@tangle-network/agent-eval 0.183.0 → 0.185.0

package/dist/types-C8VBbiw6.js

@@ -0,0 +1,91 @@
+import { createHash } from "node:crypto";
+//#region src/ledger-core/deep-freeze.ts
+/** Freeze a detached canonical-JSON graph. Canonicalization has already ruled out cycles.
+*
+* Lives outside canonical.ts so the analyst-benchmark implementation digest,
+* which covers canonical.ts, stays bound to the published benchmark evidence. */
+function deepFreezeCanonicalJson(value) {
+	if (value && typeof value === "object" && !Object.isFrozen(value)) {
+		Object.freeze(value);
+		for (const nested of Object.values(value)) deepFreezeCanonicalJson(nested);
+	}
+	return value;
+}
+//#endregion
+//#region src/analyst/types.ts
+/**
+* Analyst contract — the missing orchestration layer over agent-eval's
+* existing analyzers (analyzeTraces, MultiLayerVerifier,
+* SemanticConceptJudge, JudgeFn, ...).
+*
+* Each existing primitive returns its own output shape. The Analyst
+* contract is the single envelope every primitive lifts into, so a
+* registry can run N analysts against a run and a single renderer can
+* compose findings without knowing which analyzer produced them.
+*
+* The contract is intentionally domain-agnostic: nothing here knows
+* about code, voice, RAG, or any particular agent stack. Analysts
+* declare what INPUT KIND they need (a trace store, an artifact dir,
+* a RunRecord, a JudgeInput, or `custom`), and the registry routes
+* the matching input from `AnalystRunInputs`.
+*/
+/** Finding severity. `AnalystSeverity` derives from this array, so the type
+*  and the schema that validates a finding cannot name different levels. */
+const ANALYST_SEVERITIES = [
+	"critical",
+	"high",
+	"medium",
+	"low",
+	"info"
+];
+/**
+* Compute the stable finding_id from the identity-defining fields.
+* Default implementation hashes {analyst_id, area, subject, normalized claim}.
+* Analysts that emit findings whose claim text varies per run (timestamps,
+* counts) SHOULD either: (a) pass an explicit `id_basis` to fix the hash,
+* or (b) move the variable part into `rationale`/`metadata` and keep the
+* `claim` static.
+*/
+function computeFindingId(input) {
+	const basis = JSON.stringify({
+		a: input.analyst_id,
+		r: input.area,
+		s: input.subject ?? "",
+		c: normalizeClaim(input.id_basis ?? input.claim)
+	});
+	return `f_${createHash("sha256").update(basis).digest("hex").slice(0, 20)}`;
+}
+function normalizeClaim(c) {
+	return c.toLowerCase().replace(/\s+/g, " ").replace(/[.!?;:,]+$/g, "").trim();
+}
+/**
+* Convenience factory: produce a fully-formed AnalystFinding with the
+* id computed automatically. Analyst code stays terse.
+*/
+function makeFinding(init) {
+	const { id_basis, produced_at, ...rest } = init;
+	return {
+		schema_version: "1.0.0",
+		finding_id: computeFindingId({
+			analyst_id: rest.analyst_id,
+			area: rest.area,
+			subject: rest.subject,
+			claim: rest.claim,
+			id_basis
+		}),
+		produced_at: produced_at ?? (/* @__PURE__ */ new Date()).toISOString(),
+		...rest
+	};
+}
+/** Build a finding whose source is explicitly allowed during candidate generation. */
+function makeProposalFinding(init) {
+	const { proposal_origin, ...finding } = init;
+	return {
+		...makeFinding(finding),
+		proposal_origin
+	};
+}
+//#endregion
+export { deepFreezeCanonicalJson as a, makeProposalFinding as i, computeFindingId as n, makeFinding as r, ANALYST_SEVERITIES as t };
+
+//# sourceMappingURL=types-C8VBbiw6.js.map

package/dist/types-C8VBbiw6.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"types-C8VBbiw6.js","names":[],"sources":["../src/ledger-core/deep-freeze.ts","../src/analyst/types.ts"],"sourcesContent":["/** Freeze a detached canonical-JSON graph. Canonicalization has already ruled out cycles.\n *\n * Lives outside canonical.ts so the analyst-benchmark implementation digest,\n * which covers canonical.ts, stays bound to the published benchmark evidence. */\nexport function deepFreezeCanonicalJson<T>(value: T): T {\n  if (value && typeof value === 'object' && !Object.isFrozen(value)) {\n    Object.freeze(value)\n    for (const nested of Object.values(value)) deepFreezeCanonicalJson(nested)\n  }\n  return value\n}\n","/**\n * Analyst contract — the missing orchestration layer over agent-eval's\n * existing analyzers (analyzeTraces, MultiLayerVerifier,\n * SemanticConceptJudge, JudgeFn, ...).\n *\n * Each existing primitive returns its own output shape. The Analyst\n * contract is the single envelope every primitive lifts into, so a\n * registry can run N analysts against a run and a single renderer can\n * compose findings without knowing which analyzer produced them.\n *\n * The contract is intentionally domain-agnostic: nothing here knows\n * about code, voice, RAG, or any particular agent stack. Analysts\n * declare what INPUT KIND they need (a trace store, an artifact dir,\n * a RunRecord, a JudgeInput, or `custom`), and the registry routes\n * the matching input from `AnalystRunInputs`.\n */\n\nimport { createHash } from 'node:crypto'\nimport type { CostLedgerHandle } from '../cost-ledger'\nimport type { RunCostProvenance, RunRecord, RunTokenUsage } from '../run-record'\nimport type { TraceAnalysisStore } from '../trace-analyst/store'\nimport type { JudgeInput } from '../types'\nimport type { ChatClient } from './chat-client'\n\n/**\n * Unified envelope every analyst emits. Schema-versioned so renderers\n * and time-series diffs survive future field additions.\n */\nexport interface AnalystFinding {\n  schema_version: '1.0.0'\n  /**\n   * Stable hash over identity-defining fields (analyst_id + canonical\n   * claim + area + optional subject). Two findings from two runs that\n   * \"are the same finding\" share this id — that's what `diffFindings`\n   * uses to compute appeared/disappeared sets across runs.\n   */\n  finding_id: string\n  analyst_id: string\n  produced_at: string\n  severity: AnalystSeverity\n  /**\n   * Coarse classification. Renderers group by this. Free-form so\n   * domain-specific analysts can introduce categories without a\n   * schema change ('agent-reasoning', 'verification', 'cost',\n   * 'tool-use', 'safety', 'latency', 'data-quality', ...).\n   */\n  area: string\n  claim: string\n  rationale?: string\n  evidence_refs: EvidenceRef[]\n  recommended_action?: string\n  validation_plan?: string\n  /** 0..1 — the analyst's own confidence. Not calibrated across analysts. */\n  confidence: number\n  /**\n   * Optional subject the finding is about — leaf id, agent id, request\n   * id. Included in finding_id when present so per-subject findings\n   * diff cleanly across runs.\n   */\n  subject?: string\n  /** True when this finding was lifted from a judge result rather than observed\n   *  directly in a trace or artifact. Descriptive only: proposal access is\n   *  controlled by `ProposalFinding.proposal_origin`. */\n  derived_from_judge?: boolean\n  /** Analyst-private extras; renderers ignore unless they know the analyst. */\n  metadata?: Record<string, unknown>\n}\n\n/** Finding severity. `AnalystSeverity` derives from this array, so the type\n *  and the schema that validates a finding cannot name different levels. */\nexport const ANALYST_SEVERITIES = ['critical', 'high', 'medium', 'low', 'info'] as const\n\nexport type AnalystSeverity = (typeof ANALYST_SEVERITIES)[number]\n\n/** Data sources that candidate generation may intentionally learn from. */\nexport type ProposalFindingOrigin = 'search' | 'production'\n\n/** A finding explicitly admitted as candidate-generation input. */\nexport type ProposalFinding = AnalystFinding & {\n  readonly proposal_origin: ProposalFindingOrigin\n}\n\nexport interface EvidenceRef {\n  /**\n   * Where the evidence lives. `span` and `event` refer to OTLP trace\n   * elements; `artifact` to a file inside the run's artifact tree;\n   * `finding` to another AnalystFinding (cross-analyst chaining);\n   * `metric` to a named scalar reading the renderer knows how to read.\n   */\n  kind: 'span' | 'event' | 'artifact' | 'finding' | 'metric'\n  uri: string\n  excerpt?: string\n}\n\n// ── Analyst contract ─────────────────────────────────────────────────\n\n/**\n * The discriminator the registry uses to pass the right input.\n * `custom` is the escape hatch — analysts that need something else\n * (e.g. an embedding cache, a partner SDK handle) read it from\n * `AnalystRunInputs.custom[<analyst id>]`.\n */\nexport type AnalystInputKind =\n  | 'trace-store'\n  | 'artifact-dir'\n  | 'run-record'\n  | 'judge-input'\n  | 'custom'\n\nexport interface AnalystCost {\n  /** `deterministic` analysts MUST NOT call the LLM. */\n  kind: 'deterministic' | 'llm'\n  /** Optional declared upper bound; the registry can enforce a budget. */\n  est_usd_per_run?: number\n  /** Models the analyst expects to use (informational). */\n  models?: string[]\n  /** Maximum post-cancellation wait for provider usage. Model analysts default to 5 seconds. */\n  settlement_timeout_ms?: number\n}\n\nexport interface AnalystRequirements {\n  /** Min number of shots / samples the analyst needs to produce signal. */\n  min_shots?: number\n  /** Capabilities the runtime must supply (e.g. ['network', 'gpu']). */\n  capabilities?: string[]\n}\n\n/**\n * What's passed to every analyst call. The registry resolves which\n * field the analyst's `inputKind` selects and asserts it's present.\n */\nexport interface AnalystRunInputs {\n  traceStore?: TraceAnalysisStore\n  artifactDir?: string\n  runRecord?: RunRecord\n  judgeInput?: JudgeInput\n  /** Keyed by analyst id; populated by callers that registered custom analysts. */\n  custom?: Record<string, unknown>\n}\n\nexport interface AnalystContext {\n  runId: string\n  /** Stable correlation id so logs from a single registry.run() share a tag. */\n  correlationId: string\n  /** Enforced wall-clock deadline (epoch ms). */\n  deadlineMs?: number\n  /** Per-analyst USD budget. Analysts MAY check before issuing LLM calls. */\n  budgetUsd?: number\n  /** Shared paid-call account when the analyst runs inside a larger campaign. */\n  costLedger?: CostLedgerHandle\n  /** Attribution phase used when writing to the shared paid-call account. */\n  costPhase?: string\n  /**\n   * Shared chat client. Analysts that call an LLM go through this so\n   * the operator picks transport (sandbox-sdk | router | cli-bridge |\n   * direct-provider | mock) at the registry boundary without touching\n   * analyst code.\n   */\n  chat?: ChatClient\n  /**\n   * Findings from a prior run the operator wants the analyst to see as\n   * retrieval context. Kinds that take advantage of cross-run memory\n   * (failure-mode \"I saw this cluster last run\", knowledge-gap \"the wiki\n   * page I asked for is still missing\") render these into the actor's\n   * working set. Filtering is the operator's job: pass the slice that\n   * matches the analyst's id, or pass everything and let the kind\n   * filter. Empty / absent means no cross-run context.\n   */\n  priorFindings?: ReadonlyArray<AnalystFinding>\n  /**\n   * Findings emitted by analysts that completed earlier in this registry run.\n   * This is separate from `priorFindings`: upstream findings are dependency\n   * context for the current pass, while prior findings are cross-run memory.\n   * The registry populates this only when `RegistryRunOpts.chainFindings` is on.\n   */\n  upstreamFindings?: ReadonlyArray<AnalystFinding>\n  /**\n   * Report metered work independently of findings. This keeps an empty finding\n   * set from erasing token/cost telemetry. Multiple receipts are accumulated.\n   */\n  recordUsage?: (receipt: AnalystUsageReceipt) => void\n  /** Free-form runtime tags (env, host, op). Findings can echo these into metadata. */\n  tags?: Record<string, string>\n  /** Logger callback — analysts SHOULD prefer this over console.* for testability. */\n  log?: (msg: string, fields?: Record<string, unknown>) => void\n  /** Optional abort signal. Analysts SHOULD pass it through to LLM calls. */\n  signal?: AbortSignal\n  /**\n   * Optional live-execution port. A runtime that owns a sandbox or checkout\n   * fills it so an analyst can execute a bounded probe against the run's\n   * produced state instead of reasoning about it from the trace alone. This\n   * package defines only the port: no field here reaches for an agent loop,\n   * and an absent probe means the analyst works from recorded evidence.\n   */\n  probe?: ExecutionProbe\n}\n\n// ── Live-execution port ─────────────────────────────────────────────\n\n/** One bounded command an analyst asks the probe to run. */\nexport interface ExecutionProbeRequest {\n  command: string\n  /** Working directory inside the probed environment. */\n  cwd?: string\n  /** Hard wall-clock deadline for this one execution. */\n  timeoutMs: number\n  /** Bytes of combined output retained; the prober truncates beyond it. */\n  maxOutputBytes?: number\n  signal?: AbortSignal\n}\n\n/**\n * Typed outcome of one probe execution. `succeeded: false` is a PROBE failure\n * (the environment could not run the command); a command that ran and exited\n * non-zero is a successful observation with a non-zero `exitCode`.\n */\nexport type ExecutionProbeOutcome =\n  | {\n      succeeded: true\n      exitCode: number\n      stdout: string\n      stderr: string\n      durationMs: number\n      /** True when output was cut at `maxOutputBytes`. */\n      truncated: boolean\n    }\n  | { succeeded: false; error: { class: string; message: string } }\n\n/**\n * The seam a runtime fills to let analysts observe produced state live.\n * Implementations own sandboxing, credentials, and cleanup; analysts only\n * submit bounded requests and read typed outcomes.\n */\nexport interface ExecutionProbe {\n  /** One plain sentence naming what is being probed (e.g. a sandbox id). */\n  readonly description: string\n  execute(request: ExecutionProbeRequest): Promise<ExecutionProbeOutcome>\n}\n\n/**\n * The minimal contract. Concrete analysts can refine `TInput` so\n * implementations stay type-safe (e.g. a trace analyst's `TInput` is\n * `TraceAnalysisStore`); the registry passes the right field from\n * `AnalystRunInputs` based on `inputKind`.\n */\nexport interface Analyst<TInput = unknown> {\n  /** Stable identifier — appears in finding_id, telemetry, and registry exclusion lists. */\n  readonly id: string\n  /** Human-readable. One sentence. */\n  readonly description: string\n  readonly inputKind: AnalystInputKind\n  readonly cost: AnalystCost\n  readonly requires?: AnalystRequirements\n  /** Bump on breaking changes to claim wording or area so old finding_ids don't collide. */\n  readonly version: string\n  analyze(input: TInput, ctx: AnalystContext): Promise<AnalystFinding[]>\n}\n\n/** Metered work performed by one analyst call. */\nexport interface AnalystUsageReceipt {\n  /** Number of model-usage records observed at the provider boundary. */\n  calls: number | null\n  /** Null when the provider did not return token accounting. */\n  tokens: RunTokenUsage | null\n  /** Observed, estimated, or explicitly uncaptured dollar cost. */\n  cost: RunCostProvenance\n  /** Known lower bound when one or more calls have uncaptured cost. */\n  knownCostUsd?: number\n  /**\n   * Token counts the provider reported on only one side. Present exactly when\n   * `tokens` is null and at least one side WAS reported: `RunTokenUsage` has no\n   * nullable side, so a one-sided count cannot live in `tokens` without writing\n   * a zero nobody measured. Read it as a lower bound, never as a total — the\n   * field exists so a null `tokens` cannot hide a real count.\n   */\n  partialTokens?: { input: number | null; output: number | null }\n  /**\n   * True when the token counts were DERIVED by the transport (from character\n   * lengths, say) rather than measured by the model provider. `cost.kind` is\n   * `estimated` both for a rate estimate over exact tokens and for one over\n   * derived tokens; this is the field that separates them.\n   */\n  tokensEstimated?: boolean\n}\n\n// ── finding_id stability ─────────────────────────────────────────────\n\n/**\n * Compute the stable finding_id from the identity-defining fields.\n * Default implementation hashes {analyst_id, area, subject, normalized claim}.\n * Analysts that emit findings whose claim text varies per run (timestamps,\n * counts) SHOULD either: (a) pass an explicit `id_basis` to fix the hash,\n * or (b) move the variable part into `rationale`/`metadata` and keep the\n * `claim` static.\n */\nexport function computeFindingId(input: {\n  analyst_id: string\n  area: string\n  subject?: string\n  claim: string\n  /** Override the claim for hashing — use when the displayed claim has run-specific bits. */\n  id_basis?: string\n}): string {\n  const basis = JSON.stringify({\n    a: input.analyst_id,\n    r: input.area,\n    s: input.subject ?? '',\n    c: normalizeClaim(input.id_basis ?? input.claim),\n  })\n  return `f_${createHash('sha256').update(basis).digest('hex').slice(0, 20)}`\n}\n\nfunction normalizeClaim(c: string): string {\n  // Lowercase, collapse whitespace, strip trailing punctuation. Goal:\n  // \"Leaf X failed install\" and \"Leaf X failed install.\" hash the same.\n  return c\n    .toLowerCase()\n    .replace(/\\s+/g, ' ')\n    .replace(/[.!?;:,]+$/g, '')\n    .trim()\n}\n\n/**\n * Convenience factory: produce a fully-formed AnalystFinding with the\n * id computed automatically. Analyst code stays terse.\n */\nexport function makeFinding(\n  init: Omit<AnalystFinding, 'schema_version' | 'finding_id' | 'produced_at'> & {\n    id_basis?: string\n    produced_at?: string\n  },\n): AnalystFinding {\n  const { id_basis, produced_at, ...rest } = init\n  return {\n    schema_version: '1.0.0',\n    finding_id: computeFindingId({\n      analyst_id: rest.analyst_id,\n      area: rest.area,\n      subject: rest.subject,\n      claim: rest.claim,\n      id_basis,\n    }),\n    produced_at: produced_at ?? new Date().toISOString(),\n    ...rest,\n  }\n}\n\n/** Build a finding whose source is explicitly allowed during candidate generation. */\nexport function makeProposalFinding(\n  init: Omit<ProposalFinding, 'schema_version' | 'finding_id' | 'produced_at'> & {\n    id_basis?: string\n    produced_at?: string\n  },\n): ProposalFinding {\n  const { proposal_origin, ...finding } = init\n  return { ...makeFinding(finding), proposal_origin }\n}\n\n// ── Registry result envelope ────────────────────────────────────────\n\nexport interface AnalystRunSummary {\n  analyst_id: string\n  status: 'ok' | 'skipped' | 'failed'\n  /** Why skipped — missing input, budget exceeded, capability unmet. */\n  reason?: string\n  findings_count: number\n  latency_ms: number\n  /** Additive model usage and cost provenance for this analyst. */\n  usage: AnalystUsageReceipt\n  /** When `status='failed'`: the error class + message, never the full stack. */\n  error?: { class: string; message: string }\n}\n\nexport interface AnalystRunResult {\n  run_id: string\n  correlation_id: string\n  started_at: string\n  ended_at: string\n  findings: AnalystFinding[]\n  per_analyst: AnalystRunSummary[]\n  /** Total LLM cost in USD across all analysts in this registry.run(). */\n  total_cost_usd: number\n  /**\n   * Provenance for `total_cost_usd`. When uncaptured, the numeric field is only\n   * the known subtotal and must not be treated as the run's total spend.\n   */\n  total_cost_provenance?: RunCostProvenance\n}\n\n// ── Streaming event envelope ────────────────────────────────────────\n\n/**\n * Events emitted by `AnalystRegistry.runStream(...)` in real time as\n * the registry executes. UIs subscribe via `for await (const ev of\n * registry.runStream(...))`; `registry.run(...)` is a thin collector\n * over the same stream, so the two surfaces share their invariants.\n *\n * Per-finding events are intentionally omitted — analyzers are batch\n * operations (a recursive engine returns the full `findings:json[]` at the\n * end of the responder), so streaming inside one analyst would only\n * emit partial JSON consumers can't render. The kind-completion event\n * is the right granularity; subscribers wanting per-finding rendering\n * iterate `event.findings` themselves.\n */\nexport type AnalystRunEvent =\n  | {\n      type: 'run-started'\n      run_id: string\n      correlation_id: string\n      started_at: string\n      /** The ordered list of analyst ids the registry will run. */\n      analyst_ids: ReadonlyArray<string>\n    }\n  | {\n      type: 'analyst-skipped'\n      summary: AnalystRunSummary\n    }\n  | {\n      type: 'analyst-started'\n      analyst_id: string\n      started_at: string\n    }\n  | {\n      type: 'analyst-completed'\n      /** `summary.status` is `'ok'` for clean completion or `'failed'` for thrown analysts. */\n      summary: AnalystRunSummary\n      findings: ReadonlyArray<AnalystFinding>\n    }\n  | {\n      type: 'run-completed'\n      result: AnalystRunResult\n    }\n"],"mappings":";;;;;;AAIA,SAAgB,wBAA2B,OAAa;CACtD,IAAI,SAAS,OAAO,UAAU,YAAY,CAAC,OAAO,SAAS,KAAK,GAAG;EACjE,OAAO,OAAO,KAAK;EACnB,KAAK,MAAM,UAAU,OAAO,OAAO,KAAK,GAAG,wBAAwB,MAAM;CAC3E;CACA,OAAO;AACT;;;;;;;;;;;;;;;;;;;;;AC4DA,MAAa,qBAAqB;CAAC;CAAY;CAAQ;CAAU;CAAO;AAAM;;;;;;;;;AAiO9E,SAAgB,iBAAiB,OAOtB;CACT,MAAM,QAAQ,KAAK,UAAU;EAC3B,GAAG,MAAM;EACT,GAAG,MAAM;EACT,GAAG,MAAM,WAAW;EACpB,GAAG,eAAe,MAAM,YAAY,MAAM,KAAK;CACjD,CAAC;CACD,OAAO,KAAK,WAAW,QAAQ,CAAC,CAAC,OAAO,KAAK,CAAC,CAAC,OAAO,KAAK,CAAC,CAAC,MAAM,GAAG,EAAE;AAC1E;AAEA,SAAS,eAAe,GAAmB;CAGzC,OAAO,EACJ,YAAY,CAAC,CACb,QAAQ,QAAQ,GAAG,CAAC,CACpB,QAAQ,eAAe,EAAE,CAAC,CAC1B,KAAK;AACV;;;;;AAMA,SAAgB,YACd,MAIgB;CAChB,MAAM,EAAE,UAAU,aAAa,GAAG,SAAS;CAC3C,OAAO;EACL,gBAAgB;EAChB,YAAY,iBAAiB;GAC3B,YAAY,KAAK;GACjB,MAAM,KAAK;GACX,SAAS,KAAK;GACd,OAAO,KAAK;GACZ;EACF,CAAC;EACD,aAAa,gCAAe,IAAI,KAAK,EAAA,CAAE,YAAY;EACnD,GAAG;CACL;AACF;;AAGA,SAAgB,oBACd,MAIiB;CACjB,MAAM,EAAE,iBAAiB,GAAG,YAAY;CACxC,OAAO;EAAE,GAAG,YAAY,OAAO;EAAG;CAAgB;AACpD"}

package/dist/{types-BvZoPTGa.d.ts → types-c3ZRTbw9.d.ts}

@@ -1,5 +1,5 @@
-import { s as RunSplitTag } from "./run-record-BiTWauyO.js";
-import { R as Scenario, d as DispatchContext } from "./types-CS0qk_Yp.js";
+import { s as RunSplitTag } from "./run-record-D9wS35fK.js";
+import { R as Scenario, d as DispatchContext } from "./types-BTRcIqFr.js";
 //#region src/benchmarks/types.d.ts
 type BenchmarkTaskKind = 'retrieval' | 'rag-answer' | 'hallucination' | 'kb-improvement' | 'routing' | 'custom';
 type BenchmarkFamily = 'beir' | 'mteb-retrieval' | 'msmarco' | 'trec-dl' | 'miracl' | 'lotte' | 'bright' | 'crag' | 'hotpotqa' | 'kilt' | 'ragtruth' | 'faithbench' | 'first-party' | 'custom';
@@ -90,4 +90,4 @@ declare const BENCHMARK_SPLIT_SEED = "agent-eval-v1";
 declare function deterministicSplit(itemId: string, seed?: string): RunSplitTag;
 //#endregion
 export { BenchmarkFamily as a, BenchmarkSource as c, BenchmarkEvaluation as i, BenchmarkTaskKind as l, BenchmarkAdapter as n, BenchmarkResponder as o, BenchmarkDatasetItem as r, BenchmarkScenario as s, BENCHMARK_SPLIT_SEED as t, deterministicSplit as u };
-//# sourceMappingURL=types-BvZoPTGa.d.ts.map
+//# sourceMappingURL=types-c3ZRTbw9.d.ts.map

package/dist/{types-BvZoPTGa.d.ts.map → types-c3ZRTbw9.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"types-BvZoPTGa.d.ts","names":[],"sources":["../src/benchmarks/types.ts"],"mappings":";;;KASY;KAQA;UAgBK,qBAAqB;;;EAGpC;;EAEA,SAAS;;EAET,QAAQ;;EAER,SAAS;;EAET,WAAW;;EAEX;;EAEA,SAAS;EACT,WAAW;;UAGI;;;;EAIf;;EAEA;;EAEA,aAAa;;;EAGb,KAAK;EACL;;UAGe;EACf;EACA;EACA;EACA;EACA;;;UAQe,iBAAiB,kBAAkB,oBAAoB;;EAEtE;EACA,SAAS;EACT,WAAW;EACX;EACA,SAAS;EACT;;;;;EAKA,YAAY,OAAO,cAAc,QAAQ,qBAAqB;;EAE9D,SAAS,MAAM,qBAAqB,WAAW,UAAU,YAAY,QAAQ;;;;EAI7E,YAAY,iBAAiB;;UAGd,kBAAkB,4BAA4B;EAC7D;EACA;EACA,QAAQ;EACR,UAAU;EACV,UAAU;EACV,MAAM,qBAAqB;;KAGjB,mBAAmB,oBAAoB,uBAAuB;EACxE,UAAU,kBAAkB;EAC5B,MAAM,qBAAqB;EAC3B,SAAS;MACL,QAAQ,aAAa;;;cAoBd;;;;;;;;;iBAUG,mBACd,gBACA,gBACC"}
+{"version":3,"file":"types-c3ZRTbw9.d.ts","names":[],"sources":["../src/benchmarks/types.ts"],"mappings":";;;KASY;KAQA;UAgBK,qBAAqB;;;EAGpC;;EAEA,SAAS;;EAET,QAAQ;;EAER,SAAS;;EAET,WAAW;;EAEX;;EAEA,SAAS;EACT,WAAW;;UAGI;;;;EAIf;;EAEA;;EAEA,aAAa;;;EAGb,KAAK;EACL;;UAGe;EACf;EACA;EACA;EACA;EACA;;;UAQe,iBAAiB,kBAAkB,oBAAoB;;EAEtE;EACA,SAAS;EACT,WAAW;EACX;EACA,SAAS;EACT;;;;;EAKA,YAAY,OAAO,cAAc,QAAQ,qBAAqB;;EAE9D,SAAS,MAAM,qBAAqB,WAAW,UAAU,YAAY,QAAQ;;;;EAI7E,YAAY,iBAAiB;;UAGd,kBAAkB,4BAA4B;EAC7D;EACA;EACA,QAAQ;EACR,UAAU;EACV,UAAU;EACV,MAAM,qBAAqB;;KAGjB,mBAAmB,oBAAoB,uBAAuB;EACxE,UAAU,kBAAkB;EAC5B,MAAM,qBAAqB;EAC3B,SAAS;MACL,QAAQ,aAAa;;;cAoBd;;;;;;;;;iBAUG,mBACd,gBACA,gBACC"}

package/dist/{types-DQ0e2E7y.js → usage-receipt-Cx3xNeyB.js}

@@ -1,16 +1,3 @@
-import { createHash } from "node:crypto";
-//#region src/ledger-core/deep-freeze.ts
-/** Freeze a detached canonical-JSON graph. Canonicalization has already ruled out cycles.
-*
-* Lives outside canonical.ts so the analyst-benchmark implementation digest,
-* which covers canonical.ts, stays bound to the published benchmark evidence. */
-function deepFreezeCanonicalJson(value) {
-	if (value && typeof value === "object" && !Object.isFrozen(value)) {
-		Object.freeze(value);
-		for (const nested of Object.values(value)) deepFreezeCanonicalJson(nested);
-	}
-	return value;
-}
 /** Convert one ledger channel's complete call set into one analyst receipt. */
 function usageReceiptFromCostLedger(ledger, filter = "analyst") {
 	const resolvedFilter = typeof filter === "string" ? { channel: filter } : filter;
@@ -84,80 +71,6 @@ function assertNonNegativeFinite(value, field, context) {
 	if (!Number.isFinite(value) || value < 0) throw new Error(`${context}: ${field} must be a non-negative finite number`);
 }
 //#endregion
-//#region src/analyst/types.ts
-/**
-* Analyst contract — the missing orchestration layer over agent-eval's
-* existing analyzers (analyzeTraces, MultiLayerVerifier,
-* SemanticConceptJudge, JudgeFn, ...).
-*
-* Each existing primitive returns its own output shape. The Analyst
-* contract is the single envelope every primitive lifts into, so a
-* registry can run N analysts against a run and a single renderer can
-* compose findings without knowing which analyzer produced them.
-*
-* The contract is intentionally domain-agnostic: nothing here knows
-* about code, voice, RAG, or any particular agent stack. Analysts
-* declare what INPUT KIND they need (a trace store, an artifact dir,
-* a RunRecord, a JudgeInput, or `custom`), and the registry routes
-* the matching input from `AnalystRunInputs`.
-*/
-/** Finding severity. `AnalystSeverity` derives from this array, so the type
-*  and the schema that validates a finding cannot name different levels. */
-const ANALYST_SEVERITIES = [
-	"critical",
-	"high",
-	"medium",
-	"low",
-	"info"
-];
-/**
-* Compute the stable finding_id from the identity-defining fields.
-* Default implementation hashes {analyst_id, area, subject, normalized claim}.
-* Analysts that emit findings whose claim text varies per run (timestamps,
-* counts) SHOULD either: (a) pass an explicit `id_basis` to fix the hash,
-* or (b) move the variable part into `rationale`/`metadata` and keep the
-* `claim` static.
-*/
-function computeFindingId(input) {
-	const basis = JSON.stringify({
-		a: input.analyst_id,
-		r: input.area,
-		s: input.subject ?? "",
-		c: normalizeClaim(input.id_basis ?? input.claim)
-	});
-	return `f_${createHash("sha256").update(basis).digest("hex").slice(0, 20)}`;
-}
-function normalizeClaim(c) {
-	return c.toLowerCase().replace(/\s+/g, " ").replace(/[.!?;:,]+$/g, "").trim();
-}
-/**
-* Convenience factory: produce a fully-formed AnalystFinding with the
-* id computed automatically. Analyst code stays terse.
-*/
-function makeFinding(init) {
-	const { id_basis, produced_at, ...rest } = init;
-	return {
-		schema_version: "1.0.0",
-		finding_id: computeFindingId({
-			analyst_id: rest.analyst_id,
-			area: rest.area,
-			subject: rest.subject,
-			claim: rest.claim,
-			id_basis
-		}),
-		produced_at: produced_at ?? (/* @__PURE__ */ new Date()).toISOString(),
-		...rest
-	};
-}
-/** Build a finding whose source is explicitly allowed during candidate generation. */
-function makeProposalFinding(init) {
-	const { proposal_origin, ...finding } = init;
-	return {
-		...makeFinding(finding),
-		proposal_origin
-	};
-}
-//#endregion
-export { assertValidAnalystUsageReceipt as a, validateUsageSettlementTimeout as c, makeProposalFinding as i, deepFreezeCanonicalJson as l, computeFindingId as n, settleUsageReceiptFromCostLedger as o, makeFinding as r, usageReceiptFromCostLedger as s, ANALYST_SEVERITIES as t };
+export { validateUsageSettlementTimeout as i, settleUsageReceiptFromCostLedger as n, usageReceiptFromCostLedger as r, assertValidAnalystUsageReceipt as t };
 
-//# sourceMappingURL=types-DQ0e2E7y.js.map
+//# sourceMappingURL=usage-receipt-Cx3xNeyB.js.map

package/dist/usage-receipt-Cx3xNeyB.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"usage-receipt-Cx3xNeyB.js","names":[],"sources":["../src/analyst/usage-receipt.ts"],"sourcesContent":["import type { CostChannel, CostLedgerFilter, CostLedgerHandle } from '../cost-ledger'\nimport type { AnalystUsageReceipt } from './types'\n\nexport const DEFAULT_USAGE_SETTLEMENT_TIMEOUT_MS = 5_000\n\n/** Convert one ledger channel's complete call set into one analyst receipt. */\nexport function usageReceiptFromCostLedger(\n  ledger: CostLedgerHandle,\n  filter: CostChannel | CostLedgerFilter = 'analyst',\n): AnalystUsageReceipt {\n  const resolvedFilter = typeof filter === 'string' ? { channel: filter } : filter\n  const summary = ledger.summary(resolvedFilter)\n  const receipts = ledger.list(resolvedFilter)\n  const hasReasoningUsage = receipts.some((receipt) => receipt.reasoningTokens !== undefined)\n  const hasCacheWriteUsage = receipts.some((receipt) => receipt.cacheWriteTokens !== undefined)\n  const cost = summary.costProvenance\n  return {\n    calls: summary.totalCalls + summary.pendingCalls,\n    tokens: summary.usageComplete\n      ? {\n          input: summary.inputTokens,\n          output: summary.outputTokens,\n          ...(hasReasoningUsage ? { reasoning: summary.reasoningTokens ?? 0 } : {}),\n          ...(summary.cachedTokens > 0 ? { cached: summary.cachedTokens } : {}),\n          ...(hasCacheWriteUsage ? { cacheWrite: summary.cacheWriteTokens ?? 0 } : {}),\n        }\n      : null,\n    cost,\n    ...(cost.kind === 'uncaptured' ? { knownCostUsd: summary.totalCostUsd } : {}),\n  }\n}\n\nexport interface SettledUsageReceipt {\n  settled: boolean\n  pendingCalls: number\n  receipt: AnalystUsageReceipt\n}\n\n/** Wait a bounded time for late provider receipts, then take one immutable snapshot. */\nexport async function settleUsageReceiptFromCostLedger(\n  ledger: CostLedgerHandle,\n  options: CostLedgerFilter & { timeoutMs?: number } = {},\n): Promise<SettledUsageReceipt> {\n  const { timeoutMs: requestedTimeoutMs, ...requestedFilter } = options\n  const filter: CostLedgerFilter = {\n    channel: requestedFilter.channel ?? 'analyst',\n    ...(requestedFilter.phase === undefined ? {} : { phase: requestedFilter.phase }),\n    ...(requestedFilter.tags === undefined ? {} : { tags: requestedFilter.tags }),\n  }\n  const timeoutMs = validateUsageSettlementTimeout(requestedTimeoutMs)\n  const initial = ledger.summary(filter)\n  const waitResult =\n    initial.pendingCalls === 0\n      ? true\n      : ledger.waitForIdle\n        ? await ledger.waitForIdle({ timeoutMs })\n        : false\n  const pendingCalls = ledger.summary(filter).pendingCalls\n  return {\n    settled: waitResult && pendingCalls === 0,\n    pendingCalls,\n    receipt: usageReceiptFromCostLedger(ledger, filter),\n  }\n}\n\nexport function validateUsageSettlementTimeout(timeoutMs?: number): number {\n  const resolved = timeoutMs ?? DEFAULT_USAGE_SETTLEMENT_TIMEOUT_MS\n  if (!Number.isSafeInteger(resolved) || resolved < 0 || resolved > 2_147_483_647) {\n    throw new TypeError(\n      'settlementTimeoutMs must be a non-negative safe integer no greater than 2147483647',\n    )\n  }\n  return resolved\n}\n\nexport function assertValidAnalystUsageReceipt(\n  receipt: AnalystUsageReceipt,\n  context = 'AnalystContext.recordUsage',\n): void {\n  if (receipt.calls !== null && (!Number.isSafeInteger(receipt.calls) || receipt.calls < 0)) {\n    throw new Error(`${context}: calls must be a non-negative safe integer or null`)\n  }\n  if (receipt.tokens) {\n    assertNonNegativeSafeInteger(receipt.tokens.input, 'tokens.input', context)\n    assertNonNegativeSafeInteger(receipt.tokens.output, 'tokens.output', context)\n    if (receipt.tokens.reasoning !== undefined) {\n      assertNonNegativeSafeInteger(receipt.tokens.reasoning, 'tokens.reasoning', context)\n      if (receipt.tokens.reasoning > receipt.tokens.output) {\n        throw new Error(`${context}: tokens.reasoning must not exceed tokens.output`)\n      }\n    }\n    if (receipt.tokens.cached !== undefined) {\n      assertNonNegativeSafeInteger(receipt.tokens.cached, 'tokens.cached', context)\n    }\n    if (receipt.tokens.cacheWrite !== undefined) {\n      assertNonNegativeSafeInteger(receipt.tokens.cacheWrite, 'tokens.cacheWrite', context)\n    }\n  }\n  if (receipt.cost.kind !== 'uncaptured') {\n    assertNonNegativeFinite(receipt.cost.usd, 'cost.usd', context)\n  } else if (receipt.cost.usd !== null) {\n    throw new Error(`${context}: uncaptured cost.usd must be null`)\n  }\n  if (receipt.knownCostUsd !== undefined) {\n    assertNonNegativeFinite(receipt.knownCostUsd, 'knownCostUsd', context)\n  }\n  if (receipt.partialTokens) {\n    const { input, output } = receipt.partialTokens\n    if (receipt.tokens) {\n      throw new Error(`${context}: partialTokens must be absent when tokens is complete`)\n    }\n    if (input === null && output === null) {\n      throw new Error(`${context}: partialTokens must carry at least one reported side`)\n    }\n    if (input !== null) assertNonNegativeSafeInteger(input, 'partialTokens.input', context)\n    if (output !== null) assertNonNegativeSafeInteger(output, 'partialTokens.output', context)\n  }\n}\n\nfunction assertNonNegativeSafeInteger(value: number, field: string, context: string): void {\n  if (!Number.isSafeInteger(value) || value < 0) {\n    throw new Error(`${context}: ${field} must be a non-negative safe integer`)\n  }\n}\n\nfunction assertNonNegativeFinite(value: number, field: string, context: string): void {\n  if (!Number.isFinite(value) || value < 0) {\n    throw new Error(`${context}: ${field} must be a non-negative finite number`)\n  }\n}\n"],"mappings":";AAMA,SAAgB,2BACd,QACA,SAAyC,WACpB;CACrB,MAAM,iBAAiB,OAAO,WAAW,WAAW,EAAE,SAAS,OAAO,IAAI;CAC1E,MAAM,UAAU,OAAO,QAAQ,cAAc;CAC7C,MAAM,WAAW,OAAO,KAAK,cAAc;CAC3C,MAAM,oBAAoB,SAAS,MAAM,YAAY,QAAQ,oBAAoB,KAAA,CAAS;CAC1F,MAAM,qBAAqB,SAAS,MAAM,YAAY,QAAQ,qBAAqB,KAAA,CAAS;CAC5F,MAAM,OAAO,QAAQ;CACrB,OAAO;EACL,OAAO,QAAQ,aAAa,QAAQ;EACpC,QAAQ,QAAQ,gBACZ;GACE,OAAO,QAAQ;GACf,QAAQ,QAAQ;GAChB,GAAI,oBAAoB,EAAE,WAAW,QAAQ,mBAAmB,EAAE,IAAI,CAAC;GACvE,GAAI,QAAQ,eAAe,IAAI,EAAE,QAAQ,QAAQ,aAAa,IAAI,CAAC;GACnE,GAAI,qBAAqB,EAAE,YAAY,QAAQ,oBAAoB,EAAE,IAAI,CAAC;EAC5E,IACA;EACJ;EACA,GAAI,KAAK,SAAS,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;CAC7E;AACF;;AASA,eAAsB,iCACpB,QACA,UAAqD,CAAC,GACxB;CAC9B,MAAM,EAAE,WAAW,oBAAoB,GAAG,oBAAoB;CAC9D,MAAM,SAA2B;EAC/B,SAAS,gBAAgB,WAAW;EACpC,GAAI,gBAAgB,UAAU,KAAA,IAAY,CAAC,IAAI,EAAE,OAAO,gBAAgB,MAAM;EAC9E,GAAI,gBAAgB,SAAS,KAAA,IAAY,CAAC,IAAI,EAAE,MAAM,gBAAgB,KAAK;CAC7E;CACA,MAAM,YAAY,+BAA+B,kBAAkB;CAEnE,MAAM,aADU,OAAO,QAAQ,MAEvB,CAAC,CAAC,iBAAiB,IACrB,OACA,OAAO,cACL,MAAM,OAAO,YAAY,EAAE,UAAU,CAAC,IACtC;CACR,MAAM,eAAe,OAAO,QAAQ,MAAM,CAAC,CAAC;CAC5C,OAAO;EACL,SAAS,cAAc,iBAAiB;EACxC;EACA,SAAS,2BAA2B,QAAQ,MAAM;CACpD;AACF;AAEA,SAAgB,+BAA+B,WAA4B;CACzE,MAAM,WAAW,aAAA;CACjB,IAAI,CAAC,OAAO,cAAc,QAAQ,KAAK,WAAW,KAAK,WAAW,YAChE,MAAM,IAAI,UACR,oFACF;CAEF,OAAO;AACT;AAEA,SAAgB,+BACd,SACA,UAAU,8BACJ;CACN,IAAI,QAAQ,UAAU,SAAS,CAAC,OAAO,cAAc,QAAQ,KAAK,KAAK,QAAQ,QAAQ,IACrF,MAAM,IAAI,MAAM,GAAG,QAAQ,oDAAoD;CAEjF,IAAI,QAAQ,QAAQ;EAClB,6BAA6B,QAAQ,OAAO,OAAO,gBAAgB,OAAO;EAC1E,6BAA6B,QAAQ,OAAO,QAAQ,iBAAiB,OAAO;EAC5E,IAAI,QAAQ,OAAO,cAAc,KAAA,GAAW;GAC1C,6BAA6B,QAAQ,OAAO,WAAW,oBAAoB,OAAO;GAClF,IAAI,QAAQ,OAAO,YAAY,QAAQ,OAAO,QAC5C,MAAM,IAAI,MAAM,GAAG,QAAQ,iDAAiD;EAEhF;EACA,IAAI,QAAQ,OAAO,WAAW,KAAA,GAC5B,6BAA6B,QAAQ,OAAO,QAAQ,iBAAiB,OAAO;EAE9E,IAAI,QAAQ,OAAO,eAAe,KAAA,GAChC,6BAA6B,QAAQ,OAAO,YAAY,qBAAqB,OAAO;CAExF;CACA,IAAI,QAAQ,KAAK,SAAS,cACxB,wBAAwB,QAAQ,KAAK,KAAK,YAAY,OAAO;MACxD,IAAI,QAAQ,KAAK,QAAQ,MAC9B,MAAM,IAAI,MAAM,GAAG,QAAQ,mCAAmC;CAEhE,IAAI,QAAQ,iBAAiB,KAAA,GAC3B,wBAAwB,QAAQ,cAAc,gBAAgB,OAAO;CAEvE,IAAI,QAAQ,eAAe;EACzB,MAAM,EAAE,OAAO,WAAW,QAAQ;EAClC,IAAI,QAAQ,QACV,MAAM,IAAI,MAAM,GAAG,QAAQ,uDAAuD;EAEpF,IAAI,UAAU,QAAQ,WAAW,MAC/B,MAAM,IAAI,MAAM,GAAG,QAAQ,sDAAsD;EAEnF,IAAI,UAAU,MAAM,6BAA6B,OAAO,uBAAuB,OAAO;EACtF,IAAI,WAAW,MAAM,6BAA6B,QAAQ,wBAAwB,OAAO;CAC3F;AACF;AAEA,SAAS,6BAA6B,OAAe,OAAe,SAAuB;CACzF,IAAI,CAAC,OAAO,cAAc,KAAK,KAAK,QAAQ,GAC1C,MAAM,IAAI,MAAM,GAAG,QAAQ,IAAI,MAAM,qCAAqC;AAE9E;AAEA,SAAS,wBAAwB,OAAe,OAAe,SAAuB;CACpF,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,QAAQ,GACrC,MAAM,IAAI,MAAM,GAAG,QAAQ,IAAI,MAAM,sCAAsC;AAE/E"}

package/dist/{verdict-BQ3pCFf8.js → verdict-DhPoXApk.js}

@@ -1,4 +1,4 @@
-import { n as contentHash } from "./verdict-cache-B3eCVQtY.js";
+import { n as contentHash } from "./verdict-cache-BQEsVy9n.js";
 //#region src/verdict.ts
 /**
 * Validator-output verdict — substrate primitive for "did this output pass,
@@ -58,4 +58,4 @@ function equivalenceVerdict(record) {
 //#endregion
 export { equivalenceVerdict as n, certificationEvidenceDigest as t };
 
-//# sourceMappingURL=verdict-BQ3pCFf8.js.map
+//# sourceMappingURL=verdict-DhPoXApk.js.map

package/dist/{verdict-BQ3pCFf8.js.map → verdict-DhPoXApk.js.map}

@@ -1 +1 @@
-{"version":3,"file":"verdict-BQ3pCFf8.js","names":[],"sources":["../src/verdict.ts"],"sourcesContent":["/**\n * Validator-output verdict — substrate primitive for \"did this output pass,\n * and how well?\"\n *\n * Used by:\n *   - `@tangle-network/agent-eval/matrix` — verdict per cell in the cartesian.\n *   - `@tangle-network/agent-runtime` — Validator<Output, Verdict = DefaultVerdict>.\n *     Runtime keeps `Validator` because it's coupled to runtime-shaped\n *     `ValidationCtx` (iteration, signal, traceEmitter); the verdict TYPE\n *     itself is a substrate concept and lives here.\n *\n * Repo layering: agent-eval is the substrate (no upward deps). Both\n * agent-runtime and agent-knowledge consume this type FROM agent-eval —\n * never the other way around. See CLAUDE.md \"Repo layering\" for the rule.\n */\n\nimport { contentHash } from './verdict-cache'\nimport type {\n  CheckerIdentity,\n  EquivalenceRecord,\n  VerificationStrategySource,\n} from './verification-strategy'\n\n/**\n * What certified a verdict — the epistemics a bare `valid` + `score` pair\n * cannot carry. A kernel-checked proof and an LLM judge can produce the\n * same `{ valid: true, score: 1 }`; this record is what tells them apart.\n *\n * Every field is required on purpose. A certification without a checker\n * identity or an evidence digest is exactly the unverifiable claim this\n * record exists to kill.\n */\nexport interface VerdictCertification {\n  /**\n   * Which verification-strategy member produced the certificate. Each\n   * member's failure mode is documented on the family type and in\n   * `VERIFICATION_STRATEGIES` (src/verification-strategy.ts).\n   */\n  strategy: VerificationStrategySource\n  /**\n   * Exact checker identity, e.g. `{ name: 'lean4', version: '4.33.0',\n   * pins: { mathlib: 'db584cd6d46c' } }` for a kernel, or a judge model\n   * id + snapshot for a judge.\n   */\n  checker: CheckerIdentity\n  /**\n   * Every step the certificate rests on that the checker did NOT verify,\n   * e.g. 'diagonal embedding argued in prose' or 'agreement arms share\n   * training-corpus blind spots'. An empty array is the producer's\n   * explicit, auditable claim that there are none — it is not a default.\n   */\n  assumptions: string[]\n  /** Digest (e.g. sha256) of the evidence artifact the checker consumed or produced. */\n  evidenceDigest: string\n}\n\n/**\n * Minimal verdict shape — `valid` + `score` are required; `scores` +\n * `notes` are optional surface. Validators that need richer shapes\n * parameterise `Validator<Output, MyVerdict>` with their own type.\n *\n * Need structured extras? Extend DefaultVerdict with typed fields — never\n * serialize extras into `notes`.\n */\nexport interface DefaultVerdict {\n  /** Whether the output meets the validator's pass criteria. */\n  valid: boolean\n  /** Aggregate score in [0, 1]. Drivers use this for winner selection. */\n  score: number\n  /** Per-dimension scores. Free-form; weighted into `score` by the validator. */\n  scores?: Record<string, number>\n  /** Human-readable rationale; surfaces in trace + final-result `winner.verdict`. */\n  notes?: string\n  /**\n   * What certified this verdict, when anything did. Absent = uncertified:\n   * scored, but no strategy vouches for it. A consumer that cares reads\n   * `certification.strategy` to tell a kernel-checked verdict from a\n   * judge-scored one; a consumer that does not read it sees the exact\n   * verdict it always saw.\n   */\n  certification?: VerdictCertification\n}\n\n/**\n * Digest a certification's evidence artifact: sha-256 over the canonical\n * JSON of the evidence AS SERIALIZED. The JSON round-trip drops\n * `undefined`-valued fields on purpose — certified verdicts persist as\n * JSON, so the digest must cover exactly what a reader of the persisted\n * artifact can re-digest, nothing the wire cannot carry. Throws when the\n * evidence has no JSON form at all.\n */\nexport function certificationEvidenceDigest(evidence: unknown): string {\n  const serialized = JSON.stringify(evidence)\n  if (serialized === undefined) {\n    throw new Error(\n      'certificationEvidenceDigest: evidence has no JSON form — a certificate over nothing is the unverifiable claim this record exists to kill',\n    )\n  }\n  return contentHash(JSON.parse(serialized))\n}\n\n/**\n * Land a finished statement-equivalence record in the shared verdict\n * spine. `valid` answers the protocol's question — are the two blind\n * statements equivalent — so a refutation is `valid: false` even though a\n * separating witness in hand is a successful protocol outcome.\n *\n * Certification is present exactly when the obligation was discharged\n * (proved or refuted): the checker vouches for the discharge either way.\n * An unresolved obligation certifies nothing and keeps its reason in\n * `notes`.\n */\nexport function equivalenceVerdict(record: EquivalenceRecord): DefaultVerdict {\n  const { obligation, spec, arms } = record\n  const discharged = obligation.status !== 'unresolved'\n  const notes =\n    obligation.status === 'proved'\n      ? `statements equivalent: ${spec.artifact}`\n      : obligation.status === 'refuted-with-separating-witness'\n        ? `formalization gap witnessed: ${obligation.separatingWitness}`\n        : `obligation undischarged: ${obligation.unresolvedReason}`\n  return {\n    valid: obligation.status === 'proved',\n    score: obligation.status === 'proved' ? 1 : 0,\n    notes,\n    ...(discharged && obligation.evidenceDigest !== undefined\n      ? {\n          certification: {\n            strategy: spec.source,\n            checker: obligation.checker,\n            // The arms' blindness declarations are attested by the arms,\n            // never verified by the checker — the one step every\n            // equivalence certificate rests on.\n            assumptions: [\n              `arm '${arms[0].armId}' and arm '${arms[1].armId}' self-declare blind derivation`,\n            ],\n            evidenceDigest: obligation.evidenceDigest,\n          } satisfies VerdictCertification,\n        }\n      : {}),\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;AA2FA,SAAgB,4BAA4B,UAA2B;CACrE,MAAM,aAAa,KAAK,UAAU,QAAQ;CAC1C,IAAI,eAAe,KAAA,GACjB,MAAM,IAAI,MACR,0IACF;CAEF,OAAO,YAAY,KAAK,MAAM,UAAU,CAAC;AAC3C;;;;;;;;;;;;AAaA,SAAgB,mBAAmB,QAA2C;CAC5E,MAAM,EAAE,YAAY,MAAM,SAAS;CACnC,MAAM,aAAa,WAAW,WAAW;CACzC,MAAM,QACJ,WAAW,WAAW,WAClB,0BAA0B,KAAK,aAC/B,WAAW,WAAW,oCACpB,gCAAgC,WAAW,sBAC3C,4BAA4B,WAAW;CAC/C,OAAO;EACL,OAAO,WAAW,WAAW;EAC7B,OAAO,WAAW,WAAW,WAAW,IAAI;EAC5C;EACA,GAAI,cAAc,WAAW,mBAAmB,KAAA,IAC5C,EACE,eAAe;GACb,UAAU,KAAK;GACf,SAAS,WAAW;GAIpB,aAAa,CACX,QAAQ,KAAK,EAAE,CAAC,MAAM,aAAa,KAAK,EAAE,CAAC,MAAM,gCACnD;GACA,gBAAgB,WAAW;EAC7B,EACF,IACA,CAAC;CACP;AACF"}
+{"version":3,"file":"verdict-DhPoXApk.js","names":[],"sources":["../src/verdict.ts"],"sourcesContent":["/**\n * Validator-output verdict — substrate primitive for \"did this output pass,\n * and how well?\"\n *\n * Used by:\n *   - `@tangle-network/agent-eval/matrix` — verdict per cell in the cartesian.\n *   - `@tangle-network/agent-runtime` — Validator<Output, Verdict = DefaultVerdict>.\n *     Runtime keeps `Validator` because it's coupled to runtime-shaped\n *     `ValidationCtx` (iteration, signal, traceEmitter); the verdict TYPE\n *     itself is a substrate concept and lives here.\n *\n * Repo layering: agent-eval is the substrate (no upward deps). Both\n * agent-runtime and agent-knowledge consume this type FROM agent-eval —\n * never the other way around. See CLAUDE.md \"Repo layering\" for the rule.\n */\n\nimport { contentHash } from './verdict-cache'\nimport type {\n  CheckerIdentity,\n  EquivalenceRecord,\n  VerificationStrategySource,\n} from './verification-strategy'\n\n/**\n * What certified a verdict — the epistemics a bare `valid` + `score` pair\n * cannot carry. A kernel-checked proof and an LLM judge can produce the\n * same `{ valid: true, score: 1 }`; this record is what tells them apart.\n *\n * Every field is required on purpose. A certification without a checker\n * identity or an evidence digest is exactly the unverifiable claim this\n * record exists to kill.\n */\nexport interface VerdictCertification {\n  /**\n   * Which verification-strategy member produced the certificate. Each\n   * member's failure mode is documented on the family type and in\n   * `VERIFICATION_STRATEGIES` (src/verification-strategy.ts).\n   */\n  strategy: VerificationStrategySource\n  /**\n   * Exact checker identity, e.g. `{ name: 'lean4', version: '4.33.0',\n   * pins: { mathlib: 'db584cd6d46c' } }` for a kernel, or a judge model\n   * id + snapshot for a judge.\n   */\n  checker: CheckerIdentity\n  /**\n   * Every step the certificate rests on that the checker did NOT verify,\n   * e.g. 'diagonal embedding argued in prose' or 'agreement arms share\n   * training-corpus blind spots'. An empty array is the producer's\n   * explicit, auditable claim that there are none — it is not a default.\n   */\n  assumptions: string[]\n  /** Digest (e.g. sha256) of the evidence artifact the checker consumed or produced. */\n  evidenceDigest: string\n}\n\n/**\n * Minimal verdict shape — `valid` + `score` are required; `scores` +\n * `notes` are optional surface. Validators that need richer shapes\n * parameterise `Validator<Output, MyVerdict>` with their own type.\n *\n * Need structured extras? Extend DefaultVerdict with typed fields — never\n * serialize extras into `notes`.\n */\nexport interface DefaultVerdict {\n  /** Whether the output meets the validator's pass criteria. */\n  valid: boolean\n  /** Aggregate score in [0, 1]. Drivers use this for winner selection. */\n  score: number\n  /** Per-dimension scores. Free-form; weighted into `score` by the validator. */\n  scores?: Record<string, number>\n  /** Human-readable rationale; surfaces in trace + final-result `winner.verdict`. */\n  notes?: string\n  /**\n   * What certified this verdict, when anything did. Absent = uncertified:\n   * scored, but no strategy vouches for it. A consumer that cares reads\n   * `certification.strategy` to tell a kernel-checked verdict from a\n   * judge-scored one; a consumer that does not read it sees the exact\n   * verdict it always saw.\n   */\n  certification?: VerdictCertification\n}\n\n/**\n * Digest a certification's evidence artifact: sha-256 over the canonical\n * JSON of the evidence AS SERIALIZED. The JSON round-trip drops\n * `undefined`-valued fields on purpose — certified verdicts persist as\n * JSON, so the digest must cover exactly what a reader of the persisted\n * artifact can re-digest, nothing the wire cannot carry. Throws when the\n * evidence has no JSON form at all.\n */\nexport function certificationEvidenceDigest(evidence: unknown): string {\n  const serialized = JSON.stringify(evidence)\n  if (serialized === undefined) {\n    throw new Error(\n      'certificationEvidenceDigest: evidence has no JSON form — a certificate over nothing is the unverifiable claim this record exists to kill',\n    )\n  }\n  return contentHash(JSON.parse(serialized))\n}\n\n/**\n * Land a finished statement-equivalence record in the shared verdict\n * spine. `valid` answers the protocol's question — are the two blind\n * statements equivalent — so a refutation is `valid: false` even though a\n * separating witness in hand is a successful protocol outcome.\n *\n * Certification is present exactly when the obligation was discharged\n * (proved or refuted): the checker vouches for the discharge either way.\n * An unresolved obligation certifies nothing and keeps its reason in\n * `notes`.\n */\nexport function equivalenceVerdict(record: EquivalenceRecord): DefaultVerdict {\n  const { obligation, spec, arms } = record\n  const discharged = obligation.status !== 'unresolved'\n  const notes =\n    obligation.status === 'proved'\n      ? `statements equivalent: ${spec.artifact}`\n      : obligation.status === 'refuted-with-separating-witness'\n        ? `formalization gap witnessed: ${obligation.separatingWitness}`\n        : `obligation undischarged: ${obligation.unresolvedReason}`\n  return {\n    valid: obligation.status === 'proved',\n    score: obligation.status === 'proved' ? 1 : 0,\n    notes,\n    ...(discharged && obligation.evidenceDigest !== undefined\n      ? {\n          certification: {\n            strategy: spec.source,\n            checker: obligation.checker,\n            // The arms' blindness declarations are attested by the arms,\n            // never verified by the checker — the one step every\n            // equivalence certificate rests on.\n            assumptions: [\n              `arm '${arms[0].armId}' and arm '${arms[1].armId}' self-declare blind derivation`,\n            ],\n            evidenceDigest: obligation.evidenceDigest,\n          } satisfies VerdictCertification,\n        }\n      : {}),\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;AA2FA,SAAgB,4BAA4B,UAA2B;CACrE,MAAM,aAAa,KAAK,UAAU,QAAQ;CAC1C,IAAI,eAAe,KAAA,GACjB,MAAM,IAAI,MACR,0IACF;CAEF,OAAO,YAAY,KAAK,MAAM,UAAU,CAAC;AAC3C;;;;;;;;;;;;AAaA,SAAgB,mBAAmB,QAA2C;CAC5E,MAAM,EAAE,YAAY,MAAM,SAAS;CACnC,MAAM,aAAa,WAAW,WAAW;CACzC,MAAM,QACJ,WAAW,WAAW,WAClB,0BAA0B,KAAK,aAC/B,WAAW,WAAW,oCACpB,gCAAgC,WAAW,sBAC3C,4BAA4B,WAAW;CAC/C,OAAO;EACL,OAAO,WAAW,WAAW;EAC7B,OAAO,WAAW,WAAW,WAAW,IAAI;EAC5C;EACA,GAAI,cAAc,WAAW,mBAAmB,KAAA,IAC5C,EACE,eAAe;GACb,UAAU,KAAK;GACf,SAAS,WAAW;GAIpB,aAAa,CACX,QAAQ,KAAK,EAAE,CAAC,MAAM,aAAa,KAAK,EAAE,CAAC,MAAM,gCACnD;GACA,gBAAgB,WAAW;EAC7B,EACF,IACA,CAAC;CACP;AACF"}

package/dist/{verdict-cache-B3eCVQtY.js → verdict-cache-BQEsVy9n.js}

@@ -1,4 +1,4 @@
-import { a as hashCanonical, r as canonicalString } from "./canonical-DPyQ_rpt.js";
+import { a as hashCanonical, r as canonicalString } from "./canonical-BsZVgBhW.js";
 import { appendFileSync, existsSync, readFileSync } from "node:fs";
 //#region src/verdict-cache.ts
 /**
@@ -85,4 +85,4 @@ function fileVerdictCache(path) {
 //#endregion
 export { contentHash as n, fileVerdictCache as r, canonicalJson as t };
 
-//# sourceMappingURL=verdict-cache-B3eCVQtY.js.map
+//# sourceMappingURL=verdict-cache-BQEsVy9n.js.map

package/dist/{verdict-cache-B3eCVQtY.js.map → verdict-cache-BQEsVy9n.js.map}

@@ -1 +1 @@
-{"version":3,"file":"verdict-cache-B3eCVQtY.js","names":[],"sources":["../src/verdict-cache.ts"],"sourcesContent":["/**\n * Content-addressed judge-verdict caching.\n *\n * LAW: cache JUDGE VERDICTS only — judging the same artifact with the same\n * judge+rubric is pure. NEVER cache agent rollouts. (A router that cached\n * identical fanout prompts silently destroyed best-of-N diversity; rollout\n * caching reintroduces that failure class. Judging has no diversity to\n * destroy — same artifact + same rubric ⇒ same verdict is the desired\n * property, not a bug.)\n *\n * The cache key is `v2:` + a sha-256 over the canonical JSON of everything\n * that can change a verdict: the artifact content, the scenario id, the judge\n * name, the full dimension list (key + description — the description IS the\n * rubric text shown to the judge), and a caller-supplied `judgeVersion`.\n * `judgeVersion` is REQUIRED: a judge whose prompt/model/ensemble changes\n * without a version bump would otherwise silently serve stale verdicts.\n * The `v2:` prefix names the key scheme; a store written under another scheme\n * misses once and repopulates.\n *\n * Canonical JSON is RFC 8785 from `ledger-core/canonical` — the package's one\n * identity scheme. It refuses `undefined`, functions, symbols, non-finite\n * numbers, class instances (including `Date`), `Map`, `Set`, cycles, and\n * unpaired surrogates: an artifact that cannot be unambiguously serialized\n * cannot be content-addressed, and coercing it would let two different\n * artifacts collide on one key. Pass timestamps as ISO strings.\n */\n\nimport { appendFileSync, existsSync, readFileSync } from 'node:fs'\nimport type { JudgeConfig, JudgeScore, Scenario } from './campaign/types'\nimport { canonicalString, hashCanonical } from './ledger-core/canonical'\n\n// ── canonical JSON + content hash ─────────────────────────────────────────\n\n/** Key scheme of every entry `cachedJudge` writes. */\nexport const VERDICT_CACHE_KEY_SCHEME = 'v2'\n\n/**\n * Canonical JSON (RFC 8785): object keys sorted recursively, so two\n * semantically equal values produce byte-identical output regardless of key\n * insertion order. Throws `LedgerCanonicalizationError` on any value with no\n * faithful JSON form — see `ledger-core/canonical`.\n */\nexport function canonicalJson(value: unknown): string {\n  return canonicalString(value)\n}\n\n/** Hex sha-256 over `canonicalJson(value)`. The content address used by the\n *  verdict cache and report attestation. Equals `hashCanonical(value)` with\n *  the `sha256:` prefix removed. */\nexport function contentHash(value: unknown): string {\n  return hashCanonical(value).slice('sha256:'.length)\n}\n\n// ── store contract ─────────────────────────────────────────────────────────\n\n/** Pluggable verdict store. Sync or async on both legs — `cachedJudge`\n *  awaits the results either way. */\nexport interface VerdictCacheStore {\n  get(key: string): Promise<JudgeScore | undefined> | JudgeScore | undefined\n  set(key: string, score: JudgeScore): Promise<void> | void\n}\n\n/** Process-local Map-backed store. */\nexport function inMemoryVerdictCache(): VerdictCacheStore {\n  const entries = new Map<string, JudgeScore>()\n  return {\n    get: (key) => entries.get(key),\n    set: (key, score) => {\n      entries.set(key, score)\n    },\n  }\n}\n\ninterface VerdictCacheLine {\n  key: string\n  score: JudgeScore\n}\n\nfunction parseCacheLine(line: string, path: string, lineNo: number): VerdictCacheLine {\n  let parsed: unknown\n  try {\n    parsed = JSON.parse(line)\n  } catch (err) {\n    throw new Error(\n      `fileVerdictCache: corrupt JSONL at ${path}:${lineNo} — ${err instanceof Error ? err.message : String(err)}`,\n    )\n  }\n  const rec = parsed as Partial<VerdictCacheLine>\n  if (\n    typeof rec !== 'object' ||\n    rec === null ||\n    typeof rec.key !== 'string' ||\n    typeof rec.score !== 'object' ||\n    rec.score === null ||\n    typeof rec.score.composite !== 'number' ||\n    typeof rec.score.dimensions !== 'object'\n  ) {\n    throw new Error(\n      `fileVerdictCache: invalid record shape at ${path}:${lineNo} — expected {key, score:{dimensions, composite, notes}}`,\n    )\n  }\n  return rec as VerdictCacheLine\n}\n\n/**\n * JSONL-file-backed store: the full file is loaded into an in-memory index at\n * construction; every `set` appends one line synchronously (durable before\n * the verdict is returned). A corrupt or malformed line throws at load with\n * file:line — a skipped line would silently re-judge (cost) or, worse, mask\n * a half-written file that needs operator attention.\n */\nexport function fileVerdictCache(path: string): VerdictCacheStore {\n  const entries = new Map<string, JudgeScore>()\n  if (existsSync(path)) {\n    const lines = readFileSync(path, 'utf8').split('\\n')\n    for (let i = 0; i < lines.length; i++) {\n      const line = lines[i]\n      if (line === undefined || line.trim() === '') continue\n      const rec = parseCacheLine(line, path, i + 1)\n      entries.set(rec.key, rec.score)\n    }\n  }\n  return {\n    get: (key) => entries.get(key),\n    set: (key, score) => {\n      appendFileSync(path, `${JSON.stringify({ key, score })}\\n`, 'utf8')\n      entries.set(key, score)\n    },\n  }\n}\n\n// ── cached judge wrapper ───────────────────────────────────────────────────\n\nexport interface VerdictCacheStats {\n  hits: number\n  misses: number\n}\n\nexport interface CachedJudgeOptions {\n  /** REQUIRED — part of the cache key. Bump on any change to the judge's\n   *  prompt, model, ensemble, or scoring logic; silent judge upgrades must\n   *  never serve stale verdicts. */\n  judgeVersion: string\n}\n\n/** The wrapped judge: same `JudgeConfig` seam, plus hit/miss observability. */\nexport type CachedJudge<TArtifact, TScenario extends Scenario = Scenario> = JudgeConfig<\n  TArtifact,\n  TScenario\n> & {\n  stats(): VerdictCacheStats\n}\n\n/**\n * Wrap a `JudgeConfig` so repeat judgments of the same artifact are served\n * from the store instead of re-invoking `score()`. The wrapper is generic\n * over the judge's own type parameters and preserves `appliesTo` — it is a\n * drop-in replacement anywhere a `JudgeConfig` is accepted.\n *\n * A judge that throws is NOT cached: the error propagates and the next\n * attempt re-judges (caching a failure would pin a transient outage forever).\n */\nexport function cachedJudge<TArtifact, TScenario extends Scenario = Scenario>(\n  judge: JudgeConfig<TArtifact, TScenario>,\n  store: VerdictCacheStore,\n  options: CachedJudgeOptions,\n): CachedJudge<TArtifact, TScenario> {\n  if (typeof options.judgeVersion !== 'string' || options.judgeVersion.trim() === '') {\n    throw new Error('cachedJudge: judgeVersion is required and must be a non-empty string')\n  }\n  const stats: VerdictCacheStats = { hits: 0, misses: 0 }\n  const wrapped: CachedJudge<TArtifact, TScenario> = {\n    name: judge.name,\n    dimensions: judge.dimensions,\n    judgeVersion: options.judgeVersion,\n    async score(input) {\n      const key = `${VERDICT_CACHE_KEY_SCHEME}:${contentHash({\n        artifact: canonicalJson(input.artifact),\n        scenarioId: input.scenario.id,\n        judgeName: judge.name,\n        dimensions: judge.dimensions,\n        judgeVersion: options.judgeVersion,\n      })}`\n      const cached = await store.get(key)\n      if (cached !== undefined) {\n        stats.hits += 1\n        return cached\n      }\n      const score = await judge.score(input)\n      await store.set(key, score)\n      stats.misses += 1\n      return score\n    },\n    stats: () => ({ ...stats }),\n  }\n  if (judge.appliesTo) wrapped.appliesTo = judge.appliesTo\n  return wrapped\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA0CA,SAAgB,cAAc,OAAwB;CACpD,OAAO,gBAAgB,KAAK;AAC9B;;;;AAKA,SAAgB,YAAY,OAAwB;CAClD,OAAO,cAAc,KAAK,CAAC,CAAC,MAAM,CAAgB;AACpD;AA2BA,SAAS,eAAe,MAAc,MAAc,QAAkC;CACpF,IAAI;CACJ,IAAI;EACF,SAAS,KAAK,MAAM,IAAI;CAC1B,SAAS,KAAK;EACZ,MAAM,IAAI,MACR,sCAAsC,KAAK,GAAG,OAAO,KAAK,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,GAC3G;CACF;CACA,MAAM,MAAM;CACZ,IACE,OAAO,QAAQ,YACf,QAAQ,QACR,OAAO,IAAI,QAAQ,YACnB,OAAO,IAAI,UAAU,YACrB,IAAI,UAAU,QACd,OAAO,IAAI,MAAM,cAAc,YAC/B,OAAO,IAAI,MAAM,eAAe,UAEhC,MAAM,IAAI,MACR,6CAA6C,KAAK,GAAG,OAAO,wDAC9D;CAEF,OAAO;AACT;;;;;;;;AASA,SAAgB,iBAAiB,MAAiC;CAChE,MAAM,0BAAU,IAAI,IAAwB;CAC5C,IAAI,WAAW,IAAI,GAAG;EACpB,MAAM,QAAQ,aAAa,MAAM,MAAM,CAAC,CAAC,MAAM,IAAI;EACnD,KAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;GACrC,MAAM,OAAO,MAAM;GACnB,IAAI,SAAS,KAAA,KAAa,KAAK,KAAK,MAAM,IAAI;GAC9C,MAAM,MAAM,eAAe,MAAM,MAAM,IAAI,CAAC;GAC5C,QAAQ,IAAI,IAAI,KAAK,IAAI,KAAK;EAChC;CACF;CACA,OAAO;EACL,MAAM,QAAQ,QAAQ,IAAI,GAAG;EAC7B,MAAM,KAAK,UAAU;GACnB,eAAe,MAAM,GAAG,KAAK,UAAU;IAAE;IAAK;GAAM,CAAC,EAAE,KAAK,MAAM;GAClE,QAAQ,IAAI,KAAK,KAAK;EACxB;CACF;AACF"}
+{"version":3,"file":"verdict-cache-BQEsVy9n.js","names":[],"sources":["../src/verdict-cache.ts"],"sourcesContent":["/**\n * Content-addressed judge-verdict caching.\n *\n * LAW: cache JUDGE VERDICTS only — judging the same artifact with the same\n * judge+rubric is pure. NEVER cache agent rollouts. (A router that cached\n * identical fanout prompts silently destroyed best-of-N diversity; rollout\n * caching reintroduces that failure class. Judging has no diversity to\n * destroy — same artifact + same rubric ⇒ same verdict is the desired\n * property, not a bug.)\n *\n * The cache key is `v2:` + a sha-256 over the canonical JSON of everything\n * that can change a verdict: the artifact content, the scenario id, the judge\n * name, the full dimension list (key + description — the description IS the\n * rubric text shown to the judge), and a caller-supplied `judgeVersion`.\n * `judgeVersion` is REQUIRED: a judge whose prompt/model/ensemble changes\n * without a version bump would otherwise silently serve stale verdicts.\n * The `v2:` prefix names the key scheme; a store written under another scheme\n * misses once and repopulates.\n *\n * Canonical JSON is RFC 8785 from `ledger-core/canonical` — the package's one\n * identity scheme. It refuses `undefined`, functions, symbols, non-finite\n * numbers, class instances (including `Date`), `Map`, `Set`, cycles, and\n * unpaired surrogates: an artifact that cannot be unambiguously serialized\n * cannot be content-addressed, and coercing it would let two different\n * artifacts collide on one key. Pass timestamps as ISO strings.\n */\n\nimport { appendFileSync, existsSync, readFileSync } from 'node:fs'\nimport type { JudgeConfig, JudgeScore, Scenario } from './campaign/types'\nimport { canonicalString, hashCanonical } from './ledger-core/canonical'\n\n// ── canonical JSON + content hash ─────────────────────────────────────────\n\n/** Key scheme of every entry `cachedJudge` writes. */\nexport const VERDICT_CACHE_KEY_SCHEME = 'v2'\n\n/**\n * Canonical JSON (RFC 8785): object keys sorted recursively, so two\n * semantically equal values produce byte-identical output regardless of key\n * insertion order. Throws `LedgerCanonicalizationError` on any value with no\n * faithful JSON form — see `ledger-core/canonical`.\n */\nexport function canonicalJson(value: unknown): string {\n  return canonicalString(value)\n}\n\n/** Hex sha-256 over `canonicalJson(value)`. The content address used by the\n *  verdict cache and report attestation. Equals `hashCanonical(value)` with\n *  the `sha256:` prefix removed. */\nexport function contentHash(value: unknown): string {\n  return hashCanonical(value).slice('sha256:'.length)\n}\n\n// ── store contract ─────────────────────────────────────────────────────────\n\n/** Pluggable verdict store. Sync or async on both legs — `cachedJudge`\n *  awaits the results either way. */\nexport interface VerdictCacheStore {\n  get(key: string): Promise<JudgeScore | undefined> | JudgeScore | undefined\n  set(key: string, score: JudgeScore): Promise<void> | void\n}\n\n/** Process-local Map-backed store. */\nexport function inMemoryVerdictCache(): VerdictCacheStore {\n  const entries = new Map<string, JudgeScore>()\n  return {\n    get: (key) => entries.get(key),\n    set: (key, score) => {\n      entries.set(key, score)\n    },\n  }\n}\n\ninterface VerdictCacheLine {\n  key: string\n  score: JudgeScore\n}\n\nfunction parseCacheLine(line: string, path: string, lineNo: number): VerdictCacheLine {\n  let parsed: unknown\n  try {\n    parsed = JSON.parse(line)\n  } catch (err) {\n    throw new Error(\n      `fileVerdictCache: corrupt JSONL at ${path}:${lineNo} — ${err instanceof Error ? err.message : String(err)}`,\n    )\n  }\n  const rec = parsed as Partial<VerdictCacheLine>\n  if (\n    typeof rec !== 'object' ||\n    rec === null ||\n    typeof rec.key !== 'string' ||\n    typeof rec.score !== 'object' ||\n    rec.score === null ||\n    typeof rec.score.composite !== 'number' ||\n    typeof rec.score.dimensions !== 'object'\n  ) {\n    throw new Error(\n      `fileVerdictCache: invalid record shape at ${path}:${lineNo} — expected {key, score:{dimensions, composite, notes}}`,\n    )\n  }\n  return rec as VerdictCacheLine\n}\n\n/**\n * JSONL-file-backed store: the full file is loaded into an in-memory index at\n * construction; every `set` appends one line synchronously (durable before\n * the verdict is returned). A corrupt or malformed line throws at load with\n * file:line — a skipped line would silently re-judge (cost) or, worse, mask\n * a half-written file that needs operator attention.\n */\nexport function fileVerdictCache(path: string): VerdictCacheStore {\n  const entries = new Map<string, JudgeScore>()\n  if (existsSync(path)) {\n    const lines = readFileSync(path, 'utf8').split('\\n')\n    for (let i = 0; i < lines.length; i++) {\n      const line = lines[i]\n      if (line === undefined || line.trim() === '') continue\n      const rec = parseCacheLine(line, path, i + 1)\n      entries.set(rec.key, rec.score)\n    }\n  }\n  return {\n    get: (key) => entries.get(key),\n    set: (key, score) => {\n      appendFileSync(path, `${JSON.stringify({ key, score })}\\n`, 'utf8')\n      entries.set(key, score)\n    },\n  }\n}\n\n// ── cached judge wrapper ───────────────────────────────────────────────────\n\nexport interface VerdictCacheStats {\n  hits: number\n  misses: number\n}\n\nexport interface CachedJudgeOptions {\n  /** REQUIRED — part of the cache key. Bump on any change to the judge's\n   *  prompt, model, ensemble, or scoring logic; silent judge upgrades must\n   *  never serve stale verdicts. */\n  judgeVersion: string\n}\n\n/** The wrapped judge: same `JudgeConfig` seam, plus hit/miss observability. */\nexport type CachedJudge<TArtifact, TScenario extends Scenario = Scenario> = JudgeConfig<\n  TArtifact,\n  TScenario\n> & {\n  stats(): VerdictCacheStats\n}\n\n/**\n * Wrap a `JudgeConfig` so repeat judgments of the same artifact are served\n * from the store instead of re-invoking `score()`. The wrapper is generic\n * over the judge's own type parameters and preserves `appliesTo` — it is a\n * drop-in replacement anywhere a `JudgeConfig` is accepted.\n *\n * A judge that throws is NOT cached: the error propagates and the next\n * attempt re-judges (caching a failure would pin a transient outage forever).\n */\nexport function cachedJudge<TArtifact, TScenario extends Scenario = Scenario>(\n  judge: JudgeConfig<TArtifact, TScenario>,\n  store: VerdictCacheStore,\n  options: CachedJudgeOptions,\n): CachedJudge<TArtifact, TScenario> {\n  if (typeof options.judgeVersion !== 'string' || options.judgeVersion.trim() === '') {\n    throw new Error('cachedJudge: judgeVersion is required and must be a non-empty string')\n  }\n  const stats: VerdictCacheStats = { hits: 0, misses: 0 }\n  const wrapped: CachedJudge<TArtifact, TScenario> = {\n    name: judge.name,\n    dimensions: judge.dimensions,\n    judgeVersion: options.judgeVersion,\n    async score(input) {\n      const key = `${VERDICT_CACHE_KEY_SCHEME}:${contentHash({\n        artifact: canonicalJson(input.artifact),\n        scenarioId: input.scenario.id,\n        judgeName: judge.name,\n        dimensions: judge.dimensions,\n        judgeVersion: options.judgeVersion,\n      })}`\n      const cached = await store.get(key)\n      if (cached !== undefined) {\n        stats.hits += 1\n        return cached\n      }\n      const score = await judge.score(input)\n      await store.set(key, score)\n      stats.misses += 1\n      return score\n    },\n    stats: () => ({ ...stats }),\n  }\n  if (judge.appliesTo) wrapped.appliesTo = judge.appliesTo\n  return wrapped\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA0CA,SAAgB,cAAc,OAAwB;CACpD,OAAO,gBAAgB,KAAK;AAC9B;;;;AAKA,SAAgB,YAAY,OAAwB;CAClD,OAAO,cAAc,KAAK,CAAC,CAAC,MAAM,CAAgB;AACpD;AA2BA,SAAS,eAAe,MAAc,MAAc,QAAkC;CACpF,IAAI;CACJ,IAAI;EACF,SAAS,KAAK,MAAM,IAAI;CAC1B,SAAS,KAAK;EACZ,MAAM,IAAI,MACR,sCAAsC,KAAK,GAAG,OAAO,KAAK,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,GAC3G;CACF;CACA,MAAM,MAAM;CACZ,IACE,OAAO,QAAQ,YACf,QAAQ,QACR,OAAO,IAAI,QAAQ,YACnB,OAAO,IAAI,UAAU,YACrB,IAAI,UAAU,QACd,OAAO,IAAI,MAAM,cAAc,YAC/B,OAAO,IAAI,MAAM,eAAe,UAEhC,MAAM,IAAI,MACR,6CAA6C,KAAK,GAAG,OAAO,wDAC9D;CAEF,OAAO;AACT;;;;;;;;AASA,SAAgB,iBAAiB,MAAiC;CAChE,MAAM,0BAAU,IAAI,IAAwB;CAC5C,IAAI,WAAW,IAAI,GAAG;EACpB,MAAM,QAAQ,aAAa,MAAM,MAAM,CAAC,CAAC,MAAM,IAAI;EACnD,KAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;GACrC,MAAM,OAAO,MAAM;GACnB,IAAI,SAAS,KAAA,KAAa,KAAK,KAAK,MAAM,IAAI;GAC9C,MAAM,MAAM,eAAe,MAAM,MAAM,IAAI,CAAC;GAC5C,QAAQ,IAAI,IAAI,KAAK,IAAI,KAAK;EAChC;CACF;CACA,OAAO;EACL,MAAM,QAAQ,QAAQ,IAAI,GAAG;EAC7B,MAAM,KAAK,UAAU;GACnB,eAAe,MAAM,GAAG,KAAK,UAAU;IAAE;IAAK;GAAM,CAAC,EAAE,KAAK,MAAM;GAClE,QAAQ,IAAI,KAAK,KAAK;EACxB;CACF;AACF"}

package/dist/wire/index.d.ts

@@ -1,7 +1,7 @@
 import { b as CustomTokenPricing, c as CostLedgerHandle } from "../cost-ledger-DbQdN3nO.js";
 import { x as ChatClient } from "../types-CBbLtr2J.js";
 import { s as TraceStore } from "../store-BErPvYBr.js";
-import { m as FeedbackTrajectoryStore } from "../feedback-trajectory-CXmtITBo.js";
+import { m as FeedbackTrajectoryStore } from "../feedback-trajectory-CytTR7u4.js";
 import { z } from "zod";
 import { ServerType } from "@hono/node-server";
 import { Hono } from "hono";

package/dist/wire/index.js

@@ -1,2 +1,2 @@
-import { A as WIRE_VERSION, C as RUBRIC_VERSION_SCHEME, D as TracesIngestRequestSchema, E as TraceEventSchema, O as TracesIngestResponseSchema, S as ListRubricsResponseSchema, T as RubricSchema, _ as FeedbackIngestResponseSchema, a as runRpcBatch, b as JudgeRequestSchema, c as WireError, d as handleListRubrics, f as handleTracesIngest, g as ErrorResponseSchema, h as listBuiltinRubrics, i as dispatchRpc, j as hashRubric, k as VersionResponseSchema, l as handleFeedbackIngest, m as getBuiltinRubric, n as startServer, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi, t as createApp, u as handleJudge, v as FeedbackTrajectorySchema, w as RubricDimensionSchema, x as JudgeResultSchema, y as HealthResponseSchema } from "../server-BZQR1f0C.js";
+import { A as WIRE_VERSION, C as RUBRIC_VERSION_SCHEME, D as TracesIngestRequestSchema, E as TraceEventSchema, O as TracesIngestResponseSchema, S as ListRubricsResponseSchema, T as RubricSchema, _ as FeedbackIngestResponseSchema, a as runRpcBatch, b as JudgeRequestSchema, c as WireError, d as handleListRubrics, f as handleTracesIngest, g as ErrorResponseSchema, h as listBuiltinRubrics, i as dispatchRpc, j as hashRubric, k as VersionResponseSchema, l as handleFeedbackIngest, m as getBuiltinRubric, n as startServer, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi, t as createApp, u as handleJudge, v as FeedbackTrajectorySchema, w as RubricDimensionSchema, x as JudgeResultSchema, y as HealthResponseSchema } from "../server-DmpcFiVK.js";
 export { ErrorResponseSchema, FeedbackIngestResponseSchema, FeedbackTrajectorySchema, HealthResponseSchema, JudgeRequestSchema, JudgeResultSchema, ListRubricsResponseSchema, RUBRIC_VERSION_SCHEME, RubricDimensionSchema, RubricSchema, TraceEventSchema, TracesIngestRequestSchema, TracesIngestResponseSchema, VersionResponseSchema, WIRE_VERSION, WireError, buildOpenApi, createApp, dispatchRpc, getBuiltinRubric, handleFeedbackIngest, handleJudge, handleListRubrics, handleTracesIngest, handleVersion, hashRubric, listBuiltinRubrics, runRpcBatch, runRpcOnce, startServer, startServerAsync };

package/docs/changes/evaluation-definition-identity.md

@@ -0,0 +1,27 @@
+# Unreleased: evaluation definition identity
+
+Static `jevJudge` and `jevAnalyst` versions now bind authored question order and
+Choice alternative order in addition to their canonical content. The request is
+not sorted or rewritten. Optional undefined SDK fields remain absent from the
+JSON document, so JSON round-trips retain the same identity.
+
+This intentionally changes static evaluator version hashes. A cached verdict from
+the previous identity scheme must not be treated as evidence for a reordered
+classifier definition. Existing caches remain stored under their previous versions;
+this change does not delete or migrate them.
+
+`asJudge` and `asAnalyst` capture callback references when constructed and detach
+metadata arrays/objects from the supplied options. Replacing `options.evaluate`,
+`options.map` or `options.record` later no longer changes an already-versioned
+adapter. Construct another adapter with a new version for an intentional change.
+
+This is not a sandbox or closure serializer. Callers must still version opaque
+renderer/mapping behavior and mutable state captured inside their own functions.
+Dynamic question builders remain supported; retain each actual request through
+the existing observation path. No model calls, thresholds, domain policies,
+transport defaults or execution machinery are added.
+
+Regression tests cover version collisions from reordering, preserved dispatch
+order, SDK-style undefined fields, metadata mutation and callback replacement
+through the real paid-call ledger. They establish code contracts, not classifier
+quality or deployed outcome improvements.

package/docs/changes/jev-evidence-review.md

@@ -0,0 +1,18 @@
+# Unreleased: opt-in evidence reviews on native evaluations
+
+`@tangle-network/agent-eval/jev` adds `prepareJevReview`, `assessJevReview` and
+`jevReviewFindings`. Callers retain ownership of questions, hypotheses, evidence, severity,
+thresholds and result policy. Review identity includes authored option order; request JSON is
+not sorted before dispatch. Missing or partial negative evidence cannot become a clean review.
+
+The optional `examples/jev-behavior-review.ts` recipe covers reward gaming, evaluator tampering,
+policy/authorization/data-boundary breaches, injection-following, contradicted completion,
+oversight bypass, matched-baseline anomalies and cross-run behavior. It composes the existing
+`jevAnalyst`; the taxonomy is not installed in the generic evaluator and is not a calibrated
+safety detector. See [behavioral reviews](../jev-behavior-review.md) for evidence and limits.
+
+No dependency, transport, scheduler, package entrypoint, default model, production action or
+spending policy changes. Existing evaluator, judge, analyst and protocol APIs remain unchanged.
+Use existing deterministic control/reward checks and independent final assessment alongside
+semantic reviews. Adversarial model quality and product/harness adoption remain tracked in
+#772/#771 and the cross-repository roadmap #768.

package/docs/changes/jev-product-adoption.md

@@ -0,0 +1,18 @@
+# 0.184.0: installed product adoption
+
+The new version publishes the merged typed evaluator and Jev APIs that were not in the
+previously published 0.183.0 artifact. JS and Python distribution versions remain in lockstep.
+
+`@tangle-network/agent-eval/jev/protocol` is a leaf export of the existing protocol implementation.
+Products can validate native requests/responses without importing analyst/campaign execution.
+Native extension metadata is preserved, not projected away; lossy non-JSON metadata is rejected
+before consumers persist an apparently valid but changed observation. Already-paid receipts remain
+recorded when answer validation fails.
+
+`examples/jev-decision-benchmark.ts` compares caller-defined classifier configurations with the
+existing matrix runner. It separates independent expected labels from model input, retains
+abstention, records complete observations and paid failures, and groups results by configuration.
+It is a recipe, not another optimizer or universal scoring policy.
+
+Release this version through the existing verify-and-publish workflow before merging consumers
+pinned to 0.184.0. A source manifest version is not evidence that the npm artifact is available.

package/docs/changes/probability-policy-audit.md

@@ -0,0 +1,81 @@
+# Unreleased: audit a frozen probability policy
+
+Added `auditProbabilityPolicy`, `ProbabilityPolicyAuditInput` and
+`ProbabilityPolicyAuditReport` to `@tangle-network/agent-eval/meta-eval`.
+Existing `auditEvaluator` behavior and report hashes are unchanged. No dependency,
+model call, scheduler or automatic promotion is added.
+
+Use this when a product turns a probability into accept, reject or escalation.
+The caller supplies both thresholds and independently labeled observations. The
+helper maps the observations, then uses the existing `auditEvaluator` statistics,
+independent-unit counting, development-exposure exclusion and authority checks.
+
+```ts
+import { readFile } from 'node:fs/promises'
+import { auditProbabilityPolicy } from '@tangle-network/agent-eval/meta-eval'
+
+const path = process.argv[2]
+if (!path) throw new Error('Supply the frozen audit JSON path')
+const report = auditProbabilityPolicy(JSON.parse(await readFile(path, 'utf8')))
+console.log(JSON.stringify(report, null, 2))
+```
+
+The input has the existing evaluator audit fields: `evaluatorDigest`, `population`,
+`samplingFrame`, `authority`, `policy` and `observations`. Add explicit `thresholds`
+with `rejectAtOrBelow` and `acceptAtOrAbove`. Every observation carries `id`,
+`independentUnitId`, `evidenceRef`, independently checked `expected` (`accept` or
+`reject`), `exposure` (`fresh` or `development`), and `acceptProbability` instead
+of `observed`.
+
+`acceptProbability` is P(the event the caller defines as acceptable), in [0,1], or
+null for missing output. For Jev, a caller can use a suitable Noul or a named Choice
+probability. Do not substitute Jev's distribution-confidence statistic or an
+arbitrarily scaled numeric grade. Other providers work without pretending to be Jev.
+
+Probabilities at or above `acceptAtOrAbove` map to accept. Those at or below
+`rejectAtOrBelow` map to reject. The open interval between them and null map to
+unknown. Thresholds must be finite, in [0,1], and strictly separated. Nothing
+chooses thresholds or inserts a default cutoff for the application.
+
+## Read the result correctly
+
+The result retains the existing admit/reject/inconclusive verdict, error bounds,
+coverage, exclusions, authority and mapped observations. Unknown judgments remain
+in the conservative error bounds, never measured successes. The helper does not
+quietly discard escalated or unavailable cases to improve apparent accuracy.
+`coverage.eligibleCases - coverage.unknownCases` is the number of decided eligible
+cases; the full denominator remains `eligibleCases`.
+
+`probabilityPolicy` retains the original classifier digest, thresholds and exact
+probabilities keyed by observation id. `evaluatorDigest` identifies the composed
+classifier-plus-threshold program. `inputDigest` binds the complete validated
+probability-audit input (sorted by observation id), including probability changes
+that do not change the selected action. `reportDigest` covers the complete report.
+
+The original evaluator digest must identify the actual frozen classifier, question
+and evidence-rendering definition. An arbitrary string or a model alias is not
+an independently verified program identity. Hashes establish identity, not truth.
+
+## Keep selection and final assessment separate
+
+Tune thresholds on development data, freeze them, then use fresh independent cases
+for the final audit. A development-exposed source excludes its fresh-looking
+variants too. Duplicate observation ids and identical declared author/auditor
+identities are refused. Distinct declared names do not prove real-world independence;
+the execution owner must enforce it and supply evidence.
+
+The output audits this fixed binary decision policy. It does not prove probability
+calibration, downstream workflow correctness, cost savings, causal repair quality
+or external-action authorization. Use existing calibration/comparison utilities
+and actual task outcomes for those claims. A small or entirely missing dataset is
+inconclusive rather than perfect evidence. Real datasets, task slices, cost/latency
+comparisons and deployed validation remain tracked in issue #771 under roadmap #768.
+
+## Checks
+
+`pnpm exec vitest run tests/probability-policy-audit.test.ts tests/evaluator-admission.test.ts`
+
+Regressions cover boundary decisions, missing data, invalid probabilities, overlapping
+thresholds, digest identity, input immutability, independent-unit denominators,
+development leakage, authority checks and unchanged underlying error bounds. These
+are executable arithmetic/contract tests, not live Jev quality measurements.

package/docs/diagnosis.md

@@ -0,0 +1,42 @@
+# Agent-failure diagnosis
+
+`@tangle-network/agent-eval/diagnosis` exports one engine: `diagnoseSpans(spans, context, options)`.
+It reads flat spans and returns a findings document that validates against the diagnosis kit's schema, `https://tangle.tools/schemas/diagnosis-findings-v1.json`.
+
+The engine is a plain function.
+It keeps no storage, files no issues, and retains nothing.
+In model mode its only network use is the caller's transport; in deterministic mode it makes no call.
+The caller owns where the spans come from and where the findings go, so one engine serves an internal mining run and a customer diagnosis on separate data paths.
+
+## Input
+
+Each span is one object in the flat shape that trace-adapters and the traces CLI emit:
+`{trace_id, span_id, parent_span_id, name, start_time, end_time, status | status_code, status_message, attributes, resource?}`.
+Times may be OTLP nanosecond strings, epoch milliseconds, or ISO-8601.
+
+`context` names the subject (`internal` or `customer`), a label in the owner's words, an optional `focus` question, and `contentIncluded`.
+Content is excluded by default: prompt, response, and tool payload attributes are dropped before anything reads the spans.
+Only an internal context may carry a `topology` (a run graph); a customer context with one is refused.
+
+## Order of operations
+
+1. The secret filter runs on every string: span attributes, names, status messages, the focus question, and the topology.
+   `DIAGNOSIS_SECRET_RULES` holds trace-archive's line-anchored credential-assignment pattern plus token shapes (`sk-`, `sk-ant-`, GitHub, Slack, AWS, Google, JWT, bearer headers, URL passwords, PEM blocks, and JSON or inline `...key=` values).
+   A match becomes `[REDACTED:<rule>]`, and the document's `coverage.redaction` counts matches by rule.
+2. The deterministic pass reports the trace contract's capability table with the trace's own reasons, execution facts with token and cost accounting, and `observed` findings.
+   Each observed finding carries a measure with its denominator and cites the spans it counted.
+3. Model mode reads the runs with the most errors through the prime protocol (`runPrimeExchange`).
+   A run larger than the inline budget is read as whole segments, most errors first; the segments not read are listed in `coverage.skipped`.
+4. Every cited span id must exist in the input and belong to one run.
+   A model row that cites an unknown or shared id is rejected with its reason in `result.rejected`.
+
+## Output
+
+`result.document` is the only part a customer report renders.
+Every finding has `confidence`: `observed` for a fact read off the spans, `inferred` for the model's reading.
+`result.questions`, `result.notes` (operator critique and topology, internal only), `result.facts`, `result.model.usage`, and `result.rejected` support the report without entering the document.
+
+## Model transport
+
+`options.model.transport` is a `PrimeBridgeTransport`: an HTTP bridge (`nodeHttpPrimeBridgeTransport`) or any local runner with the same shape.
+tangle-tools `trace-insights/mine` runs prime-agent on the Tangle router with `deepseek/deepseek-v4.1-flash`, in a throwaway home that is deleted after each call.