@tangle-network/agent-eval 0.86.0 → 0.90.0

package/CHANGELOG.md

@@ -4,7 +4,15 @@ All notable changes to `@tangle-network/agent-eval` and its sibling `agent-eval-
 
 ---
 
-## [0.86.0] — 2026-06-09 — fleet-rebuilt eval primitives
+## [0.90.0] — 2026-06-10 — infra perf-benchmark substrate (`/perf`)
+
+Domain-agnostic infra-performance benchmarking: a journeys × axes scenario matrix, record-integrity contracts over flat metric records, and a percentile ratchet. Complements the judge-panel `BenchmarkRunner` (root) — that one scores QUALITY via judges; `/perf` scores LATENCY / RELIABILITY. All additive — no existing export changed.
+
+### Added
+
+- **`JourneySpec` + `expandMatrix` + `scenarioKey` (`/perf` + root).** A journey is one measurable user path (`provision.cold`, `chat.ttft`) carrying its own data contract: `requiredFields` (must be non-null on a passing record), `minimums` (numeric floors, e.g. `event_count ≥ 1` for streaming), `phaseFields` (per-phase breakdown, reported separately), and `requiresLLM` (nightly vs per-PR scheduling). `expandMatrix` does the cartesian expansion over free-form `ScenarioAxes` (driver × region × …) with a `filter` for invalid combos; scenario keys are `journeyId|dim=value|…` with dims sorted, so the key is stable across axes-object insertion order.
+- **`checkRecordIntegrity` + `assertRecordIntegrity` (`/perf` + root).** A record claiming `pass === true` must actually carry its journey's required measurements — a "passing" run with a null `total_ms` is an integrity violation (`null-required-field` / `below-minimum`), not a pass. Failed records are exempt (an errored run legitimately has nulls); `resolveJourney` returning null skips the record. The assert variant throws listing every violation.
+- **`summarizeRecords` + `gatePerf` (`/perf` + root).** Percentile ratchet: fold flat records into per-scenario `PerfStat` (`p50` / `p90` / `n`, nearest-rank on sorted values), then gate a current `PerfBaseline` against a committed one. Null / non-numeric metric values are excluded from `n` and a zero-sample field is omitted — no fake zeros. Regressions trip when p50 OR p90 exceed `tolerancePct` (default 10) over baseline; strict improvements are reported with negative `overBy`; scenarios under `minSamples` (default 3) in current are surfaced in `missingScenarios` and never gated; baseline/current key drift lands in `missingScenarios` / `newScenarios`.
 
 One clean, canonical version of five generic patterns the fleet kept hand-rolling across 2–4 product agents each. All additive — no existing export changed.
 

package/dist/adapters/http.d.ts

@@ -1,6 +1,6 @@
-import { S as Scenario, D as DispatchFn, b as DispatchContext } from '../types-D7lLRYe9.js';
-import '../run-record-De9VarXR.js';
-import '../errors-Dwqw-T_m.js';
+import { S as Scenario, D as DispatchFn, b as DispatchContext } from '../types-BU-7W85F.js';
+import '../run-record-e7vj1uZQ.js';
+import '../errors-CzMUYo7b.js';
 import '../schema-m0gsnbt3.js';
 
 /**

package/dist/adapters/langchain.d.ts

@@ -1,6 +1,6 @@
-import { S as Scenario, J as JudgeScore, D as DispatchFn, a as JudgeConfig } from '../types-D7lLRYe9.js';
-import '../run-record-De9VarXR.js';
-import '../errors-Dwqw-T_m.js';
+import { S as Scenario, J as JudgeScore, D as DispatchFn, a as JudgeConfig } from '../types-BU-7W85F.js';
+import '../run-record-e7vj1uZQ.js';
+import '../errors-CzMUYo7b.js';
 import '../schema-m0gsnbt3.js';
 
 /**

package/dist/adapters/otel.d.ts

@@ -1,13 +1,13 @@
 import { TraceSpanEvent, HostedClient } from '../hosted/index.js';
-import '../types-D7lLRYe9.js';
-import '../run-record-De9VarXR.js';
-import '../errors-Dwqw-T_m.js';
+import '../types-BU-7W85F.js';
+import '../run-record-e7vj1uZQ.js';
+import '../errors-CzMUYo7b.js';
 import '../schema-m0gsnbt3.js';
-import '../insight-report-3ADTfClO.js';
-import '../summary-report-Db0dDSWP.js';
+import '../insight-report-BBwvOh6x.js';
+import '../summary-report-DGmUucwQ.js';
 import '../failure-cluster-CL7IVgkJ.js';
 import '../store-CKUAgsJz.js';
-import '../judge-calibration-DilmB3Ml.js';
+import '../judge-calibration-0p2QcWNE.js';
 
 /**
  * # `@tangle-network/agent-eval/adapters/otel` — OTel→hosted bridge.

package/dist/adversarial-DIVcDoI_.d.ts

@@ -0,0 +1,88 @@
+/**
+ * Adversarial scenario search.
+ *
+ * Capability evaluation on a fixed scenario set measures performance on
+ * the distribution someone curated. Production failure modes live in the
+ * tail — inputs the curator didn't think of, or actively avoided. The
+ * adversarial-search primitive actively looks for them: starting from a
+ * pool of scenarios where the policy already passes, it mutates them
+ * (paraphrase, edge-case substitution, compositional combination) and
+ * keeps the mutations that *break* the policy.
+ *
+ * This is not magic. It's the simplest version of the loop that AdA
+ * (Open-Ended Adaptation, DeepMind 2023), POET, and Anthropic's
+ * auto-jailbreak rigs all run: hill-climb against a failure indicator,
+ * keep the survivors, repeat. We ship the harness; consumers supply the
+ * mutation strategies and the failure detector.
+ *
+ * Why ship this in agent-eval and not as a separate red-team tool: every
+ * piece of the standard adversarial loop is already in this package
+ * (`runEvalCampaign` for the matrix run, `RawProviderSink` for capture,
+ * `assertRunCaptured` for integrity, `pairedEvalueSequence` for stop
+ * criteria). The adversarial primitive is just the *scenario-mutation
+ * meta-loop* on top of that machinery.
+ */
+interface AdversarialScenario<S> {
+    /** Stable id — used for deduplication and lineage tracking. */
+    id: string;
+    /** Generation index — 0 for seeds, 1 for first round of mutations, etc. */
+    generation: number;
+    /** Lineage — id of the parent scenario this was mutated from, if any. */
+    parentId: string | null;
+    scenario: S;
+    /** Score on the policy under test. Lower = adversarial signal. */
+    score: number | null;
+    /** Strategy that produced this mutation, for diagnostics. */
+    mutationStrategy: string | null;
+}
+interface AdversarialMutation<S> {
+    id: string;
+    /**
+     * Mutate one scenario. Return null to skip; return one or more new
+     * scenarios. The harness deduplicates by `mutateScenarioId(scenario)`.
+     */
+    mutate(parent: S, rng: () => number): Promise<S[]> | S[];
+}
+interface AdversarialSearchOptions<S> {
+    /** Initial scenarios — typically those the policy currently passes. */
+    seeds: S[];
+    /** Stable identifier extraction. */
+    mutateScenarioId: (s: S) => string;
+    /** Mutation strategies. */
+    mutations: AdversarialMutation<S>[];
+    /**
+     * Run the policy under test against one scenario, return a scalar score
+     * in [0, 1]. Lower = adversarial signal.
+     */
+    scoreFn: (s: S) => Promise<number>;
+    /**
+     * Threshold below which a scenario counts as a "failure" worth keeping.
+     * Default 0.5.
+     */
+    failureThreshold?: number;
+    /** Number of mutation rounds. Default 3. */
+    rounds?: number;
+    /** Children per parent per round. Default 4. */
+    childrenPerParent?: number;
+    /** Maximum total scenarios examined. Default Infinity. */
+    budget?: number;
+    /** Seed for the deterministic RNG. Default 1. */
+    seed?: number;
+}
+interface AdversarialSearchReport<S> {
+    scenarios: AdversarialScenario<S>[];
+    /** Discovered failures sorted by score ascending. */
+    failures: AdversarialScenario<S>[];
+    /** Round-by-round counts. */
+    byGeneration: Array<{
+        generation: number;
+        total: number;
+        failures: number;
+        meanScore: number;
+    }>;
+    /** Total scoreFn invocations consumed. */
+    scoreCalls: number;
+}
+declare function adversarialScenarioSearch<S>(opts: AdversarialSearchOptions<S>): Promise<AdversarialSearchReport<S>>;
+
+export { type AdversarialMutation as A, type AdversarialScenario as a, type AdversarialSearchOptions as b, type AdversarialSearchReport as c, adversarialScenarioSearch as d };

package/dist/analyst/index.d.ts

@@ -1,22 +1,23 @@
 import { AxAIService, AxFunction } from '@ax-llm/ax';
-import { M as MultiLayerVerifier, V as VerifyOptions, S as Severity } from '../multi-layer-verifier-DlWCXuxL.js';
+import { M as MultiLayerVerifier, V as VerifyOptions, S as Severity } from '../multi-layer-verifier-DUZXrPDA.js';
 import { c as RunCritic, a as RunTrace } from '../run-critic-BAIjX99r.js';
-import { S as SemanticConceptJudgeOptions, a as SemanticConceptJudgeInput, B as BehavioralMetrics } from '../semantic-concept-judge-DIEgr_6v.js';
-export { C as CreateAnalystAiConfig, D as DEFAULT_TRACE_ANALYST_KINDS, b as DefaultAnalystRegistryOptions, c as DiffPolicy, F as FAILURE_MODE_KIND_SPEC, d as FINDING_SUBJECT_GRAMMAR_PROMPT, e as FINDING_SUBJECT_KINDS, f as FindingSubject, g as FindingSubjectKind, h as FindingSubjectStringSchema, i as FindingsDiff, j as FindingsStore, I as IMPROVEMENT_KIND_SPEC, K as KIND_EXPECTED_SUBJECTS, k as KNOWLEDGE_GAP_KIND_SPEC, l as KNOWLEDGE_POISONING_KIND_SPEC, P as PersistedFinding, m as SKILL_USAGE_ANALYST, n as SkillUsageAnalyst, o as SkillUsageRecord, p as SkillUsageReport, q as SkillUsageScanConfig, r as buildDefaultAnalystRegistry, s as buildSkillUsageReport, t as createAnalystAi, u as defaultIsMaterial, v as diffFindings, w as emitSkillUsageFindings, x as parseFindingSubject, y as renderFindingSubject } from '../semantic-concept-judge-DIEgr_6v.js';
+import { S as SemanticConceptJudgeOptions, a as SemanticConceptJudgeInput, B as BehavioralMetrics } from '../semantic-concept-judge-Dn8Z6KEG.js';
+export { C as CreateAnalystAiConfig, D as DEFAULT_TRACE_ANALYST_KINDS, b as DiffPolicy, F as FAILURE_MODE_KIND_SPEC, c as FINDING_SUBJECT_GRAMMAR_PROMPT, d as FINDING_SUBJECT_KINDS, e as FindingSubject, f as FindingSubjectKind, g as FindingSubjectStringSchema, h as FindingsDiff, i as FindingsStore, I as IMPROVEMENT_KIND_SPEC, K as KIND_EXPECTED_SUBJECTS, j as KNOWLEDGE_GAP_KIND_SPEC, k as KNOWLEDGE_POISONING_KIND_SPEC, P as PersistedFinding, l as SKILL_USAGE_ANALYST, m as SkillUsageAnalyst, n as SkillUsageRecord, o as SkillUsageReport, p as SkillUsageScanConfig, q as buildSkillUsageReport, r as createAnalystAi, s as defaultIsMaterial, t as diffFindings, u as emitSkillUsageFindings, v as parseFindingSubject, w as renderFindingSubject } from '../semantic-concept-judge-Dn8Z6KEG.js';
 import { A as AnalyzeTracesOptions } from '../analyst-C8HHvfJp.js';
 import { T as TraceAnalysisStore } from '../store-C1YxJDEK.js';
 import { b as JudgeFn, a as JudgeInput } from '../types-Croy5h7V.js';
-import { A as Analyst, h as AnalystSeverity, c as AnalystFinding } from '../types-Cu3u_x59.js';
-export { a as AnalystContext, g as AnalystCost, i as AnalystInputKind, j as AnalystRequirements, f as AnalystRunEvent, e as AnalystRunInputs, d as AnalystRunResult, b as AnalystRunSummary, k as ChatCallOpts, C as ChatClient, l as ChatRequest, m as ChatResponse, n as ChatTransport, o as CliBridgeTransportOpts, p as CreateChatClientOpts, D as DirectProviderTransportOpts, E as EvidenceRef, M as MockTransportOpts, R as RouterTransportOpts, S as SandboxSdkTransportOpts, q as computeFindingId, r as createChatClient, s as makeFinding } from '../types-Cu3u_x59.js';
+import { A as Analyst, h as AnalystSeverity, c as AnalystFinding } from '../types-2VVIL04s.js';
+export { a as AnalystContext, g as AnalystCost, i as AnalystInputKind, j as AnalystRequirements, f as AnalystRunEvent, e as AnalystRunInputs, d as AnalystRunResult, b as AnalystRunSummary, k as ChatCallOpts, C as ChatClient, l as ChatRequest, m as ChatResponse, n as ChatTransport, o as CliBridgeTransportOpts, p as CreateChatClientOpts, D as DirectProviderTransportOpts, E as EvidenceRef, M as MockTransportOpts, R as RouterTransportOpts, S as SandboxSdkTransportOpts, q as computeFindingId, r as createChatClient, s as makeFinding } from '../types-2VVIL04s.js';
 import { TCloud } from '@tangle-network/tcloud';
-export { A as ANALYST_SEVERITIES, C as CreateTraceAnalystKindOpts, R as RAW_FINDING_SCHEMA_PROMPT, a as RawAnalystFinding, b as RawAnalystFindingSchema, T as TraceAnalystGolden, c as TraceAnalystKindSpec, d as createTraceAnalystKind, p as parseRawFinding, r as renderPriorFindings } from '../kind-factory-CVecZZG_.js';
-export { A as AnalystHooks, a as AnalystRegistry, b as AnalystRegistryOptions, B as BudgetPolicy, R as RegistryRunOpts } from '../registry-DrEQ3Luj.js';
-import { L as LlmClientOptions } from '../llm-client-CuUg2Mn3.js';
+export { a as AnalystHooks, A as AnalystRegistry, c as AnalystRegistryOptions, B as BudgetPolicy, D as DefaultAnalystRegistryOptions, R as RegistryRunOpts, b as buildDefaultAnalystRegistry } from '../default-registry-zoGHUQEH.js';
+export { A as ANALYST_SEVERITIES, C as CreateTraceAnalystKindOpts, R as RAW_FINDING_SCHEMA_PROMPT, a as RawAnalystFinding, b as RawAnalystFindingSchema, c as TraceAnalystGolden, T as TraceAnalystKindSpec, d as createTraceAnalystKind, p as parseRawFinding, r as renderPriorFindings } from '../kind-factory-5b7xXXOr.js';
+import { L as LlmClientOptions } from '../llm-client-BeEcAokY.js';
+import '../verdict-C9MlYujm.js';
 import '../schema-m0gsnbt3.js';
 import '../store-CKUAgsJz.js';
 import 'zod';
-import '../run-record-De9VarXR.js';
-import '../errors-Dwqw-T_m.js';
+import '../run-record-e7vj1uZQ.js';
+import '../errors-CzMUYo7b.js';
 import '../raw-provider-sink-C46HDghv.js';
 
 /**

package/dist/analyst/index.js

@@ -4,17 +4,19 @@ import {
   SEMANTIC_CONCEPT_JUDGE_VERSION,
   SKILL_USAGE_ANALYST,
   SkillUsageAnalyst,
-  behavioralAnalyst,
-  buildDefaultAnalystRegistry,
   buildSkillUsageReport,
   createAnalystAi,
   createChatClient,
   defaultIsMaterial,
-  deriveEfficiencyFindings,
   diffFindings,
   emitSkillUsageFindings,
   runSemanticConceptJudge
-} from "../chunk-L5G7OUKD.js";
+} from "../chunk-XY4DDNEG.js";
+import {
+  behavioralAnalyst,
+  buildDefaultAnalystRegistry,
+  deriveEfficiencyFindings
+} from "../chunk-UMMZHCPB.js";
 import {
   ANALYST_SEVERITIES,
   AnalystRegistry,
@@ -32,20 +34,23 @@ import {
   buildTraceToolsForGroup,
   coerceJson,
   coerceToFindingRows,
-  computeFindingId,
   createTraceAnalystKind,
-  makeFinding,
   parseFindingSubject,
   parseRawFinding,
   renderFindingSubject,
   renderPriorFindings,
   stripCodeFences,
   structureFindings
-} from "../chunk-VIDQF3F5.js";
+} from "../chunk-D3V5B42D.js";
 import "../chunk-CVVHBFGN.js";
+import {
+  computeFindingId,
+  makeFinding
+} from "../chunk-45EEMHTC.js";
 import {
   analyzeTraces
-} from "../chunk-VUINJM5M.js";
+} from "../chunk-UHMJT4T7.js";
+import "../chunk-QAY5UIJO.js";
 import "../chunk-PC4UYEBM.js";
 import "../chunk-3BFEG2F6.js";
 import "../chunk-PZ5AY32C.js";

package/dist/analyst/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/analyst/adapters.ts","../../src/analyst/steer-firewall.ts"],"sourcesContent":["/**\n * Adapter factories — lift each existing agent-eval primitive into the\n * Analyst contract without re-implementing it.\n *\n * Five primitives, five factories. Each one:\n *   - Builds an Analyst with a stable id (caller chooses; defaults\n *     given), a sensible default `inputKind`, a version derived from\n *     the wrapped primitive's version + an adapter revision, and an\n *     `analyze()` that calls the primitive and lifts its output to\n *     AnalystFinding[] using `makeFinding()`.\n *   - Maps severities: the existing `Severity` ('critical' | 'major' |\n *     'minor' | 'info') projects onto AnalystSeverity ('critical' |\n *     'high' | 'medium' | 'low' | 'info'); 'major' → 'high', 'minor' →\n *     'medium'. Domain analysts that want finer-grained mapping override.\n *\n * Adapters never own state. Calling the same factory twice with the\n * same primitive instance is safe.\n */\n\nimport type { AxAIService } from '@ax-llm/ax'\nimport type {\n  Finding as LayerFinding,\n  Severity as LayerSeverity,\n  MultiLayerVerifier,\n  VerifyOptions,\n} from '../multi-layer-verifier'\nimport { RunCritic, type RunTrace } from '../run-critic'\nimport {\n  runSemanticConceptJudge,\n  SEMANTIC_CONCEPT_JUDGE_VERSION,\n  type SemanticConceptJudgeInput,\n  type SemanticConceptJudgeOptions,\n} from '../semantic-concept-judge'\nimport { type AnalyzeTracesOptions, analyzeTraces } from '../trace-analyst/analyst'\nimport type { TraceAnalysisStore } from '../trace-analyst/store'\nimport type { JudgeFn, JudgeInput, JudgeScore, TCloud } from '../types'\nimport type { Analyst, AnalystFinding, AnalystSeverity } from './types'\nimport { makeFinding } from './types'\n\nconst ADAPTER_REV = '1'\n\n// ── Severity bridges ───────────────────────────────────────────────\n\nexport function liftSeverity(s: LayerSeverity): AnalystSeverity {\n  switch (s) {\n    case 'critical':\n      return 'critical'\n    case 'major':\n      return 'high'\n    case 'minor':\n      return 'medium'\n    case 'info':\n      return 'info'\n  }\n}\n\n// ── 1. analyzeTraces → Analyst ─────────────────────────────────────\n\nexport interface TraceAnalystAdapterOpts {\n  id?: string\n  area?: string\n  /** The natural-language question(s) put to the analyst. One finding per question. */\n  questions: string[]\n  /** Caller-provided AxAI service — same one trace-analyst.ts expects. */\n  ai: AxAIService\n  model?: string\n  /** Forwarded to analyzeTraces. */\n  extra?: Omit<AnalyzeTracesOptions, 'source' | 'ai' | 'model'>\n}\n\n/**\n * @deprecated Prefer `createTraceAnalystKind` + one of the failure /\n * improvement kinds from `./kinds`. This adapter wraps the legacy\n * `analyzeTraces` flow whose output is `findings:string[]` — every\n * bullet gets flat-defaulted severity `medium` / confidence `0.6`,\n * which loses the per-finding grading kinds provide via Ax structured\n * output + Zod validation. Kept for one minor while consumers migrate.\n */\nexport function createTraceAnalystAdapter(\n  opts: TraceAnalystAdapterOpts,\n): Analyst<TraceAnalysisStore> {\n  const id = opts.id ?? 'trace-analyst'\n  const area = opts.area ?? 'agent-reasoning'\n  return {\n    id,\n    description:\n      'Runs the agent-eval trace analyst over an OTLP trace store and lifts its bulleted findings.',\n    inputKind: 'trace-store',\n    cost: { kind: 'llm', models: opts.model ? [opts.model] : undefined },\n    version: `trace-analyst-${ADAPTER_REV}`,\n    async analyze(store, ctx) {\n      const out: AnalystFinding[] = []\n      for (const question of opts.questions) {\n        if (ctx.signal?.aborted) break\n        const result = await analyzeTraces(\n          { question },\n          { source: store, ai: opts.ai, model: opts.model, ...opts.extra },\n        )\n        const subject = ctx.tags?.subject ?? question.slice(0, 60)\n        // The responder produces a list of bullet strings. Each becomes\n        // one finding; the prose answer is attached as rationale on the\n        // first (so renderers that show only top-N still get context).\n        if (result.findings.length === 0) {\n          out.push(\n            makeFinding({\n              analyst_id: id,\n              area,\n              subject,\n              claim: result.answer.slice(0, 200),\n              rationale: result.answer,\n              severity: 'info',\n              confidence: 0.5,\n              evidence_refs: [],\n              metadata: {\n                actor_prompt_version: result.actorPromptVersion,\n                turns: result.turnCount,\n              },\n            }),\n          )\n          continue\n        }\n        result.findings.forEach((claim, i) => {\n          out.push(\n            makeFinding({\n              analyst_id: id,\n              area,\n              subject,\n              claim,\n              rationale: i === 0 ? result.answer : undefined,\n              severity: 'medium',\n              confidence: 0.6,\n              evidence_refs: [],\n              metadata: { question, turns: result.turnCount, finding_index: i },\n            }),\n          )\n        })\n      }\n      return out\n    },\n  }\n}\n\n// ── 2. MultiLayerVerifier → Analyst ─────────────────────────────────\n\nexport interface VerifierAdapterOpts<Env> {\n  id?: string\n  area?: string\n  verifier: MultiLayerVerifier<Env>\n  /**\n   * The verifier expects an `env` per run. Adapters take it from\n   * `AnalystRunInputs.custom[<id>]` via the registry's 'custom' routing.\n   */\n  options?: Omit<VerifyOptions<Env>, 'env'>\n}\n\nexport function createVerifierAdapter<Env>(opts: VerifierAdapterOpts<Env>): Analyst<Env> {\n  const id = opts.id ?? 'multi-layer-verifier'\n  const area = opts.area ?? 'verification'\n  return {\n    id,\n    description:\n      \"Runs a MultiLayerVerifier and lifts each layer's findings into the analyst envelope.\",\n    inputKind: 'custom',\n    cost: { kind: 'deterministic' },\n    version: `verifier-${ADAPTER_REV}`,\n    async analyze(env, ctx) {\n      const report = await opts.verifier.run({ env, ...opts.options })\n      const out: AnalystFinding[] = []\n      for (const layer of report.layers) {\n        for (const finding of layer.findings) {\n          out.push(liftLayerFinding(id, area, layer.layer, finding))\n        }\n        // Layer-level signal: a failed/error layer is itself a finding\n        // even if it didn't emit per-finding rows.\n        if (layer.status === 'fail' || layer.status === 'error' || layer.status === 'timeout') {\n          out.push(\n            makeFinding({\n              analyst_id: id,\n              area,\n              subject: layer.layer,\n              claim: `layer \"${layer.layer}\" ${layer.status}: ${layer.reason ?? 'no reason given'}`,\n              severity:\n                layer.status === 'error' ? 'high' : layer.status === 'timeout' ? 'medium' : 'high',\n              confidence: 1,\n              evidence_refs: [],\n              metadata: {\n                layer_status: layer.status,\n                duration_ms: layer.durationMs,\n                score: layer.score,\n                diagnostics: layer.diagnostics,\n              },\n            }),\n          )\n        }\n      }\n      ctx.log?.('verifier complete', {\n        layers: report.layers.length,\n        blended: report.blendedScore,\n        all_pass: report.allPass,\n      })\n      return out\n    },\n  }\n}\n\nfunction liftLayerFinding(\n  analyst_id: string,\n  area: string,\n  layer: string,\n  f: LayerFinding,\n): AnalystFinding {\n  return makeFinding({\n    analyst_id,\n    area,\n    subject: f.layer ?? layer,\n    claim: f.message,\n    severity: liftSeverity(f.severity),\n    confidence: 0.85,\n    evidence_refs: f.evidence\n      ? [{ kind: 'artifact', uri: 'inline:evidence', excerpt: f.evidence }]\n      : [],\n    metadata: f.detail,\n  })\n}\n\n// ── 3. RunCritic → Analyst ──────────────────────────────────────────\n\nexport interface RunCriticAdapterOpts {\n  id?: string\n  area?: string\n  critic?: RunCritic\n  /** Optional threshold below which a dimension is reported as a finding. Default 0.5. */\n  threshold?: number\n}\n\nexport function createRunCriticAdapter(opts: RunCriticAdapterOpts = {}): Analyst<RunTrace> {\n  const id = opts.id ?? 'run-critic'\n  const area = opts.area ?? 'run-quality'\n  const critic = opts.critic ?? new RunCritic()\n  const threshold = opts.threshold ?? 0.5\n  return {\n    id,\n    description:\n      'Scores a single run across success / grounding / drift / tool-quality and surfaces below-threshold dimensions.',\n    inputKind: 'custom',\n    cost: { kind: 'deterministic' },\n    version: `run-critic-${ADAPTER_REV}`,\n    async analyze(trace) {\n      const score = critic.scoreTrace(trace)\n      const out: AnalystFinding[] = []\n      const dims: Array<[keyof typeof score, AnalystSeverity, string]> = [\n        ['success', 'critical', 'run did not complete successfully'],\n        ['goalProgress', 'high', 'goal progress is low'],\n        ['repoGroundedness', 'high', 'output is poorly grounded in the repository'],\n        ['toolUseQuality', 'medium', 'tool use quality is low'],\n        ['patchQuality', 'medium', 'no real patch/edit evidence'],\n        ['testReality', 'high', 'no real test/build evidence'],\n        ['finalGate', 'critical', 'final gate is blocking'],\n      ]\n      for (const [dim, sev, msg] of dims) {\n        const value = score[dim] as number\n        if (typeof value === 'number' && value < threshold) {\n          out.push(\n            makeFinding({\n              analyst_id: id,\n              area,\n              subject: dim,\n              claim: msg,\n              rationale: `${dim}=${value.toFixed(2)} below threshold ${threshold}`,\n              severity: sev,\n              confidence: 1,\n              evidence_refs: [],\n              metadata: { dimension: dim, value, threshold, run_id: trace.run.runId },\n            }),\n          )\n        }\n      }\n      // Drift penalty is high → surface as a finding (inverse threshold).\n      if (score.driftPenalty > 1 - threshold) {\n        out.push(\n          makeFinding({\n            analyst_id: id,\n            area,\n            subject: 'drift',\n            claim: 'agent output drifted from repository signal',\n            rationale: `driftPenalty=${score.driftPenalty.toFixed(2)}`,\n            severity: 'medium',\n            confidence: 0.9,\n            evidence_refs: [],\n            metadata: { drift_penalty: score.driftPenalty, notes: score.notes },\n          }),\n        )\n      }\n      return out\n    },\n  }\n}\n\n// ── 4. JudgeFn → Analyst ────────────────────────────────────────────\n\nexport interface JudgeAdapterOpts {\n  id?: string\n  area?: string\n  judge: JudgeFn\n  /** TCloud handle the JudgeFn calls. */\n  tcloud: TCloud\n  /** Optional cost classification — most judges call an LLM. */\n  cost?: Analyst['cost']\n  /** Optional threshold below which a JudgeScore becomes a finding. Default 6 (on 0-10 scale). */\n  threshold?: number\n}\n\nexport function createJudgeAdapter(opts: JudgeAdapterOpts): Analyst<JudgeInput> {\n  const id = opts.id ?? 'judge'\n  const area = opts.area ?? 'judge'\n  const threshold = opts.threshold ?? 6\n  return {\n    id,\n    description:\n      'Wraps an agent-eval JudgeFn into an analyst; below-threshold dimensions surface as findings.',\n    inputKind: 'judge-input',\n    cost: opts.cost ?? { kind: 'llm' },\n    version: `judge-${ADAPTER_REV}`,\n    async analyze(input) {\n      const scores = await opts.judge(opts.tcloud, input)\n      return scores\n        .filter((s) => normalize10(s.score) < threshold)\n        .map((s) => liftJudgeScore(id, area, s))\n    },\n  }\n}\n\nfunction normalize10(s: number): number {\n  // JudgeScore convention is 0-10 but some judges emit 0-1. Coerce to 0-10.\n  return s <= 1 ? s * 10 : s\n}\n\nfunction liftJudgeScore(analyst_id: string, area: string, s: JudgeScore): AnalystFinding {\n  const score10 = normalize10(s.score)\n  const severity: AnalystSeverity =\n    score10 < 3 ? 'critical' : score10 < 5 ? 'high' : score10 < 7 ? 'medium' : 'low'\n  return makeFinding({\n    analyst_id,\n    area,\n    subject: s.dimension,\n    claim: `${s.judgeName}/${s.dimension} scored ${score10.toFixed(1)}/10`,\n    rationale: s.reasoning,\n    severity,\n    confidence: 0.8,\n    evidence_refs: s.evidence\n      ? [{ kind: 'artifact', uri: 'inline:evidence', excerpt: s.evidence }]\n      : [],\n    // Provenance: this finding IS a judge verdict (an acceptance score), not an\n    // observation of behavior. The steer firewall (assertNoJudgeVerdict) rejects\n    // it from steering — even when it cites an artifact above — because letting a\n    // verdict steer the next attempt is the held-out judge leaking into the loop.\n    derived_from_judge: true,\n    metadata: { judge_name: s.judgeName, dimension: s.dimension, score_10: score10 },\n  })\n}\n\n// ── 5. SemanticConceptJudge → Analyst ──────────────────────────────\n\nexport interface SemanticConceptJudgeAdapterOpts {\n  id?: string\n  area?: string\n  options?: SemanticConceptJudgeOptions\n}\n\nexport function createSemanticConceptJudgeAdapter(\n  opts: SemanticConceptJudgeAdapterOpts = {},\n): Analyst<SemanticConceptJudgeInput> {\n  const id = opts.id ?? 'semantic-concept-judge'\n  const area = opts.area ?? 'concept-coverage'\n  return {\n    id,\n    description:\n      'Runs the semantic-concept judge and surfaces missing / weak concepts as findings.',\n    inputKind: 'custom',\n    cost: { kind: 'llm', models: opts.options?.model ? [opts.options.model] : undefined },\n    version: `${SEMANTIC_CONCEPT_JUDGE_VERSION}-adapter-${ADAPTER_REV}`,\n    async analyze(input) {\n      const result = await runSemanticConceptJudge(input, opts.options)\n      if (!result.available) {\n        return [\n          makeFinding({\n            analyst_id: id,\n            area,\n            claim: 'semantic-concept judge unavailable',\n            rationale: result.error,\n            severity: 'info',\n            confidence: 1,\n            evidence_refs: [],\n            metadata: { reason: result.error },\n          }),\n        ]\n      }\n      const out: AnalystFinding[] = []\n      for (const f of result.findings) {\n        // Only surface gaps: missing concepts or low scores. Concepts at\n        // 7+/10 with present=true are not findings — they're successes.\n        if (f.present && f.score >= 7) continue\n        out.push(\n          makeFinding({\n            analyst_id: id,\n            area,\n            subject: f.concept,\n            claim: f.present\n              ? `concept \"${f.concept}\" is weak (${f.score}/10)`\n              : `concept \"${f.concept}\" is missing`,\n            rationale: f.evidence,\n            severity: liftSeverity(f.severity),\n            confidence: 0.85,\n            evidence_refs: [{ kind: 'artifact', uri: 'inline:evidence', excerpt: f.evidence }],\n            metadata: {\n              concept: f.concept,\n              present: f.present,\n              score_10: f.score,\n              cost_usd: result.costUsd ?? undefined,\n            },\n          }),\n        )\n      }\n      return out\n    },\n  }\n}\n","// The realness-oracle firewall (docs/learning-flywheel.md, \"The steer is f(trace)\").\n//\n// A realness/authenticity signal has TWO legitimate roles that must stay\n// separated by a firewall:\n//   (a) anchor judge J — write-only: scores the chosen output, gates promotion,\n//       NEVER seen by the worker/optimizer mid-run (else the loop games it).\n//   (b) steer f(trace) — an analyst observes the agent's OWN behavior in the\n//       trace (\"imported a stub\", \"used a non-crypto PRNG where encryption was\n//       required\") and steers the next attempt. Legitimate, because it is derived\n//       from OBSERVABLE BEHAVIOR, not from J's held-out verdict.\n//\n// The correct discriminator is PROVENANCE, not evidence presence. A judge verdict\n// lifted into a finding (createJudgeAdapter → liftJudgeScore) is a verdict even\n// when it cites an artifact; an evidence-less trace-analyst bullet is an\n// observation even though it cites nothing. So the firewall keys on\n// `AnalystFinding.derived_from_judge` (set at the judge lift site), NOT on whether\n// evidence_refs is populated. The instant a verdict steers the next attempt it is\n// a back-channel for J and the loop Goodharts realness exactly as it would\n// Goodhart pass-rate.\n\nimport type { AnalystFinding, EvidenceRef } from './types'\n\n/** Evidence grounded in the agent's OWN execution: OTLP trace elements\n *  (`span`/`event`) or the artifact it produced (`artifact`). */\nconst OBSERVABLE_KINDS: ReadonlySet<EvidenceRef['kind']> = new Set<EvidenceRef['kind']>([\n  'span',\n  'event',\n  'artifact',\n])\n\n/** DESCRIPTIVE predicate: does the finding cite at least one observable\n *  (span/event/artifact) evidence ref. Useful for ranking evidence quality or\n *  rendering — it is NOT the steer gate. Evidence presence is the WRONG\n *  discriminator for steering: a legitimate trace-analyst observation may cite\n *  nothing (it would be wrongly rejected), and a judge verdict may cite an\n *  artifact (it would be wrongly admitted). Use `assertNoJudgeVerdict` to gate\n *  steering; use this only where \"is this grounded in observable evidence\" is the\n *  literal question. */\nexport function isTraceObservable(finding: AnalystFinding): boolean {\n  return finding.evidence_refs.some((ref) => OBSERVABLE_KINDS.has(ref.kind))\n}\n\n/** True iff the finding is a JUDGE VERDICT (an acceptance score lifted into a\n *  finding), identified by provenance set at the lift site — independent of\n *  whatever evidence it cites. */\nexport function isJudgeVerdict(finding: AnalystFinding): boolean {\n  return finding.derived_from_judge === true\n}\n\n/**\n * THE steer firewall. Fail-loud guard for any path that admits analyst findings\n * as STEERING input (the `f(trace)` role): rejects — naming the offenders — any\n * finding whose provenance is a judge verdict, rather than let `J` leak into the\n * loop. Returns the findings unchanged for chaining.\n *\n * Call this at the chokepoint where a detector that ALSO scores/gates has its\n * findings turned into a steer (the judge-and-steer dual-role case). It keys on\n * provenance, so it correctly admits evidence-less trace-analyst observations and\n * correctly rejects an artifact-citing judge verdict — the cases an evidence\n * check gets backwards.\n *\n * It is necessary, not sufficient: it stops PROVENANCE-tagged verdicts. A judge\n * whose output is laundered through a hand-built finding with no provenance flag\n * is out of its reach — provenance must be honestly set at every judge→finding\n * lift (today: createJudgeAdapter). That is why the integrity rule lives at the\n * lift site, and why ProposeContext.judgeScores?: never is the complementary\n * compile-time tripwire on the obvious direct channel.\n */\nexport function assertNoJudgeVerdict(\n  findings: ReadonlyArray<AnalystFinding>,\n  context = 'steer',\n): ReadonlyArray<AnalystFinding> {\n  const leaks = findings.filter(isJudgeVerdict)\n  if (leaks.length > 0) {\n    throw new Error(\n      `${context}: a judge verdict cannot be admitted as steering input — that is the ` +\n        `held-out judge leaking into the loop. Offending judge-derived findings: [${leaks\n          .map((f) => f.finding_id)\n          .join(', ')}]. Steering consumes observations of behavior, never acceptance verdicts.`,\n    )\n  }\n  return findings\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuCA,IAAM,cAAc;AAIb,SAAS,aAAa,GAAmC;AAC9D,UAAQ,GAAG;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,EACX;AACF;AAwBO,SAAS,0BACd,MAC6B;AAC7B,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,OAAO,QAAQ,KAAK,QAAQ,CAAC,KAAK,KAAK,IAAI,OAAU;AAAA,IACnE,SAAS,iBAAiB,WAAW;AAAA,IACrC,MAAM,QAAQ,OAAO,KAAK;AACxB,YAAM,MAAwB,CAAC;AAC/B,iBAAW,YAAY,KAAK,WAAW;AACrC,YAAI,IAAI,QAAQ,QAAS;AACzB,cAAM,SAAS,MAAM;AAAA,UACnB,EAAE,SAAS;AAAA,UACX,EAAE,QAAQ,OAAO,IAAI,KAAK,IAAI,OAAO,KAAK,OAAO,GAAG,KAAK,MAAM;AAAA,QACjE;AACA,cAAM,UAAU,IAAI,MAAM,WAAW,SAAS,MAAM,GAAG,EAAE;AAIzD,YAAI,OAAO,SAAS,WAAW,GAAG;AAChC,cAAI;AAAA,YACF,YAAY;AAAA,cACV,YAAY;AAAA,cACZ;AAAA,cACA;AAAA,cACA,OAAO,OAAO,OAAO,MAAM,GAAG,GAAG;AAAA,cACjC,WAAW,OAAO;AAAA,cAClB,UAAU;AAAA,cACV,YAAY;AAAA,cACZ,eAAe,CAAC;AAAA,cAChB,UAAU;AAAA,gBACR,sBAAsB,OAAO;AAAA,gBAC7B,OAAO,OAAO;AAAA,cAChB;AAAA,YACF,CAAC;AAAA,UACH;AACA;AAAA,QACF;AACA,eAAO,SAAS,QAAQ,CAAC,OAAO,MAAM;AACpC,cAAI;AAAA,YACF,YAAY;AAAA,cACV,YAAY;AAAA,cACZ;AAAA,cACA;AAAA,cACA;AAAA,cACA,WAAW,MAAM,IAAI,OAAO,SAAS;AAAA,cACrC,UAAU;AAAA,cACV,YAAY;AAAA,cACZ,eAAe,CAAC;AAAA,cAChB,UAAU,EAAE,UAAU,OAAO,OAAO,WAAW,eAAe,EAAE;AAAA,YAClE,CAAC;AAAA,UACH;AAAA,QACF,CAAC;AAAA,MACH;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAeO,SAAS,sBAA2B,MAA8C;AACvF,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,gBAAgB;AAAA,IAC9B,SAAS,YAAY,WAAW;AAAA,IAChC,MAAM,QAAQ,KAAK,KAAK;AACtB,YAAM,SAAS,MAAM,KAAK,SAAS,IAAI,EAAE,KAAK,GAAG,KAAK,QAAQ,CAAC;AAC/D,YAAM,MAAwB,CAAC;AAC/B,iBAAW,SAAS,OAAO,QAAQ;AACjC,mBAAW,WAAW,MAAM,UAAU;AACpC,cAAI,KAAK,iBAAiB,IAAI,MAAM,MAAM,OAAO,OAAO,CAAC;AAAA,QAC3D;AAGA,YAAI,MAAM,WAAW,UAAU,MAAM,WAAW,WAAW,MAAM,WAAW,WAAW;AACrF,cAAI;AAAA,YACF,YAAY;AAAA,cACV,YAAY;AAAA,cACZ;AAAA,cACA,SAAS,MAAM;AAAA,cACf,OAAO,UAAU,MAAM,KAAK,KAAK,MAAM,MAAM,KAAK,MAAM,UAAU,iBAAiB;AAAA,cACnF,UACE,MAAM,WAAW,UAAU,SAAS,MAAM,WAAW,YAAY,WAAW;AAAA,cAC9E,YAAY;AAAA,cACZ,eAAe,CAAC;AAAA,cAChB,UAAU;AAAA,gBACR,cAAc,MAAM;AAAA,gBACpB,aAAa,MAAM;AAAA,gBACnB,OAAO,MAAM;AAAA,gBACb,aAAa,MAAM;AAAA,cACrB;AAAA,YACF,CAAC;AAAA,UACH;AAAA,QACF;AAAA,MACF;AACA,UAAI,MAAM,qBAAqB;AAAA,QAC7B,QAAQ,OAAO,OAAO;AAAA,QACtB,SAAS,OAAO;AAAA,QAChB,UAAU,OAAO;AAAA,MACnB,CAAC;AACD,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAEA,SAAS,iBACP,YACA,MACA,OACA,GACgB;AAChB,SAAO,YAAY;AAAA,IACjB;AAAA,IACA;AAAA,IACA,SAAS,EAAE,SAAS;AAAA,IACpB,OAAO,EAAE;AAAA,IACT,UAAU,aAAa,EAAE,QAAQ;AAAA,IACjC,YAAY;AAAA,IACZ,eAAe,EAAE,WACb,CAAC,EAAE,MAAM,YAAY,KAAK,mBAAmB,SAAS,EAAE,SAAS,CAAC,IAClE,CAAC;AAAA,IACL,UAAU,EAAE;AAAA,EACd,CAAC;AACH;AAYO,SAAS,uBAAuB,OAA6B,CAAC,GAAsB;AACzF,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,SAAS,KAAK,UAAU,IAAI,UAAU;AAC5C,QAAM,YAAY,KAAK,aAAa;AACpC,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,gBAAgB;AAAA,IAC9B,SAAS,cAAc,WAAW;AAAA,IAClC,MAAM,QAAQ,OAAO;AACnB,YAAM,QAAQ,OAAO,WAAW,KAAK;AACrC,YAAM,MAAwB,CAAC;AAC/B,YAAM,OAA6D;AAAA,QACjE,CAAC,WAAW,YAAY,mCAAmC;AAAA,QAC3D,CAAC,gBAAgB,QAAQ,sBAAsB;AAAA,QAC/C,CAAC,oBAAoB,QAAQ,6CAA6C;AAAA,QAC1E,CAAC,kBAAkB,UAAU,yBAAyB;AAAA,QACtD,CAAC,gBAAgB,UAAU,6BAA6B;AAAA,QACxD,CAAC,eAAe,QAAQ,6BAA6B;AAAA,QACrD,CAAC,aAAa,YAAY,wBAAwB;AAAA,MACpD;AACA,iBAAW,CAAC,KAAK,KAAK,GAAG,KAAK,MAAM;AAClC,cAAM,QAAQ,MAAM,GAAG;AACvB,YAAI,OAAO,UAAU,YAAY,QAAQ,WAAW;AAClD,cAAI;AAAA,YACF,YAAY;AAAA,cACV,YAAY;AAAA,cACZ;AAAA,cACA,SAAS;AAAA,cACT,OAAO;AAAA,cACP,WAAW,GAAG,GAAG,IAAI,MAAM,QAAQ,CAAC,CAAC,oBAAoB,SAAS;AAAA,cAClE,UAAU;AAAA,cACV,YAAY;AAAA,cACZ,eAAe,CAAC;AAAA,cAChB,UAAU,EAAE,WAAW,KAAK,OAAO,WAAW,QAAQ,MAAM,IAAI,MAAM;AAAA,YACxE,CAAC;AAAA,UACH;AAAA,QACF;AAAA,MACF;AAEA,UAAI,MAAM,eAAe,IAAI,WAAW;AACtC,YAAI;AAAA,UACF,YAAY;AAAA,YACV,YAAY;AAAA,YACZ;AAAA,YACA,SAAS;AAAA,YACT,OAAO;AAAA,YACP,WAAW,gBAAgB,MAAM,aAAa,QAAQ,CAAC,CAAC;AAAA,YACxD,UAAU;AAAA,YACV,YAAY;AAAA,YACZ,eAAe,CAAC;AAAA,YAChB,UAAU,EAAE,eAAe,MAAM,cAAc,OAAO,MAAM,MAAM;AAAA,UACpE,CAAC;AAAA,QACH;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAgBO,SAAS,mBAAmB,MAA6C;AAC9E,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,YAAY,KAAK,aAAa;AACpC,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,KAAK,QAAQ,EAAE,MAAM,MAAM;AAAA,IACjC,SAAS,SAAS,WAAW;AAAA,IAC7B,MAAM,QAAQ,OAAO;AACnB,YAAM,SAAS,MAAM,KAAK,MAAM,KAAK,QAAQ,KAAK;AAClD,aAAO,OACJ,OAAO,CAAC,MAAM,YAAY,EAAE,KAAK,IAAI,SAAS,EAC9C,IAAI,CAAC,MAAM,eAAe,IAAI,MAAM,CAAC,CAAC;AAAA,IAC3C;AAAA,EACF;AACF;AAEA,SAAS,YAAY,GAAmB;AAEtC,SAAO,KAAK,IAAI,IAAI,KAAK;AAC3B;AAEA,SAAS,eAAe,YAAoB,MAAc,GAA+B;AACvF,QAAM,UAAU,YAAY,EAAE,KAAK;AACnC,QAAM,WACJ,UAAU,IAAI,aAAa,UAAU,IAAI,SAAS,UAAU,IAAI,WAAW;AAC7E,SAAO,YAAY;AAAA,IACjB;AAAA,IACA;AAAA,IACA,SAAS,EAAE;AAAA,IACX,OAAO,GAAG,EAAE,SAAS,IAAI,EAAE,SAAS,WAAW,QAAQ,QAAQ,CAAC,CAAC;AAAA,IACjE,WAAW,EAAE;AAAA,IACb;AAAA,IACA,YAAY;AAAA,IACZ,eAAe,EAAE,WACb,CAAC,EAAE,MAAM,YAAY,KAAK,mBAAmB,SAAS,EAAE,SAAS,CAAC,IAClE,CAAC;AAAA;AAAA;AAAA;AAAA;AAAA,IAKL,oBAAoB;AAAA,IACpB,UAAU,EAAE,YAAY,EAAE,WAAW,WAAW,EAAE,WAAW,UAAU,QAAQ;AAAA,EACjF,CAAC;AACH;AAUO,SAAS,kCACd,OAAwC,CAAC,GACL;AACpC,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,OAAO,QAAQ,KAAK,SAAS,QAAQ,CAAC,KAAK,QAAQ,KAAK,IAAI,OAAU;AAAA,IACpF,SAAS,GAAG,8BAA8B,YAAY,WAAW;AAAA,IACjE,MAAM,QAAQ,OAAO;AACnB,YAAM,SAAS,MAAM,wBAAwB,OAAO,KAAK,OAAO;AAChE,UAAI,CAAC,OAAO,WAAW;AACrB,eAAO;AAAA,UACL,YAAY;AAAA,YACV,YAAY;AAAA,YACZ;AAAA,YACA,OAAO;AAAA,YACP,WAAW,OAAO;AAAA,YAClB,UAAU;AAAA,YACV,YAAY;AAAA,YACZ,eAAe,CAAC;AAAA,YAChB,UAAU,EAAE,QAAQ,OAAO,MAAM;AAAA,UACnC,CAAC;AAAA,QACH;AAAA,MACF;AACA,YAAM,MAAwB,CAAC;AAC/B,iBAAW,KAAK,OAAO,UAAU;AAG/B,YAAI,EAAE,WAAW,EAAE,SAAS,EAAG;AAC/B,YAAI;AAAA,UACF,YAAY;AAAA,YACV,YAAY;AAAA,YACZ;AAAA,YACA,SAAS,EAAE;AAAA,YACX,OAAO,EAAE,UACL,YAAY,EAAE,OAAO,cAAc,EAAE,KAAK,SAC1C,YAAY,EAAE,OAAO;AAAA,YACzB,WAAW,EAAE;AAAA,YACb,UAAU,aAAa,EAAE,QAAQ;AAAA,YACjC,YAAY;AAAA,YACZ,eAAe,CAAC,EAAE,MAAM,YAAY,KAAK,mBAAmB,SAAS,EAAE,SAAS,CAAC;AAAA,YACjF,UAAU;AAAA,cACR,SAAS,EAAE;AAAA,cACX,SAAS,EAAE;AAAA,cACX,UAAU,EAAE;AAAA,cACZ,UAAU,OAAO,WAAW;AAAA,YAC9B;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;;;AClZA,IAAM,mBAAqD,oBAAI,IAAyB;AAAA,EACtF;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAUM,SAAS,kBAAkB,SAAkC;AAClE,SAAO,QAAQ,cAAc,KAAK,CAAC,QAAQ,iBAAiB,IAAI,IAAI,IAAI,CAAC;AAC3E;AAKO,SAAS,eAAe,SAAkC;AAC/D,SAAO,QAAQ,uBAAuB;AACxC;AAqBO,SAAS,qBACd,UACA,UAAU,SACqB;AAC/B,QAAM,QAAQ,SAAS,OAAO,cAAc;AAC5C,MAAI,MAAM,SAAS,GAAG;AACpB,UAAM,IAAI;AAAA,MACR,GAAG,OAAO,sJACoE,MACzE,IAAI,CAAC,MAAM,EAAE,UAAU,EACvB,KAAK,IAAI,CAAC;AAAA,IACjB;AAAA,EACF;AACA,SAAO;AACT;","names":[]}
+{"version":3,"sources":["../../src/analyst/adapters.ts","../../src/analyst/steer-firewall.ts"],"sourcesContent":["/**\n * Adapter factories — lift each existing agent-eval primitive into the\n * Analyst contract without re-implementing it.\n *\n * Five primitives, five factories. Each one:\n *   - Builds an Analyst with a stable id (caller chooses; defaults\n *     given), a sensible default `inputKind`, a version derived from\n *     the wrapped primitive's version + an adapter revision, and an\n *     `analyze()` that calls the primitive and lifts its output to\n *     AnalystFinding[] using `makeFinding()`.\n *   - Maps severities: the existing `Severity` ('critical' | 'major' |\n *     'minor' | 'info') projects onto AnalystSeverity ('critical' |\n *     'high' | 'medium' | 'low' | 'info'); 'major' → 'high', 'minor' →\n *     'medium'. Domain analysts that want finer-grained mapping override.\n *\n * Adapters never own state. Calling the same factory twice with the\n * same primitive instance is safe.\n */\n\nimport type { AxAIService } from '@ax-llm/ax'\nimport type {\n  Finding as LayerFinding,\n  Severity as LayerSeverity,\n  MultiLayerVerifier,\n  VerifyOptions,\n} from '../multi-layer-verifier'\nimport { RunCritic, type RunTrace } from '../run-critic'\nimport {\n  runSemanticConceptJudge,\n  SEMANTIC_CONCEPT_JUDGE_VERSION,\n  type SemanticConceptJudgeInput,\n  type SemanticConceptJudgeOptions,\n} from '../semantic-concept-judge'\nimport { type AnalyzeTracesOptions, analyzeTraces } from '../trace-analyst/analyst'\nimport type { TraceAnalysisStore } from '../trace-analyst/store'\nimport type { JudgeFn, JudgeInput, JudgeScore, TCloud } from '../types'\nimport type { Analyst, AnalystFinding, AnalystSeverity } from './types'\nimport { makeFinding } from './types'\n\nconst ADAPTER_REV = '1'\n\n// ── Severity bridges ───────────────────────────────────────────────\n\nexport function liftSeverity(s: LayerSeverity): AnalystSeverity {\n  switch (s) {\n    case 'critical':\n      return 'critical'\n    case 'major':\n      return 'high'\n    case 'minor':\n      return 'medium'\n    case 'info':\n      return 'info'\n  }\n}\n\n// ── 1. analyzeTraces → Analyst ─────────────────────────────────────\n\nexport interface TraceAnalystAdapterOpts {\n  id?: string\n  area?: string\n  /** The natural-language question(s) put to the analyst. One finding per question. */\n  questions: string[]\n  /** Caller-provided AxAI service — same one trace-analyst.ts expects. */\n  ai: AxAIService\n  model?: string\n  /** Forwarded to analyzeTraces. */\n  extra?: Omit<AnalyzeTracesOptions, 'source' | 'ai' | 'model'>\n}\n\n/**\n * @deprecated Prefer `createTraceAnalystKind` + one of the failure /\n * improvement kinds from `./kinds`. This adapter wraps the legacy\n * `analyzeTraces` flow whose output is `findings:string[]` — every\n * bullet gets flat-defaulted severity `medium` / confidence `0.6`,\n * which loses the per-finding grading kinds provide via Ax structured\n * output + Zod validation. Kept for one minor while consumers migrate.\n */\nexport function createTraceAnalystAdapter(\n  opts: TraceAnalystAdapterOpts,\n): Analyst<TraceAnalysisStore> {\n  const id = opts.id ?? 'trace-analyst'\n  const area = opts.area ?? 'agent-reasoning'\n  return {\n    id,\n    description:\n      'Runs the agent-eval trace analyst over an OTLP trace store and lifts its bulleted findings.',\n    inputKind: 'trace-store',\n    cost: { kind: 'llm', models: opts.model ? [opts.model] : undefined },\n    version: `trace-analyst-${ADAPTER_REV}`,\n    async analyze(store, ctx) {\n      const out: AnalystFinding[] = []\n      for (const question of opts.questions) {\n        if (ctx.signal?.aborted) break\n        const result = await analyzeTraces(\n          { question },\n          { source: store, ai: opts.ai, model: opts.model, ...opts.extra },\n        )\n        const subject = ctx.tags?.subject ?? question.slice(0, 60)\n        // The responder produces a list of bullet strings. Each becomes\n        // one finding; the prose answer is attached as rationale on the\n        // first (so renderers that show only top-N still get context).\n        if (result.findings.length === 0) {\n          out.push(\n            makeFinding({\n              analyst_id: id,\n              area,\n              subject,\n              claim: result.answer.slice(0, 200),\n              rationale: result.answer,\n              severity: 'info',\n              confidence: 0.5,\n              evidence_refs: [],\n              metadata: {\n                actor_prompt_version: result.actorPromptVersion,\n                turns: result.turnCount,\n              },\n            }),\n          )\n          continue\n        }\n        result.findings.forEach((claim, i) => {\n          out.push(\n            makeFinding({\n              analyst_id: id,\n              area,\n              subject,\n              claim,\n              rationale: i === 0 ? result.answer : undefined,\n              severity: 'medium',\n              confidence: 0.6,\n              evidence_refs: [],\n              metadata: { question, turns: result.turnCount, finding_index: i },\n            }),\n          )\n        })\n      }\n      return out\n    },\n  }\n}\n\n// ── 2. MultiLayerVerifier → Analyst ─────────────────────────────────\n\nexport interface VerifierAdapterOpts<Env> {\n  id?: string\n  area?: string\n  verifier: MultiLayerVerifier<Env>\n  /**\n   * The verifier expects an `env` per run. Adapters take it from\n   * `AnalystRunInputs.custom[<id>]` via the registry's 'custom' routing.\n   */\n  options?: Omit<VerifyOptions<Env>, 'env'>\n}\n\nexport function createVerifierAdapter<Env>(opts: VerifierAdapterOpts<Env>): Analyst<Env> {\n  const id = opts.id ?? 'multi-layer-verifier'\n  const area = opts.area ?? 'verification'\n  return {\n    id,\n    description:\n      \"Runs a MultiLayerVerifier and lifts each layer's findings into the analyst envelope.\",\n    inputKind: 'custom',\n    cost: { kind: 'deterministic' },\n    version: `verifier-${ADAPTER_REV}`,\n    async analyze(env, ctx) {\n      const report = await opts.verifier.run({ env, ...opts.options })\n      const out: AnalystFinding[] = []\n      for (const layer of report.layers) {\n        for (const finding of layer.findings) {\n          out.push(liftLayerFinding(id, area, layer.layer, finding))\n        }\n        // Layer-level signal: a failed/error layer is itself a finding\n        // even if it didn't emit per-finding rows.\n        if (layer.status === 'fail' || layer.status === 'error' || layer.status === 'timeout') {\n          out.push(\n            makeFinding({\n              analyst_id: id,\n              area,\n              subject: layer.layer,\n              claim: `layer \"${layer.layer}\" ${layer.status}: ${layer.reason ?? 'no reason given'}`,\n              severity:\n                layer.status === 'error' ? 'high' : layer.status === 'timeout' ? 'medium' : 'high',\n              confidence: 1,\n              evidence_refs: [],\n              metadata: {\n                layer_status: layer.status,\n                duration_ms: layer.durationMs,\n                score: layer.score,\n                diagnostics: layer.diagnostics,\n              },\n            }),\n          )\n        }\n      }\n      ctx.log?.('verifier complete', {\n        layers: report.layers.length,\n        blended: report.blendedScore,\n        all_pass: report.allPass,\n      })\n      return out\n    },\n  }\n}\n\nfunction liftLayerFinding(\n  analyst_id: string,\n  area: string,\n  layer: string,\n  f: LayerFinding,\n): AnalystFinding {\n  return makeFinding({\n    analyst_id,\n    area,\n    subject: f.layer ?? layer,\n    claim: f.message,\n    severity: liftSeverity(f.severity),\n    confidence: 0.85,\n    evidence_refs: f.evidence\n      ? [{ kind: 'artifact', uri: 'inline:evidence', excerpt: f.evidence }]\n      : [],\n    metadata: f.detail,\n  })\n}\n\n// ── 3. RunCritic → Analyst ──────────────────────────────────────────\n\nexport interface RunCriticAdapterOpts {\n  id?: string\n  area?: string\n  critic?: RunCritic\n  /** Optional threshold below which a dimension is reported as a finding. Default 0.5. */\n  threshold?: number\n}\n\nexport function createRunCriticAdapter(opts: RunCriticAdapterOpts = {}): Analyst<RunTrace> {\n  const id = opts.id ?? 'run-critic'\n  const area = opts.area ?? 'run-quality'\n  const critic = opts.critic ?? new RunCritic()\n  const threshold = opts.threshold ?? 0.5\n  return {\n    id,\n    description:\n      'Scores a single run across success / grounding / drift / tool-quality and surfaces below-threshold dimensions.',\n    inputKind: 'custom',\n    cost: { kind: 'deterministic' },\n    version: `run-critic-${ADAPTER_REV}`,\n    async analyze(trace) {\n      const score = critic.scoreTrace(trace)\n      const out: AnalystFinding[] = []\n      const dims: Array<[keyof typeof score, AnalystSeverity, string]> = [\n        ['success', 'critical', 'run did not complete successfully'],\n        ['goalProgress', 'high', 'goal progress is low'],\n        ['repoGroundedness', 'high', 'output is poorly grounded in the repository'],\n        ['toolUseQuality', 'medium', 'tool use quality is low'],\n        ['patchQuality', 'medium', 'no real patch/edit evidence'],\n        ['testReality', 'high', 'no real test/build evidence'],\n        ['finalGate', 'critical', 'final gate is blocking'],\n      ]\n      for (const [dim, sev, msg] of dims) {\n        const value = score[dim] as number\n        if (typeof value === 'number' && value < threshold) {\n          out.push(\n            makeFinding({\n              analyst_id: id,\n              area,\n              subject: dim,\n              claim: msg,\n              rationale: `${dim}=${value.toFixed(2)} below threshold ${threshold}`,\n              severity: sev,\n              confidence: 1,\n              evidence_refs: [],\n              metadata: { dimension: dim, value, threshold, run_id: trace.run.runId },\n            }),\n          )\n        }\n      }\n      // Drift penalty is high → surface as a finding (inverse threshold).\n      if (score.driftPenalty > 1 - threshold) {\n        out.push(\n          makeFinding({\n            analyst_id: id,\n            area,\n            subject: 'drift',\n            claim: 'agent output drifted from repository signal',\n            rationale: `driftPenalty=${score.driftPenalty.toFixed(2)}`,\n            severity: 'medium',\n            confidence: 0.9,\n            evidence_refs: [],\n            metadata: { drift_penalty: score.driftPenalty, notes: score.notes },\n          }),\n        )\n      }\n      return out\n    },\n  }\n}\n\n// ── 4. JudgeFn → Analyst ────────────────────────────────────────────\n\nexport interface JudgeAdapterOpts {\n  id?: string\n  area?: string\n  judge: JudgeFn\n  /** TCloud handle the JudgeFn calls. */\n  tcloud: TCloud\n  /** Optional cost classification — most judges call an LLM. */\n  cost?: Analyst['cost']\n  /** Optional threshold below which a JudgeScore becomes a finding. Default 6 (on 0-10 scale). */\n  threshold?: number\n}\n\nexport function createJudgeAdapter(opts: JudgeAdapterOpts): Analyst<JudgeInput> {\n  const id = opts.id ?? 'judge'\n  const area = opts.area ?? 'judge'\n  const threshold = opts.threshold ?? 6\n  return {\n    id,\n    description:\n      'Wraps an agent-eval JudgeFn into an analyst; below-threshold dimensions surface as findings.',\n    inputKind: 'judge-input',\n    cost: opts.cost ?? { kind: 'llm' },\n    version: `judge-${ADAPTER_REV}`,\n    async analyze(input) {\n      const scores = await opts.judge(opts.tcloud, input)\n      return scores\n        .filter((s) => normalize10(s.score) < threshold)\n        .map((s) => liftJudgeScore(id, area, s))\n    },\n  }\n}\n\nfunction normalize10(s: number): number {\n  // JudgeScore convention is 0-10 but some judges emit 0-1. Coerce to 0-10.\n  return s <= 1 ? s * 10 : s\n}\n\nfunction liftJudgeScore(analyst_id: string, area: string, s: JudgeScore): AnalystFinding {\n  const score10 = normalize10(s.score)\n  const severity: AnalystSeverity =\n    score10 < 3 ? 'critical' : score10 < 5 ? 'high' : score10 < 7 ? 'medium' : 'low'\n  return makeFinding({\n    analyst_id,\n    area,\n    subject: s.dimension,\n    claim: `${s.judgeName}/${s.dimension} scored ${score10.toFixed(1)}/10`,\n    rationale: s.reasoning,\n    severity,\n    confidence: 0.8,\n    evidence_refs: s.evidence\n      ? [{ kind: 'artifact', uri: 'inline:evidence', excerpt: s.evidence }]\n      : [],\n    // Provenance: this finding IS a judge verdict (an acceptance score), not an\n    // observation of behavior. The steer firewall (assertNoJudgeVerdict) rejects\n    // it from steering — even when it cites an artifact above — because letting a\n    // verdict steer the next attempt is the held-out judge leaking into the loop.\n    derived_from_judge: true,\n    metadata: { judge_name: s.judgeName, dimension: s.dimension, score_10: score10 },\n  })\n}\n\n// ── 5. SemanticConceptJudge → Analyst ──────────────────────────────\n\nexport interface SemanticConceptJudgeAdapterOpts {\n  id?: string\n  area?: string\n  options?: SemanticConceptJudgeOptions\n}\n\nexport function createSemanticConceptJudgeAdapter(\n  opts: SemanticConceptJudgeAdapterOpts = {},\n): Analyst<SemanticConceptJudgeInput> {\n  const id = opts.id ?? 'semantic-concept-judge'\n  const area = opts.area ?? 'concept-coverage'\n  return {\n    id,\n    description:\n      'Runs the semantic-concept judge and surfaces missing / weak concepts as findings.',\n    inputKind: 'custom',\n    cost: { kind: 'llm', models: opts.options?.model ? [opts.options.model] : undefined },\n    version: `${SEMANTIC_CONCEPT_JUDGE_VERSION}-adapter-${ADAPTER_REV}`,\n    async analyze(input) {\n      const result = await runSemanticConceptJudge(input, opts.options)\n      if (!result.available) {\n        return [\n          makeFinding({\n            analyst_id: id,\n            area,\n            claim: 'semantic-concept judge unavailable',\n            rationale: result.error,\n            severity: 'info',\n            confidence: 1,\n            evidence_refs: [],\n            metadata: { reason: result.error },\n          }),\n        ]\n      }\n      const out: AnalystFinding[] = []\n      for (const f of result.findings) {\n        // Only surface gaps: missing concepts or low scores. Concepts at\n        // 7+/10 with present=true are not findings — they're successes.\n        if (f.present && f.score >= 7) continue\n        out.push(\n          makeFinding({\n            analyst_id: id,\n            area,\n            subject: f.concept,\n            claim: f.present\n              ? `concept \"${f.concept}\" is weak (${f.score}/10)`\n              : `concept \"${f.concept}\" is missing`,\n            rationale: f.evidence,\n            severity: liftSeverity(f.severity),\n            confidence: 0.85,\n            evidence_refs: [{ kind: 'artifact', uri: 'inline:evidence', excerpt: f.evidence }],\n            metadata: {\n              concept: f.concept,\n              present: f.present,\n              score_10: f.score,\n              cost_usd: result.costUsd ?? undefined,\n            },\n          }),\n        )\n      }\n      return out\n    },\n  }\n}\n","// The realness-oracle firewall (docs/learning-flywheel.md, \"The steer is f(trace)\").\n//\n// A realness/authenticity signal has TWO legitimate roles that must stay\n// separated by a firewall:\n//   (a) anchor judge J — write-only: scores the chosen output, gates promotion,\n//       NEVER seen by the worker/optimizer mid-run (else the loop games it).\n//   (b) steer f(trace) — an analyst observes the agent's OWN behavior in the\n//       trace (\"imported a stub\", \"used a non-crypto PRNG where encryption was\n//       required\") and steers the next attempt. Legitimate, because it is derived\n//       from OBSERVABLE BEHAVIOR, not from J's held-out verdict.\n//\n// The correct discriminator is PROVENANCE, not evidence presence. A judge verdict\n// lifted into a finding (createJudgeAdapter → liftJudgeScore) is a verdict even\n// when it cites an artifact; an evidence-less trace-analyst bullet is an\n// observation even though it cites nothing. So the firewall keys on\n// `AnalystFinding.derived_from_judge` (set at the judge lift site), NOT on whether\n// evidence_refs is populated. The instant a verdict steers the next attempt it is\n// a back-channel for J and the loop Goodharts realness exactly as it would\n// Goodhart pass-rate.\n\nimport type { AnalystFinding, EvidenceRef } from './types'\n\n/** Evidence grounded in the agent's OWN execution: OTLP trace elements\n *  (`span`/`event`) or the artifact it produced (`artifact`). */\nconst OBSERVABLE_KINDS: ReadonlySet<EvidenceRef['kind']> = new Set<EvidenceRef['kind']>([\n  'span',\n  'event',\n  'artifact',\n])\n\n/** DESCRIPTIVE predicate: does the finding cite at least one observable\n *  (span/event/artifact) evidence ref. Useful for ranking evidence quality or\n *  rendering — it is NOT the steer gate. Evidence presence is the WRONG\n *  discriminator for steering: a legitimate trace-analyst observation may cite\n *  nothing (it would be wrongly rejected), and a judge verdict may cite an\n *  artifact (it would be wrongly admitted). Use `assertNoJudgeVerdict` to gate\n *  steering; use this only where \"is this grounded in observable evidence\" is the\n *  literal question. */\nexport function isTraceObservable(finding: AnalystFinding): boolean {\n  return finding.evidence_refs.some((ref) => OBSERVABLE_KINDS.has(ref.kind))\n}\n\n/** True iff the finding is a JUDGE VERDICT (an acceptance score lifted into a\n *  finding), identified by provenance set at the lift site — independent of\n *  whatever evidence it cites. */\nexport function isJudgeVerdict(finding: AnalystFinding): boolean {\n  return finding.derived_from_judge === true\n}\n\n/**\n * THE steer firewall. Fail-loud guard for any path that admits analyst findings\n * as STEERING input (the `f(trace)` role): rejects — naming the offenders — any\n * finding whose provenance is a judge verdict, rather than let `J` leak into the\n * loop. Returns the findings unchanged for chaining.\n *\n * Call this at the chokepoint where a detector that ALSO scores/gates has its\n * findings turned into a steer (the judge-and-steer dual-role case). It keys on\n * provenance, so it correctly admits evidence-less trace-analyst observations and\n * correctly rejects an artifact-citing judge verdict — the cases an evidence\n * check gets backwards.\n *\n * It is necessary, not sufficient: it stops PROVENANCE-tagged verdicts. A judge\n * whose output is laundered through a hand-built finding with no provenance flag\n * is out of its reach — provenance must be honestly set at every judge→finding\n * lift (today: createJudgeAdapter). That is why the integrity rule lives at the\n * lift site, and why ProposeContext.judgeScores?: never is the complementary\n * compile-time tripwire on the obvious direct channel.\n */\nexport function assertNoJudgeVerdict(\n  findings: ReadonlyArray<AnalystFinding>,\n  context = 'steer',\n): ReadonlyArray<AnalystFinding> {\n  const leaks = findings.filter(isJudgeVerdict)\n  if (leaks.length > 0) {\n    throw new Error(\n      `${context}: a judge verdict cannot be admitted as steering input — that is the ` +\n        `held-out judge leaking into the loop. Offending judge-derived findings: [${leaks\n          .map((f) => f.finding_id)\n          .join(', ')}]. Steering consumes observations of behavior, never acceptance verdicts.`,\n    )\n  }\n  return findings\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuCA,IAAM,cAAc;AAIb,SAAS,aAAa,GAAmC;AAC9D,UAAQ,GAAG;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,EACX;AACF;AAwBO,SAAS,0BACd,MAC6B;AAC7B,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,OAAO,QAAQ,KAAK,QAAQ,CAAC,KAAK,KAAK,IAAI,OAAU;AAAA,IACnE,SAAS,iBAAiB,WAAW;AAAA,IACrC,MAAM,QAAQ,OAAO,KAAK;AACxB,YAAM,MAAwB,CAAC;AAC/B,iBAAW,YAAY,KAAK,WAAW;AACrC,YAAI,IAAI,QAAQ,QAAS;AACzB,cAAM,SAAS,MAAM;AAAA,UACnB,EAAE,SAAS;AAAA,UACX,EAAE,QAAQ,OAAO,IAAI,KAAK,IAAI,OAAO,KAAK,OAAO,GAAG,KAAK,MAAM;AAAA,QACjE;AACA,cAAM,UAAU,IAAI,MAAM,WAAW,SAAS,MAAM,GAAG,EAAE;AAIzD,YAAI,OAAO,SAAS,WAAW,GAAG;AAChC,cAAI;AAAA,YACF,YAAY;AAAA,cACV,YAAY;AAAA,cACZ;AAAA,cACA;AAAA,cACA,OAAO,OAAO,OAAO,MAAM,GAAG,GAAG;AAAA,cACjC,WAAW,OAAO;AAAA,cAClB,UAAU;AAAA,cACV,YAAY;AAAA,cACZ,eAAe,CAAC;AAAA,cAChB,UAAU;AAAA,gBACR,sBAAsB,OAAO;AAAA,gBAC7B,OAAO,OAAO;AAAA,cAChB;AAAA,YACF,CAAC;AAAA,UACH;AACA;AAAA,QACF;AACA,eAAO,SAAS,QAAQ,CAAC,OAAO,MAAM;AACpC,cAAI;AAAA,YACF,YAAY;AAAA,cACV,YAAY;AAAA,cACZ;AAAA,cACA;AAAA,cACA;AAAA,cACA,WAAW,MAAM,IAAI,OAAO,SAAS;AAAA,cACrC,UAAU;AAAA,cACV,YAAY;AAAA,cACZ,eAAe,CAAC;AAAA,cAChB,UAAU,EAAE,UAAU,OAAO,OAAO,WAAW,eAAe,EAAE;AAAA,YAClE,CAAC;AAAA,UACH;AAAA,QACF,CAAC;AAAA,MACH;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAeO,SAAS,sBAA2B,MAA8C;AACvF,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,gBAAgB;AAAA,IAC9B,SAAS,YAAY,WAAW;AAAA,IAChC,MAAM,QAAQ,KAAK,KAAK;AACtB,YAAM,SAAS,MAAM,KAAK,SAAS,IAAI,EAAE,KAAK,GAAG,KAAK,QAAQ,CAAC;AAC/D,YAAM,MAAwB,CAAC;AAC/B,iBAAW,SAAS,OAAO,QAAQ;AACjC,mBAAW,WAAW,MAAM,UAAU;AACpC,cAAI,KAAK,iBAAiB,IAAI,MAAM,MAAM,OAAO,OAAO,CAAC;AAAA,QAC3D;AAGA,YAAI,MAAM,WAAW,UAAU,MAAM,WAAW,WAAW,MAAM,WAAW,WAAW;AACrF,cAAI;AAAA,YACF,YAAY;AAAA,cACV,YAAY;AAAA,cACZ;AAAA,cACA,SAAS,MAAM;AAAA,cACf,OAAO,UAAU,MAAM,KAAK,KAAK,MAAM,MAAM,KAAK,MAAM,UAAU,iBAAiB;AAAA,cACnF,UACE,MAAM,WAAW,UAAU,SAAS,MAAM,WAAW,YAAY,WAAW;AAAA,cAC9E,YAAY;AAAA,cACZ,eAAe,CAAC;AAAA,cAChB,UAAU;AAAA,gBACR,cAAc,MAAM;AAAA,gBACpB,aAAa,MAAM;AAAA,gBACnB,OAAO,MAAM;AAAA,gBACb,aAAa,MAAM;AAAA,cACrB;AAAA,YACF,CAAC;AAAA,UACH;AAAA,QACF;AAAA,MACF;AACA,UAAI,MAAM,qBAAqB;AAAA,QAC7B,QAAQ,OAAO,OAAO;AAAA,QACtB,SAAS,OAAO;AAAA,QAChB,UAAU,OAAO;AAAA,MACnB,CAAC;AACD,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAEA,SAAS,iBACP,YACA,MACA,OACA,GACgB;AAChB,SAAO,YAAY;AAAA,IACjB;AAAA,IACA;AAAA,IACA,SAAS,EAAE,SAAS;AAAA,IACpB,OAAO,EAAE;AAAA,IACT,UAAU,aAAa,EAAE,QAAQ;AAAA,IACjC,YAAY;AAAA,IACZ,eAAe,EAAE,WACb,CAAC,EAAE,MAAM,YAAY,KAAK,mBAAmB,SAAS,EAAE,SAAS,CAAC,IAClE,CAAC;AAAA,IACL,UAAU,EAAE;AAAA,EACd,CAAC;AACH;AAYO,SAAS,uBAAuB,OAA6B,CAAC,GAAsB;AACzF,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,SAAS,KAAK,UAAU,IAAI,UAAU;AAC5C,QAAM,YAAY,KAAK,aAAa;AACpC,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,gBAAgB;AAAA,IAC9B,SAAS,cAAc,WAAW;AAAA,IAClC,MAAM,QAAQ,OAAO;AACnB,YAAM,QAAQ,OAAO,WAAW,KAAK;AACrC,YAAM,MAAwB,CAAC;AAC/B,YAAM,OAA6D;AAAA,QACjE,CAAC,WAAW,YAAY,mCAAmC;AAAA,QAC3D,CAAC,gBAAgB,QAAQ,sBAAsB;AAAA,QAC/C,CAAC,oBAAoB,QAAQ,6CAA6C;AAAA,QAC1E,CAAC,kBAAkB,UAAU,yBAAyB;AAAA,QACtD,CAAC,gBAAgB,UAAU,6BAA6B;AAAA,QACxD,CAAC,eAAe,QAAQ,6BAA6B;AAAA,QACrD,CAAC,aAAa,YAAY,wBAAwB;AAAA,MACpD;AACA,iBAAW,CAAC,KAAK,KAAK,GAAG,KAAK,MAAM;AAClC,cAAM,QAAQ,MAAM,GAAG;AACvB,YAAI,OAAO,UAAU,YAAY,QAAQ,WAAW;AAClD,cAAI;AAAA,YACF,YAAY;AAAA,cACV,YAAY;AAAA,cACZ;AAAA,cACA,SAAS;AAAA,cACT,OAAO;AAAA,cACP,WAAW,GAAG,GAAG,IAAI,MAAM,QAAQ,CAAC,CAAC,oBAAoB,SAAS;AAAA,cAClE,UAAU;AAAA,cACV,YAAY;AAAA,cACZ,eAAe,CAAC;AAAA,cAChB,UAAU,EAAE,WAAW,KAAK,OAAO,WAAW,QAAQ,MAAM,IAAI,MAAM;AAAA,YACxE,CAAC;AAAA,UACH;AAAA,QACF;AAAA,MACF;AAEA,UAAI,MAAM,eAAe,IAAI,WAAW;AACtC,YAAI;AAAA,UACF,YAAY;AAAA,YACV,YAAY;AAAA,YACZ;AAAA,YACA,SAAS;AAAA,YACT,OAAO;AAAA,YACP,WAAW,gBAAgB,MAAM,aAAa,QAAQ,CAAC,CAAC;AAAA,YACxD,UAAU;AAAA,YACV,YAAY;AAAA,YACZ,eAAe,CAAC;AAAA,YAChB,UAAU,EAAE,eAAe,MAAM,cAAc,OAAO,MAAM,MAAM;AAAA,UACpE,CAAC;AAAA,QACH;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAgBO,SAAS,mBAAmB,MAA6C;AAC9E,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,YAAY,KAAK,aAAa;AACpC,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,KAAK,QAAQ,EAAE,MAAM,MAAM;AAAA,IACjC,SAAS,SAAS,WAAW;AAAA,IAC7B,MAAM,QAAQ,OAAO;AACnB,YAAM,SAAS,MAAM,KAAK,MAAM,KAAK,QAAQ,KAAK;AAClD,aAAO,OACJ,OAAO,CAAC,MAAM,YAAY,EAAE,KAAK,IAAI,SAAS,EAC9C,IAAI,CAAC,MAAM,eAAe,IAAI,MAAM,CAAC,CAAC;AAAA,IAC3C;AAAA,EACF;AACF;AAEA,SAAS,YAAY,GAAmB;AAEtC,SAAO,KAAK,IAAI,IAAI,KAAK;AAC3B;AAEA,SAAS,eAAe,YAAoB,MAAc,GAA+B;AACvF,QAAM,UAAU,YAAY,EAAE,KAAK;AACnC,QAAM,WACJ,UAAU,IAAI,aAAa,UAAU,IAAI,SAAS,UAAU,IAAI,WAAW;AAC7E,SAAO,YAAY;AAAA,IACjB;AAAA,IACA;AAAA,IACA,SAAS,EAAE;AAAA,IACX,OAAO,GAAG,EAAE,SAAS,IAAI,EAAE,SAAS,WAAW,QAAQ,QAAQ,CAAC,CAAC;AAAA,IACjE,WAAW,EAAE;AAAA,IACb;AAAA,IACA,YAAY;AAAA,IACZ,eAAe,EAAE,WACb,CAAC,EAAE,MAAM,YAAY,KAAK,mBAAmB,SAAS,EAAE,SAAS,CAAC,IAClE,CAAC;AAAA;AAAA;AAAA;AAAA;AAAA,IAKL,oBAAoB;AAAA,IACpB,UAAU,EAAE,YAAY,EAAE,WAAW,WAAW,EAAE,WAAW,UAAU,QAAQ;AAAA,EACjF,CAAC;AACH;AAUO,SAAS,kCACd,OAAwC,CAAC,GACL;AACpC,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,OAAO,QAAQ,KAAK,SAAS,QAAQ,CAAC,KAAK,QAAQ,KAAK,IAAI,OAAU;AAAA,IACpF,SAAS,GAAG,8BAA8B,YAAY,WAAW;AAAA,IACjE,MAAM,QAAQ,OAAO;AACnB,YAAM,SAAS,MAAM,wBAAwB,OAAO,KAAK,OAAO;AAChE,UAAI,CAAC,OAAO,WAAW;AACrB,eAAO;AAAA,UACL,YAAY;AAAA,YACV,YAAY;AAAA,YACZ;AAAA,YACA,OAAO;AAAA,YACP,WAAW,OAAO;AAAA,YAClB,UAAU;AAAA,YACV,YAAY;AAAA,YACZ,eAAe,CAAC;AAAA,YAChB,UAAU,EAAE,QAAQ,OAAO,MAAM;AAAA,UACnC,CAAC;AAAA,QACH;AAAA,MACF;AACA,YAAM,MAAwB,CAAC;AAC/B,iBAAW,KAAK,OAAO,UAAU;AAG/B,YAAI,EAAE,WAAW,EAAE,SAAS,EAAG;AAC/B,YAAI;AAAA,UACF,YAAY;AAAA,YACV,YAAY;AAAA,YACZ;AAAA,YACA,SAAS,EAAE;AAAA,YACX,OAAO,EAAE,UACL,YAAY,EAAE,OAAO,cAAc,EAAE,KAAK,SAC1C,YAAY,EAAE,OAAO;AAAA,YACzB,WAAW,EAAE;AAAA,YACb,UAAU,aAAa,EAAE,QAAQ;AAAA,YACjC,YAAY;AAAA,YACZ,eAAe,CAAC,EAAE,MAAM,YAAY,KAAK,mBAAmB,SAAS,EAAE,SAAS,CAAC;AAAA,YACjF,UAAU;AAAA,cACR,SAAS,EAAE;AAAA,cACX,SAAS,EAAE;AAAA,cACX,UAAU,EAAE;AAAA,cACZ,UAAU,OAAO,WAAW;AAAA,YAC9B;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;;;AClZA,IAAM,mBAAqD,oBAAI,IAAyB;AAAA,EACtF;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAUM,SAAS,kBAAkB,SAAkC;AAClE,SAAO,QAAQ,cAAc,KAAK,CAAC,QAAQ,iBAAiB,IAAI,IAAI,IAAI,CAAC;AAC3E;AAKO,SAAS,eAAe,SAAkC;AAC/D,SAAO,QAAQ,uBAAuB;AACxC;AAqBO,SAAS,qBACd,UACA,UAAU,SACqB;AAC/B,QAAM,QAAQ,SAAS,OAAO,cAAc;AAC5C,MAAI,MAAM,SAAS,GAAG;AACpB,UAAM,IAAI;AAAA,MACR,GAAG,OAAO,sJACoE,MACzE,IAAI,CAAC,MAAM,EAAE,UAAU,EACvB,KAAK,IAAI,CAAC;AAAA,IACjB;AAAA,EACF;AACA,SAAO;AACT;","names":[]}

package/dist/analyze-runs-DwCEkpO_.d.ts

@@ -0,0 +1,81 @@
+import { A as AnalystRegistry } from './default-registry-zoGHUQEH.js';
+import { a as DatasetScenario } from './dataset-BbGkaN2I.js';
+import { R as RunRecord } from './run-record-e7vj1uZQ.js';
+import { I as InsightReport } from './insight-report-BBwvOh6x.js';
+
+/**
+ * # `analyzeRuns()` — turn a set of agent runs into an actionable decision packet.
+ *
+ * Wires the substrate's statistical, calibration, clustering, Pareto, and
+ * release-confidence primitives into one `InsightReport`. Two top-level
+ * entry points use this function:
+ *
+ *   - `selfImprove()` calls it on the campaign output to attach a packet
+ *     to every run.
+ *   - Consumers with observed `RunRecord[]` (production traces, gold
+ *     corpora, approve/reject tables) call it directly via `analyzeRuns()`
+ *     for analysis without a closed loop.
+ *
+ * Every section is opt-in based on what the input data supports — the
+ * function never invents signal. If runs carry no judge scores, `judges`
+ * is empty. If there's no baseline/candidate split, `lift` is undefined.
+ * If no `analyst` is wired, `failureClusters` is undefined.
+ *
+ * The `recommendations` array is the human-readable layer; everything
+ * else is the evidence backing each recommendation.
+ */
+
+interface AnalyzeRunsOptions {
+    /** The runs to analyze. */
+    runs: RunRecord[];
+    /** Which split to score against when reading composite from RunOutcome.
+     *  Default: holdout when ANY run has a `holdoutScore`, else search. */
+    split?: 'search' | 'holdout' | 'auto';
+    /** Pairwise analysis configuration. When both `baselineCandidateId` and
+     *  `candidateCandidateId` are present, lift is computed on paired
+     *  (experimentId, seed) tuples shared between the two sides. */
+    baselineCandidateId?: string;
+    candidateCandidateId?: string;
+    /** Canary scenarios — checked against every run's raw output for
+     *  holdout contamination. */
+    canaryScenarios?: DatasetScenario[];
+    /** Analyst registry for failure clustering. When omitted, the
+     *  `failureClusters` section is left undefined. */
+    analyst?: AnalystRegistry;
+    /** Downstream outcome metric per run (e.g. engagement rate, approval
+     *  rate, downstream pass rate). When present, the report includes
+     *  `outcomeCorrelation` + a simple linear reward model fit. */
+    outcomeSignal?: {
+        metric: string;
+        valueByRunId: Record<string, number>;
+    };
+    /** Multi-rater feedback for inter-rater agreement. Each entry is one
+     *  rater's score for one run. Two or more raters → kappa + disagreement
+     *  triage list. */
+    raterScores?: Array<{
+        runId: string;
+        rater: string;
+        score: number;
+    }>;
+    /** Number of histogram bins for distributional summaries. Default 12. */
+    histogramBins?: number;
+    /** Decision threshold — the smallest composite lift the caller cares
+     *  about. Used by the recommendations engine to call ship vs hold.
+     *  Default 0.02. */
+    decisionThreshold?: number;
+    /** Optional prior-period runs. When set, the report includes
+     *  `priorPeriodComparison` with per-metric Welch-CI deltas and
+     *  recommendations fire on statistically significant regressions.
+     *  The two windows do NOT have to share scenarios — the comparison
+     *  is two-sample unpaired (the substrate's `lift` field uses paired
+     *  bootstrap on shared (experimentId, seed) tuples; this is the
+     *  shape for "this week vs last week" rather than "candidate vs
+     *  baseline within a campaign"). */
+    baselineRuns?: RunRecord[];
+    /** Human-readable label for the baseline window, e.g. "vs prior 7
+     *  days", "vs v3.1 release". Surfaces in recommendations + UI. */
+    baselineLabel?: string;
+}
+declare function analyzeRuns(opts: AnalyzeRunsOptions): Promise<InsightReport>;
+
+export { type AnalyzeRunsOptions as A, analyzeRuns as a };

package/dist/belief-state/index.d.ts

@@ -1,12 +1,12 @@
 import { c as CalibrationReport } from '../calibration-Cpr3WaX3.js';
 import { O as OffPolicyEstimate, a as OffPolicyOptions, b as OffPolicyTrajectory } from '../off-policy-DiwuKKg7.js';
-import { d as CodeAgentSessionSource, a as CodeAgentSessionIntakeOptions, c as CodeAgentSessionMetrics, C as CodeAgentSessionDiagnostic } from '../code-agent-session-BRXmavYv.js';
-import { R as RunRecord, a as RunSplitTag } from '../run-record-De9VarXR.js';
+import { d as CodeAgentSessionSource, a as CodeAgentSessionIntakeOptions, c as CodeAgentSessionMetrics, C as CodeAgentSessionDiagnostic } from '../code-agent-session-BO8nCnv3.js';
+import { R as RunRecord, a as RunSplitTag } from '../run-record-e7vj1uZQ.js';
 import { T as TraceStore } from '../store-CKUAgsJz.js';
-import { R as RuntimeTrajectoryRecord, P as ProjectRuntimeTrajectoryEvidenceOptions, a as RuntimeTrajectoryEvidenceProjection } from '../runtime-trajectory-BLRiaifm.js';
+import { R as RuntimeTrajectoryRecord, P as ProjectRuntimeTrajectoryEvidenceOptions, a as RuntimeTrajectoryEvidenceProjection } from '../runtime-trajectory-BDgfGZSr.js';
 import '../schema-m0gsnbt3.js';
 import '../outcome-store-rnXLEqSn.js';
-import '../errors-Dwqw-T_m.js';
+import '../errors-CzMUYo7b.js';
 
 declare const BELIEF_DECISION_KINDS: readonly ["continue", "verify", "ask", "retry", "stop", "memory-write", "memory-read", "tool-select", "skill-select", "workflow-select", "surface-promote"];
 type BeliefDecisionKind = (typeof BELIEF_DECISION_KINDS)[number];

package/dist/belief-state/index.js

@@ -17,7 +17,7 @@ import {
 } from "../chunk-4DIJWVUT.js";
 import {
   confidenceInterval
-} from "../chunk-IDVBLYCY.js";
+} from "../chunk-LMZQ2Z4U.js";
 import {
   ValidationError
 } from "../chunk-3BFEG2F6.js";

package/dist/benchmarks/index.d.ts

@@ -1,4 +1,4 @@
-export { B as BENCHMARK_SPLIT_SEED, a as BenchmarkAdapter, b as BenchmarkDatasetItem, c as BenchmarkEvaluation, d as deterministicSplit, e as routing } from '../index-DE3RXAXD.js';
-import '../run-record-De9VarXR.js';
-import '../errors-Dwqw-T_m.js';
+export { B as BENCHMARK_SPLIT_SEED, a as BenchmarkAdapter, b as BenchmarkDatasetItem, c as BenchmarkEvaluation, d as deterministicSplit, e as routing } from '../index-Bx3gZ8xl.js';
+import '../run-record-e7vj1uZQ.js';
+import '../errors-CzMUYo7b.js';
 import '../schema-m0gsnbt3.js';