-size:var(--fs-sm);color:var(--fg)}.fg .v code{font-family:var(--mono);font-size:var(--fs-xs);background:var(--bg);padding:0 4px}.fg .v .ref{color:var(--cyan);font-family:var(--mono);font-size:var(--fs-xs)}.fg .v .none{color:var(--pass)}.fg .v ul{margin:2px 0;padding-left:15px}.fg .v li{margin:2px 0;font-size:var(--fs-xs);font-family:var(--mono)}
+.fg .v.reason .step{color:var(--primary-soft)}.fg .v .lnk{color:var(--primary-soft);font-family:var(--mono);font-size:var(--fs-2xs);cursor:pointer}
+.na-note{font-size:var(--fs-2xs);color:var(--dim)}
+.prev{display:inline-block;background:var(--surf-3);height:8px;width:110px;vertical-align:middle}.prev i{display:block;height:100%;background:var(--fail)}
+.areview{display:flex;gap:6px;align-items:center;padding:9px 12px;border-top:1px solid var(--border-strong);background:var(--surf)}.areview .l{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--cyan)}.areview b{font-family:var(--mono);font-size:var(--fs-2xs);padding:2px 9px;border:1px solid var(--border);color:var(--fg);opacity:.85;cursor:pointer}.areview b.agree.on{border-color:var(--pass);color:var(--pass)}.areview b.refute.on{border-color:var(--fail);color:var(--fail)}.areview input{flex:1;background:var(--surf-2);border:1px solid var(--border);color:var(--fg);font-family:var(--mono);font-size:var(--fs-xs);padding:3px 7px}
+/* §4/§2 per-trajectory scorecard grid */
+.scgrid{display:grid;grid-template-columns:repeat(auto-fill,minmax(220px,1fr));gap:5px;margin-top:6px}
+.scc{border:1px solid var(--border);background:var(--surf);padding:5px 8px;font-family:var(--mono);font-size:var(--fs-2xs);display:flex;gap:6px;align-items:center}.scc.fail{border-left:2px solid var(--fail)}.scc.uncertain,.scc.indeterminate{border-left:2px solid var(--warn)}.scc .nm{overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
+/* WS-3 §5 self-eval AGGREGATE — judge-calibration surface */
+.se-h{font-size:var(--fs-md);margin:22px 0 4px;display:flex;gap:8px;align-items:baseline;flex-wrap:wrap}
+.se-num{font-family:var(--mono);font-weight:700;color:var(--primary-soft)}
+.se-tag{font-family:var(--mono);font-size:var(--fs-2xs);font-weight:700;padding:1px 6px;border:1px solid var(--border-strong);color:var(--muted);text-transform:uppercase;letter-spacing:.03em}
+.se-tag.agg{color:var(--cyan);border-color:var(--cyan)}.se-tag.gt{color:var(--pass);border-color:var(--pass)}.se-tag.proxy{color:var(--warn);border-color:var(--warn)}.se-tag.hitl{color:var(--primary-soft);border-color:var(--primary)}.se-tag.future{color:var(--dim);border-color:var(--border)}
+.se-checks{border:1px solid var(--border);border-left:3px solid var(--primary);background:var(--surf-2);margin:10px 0 4px;padding:9px 12px}
+.se-checks-h{font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;letter-spacing:.05em;color:var(--dim);margin-bottom:6px}
+.se-checks-g{display:grid;grid-template-columns:repeat(auto-fit,minmax(240px,1fr));gap:6px}
+.se-check{display:flex;gap:7px;align-items:baseline}.se-cn{color:var(--primary-soft);font-family:var(--mono);font-weight:700}.se-ck{font-weight:600;color:var(--fg-strong);font-size:var(--fs-sm);white-space:nowrap}.se-cv{color:var(--muted);font-size:var(--fs-xs);line-height:1.4}
+.se-trust{display:flex;gap:16px;align-items:center;border:1px solid var(--border-strong);background:var(--surf);margin:12px 0;padding:14px 16px;border-left:3px solid var(--primary)}
+.se-trust.good{border-left-color:var(--pass)}.se-trust.warn{border-left-color:var(--warn)}.se-trust.bad{border-left-color:var(--fail)}
+.se-trust-score{font-family:var(--mono);font-size:var(--fs-2xl);font-weight:700;color:var(--fg-strong);line-height:1}.se-trust-d{font-size:var(--fs-md);color:var(--dim)}
+.se-trust-cap{font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;letter-spacing:.05em;color:var(--dim);margin-top:3px}
+.se-trust.good .se-trust-score{color:var(--pass)}.se-trust.warn .se-trust-score{color:var(--warn)}.se-trust.bad .se-trust-score{color:var(--fail)}
+.se-trust-label{font-weight:600;color:var(--fg-strong);font-size:var(--fs-md)}.se-trust-basis{color:var(--muted);font-size:var(--fs-sm);margin-top:4px;line-height:1.5}
+.se-clusters{display:grid;grid-template-columns:repeat(auto-fit,minmax(220px,1fr));gap:9px;margin:10px 0}
+.se-cl{border:1px solid var(--border);border-top:3px solid var(--border-strong);background:var(--surf);padding:9px 11px}
+.se-cl.good{border-top-color:var(--pass)}.se-cl.bad{border-top-color:var(--fail)}.se-cl.warn{border-top-color:var(--warn)}.se-cl.neutral{border-top-color:var(--dim)}
+.se-cl-top{display:flex;gap:8px;align-items:baseline}.se-cl-n{font-family:var(--mono);font-size:var(--fs-xl);font-weight:700;color:var(--fg-strong)}.se-cl-l{font-size:var(--fs-sm);font-weight:600;color:var(--fg)}
+.se-cl-note{font-size:var(--fs-xs);color:var(--muted);margin-top:4px;line-height:1.45}
+.se-conc{border:1px solid var(--border);background:var(--surf-2);padding:9px 12px;margin:6px 0}
+.se-conc-h{font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;letter-spacing:.05em;color:var(--dim);margin-bottom:6px}
+.se-conc-r,.se-conf-r,.se-emit-r{display:flex;gap:9px;align-items:center;margin:3px 0;font-size:var(--fs-xs)}
+.se-conc-id{font-family:var(--mono);color:var(--cyan);flex:0 0 200px;overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
+.se-conc-v,.se-conf-v,.se-emit-v{font-family:var(--mono);color:var(--muted);flex:0 0 auto;white-space:nowrap}
+.sebar{flex:1 1 auto;height:11px;background:var(--bg);border:1px solid var(--border);position:relative;overflow:hidden}
+.sebar-f{position:absolute;left:0;top:0;bottom:0}.sebar-f.p{background:linear-gradient(90deg,#3a2d63,var(--primary))}.sebar-f.good{background:var(--pass)}.sebar-f.warn{background:var(--warn)}.sebar-f.bad{background:var(--fail)}
+.se-row{display:flex;gap:10px;flex-wrap:wrap;margin:8px 0}
+.se-metric{border:1px solid var(--border);background:var(--surf);padding:9px 14px;min-width:140px;border-left:3px solid var(--border-strong)}
+.se-metric.good{border-left-color:var(--pass)}.se-metric.warn{border-left-color:var(--warn)}.se-metric.bad{border-left-color:var(--fail)}
+.se-metric-v{font-family:var(--mono);font-size:var(--fs-xl);font-weight:700;color:var(--fg-strong)}
+.se-metric.good .se-metric-v{color:var(--pass)}.se-metric.warn .se-metric-v{color:var(--warn)}.se-metric.bad .se-metric-v{color:var(--fail)}
+.se-metric-l{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted);text-transform:uppercase;margin-top:3px}
+.se-kv{display:flex;gap:16px;align-items:center;font-size:var(--fs-sm);color:var(--muted)}.se-kv b{color:var(--fg-strong);font-family:var(--mono)}
+.se-note{font-size:var(--fs-sm);color:var(--muted);margin:5px 0 2px;line-height:1.5}
+.se-empty{font-family:var(--mono);font-size:var(--fs-xs);color:var(--dim);border:1px dashed var(--border-strong);padding:12px;text-align:center;margin:6px 0}
+.se-conf-b{font-family:var(--mono);font-size:var(--fs-xs);color:var(--fg);flex:0 0 64px}
+.se-table{width:100%;border-collapse:collapse;margin:8px 0;font-size:var(--fs-xs)}
+.se-table th{text-align:left;font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;color:var(--dim);border-bottom:1px solid var(--border);padding:5px 8px}
+.se-table td{padding:5px 8px;border-bottom:1px solid var(--border);font-family:var(--mono)}
+.se-table td.good{color:var(--pass)}.se-table td.bad{color:var(--fail)}.se-table td.warn{color:var(--warn)}.se-table td.dim{color:var(--dim)}
+.se-cid{color:var(--cyan);max-width:230px;overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
+.se-flag{font-size:var(--fs-2xs);padding:1px 6px;border:1px solid}.se-flag.good{color:var(--pass);border-color:var(--pass)}.se-flag.bad{color:var(--fail);border-color:var(--fail)}.se-flag.warn{color:var(--warn);border-color:var(--warn)}
+.se-vverdict{color:var(--muted)}
+.se-hitl,.se-spot-act{display:flex;gap:4px}
+.se-pick{font-family:var(--mono);font-size:var(--fs-2xs);padding:1px 7px;border:1px solid var(--border);color:var(--muted);cursor:pointer;user-select:none}
+.se-pick:hover{border-color:var(--primary);color:var(--fg)}
+.se-pick.on{border-color:var(--primary);color:var(--primary-soft);background:rgba(126,71,215,.14);font-weight:700}
+.se-blocked{border:1px solid var(--fail);background:var(--fail-bg);padding:9px 12px;margin:6px 0;font-size:var(--fs-sm);color:var(--fg);line-height:1.5}
+.se-blocked-tag{font-family:var(--mono);font-size:var(--fs-2xs);font-weight:700;color:var(--fail);border:1px solid var(--fail);padding:1px 6px;margin-right:6px}
+.se-emit{border:1px solid var(--border);background:var(--surf-2);padding:10px 12px;margin:6px 0}
+.se-emit-l{font-family:var(--mono);font-size:var(--fs-xs);color:var(--fg);flex:0 0 160px}
+.se-spots{display:flex;flex-direction:column;gap:5px;margin:6px 0}
+.se-spot{display:flex;gap:9px;align-items:center;flex-wrap:wrap;border:1px solid var(--border);border-left:3px solid var(--warn);background:var(--surf);padding:6px 10px;font-size:var(--fs-xs)}
+.se-spot-t{color:var(--cyan);flex:0 0 auto}.se-spot-c{color:var(--fg);flex:1 1 200px}.se-spot-conf{font-family:var(--mono);color:var(--muted)}.se-spot-why{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--warn)}.se-spot-act{margin-left:auto}
+.se-future{border:1px dashed var(--border-strong);background:var(--surf-2);padding:10px 13px;margin:6px 0;opacity:.92}
+.se-future-h{font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;letter-spacing:.05em;color:var(--dim);margin-bottom:5px}
+.se-future-b{font-size:var(--fs-sm);color:var(--muted);line-height:1.55}
+/* §5 self-eval health */
+.hsc{display:grid;grid-template-columns:repeat(3,1fr);gap:1px;background:var(--border);margin-top:10px}.hsc .hc{background:var(--surf);padding:10px 12px}.hsc .hc .l{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted);text-transform:uppercase}.hsc .hc .v{font-size:var(--fs-xl);font-weight:700;margin-top:3px;font-family:var(--mono)}.hsc .hc .v.good{color:var(--pass)}.hsc .hc .v.warn{color:var(--warn)}
+/* UI-12-B — honest grounded health: rate-coloured (low=critical), capture-na=muted (never green) */
+.hsc .hc .v.rate-ok{color:var(--pass)}.hsc .hc .v.rate-mid{color:var(--warn)}.hsc .hc .v.rate-low{color:var(--fail)}
+.hsc .hc .v.capture-na{color:var(--dim);font-size:var(--fs-sm);font-weight:600;cursor:help}
+.copy-md{font-family:var(--mono);font-size:var(--fs-xs);padding:4px 10px;border:1px solid var(--border);background:var(--surf-2);color:var(--fg);cursor:pointer;margin-top:8px}
+.copy-md:hover{border-color:var(--cyan);color:var(--cyan)}
+.copy-md.copied{border-color:var(--pass);color:var(--pass)}
+.copy-md.copyfail{border-color:var(--warn);color:var(--warn)}
+/* WS-4 — refined EV-051 "routed to diagnostics" handover block */
+.hov-meta{display:flex;flex-wrap:wrap;gap:7px;margin:8px 0 10px}
+.hov-meta .hm-c{display:flex;flex-direction:column;gap:1px;border:1px solid var(--border);border-left:3px solid var(--primary);background:var(--surf);padding:5px 10px;min-width:90px}
+.hov-meta .hm-c .l{font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;letter-spacing:.04em;color:var(--cyan)}
+.hov-meta .hm-c .v{font-size:var(--fs-sm);color:var(--fg-strong)}
+.hov-list{display:flex;flex-direction:column;gap:8px;margin:8px 0}
+.hov{border:1px solid var(--border);border-left:3px solid var(--fail);background:var(--surf)}
+.hov-h{display:flex;gap:8px;align-items:center;flex-wrap:wrap;padding:7px 11px;border-bottom:1px solid var(--border);background:var(--surf-2)}
+.hov-h .hov-n{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted)}
+.hov-h b{font-family:var(--mono);font-size:var(--fs-sm);color:var(--fg-strong)}
+.hov-copy{margin:0 0 0 auto;padding:2px 9px;font-size:var(--fs-2xs)}
+.hov-g{display:grid;grid-template-columns:64px 1fr;gap:5px 11px;padding:9px 11px}
+.hov-g .hk{font-family:var(--mono);font-size:var(--fs-2xs);font-weight:600;text-transform:uppercase;letter-spacing:.04em;color:var(--cyan);padding-top:2px}
+.hov-g .hv{font-size:var(--fs-sm);color:var(--fg);line-height:1.5;min-width:0;word-break:break-word}
+.hov-dim{color:var(--muted);font-family:var(--mono);font-size:var(--fs-2xs)}
+/* §9.4.4 R3 per-tab description */
+.tabdesc{border:1px solid var(--border);border-left:3px solid var(--cyan);background:var(--surf-2);padding:7px 11px;font-size:var(--fs-xs);color:var(--fg);opacity:.92;line-height:1.5;margin:8px 0 12px}
+.tabdesc .td-i{color:var(--cyan);font-weight:700;margin-right:5px}
+/* §9.4.4 R4 criterion provenance + hover + statement */
+.prov{font-family:var(--mono);font-size:var(--fs-2xs);font-weight:700;padding:1px 6px;border:1px solid;text-transform:uppercase;cursor:help}
+.prov.defined{color:var(--cyan);border-color:var(--cyan);background:rgba(69,184,204,.10)}
+.prov.source{color:var(--primary-soft);border-color:var(--primary);background:rgba(126,71,215,.12)}
+/* UI-10 — CODE vs JUDGE method chip (how a criterion is evaluated) */
+.method{font-family:var(--mono);font-size:var(--fs-2xs);font-weight:700;padding:1px 6px;border:1px solid;text-transform:uppercase;letter-spacing:.03em;cursor:help;white-space:nowrap}
+.method.code{color:var(--warn);border-color:var(--warn);background:rgba(214,158,46,.12)}
+.method.hybrid{color:var(--cyan);border-color:var(--cyan);background:rgba(69,184,204,.10)}
+.method.judge{color:var(--muted);border-color:var(--border-strong);background:var(--surf-2)}
+.ihover{color:var(--cyan);font-size:var(--fs-xs);cursor:help;border-bottom:1px dotted var(--cyan)}
+.cstmt{font-size:var(--fs-xs);color:var(--muted);margin-top:3px;line-height:1.4;white-space:normal}
+td .cstmt{font-family:var(--sans,inherit)}
+/* §9.4.4 R3/M5 detected-but-unmatched */
+.fails-h,.detected-h{font-size:var(--fs-sm);margin-top:14px}
+.detected-h{color:var(--warn)}
+.detected{border:1px solid var(--warn);border-left:3px solid var(--warn);background:var(--warn-bg);margin-top:8px;padding:7px 11px}
+.detected-top{display:flex;gap:8px;align-items:baseline;flex-wrap:wrap}.detected-top b{font-size:var(--fs-sm);color:var(--fg-strong)}
+.dkind{font-family:var(--mono);font-size:var(--fs-2xs);font-weight:700;color:var(--warn);border:1px solid var(--warn);padding:1px 6px;text-transform:uppercase}
+.dtrace{font-size:var(--fs-2xs);color:var(--muted);margin-left:auto}
+.detected-ev{font-size:var(--fs-2xs);color:var(--muted);margin-top:4px}.detected-ev .ref{color:var(--cyan)}
+/* §9.4.4 R2/M1 internal subject-profile card */
+.profile-grid{display:grid;grid-template-columns:max-content 1fr;gap:1px;background:var(--border);border:1px solid var(--border);margin:8px 0}
+.profile-grid .pk{background:var(--surf);font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted);text-transform:uppercase;padding:6px 10px}
+.profile-grid .pv{background:var(--surf-2);font-size:var(--fs-xs);color:var(--fg);padding:6px 11px}
+/* §9.4.4 R2/M4 phase-block lens */
+.phase-lens{display:grid;grid-template-columns:repeat(auto-fill,minmax(220px,1fr));gap:6px;margin:8px 0}
+.phase-blk{border:1px solid var(--border);background:var(--surf);padding:7px 9px}
+.phase-h{display:flex;gap:6px;align-items:center;margin-bottom:4px}.phase-n{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted);margin-left:auto}
+.ps-row{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--fg);opacity:.88;line-height:1.4;margin-top:3px;border-top:1px solid var(--border);padding-top:3px}
+.phase-h .k.gather{color:var(--cyan);background:rgba(69,184,204,.10)}.phase-h .k.expect{color:var(--primary-soft);background:rgba(126,71,215,.12)}
+/* §9.4.4 R2/M2+M3 reasoning example */
+.re-blk{border:1px solid var(--border);background:var(--surf-2);padding:8px 11px;margin-top:8px}
+.re-h{font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;color:var(--cyan);margin-bottom:5px}
+.re-rephrase{font-size:var(--fs-sm);color:var(--fg-strong);font-style:italic;line-height:1.5}
+.re-sub{font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;color:var(--muted);margin-top:6px}
+.re-ul{margin:3px 0;padding-left:18px}.re-ul li{font-size:var(--fs-xs);margin:2px 0;line-height:1.4}.re-rat{color:var(--muted)}
+/* UI-2 — verdict-state legend (Overview + Scorecard) */
+.vlegend{border:1px solid var(--border);border-left:3px solid var(--cyan);background:var(--surf-2);padding:8px 11px;margin:10px 0}
+.vlegend .vl-h{font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;letter-spacing:.05em;color:var(--cyan);margin-bottom:6px}
+.vlegend .vl-item{display:flex;gap:9px;align-items:baseline;margin:4px 0;line-height:1.45}
+.vlegend .vl-k{font-family:var(--mono);font-size:var(--fs-2xs);font-weight:700;padding:1px 7px;border:1px solid;min-width:108px;text-align:center;flex:0 0 auto}
+.vlegend .vl-v{font-size:var(--fs-xs);color:var(--fg);opacity:.92}
+.vlegend .vl-k.pass{color:var(--pass);border-color:var(--pass);background:var(--pass-bg)}
+.vlegend .vl-k.fail{color:var(--fail);border-color:var(--fail);background:var(--fail-bg)}
+.vlegend .vl-k.inc{color:var(--muted);border-color:var(--border-strong);background:var(--surf-3)}
+.vlegend .vl-k.indet{color:var(--warn);border-color:var(--warn);background:var(--warn-bg)}
+.vlegend .vl-k.na{color:var(--dim);border-color:var(--border);background:var(--surf)}
+/* UI-3 — heatmap horizontal-scroll container (no page overflow) */
+.hm-scroll{overflow-x:auto;max-width:100%;padding-bottom:2px}
+.hm .colh{word-break:break-word;white-space:normal}
+/* UI-5 — success-rate + key-metrics summary row under the heatmap */
+.sc-metrics{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:6px;margin:10px 0 6px}
+.scm{border:1px solid var(--border);border-top:3px solid var(--border-strong);background:var(--surf);padding:9px 11px}
+.scm-v{font-family:var(--mono);font-size:var(--fs-2xl);font-weight:700;line-height:1.05;color:var(--fg-strong)}
+.scm-l{font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;letter-spacing:.04em;color:var(--cyan);margin-top:3px}
+.scm-n{font-size:var(--fs-2xs);color:var(--dim);margin-top:3px;line-height:1.4}
+.scm.rate-ok{border-top-color:var(--pass)}.scm.rate-ok .scm-v{color:var(--pass)}
+.scm.rate-mid{border-top-color:var(--warn)}.scm.rate-mid .scm-v{color:var(--warn)}
+.scm.rate-low{border-top-color:var(--fail)}.scm.rate-low .scm-v{color:var(--fail)}
+.scm.rate-na{border-top-color:var(--border)}
+/* UI-12-B — capture-unavailable grounded tile: long honest label, not a big green number */
+.scm.pending .scm-v{font-size:var(--fs-sm);color:var(--dim);word-break:break-word}
+/* UI-7 — criterion-definition meta line */
+.nest .row.defn-meta{color:var(--cyan);font-size:var(--fs-2xs);text-transform:uppercase;letter-spacing:.03em}
+/* UI-8 — calibrate groups (mutually-exclusive radio semantics) */
+.cal .calgroup{display:inline-flex;gap:5px;align-items:center}
+/* UI-4 — judge-reasoning bands in the §2 drill (why this verdict + full why-chain) */
+.band.whyverdict{border-left:3px solid var(--primary);background:rgba(126,71,215,.08)}
+.band.whyverdict .wv-t{font-size:var(--fs-sm);color:var(--fg);line-height:1.5;margin-top:4px}
+.ctx.whychain{margin:8px 0}
+.ctx.whychain .wc-body{padding:7px 11px;display:flex;flex-direction:column;gap:2px}
+.why-note{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted);margin-top:4px;line-height:1.45}.why-note .ref{color:var(--cyan)}
+/* WS-5 — indeterminate resolution chain (§3): per-criterion applicable·pass·fail·N/A·needs-evidence */
+.rztable{display:flex;flex-direction:column;gap:8px;margin:8px 0}
+.rzrow{border:1px solid var(--border);border-left:3px solid var(--border-strong);background:var(--surf);padding:8px 11px}
+.rzrow.crit,.rzrow.high{border-left-color:var(--fail)}.rzrow.med{border-left-color:var(--warn)}.rzrow.low{border-left-color:var(--border-strong)}
+.rz-h{display:flex;gap:8px;align-items:baseline;flex-wrap:wrap;margin-bottom:5px}
+.rz-h b{font-size:var(--fs-sm);color:var(--fg-strong)}
+.rz-rate{margin-left:auto;font-family:var(--mono);font-size:var(--fs-2xs)}
+.rz-stats{display:flex;flex-wrap:wrap;gap:6px;align-items:center}
+.rz{font-family:var(--mono);font-size:var(--fs-2xs);padding:2px 7px;border:1px solid var(--border)}
+.rz.applicable{color:var(--fg-strong);border-color:var(--border-strong)}
+.rz.pass{color:var(--pass);border-color:var(--pass)}
+.rz.fail{color:var(--fail);border-color:var(--fail)}
+.rz.na{color:var(--cyan);border-color:var(--cyan);background:rgba(108,196,212,.08)}
+.rz.ne{color:var(--warn);border-color:var(--warn);background:rgba(214,158,46,.08)}
+.rz.zero{color:var(--dim);border-color:var(--border);background:none}
+.rz-reason{display:inline-block;font-size:var(--fs-2xs);color:var(--muted);font-style:italic;margin-left:4px}
+.rz-act{display:inline-flex;flex-direction:column;font-family:var(--mono);font-size:var(--fs-2xs);color:var(--warn);border:1px dashed var(--warn);padding:2px 7px}
+.rz-act .rz-reason{color:var(--muted);font-family:var(--sans,inherit)}
+/* WS-3 — one-line plain-language gloss (what it measures · why it matters), muted,
+   sits ABOVE the technical block in BOTH §3 scorecard subcards and §4 finding cards */
+.plain{font-size:var(--fs-xs);color:var(--muted);line-height:1.5;margin:6px 0 8px;padding:6px 9px;border-left:2px solid var(--cyan);background:rgba(108,196,212,.06);font-style:italic}
+/* WS-3 follow-up — per-criterion plain-language KEY directly under the §3 heatmap (ALL criteria) */
+.plain-legend{margin:8px 0 4px}
+.plain-legend .pl-h{font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;letter-spacing:.04em;color:var(--cyan);margin:0 0 5px}
+.plain-leg{display:flex;gap:8px;align-items:baseline;margin:0 0 4px;flex-wrap:wrap}
+.plain-leg .pl-id{font-style:normal;color:var(--fg-strong);font-size:var(--fs-2xs);flex:0 0 auto}
+.plain-leg .pl-tx{font-style:italic;color:var(--muted);flex:1 1 240px;min-width:0}
+/* WS-1 — "how the judge reasoned" per-criterion critique block (side-by-side + scorecard drill) */
+.cvblock{margin:8px 0}
+.cvblock .cvbody{display:flex;flex-direction:column;gap:6px;padding:7px 9px}
+.cvrow{border:1px solid var(--border);border-left:3px solid var(--border-strong);background:var(--surf);padding:7px 9px}
+.cvrow.pass{border-left-color:var(--pass)}.cvrow.fail{border-left-color:var(--fail)}
+.cvrow.uncertain,.cvrow.indeterminate{border-left-color:var(--warn)}.cvrow.na{border-left-color:var(--border)}
+.cvh{display:flex;gap:7px;align-items:center;flex-wrap:wrap;margin-bottom:3px}
+.cvb{font-family:var(--mono);font-size:var(--fs-2xs);text-transform:uppercase;letter-spacing:.03em;padding:1px 6px;border:1px solid var(--border-strong)}
+.cvb.pass{color:var(--pass);border-color:var(--pass)}.cvb.fail{color:var(--fail);border-color:var(--fail)}
+.cvb.uncertain,.cvb.indeterminate{color:var(--warn);border-color:var(--warn)}.cvb.na{color:var(--dim)}
+.cvid{font-family:var(--mono);font-size:var(--fs-xs);color:var(--fg-strong)}
+.cvconf{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--dim);margin-left:auto}
+.cvcrit{font-size:var(--fs-sm);color:var(--fg);line-height:1.5}.cvcrit.dim{color:var(--dim)}
+.cvrefs{display:flex;flex-wrap:wrap;gap:5px;margin-top:5px}
+.cvref{font-family:var(--mono);font-size:var(--fs-2xs);color:var(--cyan);border:1px solid var(--border);padding:1px 5px;word-break:break-all}
+`;
+
+// ── the embedded client JS (ledger filter/search + drill + tabs + toggles) ───
+
+/**
+ * The embedded client JS — ledger filter/search + click-row drill (side-by-side
+ * when judge_steps exist, else the per-trajectory scorecard) + tab switching +
+ * calibration/alignment toggles + copy-as-markdown. The data (ledger `L`,
+ * criteria `C`, judge-walks `J`) is injected as deterministic JSON.
+ */
+function clientScript(ledgerJson: string, critJson: string, walksJson: string, resMetaJson: string): string {
+  return `
+const L=${ledgerJson},C=${critJson},J=${walksJson},RES=${resMetaJson};let F='all';
+function resBadge(res){var m=RES[res];if(!m)return '';return '<span class="resbadge '+m.cls+'" title="'+esc(m.routing)+'">'+esc(m.badge)+'</span>';}
+function resRouting(res){var m=RES[res];if(!m)return '';return '<div class="routing '+m.cls+'"><span class="routing-k">resolution</span><span class="routing-b '+m.cls+'">'+esc(m.badge)+'</span><span class="routing-v">'+esc(m.routing)+'</span></div>';}
+function esc(s){return String(s==null?'':s).replace(/&/g,'&amp;').replace(/</g,'&lt;').replace(/>/g,'&gt;').replace(/"/g,'&quot;');}
+function wireTabs(){var btns=document.querySelectorAll('nav.tabs button'),panels=document.querySelectorAll('main .panel');
+  btns.forEach(function(b){b.addEventListener('click',function(){var key=b.getAttribute('data-tab');
+    btns.forEach(function(x){x.classList.remove('active');});panels.forEach(function(p){p.classList.remove('active');});
+    b.classList.add('active');var p=document.getElementById(key);if(p)p.classList.add('active');});});}
+function cnt(c){var p=0,f=0,n=0;for(var k in c){if(c[k]=='pass')p++;else if(c[k]=='fail')f++;else if(c[k]=='uncertain'||c[k]=='indeterminate')n++;}return[p,f,n];}
+function render(){var qi=document.getElementById('q');if(!qi)return;var q=(qi.value||'').toLowerCase();
+  var rows=L.filter(function(r){return (F=='all'||r.v==F)&&(!q||r.t.toLowerCase().indexOf(q)>=0||(r.r||'').toLowerCase().indexOf(q)>=0);});
+  var cntEl=document.getElementById('cnt');if(cntEl)cntEl.textContent=rows.length+' / '+L.length+' trajectories';
+  var cap=rows.slice(0,150);
+  document.getElementById('lrows').innerHTML=cap.map(function(r){var c=cnt(r.c),p=c[0],f=c[1],n=c[2];
+    return '<tr data-id="'+esc(r.t)+'" class="vrow '+esc(r.v)+'"><td class="tid">'+esc(r.t.slice(0,14))+(r.js?' <span class=haswalk>✦</span>':'')+'</td><td>'+esc(r.r||'all')+'</td><td><span class="mini '+esc(r.v)+'">'+esc(r.v)+'</span></td><td>'+resBadge(r.res)+'</td><td>'+p+'</td><td>'+f+'</td><td>'+n+'</td><td style="color:var(--fail)">'+esc((r.f||[]).join(', '))+'</td></tr>';}).join('');
+  var virt=document.getElementById('virt');if(virt)virt.textContent=rows.length>150?('▾ showing 150 of '+rows.length+' — filter/search to narrow'):'';
+  document.querySelectorAll('table.ledger tbody tr').forEach(function(tr){tr.addEventListener('click',function(){drill(tr.getAttribute('data-id'),tr);});});
+  /* Surface the Agent‖Judge side-by-side BY DEFAULT: on the first render, auto-open the
+     first JUDGED trace (one that has a judge walk) so the two-lane trajectory shows without
+     a click. Once only — never hijacks the drill after the user starts navigating. */
+  if(!window._autoDrilled){var jr=null;for(var i=0;i<cap.length;i++){if(cap[i].js){jr=cap[i];break;}}jr=jr||cap[0];if(jr){var atr=document.querySelector('table.ledger tbody tr[data-id="'+jr.t+'"]');if(atr){window._autoDrilled=true;drill(jr.t,atr);}}}}
+// WS-1 — the "How the Judge Reasoned" per-criterion block: result badge +
+// critique-before-verdict + grounding refs, read off the verdict file's verdicts[].
+// Shared by BOTH the side-by-side drill (walk traces) and the per-trajectory
+// scorecard drill (no-walk traces) so NO judged trace renders an empty panel.
+function cvRefs(refs){if(!refs||!refs.length)return '';
+  return '<div class="cvrefs">'+refs.map(function(rf){var s=(typeof rf==='string')?rf:[rf.obs,rf.path,rf.value].filter(Boolean).join(' · ');return s?'<span class="cvref">'+esc(s)+'</span>':'';}).join('')+'</div>';}
+function critiqueBlock(cvs){if(!cvs||!cvs.length)return '';
+  var order={fail:0,uncertain:1,indeterminate:1,pass:2,na:3};
+  var rows=cvs.slice().sort(function(a,b){return (order[a.result]==null?9:order[a.result])-(order[b.result]==null?9:order[b.result]);}).map(function(v){
+    var res=v.result||'na';var disp=res==='uncertain'?'indeterminate':res;
+    var conf=(v.confidence!=null)?'<span class="cvconf">conf '+esc(v.confidence)+(v.confidenceBand?' · '+esc(v.confidenceBand):'')+'</span>':'';
+    var crit=v.critique?'<div class="cvcrit">'+esc(v.critique)+'</div>':'<div class="cvcrit dim">— no critique recorded for this criterion</div>';
+    return '<div class="cvrow '+esc(res)+'"><div class="cvh"><span class="cvb '+esc(res)+'">'+esc(disp)+'</span><span class="cvid">'+esc(v.criterionId||'')+'</span>'+conf+'</div>'+crit+cvRefs(v.refs)+'</div>';}).join('');
+  return '<div class="ctx cvblock"><div class="ctx-h">◇ how the judge reasoned — per-criterion verdict · critique-before-verdict · grounding refs ('+cvs.length+')</div><div class="cvbody">'+rows+'</div></div>';}
+function sideBySide(d){var ctx=d.context||{};
+  // Gap A — the §2 "input + scenario" cell renders the RAW triggering INPUT (the
+  // thing that fired the agent) ABOVE the judge's scenario LABEL. Long inputs
+  // collapse into a <details> (lean, no JS) so the cell stays compact; short ones
+  // render inline (clamped). ABSENT input ⇒ the cell shows the scenario alone (or
+  // "—"). Font stays at the --fs-2xs (11px) floor + brand mono.
+  var inputCell=function(raw,scen){
+    var sc=scen?'<div class="iscn">scenario · '+esc(scen)+'</div>':'';
+    if(!raw)return (scen?'<div class="ival">'+esc(scen)+'</div>':'—');
+    var long=String(raw).length>180;
+    var body=long
+      ? '<details class="iexp"><summary>raw input · '+String(raw).length+' chars</summary><pre class="iraw">'+esc(raw)+'</pre></details>'
+      : '<pre class="iraw clamp">'+esc(raw)+'</pre>';
+    return body+sc;};
+  var refStr=function(rf){if(!rf)return '';if(typeof rf==='string')return rf;return [rf.obs,rf.path,rf.value].filter(Boolean).join(':');};
+  // UI-4 — surface the judge's REASONING (why the verdict + why the exit-states were
+  // concluded) from the EXISTING judge walk: an ordered why-chain band + the decide/bind
+  // text inlined where it explains a conclusion. No emit-contract change — read-only over
+  // the judge_steps already on the verdict file.
+  var KORD={gather:0,context:1,examine:2,detect:3,bind:4,ground:5,critique:6,decide:7,verify:8};
+  var allJs=(d.judgeSteps||[]).slice();
+  var stepText=function(kind){return allJs.filter(function(s){return s.kind===kind;}).map(function(s){return s.text||'';}).filter(Boolean).join(' · ');};
+  var decideWhy=stepText('decide')||stepText('critique');
+  var stateWhy=[stepText('bind'),stepText('gather')].filter(Boolean).join(' · ');
+  var chainSteps=allJs.slice().sort(function(a,b){var x=KORD[a.kind];var y=KORD[b.kind];return (x==null?9:x)-(y==null?9:y);});
+  var whyChain=chainSteps.length?('<div class="ctx whychain"><div class="ctx-h">◇ judge reasoning — full why-chain (gather → bind → ground → decide)</div><div class="wc-body">'+chainSteps.map(function(s){var rs=refStr(s.ref);return '<div class="jstep"><span class="k '+esc(s.kind)+'">'+esc(s.kind)+'</span><span class="t">'+esc(s.text||'')+(rs?' <span class=ref>'+esc(rs)+'</span>':'')+'</span></div>';}).join('')+'</div></div>'):'';
+  var verdictWhy=decideWhy?('<div class="band whyverdict"><div class="bh" style="color:var(--primary-soft)">◇ why this verdict</div><div class="wv-t">'+esc(decideWhy)+'</div></div>'):'';
+  // -- §2 judge lane = per-step EVAL COVERAGE --
+  // The judge lane (.step-r) used to filter judgeSteps by anchor === a.n -- but
+  // judges never emit an anchor, so EVERY step rendered a bare dash. Instead we
+  // map each agent step -> the per-criterion verdicts (d.criterionVerdicts) whose
+  // grounding refs EXAMINED that step: a precise ref.obs === step.obs match plus
+  // the tool-name fallback (ref.path === 'name' && ref.value === step.tool).
+  // Each examining criterion renders one compact entry -- result + CODE/JUDGE tag
+  // + criterionId + the judge reasoning (critique). A step no criterion references
+  // says 'not examined by any eval' (honest), never a bare dash. Any judge step
+  // that DOES carry a real anchor is still honored (future-proof).
+  var cvAll=(d.criterionVerdicts||[]);
+  var refExaminesStep=function(rf,a){
+    if(!rf||typeof rf==='string')return false;
+    if(a.obs&&rf.obs&&String(rf.obs)===String(a.obs))return true;
+    if(rf.path==='name'&&rf.value!=null&&a.tool&&String(rf.value)===String(a.tool))return true;
+    return false;};
+  var covEntry=function(v,a){
+    var res=v.result||'na';var disp=res==='uncertain'?'indeterminate':res;
+    // the router carries TWO vocabularies: matrix-derived ('deterministic') and
+    // mined ('code-based'); 'hybrid' is shared. A '[code-eval ...]'-prefixed critique
+    // is the deterministic-logic fallback when the method is unset.
+    var method=((typeof C!=='undefined'&&C[v.criterionId])||{}).m||'';
+    var isCode=method==='deterministic'||method==='code-based'||method==='hybrid'||String(v.critique||'').trim().indexOf('[code-eval')===0;
+    var tag=method==='hybrid'?'HYBRID':(isCode?'CODE':'JUDGE');
+    var matched=(v.refs||[]).filter(function(rf){return typeof rf!=='string'&&refExaminesStep(rf,a);});
+    var crit=v.critique?esc(v.critique):'<span class="dim">— no critique recorded</span>';
+    return '<div class="jcov '+esc(res)+'"><div class="jcov-h"><span class="cvb '+esc(res)+'">'+esc(disp)+'</span><span class="jm '+(isCode?'code':'judge')+'" title="'+(isCode?'deterministically checked by a code-eval (not LLM-judged)':'reasoned by the LLM judge')+'">'+esc(tag)+'</span><span class="jcid">'+esc(v.criterionId||'')+'</span></div><div class="jcrit">'+crit+'</div>'+cvRefs(matched)+'</div>';};
+  var rowsHtml='';(d.agentSteps||[]).forEach(function(a){
+    // future-proof: a judge step that carries a REAL anchor still renders.
+    var js=(d.judgeSteps||[]).filter(function(s){return s.anchor!=null&&String(s.anchor)===String(a.n)&&s.kind!=='context';});
+    var anchoredHtml=js.map(function(s){var rs=refStr(s.ref);return '<div class="jstep"><span class="k '+esc(s.kind)+'">'+esc(s.kind)+'</span><span class="t">'+esc(s.text||'')+' '+(rs?'<span class=ref>'+esc(rs)+'</span>':'')+'</span></div>';}).join('');
+    var examiners=cvAll.filter(function(v){return (v.refs||[]).some(function(rf){return refExaminesStep(rf,a);});});
+    var covHtml=examiners.map(function(v){return covEntry(v,a);}).join('');
+    var jhtml=anchoredHtml+covHtml;
+    if(!jhtml)jhtml='<div class="jstep noexam"><span class="t">— not examined by any eval</span></div>';
+    rowsHtml+='<div class="step-l"><div class="evb"><div class="top"><span class="tool">'+esc(a.tool||'')+'</span><span class="st '+esc(a.status||'')+'">'+esc(a.status||'')+'</span></div><div class="det">'+esc(a.detail||'')+'</div></div></div><div class="node '+esc(a.status||'')+'"><div class="ln"></div><div class="n">'+esc(a.n)+'</div><div class="ln"></div></div><div class="step-r"><div class="evb r">'+jhtml+'</div></div>';});
+  var h=d.health||{};
+  var sp=d.subjectProfile;
+  var spHtml=sp?'<div class="ctx"><div class="ctx-h">◇ judge · subject profile (M1) · '+esc(sp.provenance||'')+'</div><div class="ctx-g"><div class="ctx-c"><div class="l">identity</div><div class="v">'+esc(sp.identity||'—')+'</div></div><div class="ctx-c"><div class="l">purpose</div><div class="v">'+esc(sp.purpose||'—')+'</div></div><div class="ctx-c"><div class="l">scope · harness</div><div class="v">'+esc((sp.scope||'—'))+' · harness: '+esc(sp.harness||'—')+'</div></div></div></div>':'';
+  var u=d.understanding;
+  var uHtml=u?'<div class="band"><div class="bh" style="color:var(--cyan)">◇ node-0 GATHER — understanding (M2)</div><div class="re-rephrase" style="margin-top:4px">“'+esc(u.rephrase||'')+'”</div></div>':'';
+  var et=d.expectedTrajectory||[];
+  var etHtml=et.length?'<div class="band"><div class="bh" style="color:var(--primary-soft)">◇ node-0.5 EXPECTED-TRAJECTORY (M3) — how it SHOULD have acted</div><ol class="re-ul" style="margin-top:4px">'+et.map(function(s,i){return '<li><b>'+esc(s.step||i+1)+'.</b> '+esc(s.expected||'')+(s.rationale?' <span class=re-rat>— '+esc(s.rationale)+'</span>':'')+'</li>';}).join('')+'</ol></div>':'';
+  return '<div class="drillbox"><div style="display:flex;gap:9px;align-items:center;flex-wrap:wrap;margin-bottom:6px"><b class="mono">'+esc(d.traceId)+'</b><span class="chip">'+esc(d.route||'all')+'</span><span class="verd '+(d.verdict=='FAIL'?'fail':d.verdict=='PASS'?'pass':'inc')+'">'+esc(d.verdict)+'</span></div>'+
+    resRouting(d.res||'judge-walk')+
+    spHtml+
+    verdictWhy+
+    critiqueBlock(d.criterionVerdicts)+
+    '<div class="ctx"><div class="ctx-h">◇ judge · gather context</div><div class="ctx-g"><div class="ctx-c"><div class="l">harness</div><div class="v">'+esc(ctx.harness||'—')+'</div></div><div class="ctx-c"><div class="l">input + scenario</div><div class="v">'+inputCell(d.input,ctx.scenario)+'</div></div><div class="ctx-c"><div class="l">exit states</div><div class="v">'+esc(ctx.exitStates||'—')+(stateWhy?'<div class="why-note"><span class="ref">why concluded:</span> '+esc(stateWhy)+'</div>':'')+'</div></div></div></div>'+
+    whyChain+
+    uHtml+etHtml+
+    '<div class="lanehdr"><div class="a">target agent — what it did</div><div class="x">step</div><div class="j">judge — eval coverage (which criteria examined this step)</div></div>'+
+    '<div class="grid2">'+rowsHtml+'<div class="band loc"><div class="bh">↯ localize (root, not symptom)</div><div style="font-size:var(--fs-sm);margin-top:4px">'+esc(d.localize||'—')+'</div></div></div>'+
+    '<div class="health"><div class="hc"><div class="l">context</div><div class="v good">'+(h.contextGathered?'✓':'—')+'</div></div><div class="hc"><div class="l">grounded</div><div class="v good">'+esc(h.grounded||0)+'</div></div><div class="hc"><div class="l">assumed</div><div class="v '+((h.assumed||0)>0?'warn':'good')+'">'+esc(h.assumed||0)+'</div></div><div class="hc"><div class="l">root vs symptom</div><div class="v good">'+(h.stoppedAtSymptom?'symptom':'✓ root')+'</div></div></div></div>';}
+function scorecard(r){var order={fail:0,uncertain:1,indeterminate:1,pass:2,na:3};
+  var cells=Object.keys(r.c).sort(function(a,b){return (order[r.c[a]]||9)-(order[r.c[b]]||9);}).map(function(k){var v=r.c[k];var disp=v=='uncertain'?'indeterminate':v;
+    return '<div class="scc '+esc(v)+'"><span class="nm" title="'+esc((C[k]||{}).n||k)+'">'+esc(k)+'</span><span style="margin-left:auto;color:var(--muted)">'+esc(disp)+'</span></div>';}).join('');
+  return '<div class="drillbox"><div style="display:flex;gap:9px;align-items:center;margin-bottom:6px"><b class="mono">'+esc(r.t)+'</b><span class="chip">'+esc(r.r||'all')+'</span><span class="verd '+(r.v=='PASS'?'pass':r.v=='FAIL'?'fail':'inc')+'">'+esc(r.v)+'</span></div>'+
+    resRouting(r.res||'judged-walk-not-captured')+
+    (r.root?'<div class="note" style="margin:0 0 8px"><span class="tag">★ ROOT</span>&nbsp;'+esc(r.root)+'</div>':'')+
+    (r.g?'<div class="mono" style="font-size:var(--fs-xs);color:var(--muted);margin-bottom:4px">grounding: '+esc(r.g)+'</div>':'')+
+    '<div style="font-family:var(--mono);font-size:var(--fs-2xs);color:var(--muted);text-transform:uppercase;margin:6px 0">per-trajectory scorecard — '+Object.keys(r.c).length+' criteria</div>'+
+    '<div class="scgrid">'+cells+'</div>'+
+    // WS-1 — no-walk traces still surface their judge reasoning (per-criterion
+    // critique + refs) here, so a trace WITHOUT a judge_steps walk is never an
+    // empty "How the Judge Reasoned" panel when its verdict file carries verdicts.
+    critiqueBlock(r.cv)+'</div>';}
+function drill(id,el){document.querySelectorAll('table.ledger tr').forEach(function(t){t.classList.remove('sel');});if(el)el.classList.add('sel');
+  var box=document.getElementById('drill');if(!box)return;
+  var lrow=L.filter(function(x){return x.t==id;})[0];
+  if(J[id]){var d=J[id];
+    // WS-1 — graft the per-criterion verdicts (carried on the ledger row for EVERY
+    // trace) onto the walk object so the side-by-side drill also shows the per-
+    // criterion critique + refs, not just the step-anchored judge walk.
+    if(lrow&&(!d.criterionVerdicts||!d.criterionVerdicts.length))d.criterionVerdicts=lrow.cv||[];
+    if(lrow&&!d.res)d.res=lrow.res;
+    box.innerHTML=sideBySide(d);}
+  else if(lrow){box.innerHTML=scorecard(lrow);}
+  box.scrollIntoView({behavior:'smooth',block:'nearest'});}
+function wireFilters(){document.querySelectorAll('.lfilter b').forEach(function(b){b.addEventListener('click',function(){F=b.getAttribute('data-flt');document.querySelectorAll('.lfilter b').forEach(function(x){x.classList.remove('on');});b.classList.add('on');render();});});
+  var q=document.getElementById('q');if(q)q.addEventListener('input',render);}
+function wireToggles(){
+  // UI-8 — calibrate tags are MUTUALLY EXCLUSIVE within a group (radio semantics):
+  // selecting one clears the others in the same [data-calgroup]; clicking the active
+  // one again clears it. The agree/revise/refute alignment row stays free-toggle.
+  document.querySelectorAll('.calgroup').forEach(function(g){var btns=[].slice.call(g.querySelectorAll('b'));
+    btns.forEach(function(b){b.addEventListener('click',function(){var wasOn=b.classList.contains('on');
+      btns.forEach(function(x){x.classList.remove('on');});if(!wasOn)b.classList.add('on');});});});
+  document.querySelectorAll('.areview b').forEach(function(b){b.addEventListener('click',function(){b.classList.toggle('on');});});
+  // WS-3 — §5 self-eval HITL picks (5.4 keep/revise/retire · 5.8 agree/disagree):
+  // mutually-exclusive within a [data-vgroup]; click the active one again to clear.
+  var seGroups={};document.querySelectorAll('.se-pick').forEach(function(p){var g=p.getAttribute('data-vgroup');(seGroups[g]=seGroups[g]||[]).push(p);});
+  Object.keys(seGroups).forEach(function(g){var picks=seGroups[g];picks.forEach(function(p){p.addEventListener('click',function(){var wasOn=p.classList.contains('on');picks.forEach(function(x){x.classList.remove('on');});if(!wasOn)p.classList.add('on');});});});
+  document.querySelectorAll('.lnk[data-drill]').forEach(function(l){l.addEventListener('click',function(){var t2=document.querySelector('nav.tabs button[data-tab="t2"]');if(t2)t2.click();setTimeout(function(){drill(l.getAttribute('data-drill'));},60);});});
+  // WS-4 — wire EVERY copy-as-markdown button (the §5 full-handover button, each
+  // per-routed-item button, and any other): each copies its OWN data-md payload and
+  // flashes a "copied ✓" confirmation. Was querySelector (first-only) → querySelectorAll.
+  // Restricted clipboard contexts (headless / some file://) REJECT writeText — the
+  // rejection AND any synchronous throw are CAUGHT so a denied clipboard NEVER raises an
+  // uncaught page error; it just shows a "copy failed" hint and restores. ZERO errors.
+  document.querySelectorAll('.copy-md').forEach(function(cp){cp.addEventListener('click',function(){
+    var md=cp.getAttribute('data-md')||'';
+    var prev=cp.textContent;
+    var done=function(ok){cp.textContent=ok?'copied ✓':'copy failed';cp.classList.add(ok?'copied':'copyfail');
+      setTimeout(function(){cp.textContent=prev;cp.classList.remove('copied');cp.classList.remove('copyfail');},1200);};
+    try{
+      if(navigator.clipboard&&navigator.clipboard.writeText){
+        navigator.clipboard.writeText(md).then(function(){done(true);},function(){done(false);});
+      }else{done(false);}
+    }catch(e){done(false);}
+  });});}
+wireTabs();wireFilters();wireToggles();render();
+`;
+}
+
+// ── the full 5-tab report ────────────────────────────────────────────────────
+
+/**
+ * Render the 5-tab HTML eval-report (the operator-APPROVED component spec). Pure
+ * string assembly; the only I/O is reading the bundled brand assets. The injected
+ * `generatedAt` is the only non-deterministic input → masked-byte-identical.
+ */
+export function renderEvalReport(input: EvalReportInput): string {
+  const theme = readFileSync(join(BRAND_DIR, "theme.css"), "utf8");
+  const wordmark = readFileSync(join(BRAND_DIR, "wordmark.html"), "utf8");
+
+  const headerTitle = "evaluator · Eval Report";
+  const headerMeta =
+    `<span class="mk">subject</span> <span class="mv">${esc(input.subject.name)}</span>` +
+    `<span class="sep">·</span><span class="mk">generated</span> <span class="mv">${esc(input.generatedAt)}</span>`;
+  const header = wordmark.replaceAll("{{HEADER_TITLE}}", esc(headerTitle)).replaceAll("{{HEADER_META}}", headerMeta);
+
+  // deterministic embedded data for the §2 ledger + drill
+  const ledger = input.ledger ?? [];
+  const rootByTrace = new Map((input.topFindings ?? []).filter((tf) => tf.exampleTraceId).map((tf) => [tf.exampleTraceId as string, tf.root ?? ""]));
+  const compactLedger = ledger.map((r) => ({
+    t: r.trajectoryId,
+    r: r.route ?? "all",
+    v: r.verdict,
+    c: r.perCriterion,
+    f: r.failingCriteria,
+    root: rootByTrace.get(r.trajectoryId) ?? "",
+    g: r.grounding ?? "",
+    js: r.judgeSteps && r.judgeSteps.length > 0 ? 1 : 0,
+    // WS-2 — how this trajectory resolved (judge-walk · judged-walk-not-captured ·
+    // truncated) → the ledger resolution badge + the drill routing line.
+    res: r.resolution ?? "",
+    // WS-1 — the per-criterion JUDGE verdicts (result · critique · refs) for THIS
+    // trace. Carried on EVERY ledger row so the §2 drill renders "How the Judge
+    // Reasoned" for every judged trace — the walk traces graft it onto the side-by-
+    // side; the no-walk traces render it under the per-trajectory scorecard. Empty []
+    // only when the verdict file genuinely carried no per-criterion verdict.
+    cv: r.criterionVerdicts ?? [],
+  }));
+  // `m` = the code-vs-judge router (`checkMethod`) so the §2 judge lane can tag a
+  // step's eval-coverage entry CODE (deterministic) / HYBRID / JUDGE (llm).
+  const critMap: Record<string, { n: string; s: string; m: string }> = {};
+  for (const c of input.criteria) critMap[c.id] = { n: c.statement, s: c.severity, m: c.checkMethod ?? "" };
+  const walks: Record<string, unknown> = {};
+  for (const r of ledger) {
+    if (r.judgeSteps && r.judgeSteps.length > 0) {
+      walks[r.trajectoryId] = {
+        traceId: r.trajectoryId,
+        route: r.route ?? "all",
+        verdict: r.verdict,
+        // Gap A — the raw triggering INPUT (ground-truth from the trace), rendered
+        // ABOVE the judge's scenario LABEL in the §2 drill. "" ⇒ drill shows "—".
+        input: r.input ?? "",
+        context: r.context ?? {},
+        agentSteps: r.agentSteps ?? [],
+        judgeSteps: r.judgeSteps ?? [],
+        localize: r.localize ?? "",
+        health: r.health ?? {},
+        // §9.4.4 v2.2 — node-0 understanding · node-0.5 expected-trajectory · M1 profile.
+        understanding: r.understanding ?? null,
+        expectedTrajectory: r.expectedTrajectory ?? [],
+        subjectProfile: r.subjectProfile ?? input.subjectProfile ?? null,
+      };
+    }
+  }
+  const ledgerJson = JSON.stringify(compactLedger);
+  const critJson = JSON.stringify(critMap);
+  const walksJson = JSON.stringify(walks);
+  // WS-2 — the resolution badge label + routing explanation per class (server-derived,
+  // injected so the client row badge + drill routing line read ONE source of truth).
+  const resMetaJson = JSON.stringify({
+    "judge-walk": resolutionMeta("judge-walk"),
+    "judged-walk-not-captured": resolutionMeta("judged-walk-not-captured"),
+    truncated: resolutionMeta("truncated"),
+  });
+
+  return `<!DOCTYPE html>
+<html lang="en" data-theme="dark">
+<head>
+<meta charset="UTF-8">
+<meta name="viewport" content="width=device-width,initial-scale=1">
+<title>${esc(headerTitle)} — ${esc(input.subject.name)}</title>
+<link rel="preconnect" href="https://fonts.googleapis.com">
+<link href="https://fonts.googleapis.com/css2?family=Space+Grotesk:wght@400;500;600;700&family=IBM+Plex+Mono:wght@400;500;600&display=swap" rel="stylesheet">
+<style>
+${theme}
+${REPORT_CSS}
+</style>
+</head>
+<body>
+${header}
+<nav class="tabs">
+  <button class="tab-btn active" data-tab="t1">① Overview</button>
+  <button class="tab-btn" data-tab="t2">② Trajectory · Judge Behaviour</button>
+  <button class="tab-btn" data-tab="t3">③ Eval Scorecard</button>
+  <button class="tab-btn" data-tab="t4">④ Findings</button>
+  <button class="tab-btn internal" data-tab="t5">⑤ Self-Eval [INTERNAL]</button>
+</nav>
+<main>
+  <section class="panel active" id="t1">${overviewTab(input)}</section>
+  <section class="panel" id="t2">${trajectoryTab(input)}</section>
+  <section class="panel" id="t3">${scorecardTab(input)}</section>
+  <section class="panel" id="t4">${findingsTab(input)}</section>
+  <section class="panel" id="t5">${selfEvalTab(input)}</section>
+</main>
+<script data-pii="ledger">
+${clientScript(ledgerJson, critJson, walksJson, resMetaJson)}
+</script>
+</body>
+</html>
+`;
+}
+
+// ── WS-1 file-based re-render entry (models discover's writeDiscoverReportFromFiles) ──
+
+/** Injected I/O for the file-based eval-report entry (keeps the function pure of fs). */
+export interface EvalReportFilesIO {
+  readFile: (p: string) => string;
+  writeFile: (p: string, s: string) => void;
+  readDir: (p: string) => string[];
+}
+
+/** The data-completeness probe the harness reports after a re-render. */
+export interface EvalReportCompleteness {
+  tracesInVerdicts: number;
+  tracesRenderedWithJudgeSteps: number;
+  /** WS-1 — the honest completeness metric: traces whose drill surfaces ≥1 per-
+   *  criterion judge verdict (critique-before-verdict). Should == every JUDGED trace. */
+  tracesRenderedWithCritique: number;
+  /** traces whose verdict file carried renderable judge data (verdicts[] or a walk)
+   *  but the drill would render an EMPTY "How the Judge Reasoned" panel. Goal: 0. */
+  emptyUIrows: number;
+  /** WS-6 — char length of the reconstructed subject system prompt (0 ⇒ UNAVAILABLE). */
+  systemPromptChars: number;
+  /** WS-6 — true when the entity hero renders a reconstructed system prompt. */
+  systemPromptReconstructed: boolean;
+  /** WS-4 — number of true-FAIL criteria routed into the §5 EV-051 diagnostics handover. */
+  routedFailures: number;
+}
+
+/**
+ * WS-1 — re-render the §2 trajectory report DIRECTLY from on-disk artifacts (the
+ * RICH per-trajectory `verdicts/*.json` + the folded `AGG-RESULT.json` + the
+ * criteria/suite), binding the per-criterion judge reasoning (critique + refs) and
+ * the judge_steps walk (where present) to EVERY evaluated trace. This is the eval
+ * analogue of `writeDiscoverReportFromFiles`: a pure shipped-function call over the
+ * file set, with NO strict-schema gate that would drop the rich verdict files (the
+ * verdict files that carry an object `exitStates` / omit `subjectProfile.tools` —
+ * exactly the ones with the walk — survive here, where production's strict
+ * `parseMatrixVerdictFile` would have rejected them).
+ *
+ * Returns the report path + the data-completeness probe. PURE except the injected I/O.
+ */
+export function writeEvalReportFromFiles(
+  params: {
+    verdictsDir: string;
+    aggPath: string;
+    criteriaPath: string;
+    outPath: string;
+    subjectName: string;
+    subjectSource?: string;
+    generatedAt: string;
+    /** WS-6 — OPTIONAL trace batch (or a single representative trace) used ONLY to
+     *  reconstruct the subject SYSTEM PROMPT for the entity hero. The harness streams
+     *  one trace off the raw ndjson and passes it here; the system prompt is identical
+     *  across the batch so one trace suffices. ABSENT ⇒ system prompt stays UNAVAILABLE. */
+    traces?: EvalTrace[];
+  },
+  io: EvalReportFilesIO,
+): { report: string; completeness: EvalReportCompleteness } {
+  // 1) read the RICH verdict files TOLERANTLY (raw JSON.parse — never the strict
+  //    schema gate, which rejects exactly the walk-bearing files). NO trace dropped.
+  const verdictFiles = io
+    .readDir(params.verdictsDir)
+    .filter((f) => f.endsWith(".json"))
+    .map((f) => {
+      try {
+        return JSON.parse(io.readFile(join(params.verdictsDir, f))) as MatrixVerdictFile;
+      } catch {
+        return null;
+      }
+    })
+    .filter((v): v is MatrixVerdictFile => v !== null && typeof (v as { trajectoryId?: unknown }).trajectoryId === "string");
+
+  // 2) the folded scorecard (AGG-RESULT.maskedScorecard = gate + variance) + the
+  //    per-criterion folded verdicts (summary.perCriterion), with a representative
+  //    critique grafted from a matching per-trajectory verdict (for the gating rolls).
+  const agg = JSON.parse(io.readFile(params.aggPath)) as {
+    summary?: { perCriterion?: Array<{ criterionId: string; folded: string; confidence?: number }> };
+    maskedScorecard?: Scorecard;
+  };
+  const scorecard = (agg.maskedScorecard ?? { gate: { total: 0, passCount: 0, failedCriteria: [], gatedBy: [], indeterminateBy: [], passed: true, runVerdict: "pass" } }) as Scorecard;
+  const repCritique = new Map<string, { critique: string; refs?: unknown[]; traceId: string }>();
+  for (const f of verdictFiles) {
+    for (const v of (f as { verdicts?: Array<{ criterionId: string; result: string; critique?: string; refs?: unknown[] }> }).verdicts ?? []) {
+      if (v.critique && !repCritique.has(`${v.criterionId}:${v.result}`)) {
+        repCritique.set(`${v.criterionId}:${v.result}`, { critique: v.critique, refs: v.refs, traceId: (f as { trajectoryId: string }).trajectoryId });
+      }
+    }
+  }
+  const verdicts: CriterionVerdict[] = (agg.summary?.perCriterion ?? []).map((pc) => {
+    const rep = repCritique.get(`${pc.criterionId}:${pc.folded}`);
+    return {
+      criterionId: pc.criterionId,
+      traceId: rep?.traceId ?? "(suite)",
+      result: pc.folded as CriterionVerdict["result"],
+      confidence: pc.confidence ?? 0,
+      critique: rep?.critique ?? "",
+      ...(rep?.refs ? { refs: rep.refs as CriterionVerdict["refs"] } : {}),
+    };
+  });
+
+  // 3) the criteria (discover/suite shape) → ReportCriterion[].
+  const rawCrit = JSON.parse(io.readFile(params.criteriaPath)) as unknown;
+  const critArr: Array<Record<string, unknown>> = Array.isArray(rawCrit)
+    ? (rawCrit as Array<Record<string, unknown>>)
+    : ((rawCrit as { entries?: Array<Record<string, unknown>> }).entries ?? []);
+  const criteria: ReportCriterion[] = critArr.map((c) => {
+    const meta = (c.metadata ?? {}) as Record<string, unknown>;
+    const severity = String((c.severity as string) ?? (meta.severity as string) ?? "MED");
+    return {
+      id: String(c.id ?? c.criterionId ?? ""),
+      statement: String(c.statement ?? ""),
+      severity,
+      gating: GATING_SEVERITIES.has(severity),
+      ...(meta.dimension !== undefined ? { dimension: String(meta.dimension) } : {}),
+      ...(Array.isArray(c.judgeInputs) ? { judgeInputs: c.judgeInputs as string[] } : {}),
+      checkMethod: String((c.judgeKind as string) ?? (meta.check_method as string) ?? "llm-judge"),
+      provenance: { kind: "defined" as const, label: "defined eval-matrix criterion" },
+    };
+  });
+
+  // 3b) WS-4 — reconstruct the EV-051 diagnostics HANDOVER from the folded gate, using
+  //     the SHIPPED `routeFailures` (no hand-rolled routing). Each failed criterion →
+  //     a FailureRef enriched with its folded result + a representative critique/trace
+  //     (from the verdict files). `routeFailures` then partitions: only TRUE FAILs
+  //     (result === "fail") route to diagnostics — uncertain/indeterminate go to the
+  //     calibration loop and are excluded — so the §5 handover shows exactly the routed
+  //     count, matching the run's `summary.routedFailures`. NEVER invents a failure.
+  const foldedById = new Map((agg.summary?.perCriterion ?? []).map((pc) => [pc.criterionId, pc.folded]));
+  const failedCriteria = (scorecard.gate?.failedCriteria ?? []) as Array<{ criterionId: string; severity: string }>;
+  const failures: FailureRef[] = failedCriteria.map((fc) => {
+    const folded = foldedById.get(fc.criterionId) ?? "fail";
+    const rep = repCritique.get(`${fc.criterionId}:${folded}`) ?? repCritique.get(`${fc.criterionId}:fail`);
+    return {
+      criterionId: fc.criterionId,
+      severity: fc.severity,
+      flag: CriterionFlag.EvalWorthy,
+      traceId: rep?.traceId ?? "(suite)",
+      result: folded as FailureRef["result"],
+      critique: rep?.critique ?? "(no critique recorded)",
+    };
+  });
+  // routeFailures internally filters to true FAILs (partitionRouting); pass them all.
+  const diagnoseFailures = failures.filter((f) => f.result === OutcomeVerdict.Fail);
+  const handover: HandoverBundle | null =
+    diagnoseFailures.length > 0
+      ? routeFailures({
+          subject: { kind: "agent", name: params.subjectName, path: `subjects/${params.subjectName}` },
+          failures: diagnoseFailures,
+          artifacts: [],
+          producedBy: "evaluator",
+          producedAt: params.generatedAt,
+        })
+      : null;
+
+  // 4) build the rich input + render.
+  const input = buildEvalReportInput({
+    subject: { name: params.subjectName, ...(params.subjectSource ? { source: params.subjectSource } : {}) },
+    scorecard,
+    verdicts,
+    criteria,
+    matrixVerdictFiles: verdictFiles,
+    handover,
+    generatedAt: params.generatedAt,
+    // WS-6 — forward the trace(s) so the system prompt is reconstructed for the hero.
+    ...(params.traces !== undefined ? { traces: params.traces } : {}),
+  });
+  io.writeFile(params.outPath, renderEvalReport(input));
+
+  // 5) the data-completeness probe over the BOUND ledger (what the UI will render).
+  const ledger = input.ledger ?? [];
+  const completeness: EvalReportCompleteness = {
+    tracesInVerdicts: verdictFiles.length,
+    tracesRenderedWithJudgeSteps: ledger.filter((r) => (r.judgeSteps?.length ?? 0) > 0).length,
+    tracesRenderedWithCritique: ledger.filter((r) => (r.criterionVerdicts?.length ?? 0) > 0).length,
+    // an empty UI row = the verdict file carried judge data (a walk OR ≥1 per-criterion
+    // verdict) yet the drill would bind NOTHING. With per-criterion critique bound to
+    // every judged row, this is 0 unless a verdict file is genuinely empty.
+    emptyUIrows: ledger.filter(
+      (r) => (r.judgeSteps?.length ?? 0) === 0 && (r.criterionVerdicts?.length ?? 0) === 0,
+    ).length,
+    systemPromptChars: (input.subjectProfile?.systemPrompt ?? "").length,
+    systemPromptReconstructed: (input.subjectProfile?.inferredFields ?? []).includes("systemPrompt"),
+    routedFailures: input.handover?.acceptance.criteria.length ?? 0,
+  };
+  return { report: params.outPath, completeness };
+}
+
+// ── auto-open helper (built; fired by the orchestrator, NOT here) ────────────
+
+/**
+ * The cross-platform command to open a rendered report. PURE — returns the
+ * command string; it NEVER spawns a process (the orchestrator fires it
+ * post-render; a test must never auto-open). macOS `open` · Linux `xdg-open` ·
+ * Windows `start`.
+ */
+export function autoOpenCommand(platform: string, filePath: string): string {
+  if (platform === "darwin") return `open ${filePath}`;
+  if (platform === "win32") return `cmd /c start "" ${filePath}`;
+  return `xdg-open ${filePath}`;
+}
+
+// ── CLI entrypoint ──────────────────────────────────────────────────────────
+//
+// bun scripts/render-eval-report.ts <input.json> [out.html] [runId]
+// Reads an EvalReportInput (the rich, already-built shape), prints the D-1
+// terminal cards, writes the 5-tab HTML report, and prints the cross-platform
+// auto-open command (does NOT fire it — the orchestrator opens post-render).
+
+declare const Bun: { argv: string[] } | undefined;
+
+async function main(): Promise<void> {
+  const argv = typeof Bun !== "undefined" ? Bun.argv.slice(2) : process.argv.slice(2);
+  const [inputPath, outArg, runIdArg] = argv;
+  if (!inputPath) {
+    console.error("usage: render-eval-report.ts <input.json> [out.html] [runId]");
+    process.exit(2);
+    return;
+  }
+  const { readFileSync: rf, writeFileSync, mkdirSync } = await import("node:fs");
+  const { reportDir } = await import("./artifact-paths.ts");
+  const { join: pjoin } = await import("node:path");
+  const input = JSON.parse(rf(inputPath, "utf8")) as EvalReportInput;
+  // P8: a bare run DEFAULTS the report under the localized dot-root
+  // `.mutagent/evaluator/reports/<runId>/report.html` (never /tmp or .memory).
+  const runId = runIdArg ?? input.subject.name;
+  const outPath = outArg ?? pjoin(reportDir(runId), "report.html");
+  mkdirSync(pjoin(outPath, ".."), { recursive: true });
+  console.info(renderEvalCards(input));
+  writeFileSync(outPath, renderEvalReport(input));
+  const plat = typeof process !== "undefined" ? process.platform : "linux";
+  console.info(`\nreport written: ${outPath}`);
+  console.info(`auto-open (orchestrator fires this): ${autoOpenCommand(plat, outPath)}`);
+  process.exit(0);
+}
+
+const isMain =
+  typeof import.meta !== "undefined" &&
+  (import.meta as unknown as { main?: boolean }).main === true;
+if (isMain) {
+  void main();
+}