@tangle-network/agent-eval 0.133.3 → 0.134.1

package/dist/multishot/index.d.ts

@@ -1,4 +1,4 @@
-import { w as JudgeScore } from "../types-BokuXvOG.js";
+import { w as JudgeScore } from "../types-DiWLru6Z.js";
 import { o as MatrixResult } from "../index-DSC51roc2.js";
 import { AgentProfile } from "@tangle-network/agent-interface";
 //#region src/multishot/types.d.ts

package/dist/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "@tangle-network/agent-eval — wire protocol",
-    "version": "0.133.3",
+    "version": "0.134.1",
     "description": "HTTP and stdio RPC interface to agent-eval. The TypeScript runtime is the source of truth; this spec is the contract that cross-language clients (Python, Rust, Go) generate from.\n\nWire-protocol version: 1.0.0. Bumps on breaking changes to request/response schemas.",
     "contact": {
       "name": "Tangle Network",

package/dist/proposal-findings-DCawte-y.js

@@ -0,0 +1,164 @@
+import { z } from "zod";
+import { createHash } from "node:crypto";
+//#region src/analyst/types.ts
+/**
+* Analyst contract — the missing orchestration layer over agent-eval's
+* existing analyzers (analyzeTraces, MultiLayerVerifier, RunCritic,
+* 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`.
+*/
+/**
+* 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
+//#region src/abort-signal.ts
+/** Combine active cancellation sources without wrapping a single source. */
+function combineAbortSignals(...signals) {
+	const active = [...new Set(signals.filter((signal) => signal !== void 0))];
+	if (active.length === 0) return void 0;
+	if (active.length === 1) return active[0];
+	return AbortSignal.any(active);
+}
+//#endregion
+//#region src/run-score.ts
+const DEFAULT_RUN_SCORE_WEIGHTS = {
+	success: 4,
+	goalProgress: 2,
+	repoGroundedness: 1.5,
+	driftPenalty: -1.5,
+	toolUseQuality: 1,
+	patchQuality: 1.25,
+	testReality: 1.5,
+	finalGate: 3,
+	reviewerBlockers: -2,
+	costUsd: -.2,
+	wallSeconds: -.1
+};
+function aggregateRunScore(score, weights = {}) {
+	const w = {
+		...DEFAULT_RUN_SCORE_WEIGHTS,
+		...weights
+	};
+	return w.success * clamp01(score.success) + w.goalProgress * clamp01(score.goalProgress) + w.repoGroundedness * clamp01(score.repoGroundedness) + w.driftPenalty * clamp01(score.driftPenalty) + w.toolUseQuality * clamp01(score.toolUseQuality) + w.patchQuality * clamp01(score.patchQuality) + w.testReality * clamp01(score.testReality) + w.finalGate * clamp01(score.finalGate) + w.reviewerBlockers * clamp01(score.reviewerBlockers) + w.costUsd * Math.max(0, finiteOrZero(score.costUsd)) + w.wallSeconds * Math.max(0, finiteOrZero(score.wallSeconds) / 60);
+}
+function clamp01(value) {
+	if (!Number.isFinite(value)) return 0;
+	return Math.max(0, Math.min(1, value));
+}
+function finiteOrZero(value) {
+	return Number.isFinite(value) ? value : 0;
+}
+//#endregion
+//#region src/analyst/proposal-findings.ts
+const ProposalFindingSchema = z.object({
+	schema_version: z.literal("1.0.0"),
+	finding_id: z.string().min(1),
+	analyst_id: z.string().min(1),
+	produced_at: z.string().min(1),
+	severity: z.enum([
+		"critical",
+		"high",
+		"medium",
+		"low",
+		"info"
+	]),
+	area: z.string().min(1),
+	claim: z.string().min(1),
+	rationale: z.string().optional(),
+	evidence_refs: z.array(z.object({
+		kind: z.enum([
+			"span",
+			"event",
+			"artifact",
+			"finding",
+			"metric"
+		]),
+		uri: z.string().min(1),
+		excerpt: z.string().optional()
+	}).strict()),
+	recommended_action: z.string().optional(),
+	validation_plan: z.string().optional(),
+	confidence: z.number().min(0).max(1),
+	subject: z.string().optional(),
+	derived_from_judge: z.boolean().optional(),
+	metadata: z.record(z.string(), z.unknown()).optional(),
+	proposal_origin: z.enum(["search", "production"])
+}).strict();
+/** True when a finding names a source candidate generation may learn from. */
+function isProposalFinding(finding) {
+	return ProposalFindingSchema.safeParse(finding).success;
+}
+/**
+* Reject findings whose source has not been explicitly admitted for candidate
+* generation. Search feedback and observed production behavior are allowed;
+* final evaluation data has no allowed origin.
+*/
+function assertProposalFindings(findings, context = "proposal findings") {
+	if (!Array.isArray(findings)) throw new TypeError(`${context}: expected an array`);
+	const rejected = findings.flatMap((finding, index) => isProposalFinding(finding) ? [] : [findingLabel(finding, index)]);
+	if (rejected.length > 0) throw new Error(`${context}: every finding must match AnalystFinding and declare proposal_origin as search or production. Rejected findings: [${rejected.join(", ")}].`);
+	return findings;
+}
+function findingLabel(finding, index) {
+	if (typeof finding !== "object" || finding === null) return `index ${index}`;
+	const id = finding.finding_id;
+	return typeof id === "string" && id.length > 0 ? id : `index ${index}`;
+}
+//#endregion
+export { clamp01 as a, makeFinding as c, aggregateRunScore as i, makeProposalFinding as l, isProposalFinding as n, combineAbortSignals as o, DEFAULT_RUN_SCORE_WEIGHTS as r, computeFindingId as s, assertProposalFindings as t };
+
+//# sourceMappingURL=proposal-findings-DCawte-y.js.map

package/dist/proposal-findings-DCawte-y.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"proposal-findings-DCawte-y.js","names":[],"sources":["../src/analyst/types.ts","../src/abort-signal.ts","../src/run-score.ts","../src/analyst/proposal-findings.ts"],"sourcesContent":["/**\n * Analyst contract — the missing orchestration layer over agent-eval's\n * existing analyzers (analyzeTraces, MultiLayerVerifier, RunCritic,\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\nexport type AnalystSeverity = 'critical' | 'high' | 'medium' | 'low' | 'info'\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\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\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 (an Ax actor 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","/** Combine active cancellation sources without wrapping a single source. */\nexport function combineAbortSignals(\n  ...signals: Array<AbortSignal | undefined>\n): AbortSignal | undefined {\n  const active = [\n    ...new Set(signals.filter((signal): signal is AbortSignal => signal !== undefined)),\n  ]\n  if (active.length === 0) return undefined\n  if (active.length === 1) return active[0]\n  return AbortSignal.any(active)\n}\n","export interface RunScore {\n  success: number\n  goalProgress: number\n  repoGroundedness: number\n  driftPenalty: number\n  toolUseQuality: number\n  patchQuality: number\n  testReality: number\n  finalGate: number\n  reviewerBlockers: number\n  costUsd: number\n  wallSeconds: number\n  notes?: string[]\n}\n\nexport interface RunScoreWeights {\n  success: number\n  goalProgress: number\n  repoGroundedness: number\n  driftPenalty: number\n  toolUseQuality: number\n  patchQuality: number\n  testReality: number\n  finalGate: number\n  reviewerBlockers: number\n  costUsd: number\n  wallSeconds: number\n}\n\nexport const DEFAULT_RUN_SCORE_WEIGHTS: RunScoreWeights = {\n  success: 4,\n  goalProgress: 2,\n  repoGroundedness: 1.5,\n  driftPenalty: -1.5,\n  toolUseQuality: 1,\n  patchQuality: 1.25,\n  testReality: 1.5,\n  finalGate: 3,\n  reviewerBlockers: -2,\n  costUsd: -0.2,\n  wallSeconds: -0.1,\n}\n\nexport function aggregateRunScore(score: RunScore, weights: Partial<RunScoreWeights> = {}): number {\n  const w = { ...DEFAULT_RUN_SCORE_WEIGHTS, ...weights }\n  return (\n    w.success * clamp01(score.success) +\n    w.goalProgress * clamp01(score.goalProgress) +\n    w.repoGroundedness * clamp01(score.repoGroundedness) +\n    w.driftPenalty * clamp01(score.driftPenalty) +\n    w.toolUseQuality * clamp01(score.toolUseQuality) +\n    w.patchQuality * clamp01(score.patchQuality) +\n    w.testReality * clamp01(score.testReality) +\n    w.finalGate * clamp01(score.finalGate) +\n    w.reviewerBlockers * clamp01(score.reviewerBlockers) +\n    w.costUsd * Math.max(0, finiteOrZero(score.costUsd)) +\n    w.wallSeconds * Math.max(0, finiteOrZero(score.wallSeconds) / 60)\n  )\n}\n\nexport function clamp01(value: number): number {\n  if (!Number.isFinite(value)) return 0\n  return Math.max(0, Math.min(1, value))\n}\n\nfunction finiteOrZero(value: number): number {\n  return Number.isFinite(value) ? value : 0\n}\n","import { z } from 'zod'\nimport type { ProposalFinding } from './types'\n\nconst ProposalFindingSchema = z\n  .object({\n    schema_version: z.literal('1.0.0'),\n    finding_id: z.string().min(1),\n    analyst_id: z.string().min(1),\n    produced_at: z.string().min(1),\n    severity: z.enum(['critical', 'high', 'medium', 'low', 'info']),\n    area: z.string().min(1),\n    claim: z.string().min(1),\n    rationale: z.string().optional(),\n    evidence_refs: z.array(\n      z\n        .object({\n          kind: z.enum(['span', 'event', 'artifact', 'finding', 'metric']),\n          uri: z.string().min(1),\n          excerpt: z.string().optional(),\n        })\n        .strict(),\n    ),\n    recommended_action: z.string().optional(),\n    validation_plan: z.string().optional(),\n    confidence: z.number().min(0).max(1),\n    subject: z.string().optional(),\n    derived_from_judge: z.boolean().optional(),\n    metadata: z.record(z.string(), z.unknown()).optional(),\n    proposal_origin: z.enum(['search', 'production']),\n  })\n  .strict() satisfies z.ZodType<ProposalFinding>\n\n/** True when a finding names a source candidate generation may learn from. */\nexport function isProposalFinding(finding: unknown): finding is ProposalFinding {\n  return ProposalFindingSchema.safeParse(finding).success\n}\n\n/**\n * Reject findings whose source has not been explicitly admitted for candidate\n * generation. Search feedback and observed production behavior are allowed;\n * final evaluation data has no allowed origin.\n */\nexport function assertProposalFindings(\n  findings: unknown,\n  context = 'proposal findings',\n): ReadonlyArray<ProposalFinding> {\n  if (!Array.isArray(findings)) {\n    throw new TypeError(`${context}: expected an array`)\n  }\n  const rejected = findings.flatMap((finding, index) =>\n    isProposalFinding(finding) ? [] : [findingLabel(finding, index)],\n  )\n  if (rejected.length > 0) {\n    throw new Error(\n      `${context}: every finding must match AnalystFinding and declare ` +\n        `proposal_origin as search or production. ` +\n        `Rejected findings: [${rejected.join(', ')}].`,\n    )\n  }\n  return findings as ReadonlyArray<ProposalFinding>\n}\n\nfunction findingLabel(finding: unknown, index: number): string {\n  if (typeof finding !== 'object' || finding === null) return `index ${index}`\n  const id = (finding as { finding_id?: unknown }).finding_id\n  return typeof id === 'string' && id.length > 0 ? id : `index ${index}`\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;AAkOA,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;;;;AC9RA,SAAgB,oBACd,GAAG,SACsB;CACzB,MAAM,SAAS,CACb,GAAG,IAAI,IAAI,QAAQ,QAAQ,WAAkC,WAAW,KAAA,CAAS,CAAC,CACpF;CACA,IAAI,OAAO,WAAW,GAAG,OAAO,KAAA;CAChC,IAAI,OAAO,WAAW,GAAG,OAAO,OAAO;CACvC,OAAO,YAAY,IAAI,MAAM;AAC/B;;;ACmBA,MAAa,4BAA6C;CACxD,SAAS;CACT,cAAc;CACd,kBAAkB;CAClB,cAAc;CACd,gBAAgB;CAChB,cAAc;CACd,aAAa;CACb,WAAW;CACX,kBAAkB;CAClB,SAAS;CACT,aAAa;AACf;AAEA,SAAgB,kBAAkB,OAAiB,UAAoC,CAAC,GAAW;CACjG,MAAM,IAAI;EAAE,GAAG;EAA2B,GAAG;CAAQ;CACrD,OACE,EAAE,UAAU,QAAQ,MAAM,OAAO,IACjC,EAAE,eAAe,QAAQ,MAAM,YAAY,IAC3C,EAAE,mBAAmB,QAAQ,MAAM,gBAAgB,IACnD,EAAE,eAAe,QAAQ,MAAM,YAAY,IAC3C,EAAE,iBAAiB,QAAQ,MAAM,cAAc,IAC/C,EAAE,eAAe,QAAQ,MAAM,YAAY,IAC3C,EAAE,cAAc,QAAQ,MAAM,WAAW,IACzC,EAAE,YAAY,QAAQ,MAAM,SAAS,IACrC,EAAE,mBAAmB,QAAQ,MAAM,gBAAgB,IACnD,EAAE,UAAU,KAAK,IAAI,GAAG,aAAa,MAAM,OAAO,CAAC,IACnD,EAAE,cAAc,KAAK,IAAI,GAAG,aAAa,MAAM,WAAW,IAAI,EAAE;AAEpE;AAEA,SAAgB,QAAQ,OAAuB;CAC7C,IAAI,CAAC,OAAO,SAAS,KAAK,GAAG,OAAO;CACpC,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,KAAK,CAAC;AACvC;AAEA,SAAS,aAAa,OAAuB;CAC3C,OAAO,OAAO,SAAS,KAAK,IAAI,QAAQ;AAC1C;;;AChEA,MAAM,wBAAwB,EAC3B,OAAO;CACN,gBAAgB,EAAE,QAAQ,OAAO;CACjC,YAAY,EAAE,OAAO,CAAC,CAAC,IAAI,CAAC;CAC5B,YAAY,EAAE,OAAO,CAAC,CAAC,IAAI,CAAC;CAC5B,aAAa,EAAE,OAAO,CAAC,CAAC,IAAI,CAAC;CAC7B,UAAU,EAAE,KAAK;EAAC;EAAY;EAAQ;EAAU;EAAO;CAAM,CAAC;CAC9D,MAAM,EAAE,OAAO,CAAC,CAAC,IAAI,CAAC;CACtB,OAAO,EAAE,OAAO,CAAC,CAAC,IAAI,CAAC;CACvB,WAAW,EAAE,OAAO,CAAC,CAAC,SAAS;CAC/B,eAAe,EAAE,MACf,EACG,OAAO;EACN,MAAM,EAAE,KAAK;GAAC;GAAQ;GAAS;GAAY;GAAW;EAAQ,CAAC;EAC/D,KAAK,EAAE,OAAO,CAAC,CAAC,IAAI,CAAC;EACrB,SAAS,EAAE,OAAO,CAAC,CAAC,SAAS;CAC/B,CAAC,CAAC,CACD,OAAO,CACZ;CACA,oBAAoB,EAAE,OAAO,CAAC,CAAC,SAAS;CACxC,iBAAiB,EAAE,OAAO,CAAC,CAAC,SAAS;CACrC,YAAY,EAAE,OAAO,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC;CACnC,SAAS,EAAE,OAAO,CAAC,CAAC,SAAS;CAC7B,oBAAoB,EAAE,QAAQ,CAAC,CAAC,SAAS;CACzC,UAAU,EAAE,OAAO,EAAE,OAAO,GAAG,EAAE,QAAQ,CAAC,CAAC,CAAC,SAAS;CACrD,iBAAiB,EAAE,KAAK,CAAC,UAAU,YAAY,CAAC;AAClD,CAAC,CAAC,CACD,OAAO;;AAGV,SAAgB,kBAAkB,SAA8C;CAC9E,OAAO,sBAAsB,UAAU,OAAO,CAAC,CAAC;AAClD;;;;;;AAOA,SAAgB,uBACd,UACA,UAAU,qBACsB;CAChC,IAAI,CAAC,MAAM,QAAQ,QAAQ,GACzB,MAAM,IAAI,UAAU,GAAG,QAAQ,oBAAoB;CAErD,MAAM,WAAW,SAAS,SAAS,SAAS,UAC1C,kBAAkB,OAAO,IAAI,CAAC,IAAI,CAAC,aAAa,SAAS,KAAK,CAAC,CACjE;CACA,IAAI,SAAS,SAAS,GACpB,MAAM,IAAI,MACR,GAAG,QAAQ,qHAEc,SAAS,KAAK,IAAI,EAAE,GAC/C;CAEF,OAAO;AACT;AAEA,SAAS,aAAa,SAAkB,OAAuB;CAC7D,IAAI,OAAO,YAAY,YAAY,YAAY,MAAM,OAAO,SAAS;CACrE,MAAM,KAAM,QAAqC;CACjD,OAAO,OAAO,OAAO,YAAY,GAAG,SAAS,IAAI,KAAK,SAAS;AACjE"}

package/dist/{release-report-DKBtegGt.d.ts → release-report-CuULWKyk.d.ts}

@@ -1,7 +1,7 @@
 import { o as FailureClass } from "./schema-BtVldJ3T.js";
 import { a as RunRecord, s as RunSplitTag } from "./run-record-DcObtIGh.js";
 import { a as DatasetScenario, o as DatasetSplit, r as DatasetManifest } from "./dataset-BvtnC8Dc.js";
-import { b as GateDecision } from "./summary-report-DyOhItws.js";
+import { b as GateDecision } from "./summary-report-DGp0-_XO.js";
 //#region src/release-confidence.d.ts
 /** Severity of an actionable finding attached to a run/trace. */
 type AsiSeverity = 'info' | 'warning' | 'error' | 'critical';
@@ -241,4 +241,4 @@ interface RenderReleaseReportOptions {
 declare function renderReleaseReport(scorecard: ReleaseConfidenceScorecard, options?: RenderReleaseReportOptions): string;
 //#endregion
 export { ReleaseConfidenceStatus as _, JudgeReplayGateArgs as a, assertReleaseConfidence as b, judgeReplayGate as c, ReleaseConfidenceAxis as d, ReleaseConfidenceAxisName as f, ReleaseConfidenceScorecard as g, ReleaseConfidenceMetrics as h, BootstrapResult as i, ActionableSideInfo as l, ReleaseConfidenceIssue as m, renderReleaseReport as n, Verdict as o, ReleaseConfidenceInput as p, BootstrapOptions as r, bootstrapCi as s, RenderReleaseReportOptions as t, AsiSeverity as u, ReleaseConfidenceThresholds as v, evaluateReleaseConfidence as x, ReleaseTraceEvidence as y };
-//# sourceMappingURL=release-report-DKBtegGt.d.ts.map
+//# sourceMappingURL=release-report-CuULWKyk.d.ts.map

package/dist/{release-report-DKBtegGt.d.ts.map → release-report-CuULWKyk.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"release-report-DKBtegGt.d.ts","names":[],"sources":["../src/release-confidence.ts","../src/promotion-gate.ts","../src/release-report.ts"],"mappings":";;;;;;KAsBY;;UAGK;;EAEf;;EAEA;EACA,WAAW;;EAEX;;EAEA;;EAEA;;EAEA;EACA,WAAW;;KAGD;KACA;UAQK;EACf;EACA;EACA,QAAQ;EACR;EACA;EACA;EACA;EACA;;EAEA,eAAe;EACf,MAAM;EACN,WAAW;;UAGI;;EAEf;EACA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA,UAAU;EACV,qBAAqB;EACrB,gBAAgB;EAChB,kBAAkB;EAClB,eAAe;EACf,aAAa;;UAGE;EACf,MAAM;EACN,QAAQ;EACR;EACA;;UAGe;EACf,MAAM;EACN;EACA;EACA;;UAGe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,aAAa,OAAO;EACpB,cAAc;EACd,oBAAoB,QAAQ,OAAO;EACnC,0BAA0B;;;;;;;EAO1B;;UAGe;EACf;EACA;EACA;EACA,QAAQ;EACR;EACA,MAAM;EACN,QAAQ;EACR,SAAS;EACT,SAAS;EACT,cAAc;EACd;;iBAkBc,0BACd,OAAO,yBACN;iBA4Ha,wBAAwB,OAAO,yBAAyB;;;;;;;;;;;;;;;;;;;;;;;;;;;KCxR5D;UAEK;EACf;EACA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA,SAAS;;UAGM;;EAEf;;EAEA;;;;;;EAMA;;EAEA;;;;;;;;;iBAUc,YACd,oBACA,qBACA,UAAS,mBACR;;;;;;;;;;;;;;;;UA+Gc,oBAAoB;EACnC,iBAAiB;EACjB,kBAAkB;;EAElB,QAAQ,QAAQ,YAAY;EAC5B;EACA;;EAEA;;EAEA;;;;;iBAMoB,gBAAgB,SACpC,MAAM,oBAAoB,WACzB,QAAQ;EAAoB;EAAyB;;;;UCjMvC;EACf;EACA,gBAAgB;EAChB;EACA;EACA;;iBAGc,oBACd,WAAW,4BACX,UAAS"}
+{"version":3,"file":"release-report-CuULWKyk.d.ts","names":[],"sources":["../src/release-confidence.ts","../src/promotion-gate.ts","../src/release-report.ts"],"mappings":";;;;;;KAsBY;;UAGK;;EAEf;;EAEA;EACA,WAAW;;EAEX;;EAEA;;EAEA;;EAEA;EACA,WAAW;;KAGD;KACA;UAQK;EACf;EACA;EACA,QAAQ;EACR;EACA;EACA;EACA;EACA;;EAEA,eAAe;EACf,MAAM;EACN,WAAW;;UAGI;;EAEf;EACA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA,UAAU;EACV,qBAAqB;EACrB,gBAAgB;EAChB,kBAAkB;EAClB,eAAe;EACf,aAAa;;UAGE;EACf,MAAM;EACN,QAAQ;EACR;EACA;;UAGe;EACf,MAAM;EACN;EACA;EACA;;UAGe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,aAAa,OAAO;EACpB,cAAc;EACd,oBAAoB,QAAQ,OAAO;EACnC,0BAA0B;;;;;;;EAO1B;;UAGe;EACf;EACA;EACA;EACA,QAAQ;EACR;EACA,MAAM;EACN,QAAQ;EACR,SAAS;EACT,SAAS;EACT,cAAc;EACd;;iBAkBc,0BACd,OAAO,yBACN;iBA4Ha,wBAAwB,OAAO,yBAAyB;;;;;;;;;;;;;;;;;;;;;;;;;;;KCxR5D;UAEK;EACf;EACA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA,SAAS;;UAGM;;EAEf;;EAEA;;;;;;EAMA;;EAEA;;;;;;;;;iBAUc,YACd,oBACA,qBACA,UAAS,mBACR;;;;;;;;;;;;;;;;UA+Gc,oBAAoB;EACnC,iBAAiB;EACjB,kBAAkB;;EAElB,QAAQ,QAAQ,YAAY;EAC5B;EACA;;EAEA;;EAEA;;;;;iBAMoB,gBAAgB,SACpC,MAAM,oBAAoB,WACzB,QAAQ;EAAoB;EAAyB;;;;UCjMvC;EACf;EACA,gBAAgB;EAChB;EACA;EACA;;iBAGc,oBACd,WAAW,4BACX,UAAS"}

package/dist/reporting.d.ts

@@ -1,6 +1,6 @@
 import { A as benjaminiHochberg, _ as PairedBootstrapResult, ct as wilcoxonSignedRank, g as PairedBootstrapOptions, q as pairedBootstrap } from "./statistics-D_4Snl-5.js";
-import { _ as paretoChart, a as ParetoPoint, c as ResearchReportCandidate, d as ResearchReportOptions, f as ResearchReportRecommendation, g as gainHistogram, h as SummaryTableRow, i as ParetoFigureSpec, l as ResearchReportDecision, m as SummaryTableOptions, n as GainDistributionFigureSpec, o as RESEARCH_REPORT_HARD_PAIR_FLOOR, p as SummaryTable, r as GainDistributionOptions, s as ResearchReport, t as GainDistributionBin, u as ResearchReportMethodology, v as researchReport, y as summaryTable } from "./summary-report-DyOhItws.js";
-import { _ as ReleaseConfidenceStatus, a as JudgeReplayGateArgs, b as assertReleaseConfidence, c as judgeReplayGate, d as ReleaseConfidenceAxis, f as ReleaseConfidenceAxisName, g as ReleaseConfidenceScorecard, h as ReleaseConfidenceMetrics, i as BootstrapResult, m as ReleaseConfidenceIssue, n as renderReleaseReport, o as Verdict, p as ReleaseConfidenceInput, r as BootstrapOptions, s as bootstrapCi, t as RenderReleaseReportOptions, v as ReleaseConfidenceThresholds, x as evaluateReleaseConfidence, y as ReleaseTraceEvidence } from "./release-report-DKBtegGt.js";
+import { _ as paretoChart, a as ParetoPoint, c as ResearchReportCandidate, d as ResearchReportOptions, f as ResearchReportRecommendation, g as gainHistogram, h as SummaryTableRow, i as ParetoFigureSpec, l as ResearchReportDecision, m as SummaryTableOptions, n as GainDistributionFigureSpec, o as RESEARCH_REPORT_HARD_PAIR_FLOOR, p as SummaryTable, r as GainDistributionOptions, s as ResearchReport, t as GainDistributionBin, u as ResearchReportMethodology, v as researchReport, y as summaryTable } from "./summary-report-DGp0-_XO.js";
+import { _ as ReleaseConfidenceStatus, a as JudgeReplayGateArgs, b as assertReleaseConfidence, c as judgeReplayGate, d as ReleaseConfidenceAxis, f as ReleaseConfidenceAxisName, g as ReleaseConfidenceScorecard, h as ReleaseConfidenceMetrics, i as BootstrapResult, m as ReleaseConfidenceIssue, n as renderReleaseReport, o as Verdict, p as ReleaseConfidenceInput, r as BootstrapOptions, s as bootstrapCi, t as RenderReleaseReportOptions, v as ReleaseConfidenceThresholds, x as evaluateReleaseConfidence, y as ReleaseTraceEvidence } from "./release-report-CuULWKyk.js";
 import { a as PairedEvalueStep, c as pairedEvalueSequence, i as PairedEvalueSequence, n as InterimReleaseConfidenceInput, o as SequentialDecision, r as PairedEvalueOptions, s as evaluateInterimReleaseConfidence, t as InterimReleaseConfidence } from "./sequential-CYwq6Ff_.js";
 import { a as rubricPredictiveValidity, i as RubricRanking, n as RubricPredictiveValidityInput, r as RubricPredictiveValidityReport, t as RubricOutcomePair } from "./rubric-predictive-validity-C1dCLcvb.js";
 export { type BootstrapOptions, type BootstrapResult, type GainDistributionBin, type GainDistributionFigureSpec, type GainDistributionOptions, type InterimReleaseConfidence, type InterimReleaseConfidenceInput, type JudgeReplayGateArgs, type PairedBootstrapOptions, type PairedBootstrapResult, type PairedEvalueOptions, type PairedEvalueSequence, type PairedEvalueStep, type ParetoFigureSpec, type ParetoPoint, RESEARCH_REPORT_HARD_PAIR_FLOOR, type ReleaseConfidenceAxis, type ReleaseConfidenceAxisName, type ReleaseConfidenceInput, type ReleaseConfidenceIssue, type ReleaseConfidenceMetrics, type ReleaseConfidenceScorecard, type ReleaseConfidenceStatus, type ReleaseConfidenceThresholds, type ReleaseTraceEvidence, type RenderReleaseReportOptions, type ResearchReport, type ResearchReportCandidate, type ResearchReportDecision, type ResearchReportMethodology, type ResearchReportOptions, type ResearchReportRecommendation, type RubricOutcomePair, type RubricPredictiveValidityInput, type RubricPredictiveValidityReport, type RubricRanking, type SequentialDecision, type SummaryTable, type SummaryTableOptions, type SummaryTableRow, type Verdict, assertReleaseConfidence, benjaminiHochberg, bootstrapCi, evaluateInterimReleaseConfidence, evaluateReleaseConfidence, gainHistogram, judgeReplayGate, pairedBootstrap, pairedEvalueSequence, paretoChart, renderReleaseReport, researchReport, rubricPredictiveValidity, summaryTable, wilcoxonSignedRank };

package/dist/{researcher-BtD5U1Up.d.ts → researcher-DVtruQ9U.d.ts}

@@ -5,7 +5,7 @@ import { s as TraceStore } from "./store-CT9YIIve.js";
 import { i as TraceEmitter, t as RunCompleteHook } from "./emitter-DGQGoLyj.js";
 import { a as RunIntegrityReport, n as RunIntegrityExpectations } from "./integrity-rmVhXWA7.js";
 import { o as LlmClientOptions, u as LlmRouteRequirements } from "./llm-client-BiK4HW0u.js";
-import { b as GateDecision, d as ResearchReportOptions, s as ResearchReport } from "./summary-report-DyOhItws.js";
+import { b as GateDecision, d as ResearchReportOptions, s as ResearchReport } from "./summary-report-DGp0-_XO.js";
 //#region src/eval-campaign.d.ts
 interface CampaignVariant<V> {
   id: string;
@@ -312,4 +312,4 @@ declare class NoopResearcher implements Researcher {
 }
 //#endregion
 export { EvalCampaignResult as _, FailureMode as a, SteeringChange as c, CampaignRunContext as d, CampaignRunOutcome as f, EvalCampaignOptions as g, CampaignVariant as h, ExperimentResult as i, CampaignFactoryParams as l, CampaignScenario as m, CallbackResearcherOptions as n, NoopResearcher as o, CampaignRunner as p, ExperimentPlan as r, Researcher as s, CallbackResearcher as t, CampaignIntegrityPolicy as u, FailedRun as v, runEvalCampaign as y };
-//# sourceMappingURL=researcher-BtD5U1Up.d.ts.map
+//# sourceMappingURL=researcher-DVtruQ9U.d.ts.map

package/dist/{researcher-BtD5U1Up.d.ts.map → researcher-DVtruQ9U.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"researcher-BtD5U1Up.d.ts","names":[],"sources":["../src/eval-campaign.ts","../src/researcher.ts"],"mappings":";;;;;;;;;UAyEiB,gBAAgB;EAC/B;EACA,SAAS;;UAGM;EACf;;EAEA,OAAO;;UAGQ,mBAAmB;;EAElC;;EAEA;EACA,SAAS;EACT;EACA;EACA,cAAc;EACd;EACA,UAAU;;;;;;;EAOV,SAAS;EACT,OAAO;EACP,SAAS;;;;;;EAMT,SAAS;;UAGD;;EAER;;EAEA;;EAEA;;EAEA,gBAAgB;EAChB,YAAY;;EAEZ;;EAEA;;EAEA;;EAEA,MAAM;;EAEN,gBAAgB;;;;;;EAMhB,cAAc;;;;;;EAMd,eAAe,mBAAmB;;;KAIxB,qBAAqB,2BAA2B;KAEhD,eAAe,MAAM,KAAK,mBAAmB,OAAO,QAAQ;KAE5D;UAEK,oBAAoB;;;;;EAKnC;EACA,UAAU,gBAAgB;EAC1B,WAAW;;EAEX;;EAEA,WAAW;;EAEX;;;;;;EAMA,SAAS;;;;;;EAMT,oBAAoB;;;;;;EAMpB,eAAe,QAAQ,0BAA0B;;;;;;;EAOjD,kBAAkB,QAAQ,0BAA0B;;;;;EAKpD;;;;;EAKA,gBAAgB;;;;;;EAMhB,YAAY;;EAEZ,qBAAqB;;;;;EAKrB,QAAQ,eAAe;;;;;EAKvB;IAAW;MAAwB,KACjC;;;;;EAOF;;EAEA;;;;EAIA;;EAEA,SAAS,QAAQ;;;;;;;EAOjB,eACI,mBACA,0BAEE,QAAQ;IACN,SAAS;IACT,cAAc;QAGd,mBACA,wBACA,QAAQ,mBAAmB;;UAGpB;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;;EAEA;EACA;;EAEA,MAAM;;EAEN,kBAAkB;EAClB,YAAY;;EAEZ,SAAS;EACT;EACA;;iBAgBoB,gBAAgB,GACpC,MAAM,oBAAoB,KACzB,QAAQ;;;;UC5QM;;;EAGf;;EAEA;EACA;;;IAGE;;IAEA;;;;UAKa;EACf;;;;EAIA;;;EAGA;;EAEA;;;UAIe;EACf;EACA;EACA,SAAS;;;EAGT;;EAEA;IAAU;IAAkB;;;;UAIb;EACf,MAAM;EACN,MAAM;EACN,cAAc;;;;;;;;;;;;;;;;;;;UAoBC;EACf,gBAAgB,MAAM,cAAc,QAAQ;EAC5C,cAAc,UAAU,gBAAgB,QAAQ;EAChD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAC1E,eAAe,MAAM,iBAAiB,QAAQ;;UAG/B;EACf,iBAAiB;EACjB,eAAe;EACf,aAAa;EACb,gBAAgB;;;;;;cAOL,8BAA8B;mBACZ;EAA7B,YAA6B,WAAW;EAExC,gBAAgB,MAAM,cAAc,QAAQ;EAI5C,cAAc,UAAU,gBAAgB,QAAQ;EAIhD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAI1E,eAAe,MAAM,iBAAiB,QAAQ;;;;;;;;;cAYnC,0BAA0B;mBACpB;EAEjB,YAAY;EAIN,gBAAgB,OAAO,cAAc,QAAQ;EAI7C,cAAc,WAAW,gBAAgB,QAAQ;EAIjD,YACJ,UAAU,kBACV,WAAW,iBACV,QAAQ;EAIL,eAAe,OAAO,iBAAiB,QAAQ"}
+{"version":3,"file":"researcher-DVtruQ9U.d.ts","names":[],"sources":["../src/eval-campaign.ts","../src/researcher.ts"],"mappings":";;;;;;;;;UAyEiB,gBAAgB;EAC/B;EACA,SAAS;;UAGM;EACf;;EAEA,OAAO;;UAGQ,mBAAmB;;EAElC;;EAEA;EACA,SAAS;EACT;EACA;EACA,cAAc;EACd;EACA,UAAU;;;;;;;EAOV,SAAS;EACT,OAAO;EACP,SAAS;;;;;;EAMT,SAAS;;UAGD;;EAER;;EAEA;;EAEA;;EAEA,gBAAgB;EAChB,YAAY;;EAEZ;;EAEA;;EAEA;;EAEA,MAAM;;EAEN,gBAAgB;;;;;;EAMhB,cAAc;;;;;;EAMd,eAAe,mBAAmB;;;KAIxB,qBAAqB,2BAA2B;KAEhD,eAAe,MAAM,KAAK,mBAAmB,OAAO,QAAQ;KAE5D;UAEK,oBAAoB;;;;;EAKnC;EACA,UAAU,gBAAgB;EAC1B,WAAW;;EAEX;;EAEA,WAAW;;EAEX;;;;;;EAMA,SAAS;;;;;;EAMT,oBAAoB;;;;;;EAMpB,eAAe,QAAQ,0BAA0B;;;;;;;EAOjD,kBAAkB,QAAQ,0BAA0B;;;;;EAKpD;;;;;EAKA,gBAAgB;;;;;;EAMhB,YAAY;;EAEZ,qBAAqB;;;;;EAKrB,QAAQ,eAAe;;;;;EAKvB;IAAW;MAAwB,KACjC;;;;;EAOF;;EAEA;;;;EAIA;;EAEA,SAAS,QAAQ;;;;;;;EAOjB,eACI,mBACA,0BAEE,QAAQ;IACN,SAAS;IACT,cAAc;QAGd,mBACA,wBACA,QAAQ,mBAAmB;;UAGpB;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;;EAEA;EACA;;EAEA,MAAM;;EAEN,kBAAkB;EAClB,YAAY;;EAEZ,SAAS;EACT;EACA;;iBAgBoB,gBAAgB,GACpC,MAAM,oBAAoB,KACzB,QAAQ;;;;UC5QM;;;EAGf;;EAEA;EACA;;;IAGE;;IAEA;;;;UAKa;EACf;;;;EAIA;;;EAGA;;EAEA;;;UAIe;EACf;EACA;EACA,SAAS;;;EAGT;;EAEA;IAAU;IAAkB;;;;UAIb;EACf,MAAM;EACN,MAAM;EACN,cAAc;;;;;;;;;;;;;;;;;;;UAoBC;EACf,gBAAgB,MAAM,cAAc,QAAQ;EAC5C,cAAc,UAAU,gBAAgB,QAAQ;EAChD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAC1E,eAAe,MAAM,iBAAiB,QAAQ;;UAG/B;EACf,iBAAiB;EACjB,eAAe;EACf,aAAa;EACb,gBAAgB;;;;;;cAOL,8BAA8B;mBACZ;EAA7B,YAA6B,WAAW;EAExC,gBAAgB,MAAM,cAAc,QAAQ;EAI5C,cAAc,UAAU,gBAAgB,QAAQ;EAIhD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAI1E,eAAe,MAAM,iBAAiB,QAAQ;;;;;;;;;cAYnC,0BAA0B;mBACpB;EAEjB,YAAY;EAIN,gBAAgB,OAAO,cAAc,QAAQ;EAI7C,cAAc,WAAW,gBAAgB,QAAQ;EAIjD,YACJ,UAAU,kBACV,WAAW,iBACV,QAAQ;EAIL,eAAe,OAAO,iBAAiB,QAAQ"}

package/dist/rl.d.ts

@@ -2,12 +2,12 @@ import { s as VerificationReport } from "./multi-layer-verifier-BHY1gWAc.js";
 import { _ as Span } from "./schema-BtVldJ3T.js";
 import { a as RunRecord, s as RunSplitTag } from "./run-record-DcObtIGh.js";
 import { s as TraceStore } from "./store-CT9YIIve.js";
-import { a as CampaignResult } from "./types-BokuXvOG.js";
+import { a as CampaignResult } from "./types-DiWLru6Z.js";
 import { a as detectRewardHacking, c as VerifiableRewardSource, d as filterDeterministicallyRewarded, i as RewardHackingSignal, l as extractVerifiableReward, n as RewardHackingFinding, o as VerifiableReward, r as RewardHackingReport, s as VerifiableRewardExtractionOptions, t as DetectRewardHackingInput, u as extractVerifiableRewardsFromRecords } from "./reward-hacking-D-QqXvg-.js";
 import { i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "./outcome-store-BYHIuO0e.js";
 import { t as AdversarialMutation } from "./adversarial-smnADNFS.js";
 import { b as RolloutSplit, o as MintedRolloutLine } from "./schema-Cef2cFmb.js";
-import { _ as EvalCampaignResult, a as FailureMode, c as SteeringChange, g as EvalCampaignOptions, i as ExperimentResult, r as ExperimentPlan, s as Researcher, y as runEvalCampaign } from "./researcher-BtD5U1Up.js";
+import { _ as EvalCampaignResult, a as FailureMode, c as SteeringChange, g as EvalCampaignOptions, i as ExperimentResult, r as ExperimentPlan, s as Researcher, y as runEvalCampaign } from "./researcher-DVtruQ9U.js";
 import { t as InterimReleaseConfidence } from "./sequential-CYwq6Ff_.js";
 import { r as RubricPredictiveValidityReport } from "./rubric-predictive-validity-C1dCLcvb.js";
 //#region src/rl/active-curriculum.d.ts

package/dist/rl.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"rl.d.ts","names":[],"sources":["../src/rl/active-curriculum.ts","../src/rl/adaptation-eval.ts","../src/rl/compute-curves.ts","../src/rl/contamination.ts","../src/rl/rollout-input.ts","../src/rl/preferences.ts","../src/rl/process-reward.ts","../src/rl/exporters.ts","../src/rl/dataset.ts","../src/rl/corpus.ts","../src/rl/off-policy.ts","../src/rl/predictive-validity-researcher.ts","../src/rl/rl-campaign.ts","../src/rl/run-record-adapters.ts","../src/rl/sim-fidelity.ts","../src/rl/tournament.ts"],"mappings":";;;;;;;;;;;;;UAoCiB;EACf;EACA;;EAEA;;EAEA;;UAGe;EACf;EACA;;EAEA;;EAEA;;UAGe;;EAEf;;;;;;EAMA;;;;;EAKA;;;;;;;;;iBAUc,wBACd,cAAc,mBACd,gBAAgB;EAAQ;EAAmB;IAC3C,MAAM,4BACL;UAoDc;EACf;;;;;;EAMA;;EAEA;EACA;;EAEA;;;;;;;;;;;;;iBAcc,mBACd,cAAc,mBACd,gBAAgB;EAAQ;EAAmB;IAC3C,MAAM,4BACL;;iBAsDa,2BACd,MAAM,aACN;EAAQ;EAAwB;IAC/B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UChMc,iBAAiB;;;;;;EAMhC,IAAI;IAAQ,UAAU;IAAG;IAAW;MAAgB;;UAGrC,0BAA0B;EACzC,WAAW;;EAEX;;EAEA;EACA,QAAQ,iBAAiB;;EAEzB;;UAGe;EACf;EACA;EACA;EACA;EACA;;EAEA,aAAa;IAAQ;IAAoB;IAAmB;IAAgB;;;UAG7D;EACf,QAAQ;;;;;EAKR;;;;;;EAMA;;iBAGoB,mBAAmB;EAAY;GACnD,MAAM,0BAA0B,KAC/B,QAAQ;UA0DM;EACf,MAAM;IACJ;IACA;IACA;IACA;IACA;IACA;;EAEF;EACA;;EAEA;;EAEA;;;;;;;iBAQc,wBACd,GAAG,iBACH,GAAG,iBACH;EAAQ;EAAqB;EAA6B;IACzD;;iBA8Ca,WAAW,OAAO,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UC5KlC;;EAEf;;EAEA;;EAEA,OAAO;;UAGQ;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA,UAAU;;UAGK;EACf;EACA,QAAQ;;EAER;;EAEA,MAAM;;UAGS;EACf;EACA,SAAS;;;;;EAKT,cAAc,QAAQ,uBAAuB;IAC3C;IACA;IACA;IACA,UAAU;;;iBAIQ,gBAAgB,MAAM,yBAAyB,QAAQ;UAmB5D,sBAAsB;;EAErC;;EAEA,SAAS,sBAAsB,QAAQ;;EAEvC,UAAU,SAAS,MAAM;;UAGV,qBAAqB;EACpC,MAAM;EACN;EACA;EACA;;EAEA;;;iBAIoB,QAAQ,GAAG,MAAM,sBAAsB,KAAK,QAAQ,qBAAqB;UAqB9E,uBAAuB;EACtC;EACA,SAAS,sBAAsB,QAAQ;;EAEvC,YAAY,SAAS;;UAGN,sBAAsB;;EAErC;;EAEA;;EAEA,WAAW;;EAEX,gBAAgB;;EAEhB,UAAU;;;;;;iBAOU,gBAAgB,GACpC,MAAM,uBAAuB,KAC5B,QAAQ,sBAAsB;;;;;;;UAkChB;EACf;EACA;EACA;EACA;;iBAGc,eAAe,QAAQ,qBAAqB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KC3KhD;UAOK,qBAAqB;EACpC,MAAM;;EAEN,QAAQ,UAAU,MAAM,QAAQ,KAAK;;EAErC;;UAGe,wBAAwB;;EAEvC,aAAa,GAAG;;EAEhB,WAAW;;;;;EAKX,YAAY;EACZ,eAAe,qBAAqB;;;;;;EAMpC,UAAU,GAAG,MAAM;;UAGJ;;EAEf;;;;;EAKA;;;;;EAKA;;UAGe;EACf,aAAa;IACX;IACA;IACA;IACA;;IAEA;;;EAGF;IAAc;IAAW;;EACzB;EACA;EACA;EACA;;EAEA;;iBAGoB,sBAAsB,GAC1C,OAAO,wBAAwB,IAC/B,OAAM,4BACL,QAAQ;;;;;;;iBA0FK,gBAAgB;EAAY;GAC1C,uBACA,UAAS,cAAc,yBACtB,qBAAqB;;;;;;iBAoBR,aAAa;EAAY;GACvC,iBAAiB,gBAAgB,8BACjC,eACC,qBAAqB;;;;;;iBAuBR,uBAAuB;EAAY;GACjD,gBACA,iCACC,qBAAqB;;;;;;;;;;;;;UC9MP;;;;;;;;;;;EAWf,OAAO;;;UAIQ;;EAEf;;EAEA;;EAEA;;;;KCLU;UAKK;;EAEf;;EAEA;EACA;;EAEA;EACA;;EAEA;;;;;;EAMA;IAAW;IAAgB;;;EAE3B;;;;;EAKA;IACE;IACA;IACA;IACA;IACA;IACA;;;UAIa;EACf,WAAW;;;;;EAKX;;;;;;EAMA,QAAQ;;EAER;;UAGe;EACf,OAAO;;EAEP;;EAEA;;EAEA;;EAEA,UAAU;;;;;;EAMV;;;;;;;;;;;;;;;;;;;;iBAuCc,mBACd,OAAO,qBACP,OAAM,4BACL;;;;;;;;;;;;;;;;;;iBAqMmB,YACpB,SAAS,oBACT,SAAS,YACT,SAAS,qBACR,QAAQ;EAAQ;EAAgB;EAAgB;;;;;;;;;;;iBAkCnC,kBACd,SAAS,oBACT,SAAS,qBACR;EAAQ;EAAoB;EAAqB;EAAuB;;;;UChX1D;;EAEf;EACA;;EAEA;;EAEA,MAAM;;EAEN;;EAEA;;;;;EAKA;;EAEA;;EAEA;;UAGe;;EAEf,WAAW;;EAEX,MAAM,MAAM,OAAO,QAAQ,KAAK;;UAGjB;;;;;;EAMf,SAAS;;EAET,aAAa,MAAM;;iBAGC,mBACpB,OAAO,YACP,eACA,MAAM,4BACL,QAAQ;UAgCM;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;EACA;;iBAGc,yBAAyB,aAAa,eAAe;UA+CpD;;EAEf;EACA;;EAEA;EACA;;EAEA;EACA;;EAEA;EACA;;;;;;;;;;;;;;;;;iBAkBc,iBACd,kBAAkB,YAAY,eAC9B;EAAQ;EAAoB;IAC3B;;;UC3Jc;;EAEf,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;UAG3B;EACf;EACA;EACA;;EAEA;;EAEA,OAAO;;;KAIG,iBAAiB;cAEhB,yBAAyB;;;;;;;;;;;;;;;;iBAsBhB,UACpB,SAAS,oBACT,SAAS,YACT,SAAS,iBACR,QAAQ;;iBAwCK,WAAW,MAAM;UAMhB;;EAEf;;EAEA;;;;;EAKA,cAAc;;UAGC,oBACP,KAAK;;EAEb,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;UAG3B;EACf;EACA;EACA;;EAEA;EACA,OAAO;;;;;;;;;;;;;;;;;;;iBAoBa,WACpB,OAAO,qBACP,SAAS,cACR,QAAQ;iBAyDK,YAAY,MAAM;UAMjB,mBAAmB;;EAElC,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;EAE1C,YAAY,MAAM;;EAElB,WAAW,MAAM;;EAEjB;;UAGe;EACf,UAAU;IAAQ;IAAuC;;EACzD,OAAO;;;;;;;;;;;;;;;;;;;;iBAqBa,UACpB,OAAO,qBACP,SAAS,aACR,QAAQ;iBAkDK,WAAW,MAAM;UAMhB;;EAEf,WAAW,2BAA2B;;EAEtC,aAAa,eAAe,4BAA4B;;EAExD,YAAY,eAAe,uCAAuC;;UAGnD;EACf;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,OAAO;;UAGQ,uBAAuB;;;;;;;;;;EAUtC;;;;;;;;;;;;;;;;;;;;;iBAsBoB,UACpB,SAAS,qBACT,SAAS,YACT,SAAS,iBACR,QAAQ;;;;;;;;;iBA2CK,uBAAuB,MAAM,mBAAmB;cAwBnD,yBAAyB;iBA2BtB,WAAW,MAAM;UAMhB;EACf;EACA;EACA;EACA;EACA;EACA;;cAGW,iCAAiC;;;;;;;;;;;iBAiB9B,mBAAmB,aAAa,cAAc,SAAS;;;KChf3D;cACN;KACM,wBAAwB;iBAIpB,uBAAuB,iBAAiB;;;UA2BvC;EACf;EACA;;EAEA;;;EAGA;;;EAGA;IAAU,MAAM;IAAY;IAAgB;;EAC5C;EACA;EACA;;EAEA;;EAEA,UAAU;;EAEV;IACE;IACA;IACA;;;UAIa;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;;EAEA;;EAEA,QAAQ,OAAO;EACf,QAAQ;;EAER;;EAEA;EACA;EACA;IAAe;IAAe;;EAC9B;;;;;;EAMA;;UAGe,0BAA0B;EACzC,SAAS;EACT,WAAW,QAAQ,OAAO;EAC1B,OAAO;;UAGQ;EACf,UAAU;;EAEV,OAAO;;;;;;;;;iBAyDa,eACpB,OAAO,qBACP,SAAS,cAAc,YACvB,QAAQ,iBACR;EAAgB,SAAS;EAAoB,SAAS;IACrD,QAAQ;;iBA2DK,oBAAoB,GAAG;;;;;KCzO3B,eAAe;EAAc;EAAiB;;UAEzC;EACf;;;EAGA;EACA;;;;;;;;iBASc,eAAe,SAAS,gBAAgB,qBAAqB;;iBAqB7D,WAAW,qBAAqB;UAmB/B;;EAEf;;EAEA,SAAS;;EAET;;;;;;;;;;;;;;;iBAgBoB,uBACpB,oBACA,QAAQ,iBACR,OAAM,iBACL,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UCnEM;;EAEf;;EAEA;;;;;;EAMA;;;;;;EAMA;;;;;;EAMA;;;;;;;EAOA;;UAGe;;EAEf;;EAEA;;UAGe;;EAEf;;EAEA;;EAEA;;EAEA;;;;;EAKA;;EAEA,qBAAqB;;UAGN;;;;;;EAMf;;EAEA;IAAe;IAAa;;;;;;;;iBAQd,4BACd,cAAc,uBACd,OAAM,mBACL;;;;;;iBA4Ca,kCACd,cAAc,uBACd,OAAM,mBACL;;;;;;;;;;;;;;;;;;;;;iBAwDa,aACd,cAAc,uBACd,OAAM,mBACL;;;;;;iBAuEa,qBACd,cAAc,uBACd,OAAM;EACH,KAAK;EAAmB,OAAO;EAAmB,IAAI;;;;UCtQ1C;EACf,UAAU;EACV;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;EAMA,YAAY,QAAQ,0CAA0C;;;;;;cAOnD,wCAAwC;UAC3C;UACA;EAER,YAAY,MAAM;EAIZ,gBAAgB,MAAM,cAAc,QAAQ;EA0C5C,cAAc,UAAU,gBAAgB,QAAQ;EAmDhD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAS1E,eAAe,MAAM,iBAAiB,QAAQ;;;;;;EAwC9C,iBAAiB,MAAM,cAAc,QAAQ;;;;;;EAiBnD,UAAU,QAAQ;EAIlB,iBAAiB;;;;UC5KF,qBAAqB,WAAW,oBAAoB;;EAEnE,cAAc;;EAEd,mBAAmB;;EAEnB,eAAe;EACf;;EAEA;IAAe;IAAgB;IAAgB;MAAS;MAAa;;;;EAErE;IACE,MAAM;IACN,OAAO;IACP,MAAM;;;UAIO;EACf,UAAU;;EAEV,eAAe;IAAQ;IAAe,QAAQ;;;EAE9C,aAAa;;EAEb,mBAAmB;;EAEnB,eAAe;;EAEf,oBAAoB;;EAEpB;IACE,MAAM;IACN,OAAO;IACP,MAAM;;;;;EAKR;;;;EAIA;;iBAGoB,cAAc,GAAG,MAAM,qBAAqB,KAAK,QAAQ;;;UCrF9D;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,WAAW;;EAEX;;;;;;;;;;iBAWc,qBACd,UAAU,gBACV,KAAK;EAAmB;IACvB;;;;;;;;iBAyBa,8BACd,QAAQ,oBACR,KAAK;EAAmB;EAAqB;GAC7C;EAAQ;IACP;;;;;;;KCnCS,oBAAoB,QAAQ,cAAc;;;;;cAMzC;;;cAIA;;;cAIA;;;;;cAMA;;;;;;;;;;;;;;;;;;;;;;cAyBA,yBAAyB;;;;;;;;;iBAmDtB,aAAa,GAAG,wBAAwB,GAAG;;;;;;;iBAqC3C,cAAc,kBAAkB;;;iBAsBhC,YAAY,eAAe;UAU1B;;EAEf;;;EAGA;;EAEA;;UAGe;EACf;;EAEA;;EAEA,WAAW;;EAEX;;EAEA;;KAGU;UAEK;EACf,cAAc;;;;EAId;;;EAGA;;;EAGA,SAAS;;UAGM;;EAEf,WAAW;;EAEX;;;;;;;;iBASc,kBACd,WAAW,aACX,YAAY,aACZ,OAAM,qBACL;UAuHc;;;EAGf;;;EAGA;;UAGe;EACf;EACA;;EAEA;;;EAGA;;;;;;;;;iBAUc,cACd,WAAW,aACX,YAAY,aACZ,OAAM,kBACL;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UC/Xc;;EAEf;;EAEA;;;;;EAKA;;;;;;EAMA;;UAGe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;UAGe;EACf,SAAS;;EAET;;EAEA;EACA;;;;;;;;;;;iBAYc,gBACd,UAAU,mBACV;EAAQ;EAAoB;EAAwB;IACnD;;;;;;;;;UA+Gc;;EAEf;;EAEA;;iBAGc,eACd,SAAS,qBACT,SAAS,iBACT,OAAM;EACH;EAAqB;;;;;;;;UA2BT;EACf,MAAM;IACJ;;IAEA;IACA;;;;;;EAMF;;iBAGc,0BACd,OAAO,iCACN"}
+{"version":3,"file":"rl.d.ts","names":[],"sources":["../src/rl/active-curriculum.ts","../src/rl/adaptation-eval.ts","../src/rl/compute-curves.ts","../src/rl/contamination.ts","../src/rl/rollout-input.ts","../src/rl/preferences.ts","../src/rl/process-reward.ts","../src/rl/exporters.ts","../src/rl/dataset.ts","../src/rl/corpus.ts","../src/rl/off-policy.ts","../src/rl/predictive-validity-researcher.ts","../src/rl/rl-campaign.ts","../src/rl/run-record-adapters.ts","../src/rl/sim-fidelity.ts","../src/rl/tournament.ts"],"mappings":";;;;;;;;;;;;;UAoCiB;EACf;EACA;;EAEA;;EAEA;;UAGe;EACf;EACA;;EAEA;;EAEA;;UAGe;;EAEf;;;;;;EAMA;;;;;EAKA;;;;;;;;;iBAUc,wBACd,cAAc,mBACd,gBAAgB;EAAQ;EAAmB;IAC3C,MAAM,4BACL;UAoDc;EACf;;;;;;EAMA;;EAEA;EACA;;EAEA;;;;;;;;;;;;;iBAcc,mBACd,cAAc,mBACd,gBAAgB;EAAQ;EAAmB;IAC3C,MAAM,4BACL;;iBAsDa,2BACd,MAAM,aACN;EAAQ;EAAwB;IAC/B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UChMc,iBAAiB;;;;;;EAMhC,IAAI;IAAQ,UAAU;IAAG;IAAW;MAAgB;;UAGrC,0BAA0B;EACzC,WAAW;;EAEX;;EAEA;EACA,QAAQ,iBAAiB;;EAEzB;;UAGe;EACf;EACA;EACA;EACA;EACA;;EAEA,aAAa;IAAQ;IAAoB;IAAmB;IAAgB;;;UAG7D;EACf,QAAQ;;;;;EAKR;;;;;;EAMA;;iBAGoB,mBAAmB;EAAY;GACnD,MAAM,0BAA0B,KAC/B,QAAQ;UA0DM;EACf,MAAM;IACJ;IACA;IACA;IACA;IACA;IACA;;EAEF;EACA;;EAEA;;EAEA;;;;;;;iBAQc,wBACd,GAAG,iBACH,GAAG,iBACH;EAAQ;EAAqB;EAA6B;IACzD;;iBA8Ca,WAAW,OAAO,iBAAiB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UC5KlC;;EAEf;;EAEA;;EAEA,OAAO;;UAGQ;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA,UAAU;;UAGK;EACf;EACA,QAAQ;;EAER;;EAEA,MAAM;;UAGS;EACf;EACA,SAAS;;;;;EAKT,cAAc,QAAQ,uBAAuB;IAC3C;IACA;IACA;IACA,UAAU;;;iBAIQ,gBAAgB,MAAM,yBAAyB,QAAQ;UAmB5D,sBAAsB;;EAErC;;EAEA,SAAS,sBAAsB,QAAQ;;EAEvC,UAAU,SAAS,MAAM;;UAGV,qBAAqB;EACpC,MAAM;EACN;EACA;EACA;;EAEA;;;iBAIoB,QAAQ,GAAG,MAAM,sBAAsB,KAAK,QAAQ,qBAAqB;UAqB9E,uBAAuB;EACtC;EACA,SAAS,sBAAsB,QAAQ;;EAEvC,YAAY,SAAS;;UAGN,sBAAsB;;EAErC;;EAEA;;EAEA,WAAW;;EAEX,gBAAgB;;EAEhB,UAAU;;;;;;iBAOU,gBAAgB,GACpC,MAAM,uBAAuB,KAC5B,QAAQ,sBAAsB;;;;;;;UAkChB;EACf;EACA;EACA;EACA;;iBAGc,eAAe,QAAQ,qBAAqB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KC3KhD;UAOK,qBAAqB;EACpC,MAAM;;EAEN,QAAQ,UAAU,MAAM,QAAQ,KAAK;;EAErC;;UAGe,wBAAwB;;EAEvC,aAAa,GAAG;;EAEhB,WAAW;;;;;EAKX,YAAY;EACZ,eAAe,qBAAqB;;;;;;EAMpC,UAAU,GAAG,MAAM;;UAGJ;;EAEf;;;;;EAKA;;;;;EAKA;;UAGe;EACf,aAAa;IACX;IACA;IACA;IACA;;IAEA;;;EAGF;IAAc;IAAW;;EACzB;EACA;EACA;EACA;;EAEA;;iBAGoB,sBAAsB,GAC1C,OAAO,wBAAwB,IAC/B,OAAM,4BACL,QAAQ;;;;;;;iBA0FK,gBAAgB;EAAY;GAC1C,uBACA,UAAS,cAAc,yBACtB,qBAAqB;;;;;;iBAoBR,aAAa;EAAY;GACvC,iBAAiB,gBAAgB,8BACjC,eACC,qBAAqB;;;;;;iBAuBR,uBAAuB;EAAY;GACjD,gBACA,iCACC,qBAAqB;;;;;;;;;;;;;UC9MP;;;;;;;;;;;EAWf,OAAO;;;UAIQ;;EAEf;;EAEA;;EAEA;;;;KCLU;UAKK;;EAEf;;EAEA;EACA;;EAEA;EACA;;EAEA;;;;;;EAMA;IAAW;IAAgB;;;EAE3B;;;;;EAKA;IACE;IACA;IACA;IACA;IACA;IACA;;;UAIa;EACf,WAAW;;;;;EAKX;;;;;;EAMA,QAAQ;;EAER;;UAGe;EACf,OAAO;;EAEP;;EAEA;;EAEA;;EAEA,UAAU;;;;;;EAMV;;;;;;;;;;;;;;;;;;;;iBAuCc,mBACd,OAAO,qBACP,OAAM,4BACL;;;;;;;;;;;;;;;;;;iBAqMmB,YACpB,SAAS,oBACT,SAAS,YACT,SAAS,qBACR,QAAQ;EAAQ;EAAgB;EAAgB;;;;;;;;;;;iBAkCnC,kBACd,SAAS,oBACT,SAAS,qBACR;EAAQ;EAAoB;EAAqB;EAAuB;;;;UChX1D;;EAEf;EACA;;EAEA;;EAEA,MAAM;;EAEN;;EAEA;;;;;EAKA;;EAEA;;EAEA;;UAGe;;EAEf,WAAW;;EAEX,MAAM,MAAM,OAAO,QAAQ,KAAK;;UAGjB;;;;;;EAMf,SAAS;;EAET,aAAa,MAAM;;iBAGC,mBACpB,OAAO,YACP,eACA,MAAM,4BACL,QAAQ;UAgCM;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;EACA;;iBAGc,yBAAyB,aAAa,eAAe;UA+CpD;;EAEf;EACA;;EAEA;EACA;;EAEA;EACA;;EAEA;EACA;;;;;;;;;;;;;;;;;iBAkBc,iBACd,kBAAkB,YAAY,eAC9B;EAAQ;EAAoB;IAC3B;;;UC3Jc;;EAEf,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;UAG3B;EACf;EACA;EACA;;EAEA;;EAEA,OAAO;;;KAIG,iBAAiB;cAEhB,yBAAyB;;;;;;;;;;;;;;;;iBAsBhB,UACpB,SAAS,oBACT,SAAS,YACT,SAAS,iBACR,QAAQ;;iBAwCK,WAAW,MAAM;UAMhB;;EAEf;;EAEA;;;;;EAKA,cAAc;;UAGC,oBACP,KAAK;;EAEb,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;UAG3B;EACf;EACA;EACA;;EAEA;EACA,OAAO;;;;;;;;;;;;;;;;;;;iBAoBa,WACpB,OAAO,qBACP,SAAS,cACR,QAAQ;iBAyDK,YAAY,MAAM;UAMjB,mBAAmB;;EAElC,WAAW,2BAA2B;;EAEtC,eAAe,2BAA2B;;EAE1C,YAAY,MAAM;;EAElB,WAAW,MAAM;;EAEjB;;UAGe;EACf,UAAU;IAAQ;IAAuC;;EACzD,OAAO;;;;;;;;;;;;;;;;;;;;iBAqBa,UACpB,OAAO,qBACP,SAAS,aACR,QAAQ;iBAkDK,WAAW,MAAM;UAMhB;;EAEf,WAAW,2BAA2B;;EAEtC,aAAa,eAAe,4BAA4B;;EAExD,YAAY,eAAe,uCAAuC;;UAGnD;EACf;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,OAAO;;UAGQ,uBAAuB;;;;;;;;;;EAUtC;;;;;;;;;;;;;;;;;;;;;iBAsBoB,UACpB,SAAS,qBACT,SAAS,YACT,SAAS,iBACR,QAAQ;;;;;;;;;iBA2CK,uBAAuB,MAAM,mBAAmB;cAwBnD,yBAAyB;iBA2BtB,WAAW,MAAM;UAMhB;EACf;EACA;EACA;EACA;EACA;EACA;;cAGW,iCAAiC;;;;;;;;;;;iBAiB9B,mBAAmB,aAAa,cAAc,SAAS;;;KChf3D;cACN;KACM,wBAAwB;iBAIpB,uBAAuB,iBAAiB;;;UA2BvC;EACf;EACA;;EAEA;;;EAGA;;;EAGA;IAAU,MAAM;IAAY;IAAgB;;EAC5C;EACA;EACA;;EAEA;;EAEA,UAAU;;EAEV;IACE;IACA;IACA;;;UAIa;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;;EAEA;;EAEA,QAAQ,OAAO;EACf,QAAQ;;EAER;;EAEA;EACA;EACA;IAAe;IAAe;;EAC9B;;;;;;EAMA;;UAGe,0BAA0B;EACzC,SAAS;EACT,WAAW,QAAQ,OAAO;EAC1B,OAAO;;UAGQ;EACf,UAAU;;EAEV,OAAO;;;;;;;;;iBAyDa,eACpB,OAAO,qBACP,SAAS,cAAc,YACvB,QAAQ,iBACR;EAAgB,SAAS;EAAoB,SAAS;IACrD,QAAQ;;iBA2DK,oBAAoB,GAAG;;;;;KCzO3B,eAAe;EAAc;EAAiB;;UAEzC;EACf;;;EAGA;EACA;;;;;;;;iBASc,eAAe,SAAS,gBAAgB,qBAAqB;;iBAqB7D,WAAW,qBAAqB;UAmB/B;;EAEf;;EAEA,SAAS;;EAET;;;;;;;;;;;;;;;iBAgBoB,uBACpB,oBACA,QAAQ,iBACR,OAAM,iBACL,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UCnEM;;EAEf;;EAEA;;;;;;EAMA;;;;;;EAMA;;;;;;EAMA;;;;;;;EAOA;;UAGe;;EAEf;;EAEA;;UAGe;;EAEf;;EAEA;;EAEA;;EAEA;;;;;EAKA;;EAEA,qBAAqB;;UAGN;;;;;;EAMf;;EAEA;IAAe;IAAa;;;;;;;;iBAQd,4BACd,cAAc,uBACd,OAAM,mBACL;;;;;;iBA4Ca,kCACd,cAAc,uBACd,OAAM,mBACL;;;;;;;;;;;;;;;;;;;;;iBAwDa,aACd,cAAc,uBACd,OAAM,mBACL;;;;;;iBAuEa,qBACd,cAAc,uBACd,OAAM;EACH,KAAK;EAAmB,OAAO;EAAmB,IAAI;;;;UCtQ1C;EACf,UAAU;EACV;;EAEA;;EAEA;;EAEA;;EAEA;;;;;;EAMA,YAAY,QAAQ,0CAA0C;;;;;;cAOnD,wCAAwC;UAC3C;UACA;EAER,YAAY,MAAM;EAIZ,gBAAgB,MAAM,cAAc,QAAQ;EA0C5C,cAAc,UAAU,gBAAgB,QAAQ;EAmDhD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAS1E,eAAe,MAAM,iBAAiB,QAAQ;;;;;;EA4C9C,iBAAiB,MAAM,cAAc,QAAQ;;;;;;EAiBnD,UAAU,QAAQ;EAIlB,iBAAiB;;;;UChLF,qBAAqB,WAAW,oBAAoB;;EAEnE,cAAc;;EAEd,mBAAmB;;EAEnB,eAAe;EACf;;EAEA;IAAe;IAAgB;IAAgB;MAAS;MAAa;;;;EAErE;IACE,MAAM;IACN,OAAO;IACP,MAAM;;;UAIO;EACf,UAAU;;EAEV,eAAe;IAAQ;IAAe,QAAQ;;;EAE9C,aAAa;;EAEb,mBAAmB;;EAEnB,eAAe;;EAEf,oBAAoB;;EAEpB;IACE,MAAM;IACN,OAAO;IACP,MAAM;;;;;EAKR;;;;EAIA;;iBAGoB,cAAc,GAAG,MAAM,qBAAqB,KAAK,QAAQ;;;UCrF9D;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,WAAW;;EAEX;;;;;;;;;;iBAWc,qBACd,UAAU,gBACV,KAAK;EAAmB;IACvB;;;;;;;;iBAyBa,8BACd,QAAQ,oBACR,KAAK;EAAmB;EAAqB;GAC7C;EAAQ;IACP;;;;;;;KCnCS,oBAAoB,QAAQ,cAAc;;;;;cAMzC;;;cAIA;;;cAIA;;;;;cAMA;;;;;;;;;;;;;;;;;;;;;;cAyBA,yBAAyB;;;;;;;;;iBAmDtB,aAAa,GAAG,wBAAwB,GAAG;;;;;;;iBAqC3C,cAAc,kBAAkB;;;iBAsBhC,YAAY,eAAe;UAU1B;;EAEf;;;EAGA;;EAEA;;UAGe;EACf;;EAEA;;EAEA,WAAW;;EAEX;;EAEA;;KAGU;UAEK;EACf,cAAc;;;;EAId;;;EAGA;;;EAGA,SAAS;;UAGM;;EAEf,WAAW;;EAEX;;;;;;;;iBASc,kBACd,WAAW,aACX,YAAY,aACZ,OAAM,qBACL;UAuHc;;;EAGf;;;EAGA;;UAGe;EACf;EACA;;EAEA;;;EAGA;;;;;;;;;iBAUc,cACd,WAAW,aACX,YAAY,aACZ,OAAM,kBACL;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UC/Xc;;EAEf;;EAEA;;;;;EAKA;;;;;;EAMA;;UAGe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;UAGe;EACf,SAAS;;EAET;;EAEA;EACA;;;;;;;;;;;iBAYc,gBACd,UAAU,mBACV;EAAQ;EAAoB;EAAwB;IACnD;;;;;;;;;UA+Gc;;EAEf;;EAEA;;iBAGc,eACd,SAAS,qBACT,SAAS,iBACT,OAAM;EACH;EAAqB;;;;;;;;UA2BT;EACf,MAAM;IACJ;;IAEA;IACA;;;;;;EAMF;;iBAGc,0BACd,OAAO,iCACN"}

package/dist/rl.js

@@ -1566,7 +1566,9 @@ var PredictiveValidityResearcher = class {
 					baselineOverfitGap: null,
 					medianCandidateCost: null,
 					medianBaselineCost: null,
-					realnessGatedRuns: 0
+					realnessGatedRuns: 0,
+					holdoutCoverage: emptyCoverage(),
+					searchCoverage: emptyCoverage()
 				},
 				reason: "predictive-validity researcher does not execute plans; the caller is expected to run the sweep and call rubricPredictiveValidity directly with the resulting RunRecord[].",
 				rejectionCode: "few_runs"
@@ -1601,6 +1603,17 @@ var PredictiveValidityResearcher = class {
 		return this.lastReport;
 	}
 };
+/** Coverage of a split that was never dealt any work. */
+function emptyCoverage() {
+	return {
+		dealt: 0,
+		answered: 0,
+		unscoredPairs: 0,
+		candidateOnly: 0,
+		baselineOnly: 0,
+		coverage: 0
+	};
+}
 //#endregion
 //#region src/rl/preferences.ts
 /** The split each path pairs by default: training data comes from search. */