@tangle-network/agent-eval 0.119.0 → 0.120.0

package/CHANGELOG.md

@@ -4,6 +4,29 @@ All notable changes to `@tangle-network/agent-eval` and its sibling `agent-eval-
 
 ---
 
+## [0.120.0] — 2026-07-14 — exact measured provenance
+
+### Changed
+
+- Give each campaign a canonical split digest over the full scenario payload and replicate count, and require the exported benchmark to identify that exact heldout design.
+- Emit self-addressed `tangle.loop-provenance` records from one shared loop-result translator, and derive prompt or code diffs from the exact measured surfaces instead of accepting caller-authored text.
+- Retain one strict hash per scenario so consumers can verify the tested split without persisting customer task payloads.
+
+### Breaking
+
+- Require the exact baseline surface and bind portable evidence to every search, winner, heldout, and neutralized campaign; the full decision; the reconstructed receipt ledger; duration; power analysis; and validated candidate history.
+- Require every scored row to link to a successful model receipt; reject contradictory provenance, duplicate scenarios, broken cell identity or pairing, failed judges, and invalid cost, latency, or token measurements before products can publish an improvement proposal.
+- Stop optimization when a proposer returns no candidates, reject empty or non-contiguous generation records, and refuse a shipped no-op.
+- Remove the duplicate worker-record collector and caller-authored diff hook; settled receipts and exact surfaces are now the only evidence path.
+- Remove `@tangle-network/agent-eval/primeintellect`, which generated a legacy single-turn Verifiers package with unsafe substring scoring. Use `@tangle-network/agent-runtime/primeintellect` for task packaging, real runtime execution, and full trace import.
+- Verify the same npm archive implementation used by the release workflow.
+
+## [0.119.1] — 2026-07-14 — portable improvement evidence
+
+### Added
+
+- `measuredComparisonFromSelfImproveResult()` converts paired held-out quality, cost, latency, uncertainty, power, decision, and provenance into the shared `AgentImprovementMeasuredComparison` contract.
+
 ## [0.119.0] — 2026-07-14 — chained, metered trace analysis
 
 ### Added
@@ -121,8 +144,7 @@ All notable changes to `@tangle-network/agent-eval` and its sibling `agent-eval-
 - `runOptimization({ promoteTopK })` accepts only `1`; multiple concurrent incumbents were never represented by the optimizer state and now fail before dispatch.
 - `ScoredSurfaceOutcome` requires `split: 'search'` and the actual `generation` that measured the surface.
 - `llmPolicyEditProposer()` requires explicit raw-score objectives and `PolicyEditFindingInput` rows whose source is either an exact measured surface-generation pair or an explicitly global finding.
-- `LoopProvenanceRecord.schema` is now `tangle.loop-provenance.v3`.
-  Consumers of v2 records must migrate to the v3 baseline score, parent chain, coverage, eligibility, and exact surface fields; no compatibility shim is provided.
+- Loop provenance preserves complete candidate measurement lineage: baseline score, parent chain, coverage, eligibility, and exact surface fields.
 
 ## [0.115.3] — 2026-07-12 — fail-closed structured output parsing
 
@@ -186,7 +208,7 @@ Existing consumers do not need to update unless they want the new statistic.
 - `CodeSurface` is now a finalized, content-addressed code candidate.
   `gitWorktreeAdapter.finalize()` records exact base/candidate commits, the final tree, and the SHA-256 + byte length of the raw binary Git patch; `surfaceHash` and `surfaceContentHash` no longer use filesystem paths.
   Binary-patch generation runs against an isolated bare repository with fixed diff options, config, attributes, compression, and locale, so ambient repository, global, system, or environment settings cannot change the digest for identical trees.
-- `LoopProvenanceRecord.schema` is now `tangle.loop-provenance.v2`, distinguishing records that use content-addressed code-surface hashes from v1 records whose code hashes included mutable paths.
+- Code-surface provenance now uses content-addressed hashes instead of mutable paths.
 - `resolveWorktreePath()` now verifies the candidate before returning its checkout.
   Dirty or ignored files, moved refs, missing objects, wrong trees, raw byte or executable-mode mismatches, external symlinks, submodules, and patch-byte mismatches fail instead of being evaluated under stale identity.
   Raw file hashing bypasses Git clean/smudge filters so repository configuration cannot hide different executable bytes.

package/README.md

@@ -95,31 +95,6 @@ const traceRuns = fromOtelSpans({ spans: yourOtelSpans })
 await analyzeRuns({ runs: [...runs, ...traceRuns], raterScores })
 ```
 
-### 4. Export to PrimeIntellect
-
-Use this when you want PrimeIntellect Verifiers to run or train against tasks produced by Tangle evals, runtime runs, or knowledge runs.
-
-```ts
-import {
-  primeIntellectRowsFromRunRecords,
-  writePrimeIntellectEnvironmentPackage,
-} from '@tangle-network/agent-eval/primeintellect'
-
-const rows = primeIntellectRowsFromRunRecords({
-  records, // RunRecord[] with scenarioId
-  scenarios, // prompts plus answer or requiredSubstrings
-})
-
-await writePrimeIntellectEnvironmentPackage('prime-envs/refund-policy', {
-  name: 'refund-policy',
-  rows,
-  runRecords: records,
-})
-```
-
-The generated package contains `load_environment()`, `data/dataset.jsonl`, `data/run_records.jsonl`, and a manifest.
-Run it with `uv run vf-eval refund-policy` or upload it with `prime env push`.
-
 ---
 
 ## Core concepts
@@ -149,7 +124,6 @@ Each example: `README.md` + a single `index.ts` runnable via `pnpm tsx`. Prints
 |---|---|
 | `…/contract` | **The headline, frozen surface — new code starts here.** `defineAgentEval`, `selfImprove`, `analyzeRuns`, `runEval`, `runCampaign`, `runImprovementLoop`, `diffRuns`; intake adapters (`fromFeedbackTable`, `fromOtelSpans`); proposers (`gepaProposer`, `evolutionaryProposer`); gates (`defaultProductionGate`, `heldOutGate`, `paretoSignificanceGate`, `neutralizationGate`, `composeGate`); the deployment-outcome store; storage; and the five core types `Scenario` / `Dispatch` / `JudgeConfig` / `SurfaceProposer` / `Gate`. |
 | `…/hosted` | `createHostedClient` / `hostedClientFromEnv` + the wire types to ship eval-run events + trace spans to a hosted orchestrator (ours or your own implementation of the spec) |
-| `…/primeintellect` | Export validated `RunRecord`s plus scenario prompts into a PrimeIntellect Verifiers package: Python `load_environment()`, `data/dataset.jsonl`, `data/run_records.jsonl`, and export manifest |
 | `…/adapters/otel` | `createOtelBridge` — forwards OpenTelemetry-shape spans into the hosted-tier ingest, no `@opentelemetry/*` dependency |
 | `…/adapters/langchain` | Wrap any LangChain `Runnable` as a `Dispatch` (or `JudgeConfig`), no `@langchain/core` peer dep |
 | `…/adapters/http` | `httpDispatch` + `runDispatchServer` — run a campaign's worker on another machine (multi-region, remote worker execution) |

package/dist/analyst/index.d.ts

@@ -1,5 +1,5 @@
 import { TCloud } from '@tangle-network/tcloud';
-import { AxAIService, AxFunction } from '@ax-llm/ax';
+import { AxAIArgs, AxAIService, AxFunction } from '@ax-llm/ax';
 import { z } from 'zod';
 
 /**
@@ -1706,6 +1706,8 @@ interface AnalystCost {
     est_usd_per_run?: number;
     /** Models the analyst expects to use (informational). */
     models?: string[];
+    /** Maximum post-cancellation wait for provider usage. Model analysts default to 5 seconds. */
+    settlement_timeout_ms?: number;
 }
 interface AnalystRequirements {
     /** Min number of shots / samples the analyst needs to produce signal. */
@@ -1729,7 +1731,7 @@ interface AnalystContext {
     runId: string;
     /** Stable correlation id so logs from a single registry.run() share a tag. */
     correlationId: string;
-    /** Wall-clock deadline (epoch ms). Analysts SHOULD honor for graceful cancel. */
+    /** Enforced wall-clock deadline (epoch ms). */
     deadlineMs?: number;
     /** Per-analyst USD budget. Analysts MAY check before issuing LLM calls. */
     budgetUsd?: number;
@@ -1965,11 +1967,13 @@ interface CreateAnalystAiConfig {
     apiKey: string;
     /** OpenAI-compatible base URL — e.g. `https://router.tangle.tools/v1` or a
      *  cli-bridge loopback. */
-    baseUrl: string;
-    /** Model id forwarded to the analyst actor + responder. */
+    baseUrl?: string;
+    /** Additional headers required by the gateway, such as tenant or execution policy. */
+    headers?: Record<string, string>;
+    /** Model id forwarded to analyst calls. */
     model: string;
     /** Ax provider name. Defaults to the OpenAI-compatible client. */
-    provider?: 'openai' | 'anthropic';
+    provider?: AxAIArgs<unknown>['name'];
 }
 /**
  * Construct the `AxAIService` an analyst kind calls through
@@ -2145,8 +2149,8 @@ declare function parseCanonicalRawFinding(row: unknown, log?: (msg: string, fiel
  * Analyst-kind factory — the typed way to define trace analysts.
  *
  * A "kind" is a specialized analyst whose actor prompt, tool subset,
- * and Ax recursion config target one failure-mode lens (failure-mode
- * classification, knowledge gap discovery, knowledge poisoning, recursive
+ * and bounded Ax subqueries target one failure-mode lens (failure-mode
+ * classification, knowledge gap discovery, knowledge poisoning,
  * self-improvement, ...). Kinds emit findings in the typed
  * `CanonicalRawAnalystFinding` shape via a JSON-array Ax output; the factory
  * validates each row with Zod and lifts it into `AnalystFinding[]`.
@@ -2157,12 +2161,11 @@ declare function parseCanonicalRawFinding(row: unknown, log?: (msg: string, fiel
  *   - Each kind picks a narrow tool subset from `ANALYST_TOOL_GROUPS`.
  *     A kind that never needs full-trace dumps can drop `viewTrace` /
  *     `viewSpans` and stay cheap.
- *   - Each kind declares its recursion + parallelism budget. Discovery-
- *     heavy kinds (failure-mode) get higher `maxDepth`; lens kinds
- *     (poisoning) usually stay at 0 since they have a tighter brief.
+ *   - Each kind declares its subquery + parallelism budget. Discovery-heavy
+ *     kinds can fan out more bounded semantic questions than narrow lenses.
  *
  * Optimizer hook: kinds may declare `goldens` — labeled examples used
- * by `AxMiPRO` / `AxBootstrapFewShot` / `AxGEPA` to fit the actor
+ * by `AxBootstrapFewShot` / `AxGEPA` to fit the actor
  * description programmatically. Stored on the kind, not the registry,
  * because the right metric is kind-specific.
  */
@@ -2182,20 +2185,18 @@ interface TraceAnalystKindSpec {
     version: string;
     /** Actor system prompt. Must instruct the LLM to emit `findings` per the schema. */
     actorDescription: string;
-    /** Responder system prompt; falls back to a minimal "format the findings" instruction. */
-    responderDescription?: string;
     /** Tool functions the actor may call. Pick narrow subsets via `ANALYST_TOOL_GROUPS`. */
     buildTools: (store: TraceAnalysisStore) => AxFunction[];
-    /** Recursion budget. `maxDepth: 0` disables subagents. */
-    recursion?: {
-        maxDepth: number;
-        maxParallelSubagents?: number;
+    /** Bounded semantic subqueries. `maxCalls: 0` disables model fan-out. */
+    subqueries?: {
+        maxCalls: number;
+        maxParallel?: number;
     };
     /** Actor turn cap. Default 12. */
     maxTurns?: number;
     /** Runtime char cap. Default 6000. */
     maxRuntimeChars?: number;
-    /** Maximum output tokens for every actor, responder, and recursive model call. Default 4096. */
+    /** Maximum output tokens for every actor and subquery model call. Default 4096. */
     maxOutputTokens?: number;
     /** Cost classification surfaced in `registry.list()` and budget enforcement. */
     cost: AnalystCost;
@@ -2348,7 +2349,7 @@ interface RegistryRunOpts {
     skip?: string[];
     /** Budget policy — totalUsd + optional weights/allocator. Falls back to options.defaultBudget. */
     budget?: BudgetPolicy;
-    /** Wall-clock cap. Analysts SHOULD honor `ctx.deadlineMs`. */
+    /** Active-work cap for the complete registry run. Model receipt settlement may follow. */
     timeoutMs?: number;
     /** Abort signal — forwarded into every analyst's context. */
     signal?: AbortSignal;
@@ -2684,19 +2685,14 @@ declare function diffFindings(previous: PersistedFinding[], current: PersistedFi
  * findings. The actor's job is *taxonomy + evidence*, not fix-design —
  * that's the improvement-analyst's job.
  *
- * Recursion is deep (`maxDepth: 3`) because real failure-mode
- * discovery is genuinely tree-shaped: the actor splits the dataset
- * into candidate clusters, each cluster spawns a focused investigator
- * that drills into representative traces, and a deeply-recursed
- * investigator may itself split a confounded mode into two sub-modes.
- * Each level fans out 4-way, so the analyst can investigate up to
- * ~16 leaf clusters before hitting the depth ceiling.
+ * Eight bounded model subqueries let the actor compare candidate
+ * clusters in parallel after it has loaded representative evidence.
  */
 
 declare const FAILURE_MODE_KIND_SPEC: TraceAnalystKindSpec;
 
 /**
- * Improvement analyst — actionable, recursive self-improvement findings.
+ * Improvement analyst — actionable self-improvement findings.
  *
  * Brief: read findings from upstream analysts (failure-mode,
  * knowledge-gap, knowledge-poisoning) AND the trace dataset itself,
@@ -2706,14 +2702,11 @@ declare const FAILURE_MODE_KIND_SPEC: TraceAnalystKindSpec;
  * finding is one proposed edit with the locus, the diff, and the
  * expected effect.
  *
- * This is the recursive-self-improvement loop's last mile: the prior
+ * This is the self-improvement loop's last mile: the prior
  * kinds describe *what's wrong*; this kind describes *what to change*.
  *
- * Recursion is deep (`maxDepth: 3`) because real improvement proposals
- * are competitive: for each failure-mode there are usually 2-3 viable
- * fix directions (tighten prompt vs add tool vs adjust scaffolding),
- * and the actor should explore each with a focused subagent before
- * picking the highest-leverage one to recommend.
+ * Eight bounded model subqueries let the actor compare competing fix
+ * directions over the same cited evidence before recommending one.
  */
 
 declare const IMPROVEMENT_KIND_SPEC: TraceAnalystKindSpec;
@@ -2740,8 +2733,8 @@ declare const IMPROVEMENT_KIND_SPEC: TraceAnalystKindSpec;
  * knowledge-gap names the *information* whose absence (or staleness)
  * caused the break. One failure-mode often maps to several gaps.
  *
- * Recursion (`maxDepth: 2`) is enough to fan out one subagent per
- * candidate gap-source layer; each subagent runs a focused detection.
+ * Five bounded model subqueries let the actor compare candidate gaps
+ * across source layers after it has loaded the relevant excerpts.
  */
 
 declare const KNOWLEDGE_GAP_KIND_SPEC: TraceAnalystKindSpec;
@@ -2760,10 +2753,8 @@ declare const KNOWLEDGE_GAP_KIND_SPEC: TraceAnalystKindSpec;
  * surface as questions / self-correction; poisonings surface as
  * confident-but-wrong actions that downstream evidence contradicts.
  *
- * Recursion is moderate (`maxDepth: 2`) because each candidate
- * poisoning typically needs two sub-investigations: one to confirm
- * the agent acted on the false belief, one to confirm the belief
- * itself is actually false in ground truth.
+ * Eight bounded model subqueries let the actor independently assess
+ * the action and contradiction excerpts for candidate poisonings.
  */
 
 declare const KNOWLEDGE_POISONING_KIND_SPEC: TraceAnalystKindSpec;

package/dist/analyst/index.js

@@ -9,13 +9,13 @@ import {
   diffFindings,
   emitSkillUsageFindings,
   runSemanticConceptJudge
-} from "../chunk-LMJ2TGWJ.js";
+} from "../chunk-JM2SKQMS.js";
 import {
   behavioralAnalyst,
   buildDefaultAnalystRegistry,
   createChatClient,
   deriveEfficiencyFindings
-} from "../chunk-GERDAIAL.js";
+} from "../chunk-QWMPPZ3X.js";
 import "../chunk-HHWE3POT.js";
 import "../chunk-3YYRZDON.js";
 import {
@@ -72,13 +72,13 @@ import {
   validatePolicyEdit,
   validatePolicyEditCandidateRecord,
   validateUsageSettlementTimeout
-} from "../chunk-7A4LIMMY.js";
+} from "../chunk-S5TT5R3L.js";
 import "../chunk-NJC7U437.js";
 import {
   CostLedger
 } from "../chunk-JHOJHHU7.js";
 import "../chunk-VI2UW6B6.js";
-import "../chunk-6QIM2EAP.js";
+import "../chunk-PXD6ZFNY.js";
 import "../chunk-PC4UYEBM.js";
 import "../chunk-XJYR7XFV.js";
 import "../chunk-VSMTAMNK.js";
@@ -264,7 +264,11 @@ function createSemanticConceptJudgeAdapter(opts = {}) {
     id,
     description: "Runs the semantic-concept judge and surfaces missing / weak concepts as findings.",
     inputKind: "custom",
-    cost: { kind: "llm", models: opts.options?.model ? [opts.options.model] : void 0 },
+    cost: {
+      kind: "llm",
+      models: opts.options?.model ? [opts.options.model] : void 0,
+      settlement_timeout_ms: settlementTimeoutMs
+    },
     version: `${SEMANTIC_CONCEPT_JUDGE_VERSION}-adapter-${ADAPTER_REV}`,
     async analyze(input, ctx) {
       const costLedger = new CostLedger(ctx.budgetUsd);

package/dist/analyst/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/analyst/adapters.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 { CostLedger } from '../cost-ledger'\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  type SemanticConceptJudgeResult,\n} from '../semantic-concept-judge'\nimport type { JudgeFn, JudgeInput, JudgeScore, TCloud } from '../types'\nimport type { Analyst, AnalystFinding, AnalystSeverity } from './types'\nimport { makeFinding } from './types'\nimport { settleUsageReceiptFromCostLedger, validateUsageSettlementTimeout } from './usage-receipt'\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. 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}: ${layer.reason ?? 'no reason given'}`,\n              severity:\n                layer.status === 'error' ? 'high' : layer.status === 'timeout' ? 'medium' : 'high',\n              confidence: 1,\n              evidence_refs: [],\n              metadata: {\n                layer_status: layer.status,\n                duration_ms: layer.durationMs,\n                score: layer.score,\n                diagnostics: layer.diagnostics,\n              },\n            }),\n          )\n        }\n      }\n      ctx.log?.('verifier complete', {\n        layers: report.layers.length,\n        blended: report.blendedScore,\n        all_pass: report.allPass,\n      })\n      return out\n    },\n  }\n}\n\nfunction liftLayerFinding(\n  analyst_id: string,\n  area: string,\n  layer: string,\n  f: LayerFinding,\n): AnalystFinding {\n  return makeFinding({\n    analyst_id,\n    area,\n    subject: f.layer ?? layer,\n    claim: f.message,\n    severity: liftSeverity(f.severity),\n    confidence: 0.85,\n    evidence_refs: f.evidence\n      ? [{ kind: 'artifact', uri: 'inline:evidence', excerpt: f.evidence }]\n      : [],\n    metadata: f.detail,\n  })\n}\n\n// ── 2. RunCritic → Analyst ──────────────────────────────────────────\n\nexport interface RunCriticAdapterOpts {\n  id?: string\n  area?: string\n  critic?: RunCritic\n  /** Optional threshold below which a dimension is reported as a finding. Default 0.5. */\n  threshold?: number\n}\n\nexport function createRunCriticAdapter(opts: RunCriticAdapterOpts = {}): Analyst<RunTrace> {\n  const id = opts.id ?? 'run-critic'\n  const area = opts.area ?? 'run-quality'\n  const critic = opts.critic ?? new RunCritic()\n  const threshold = opts.threshold ?? 0.5\n  return {\n    id,\n    description:\n      'Scores a single run across success / grounding / drift / tool-quality and surfaces below-threshold dimensions.',\n    inputKind: 'custom',\n    cost: { kind: 'deterministic' },\n    version: `run-critic-${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// ── 3. 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 ${score10.toFixed(1)}/10`,\n    rationale: s.reasoning,\n    severity,\n    confidence: 0.8,\n    evidence_refs: s.evidence\n      ? [{ kind: 'artifact', uri: 'inline:evidence', excerpt: s.evidence }]\n      : [],\n    // Provenance: this finding IS a judge verdict (an acceptance score), not an\n    // observation of behavior. The steer firewall (assertNoJudgeVerdict) rejects\n    // it from steering — even when it cites an artifact above — because letting a\n    // verdict steer the next attempt is the held-out judge leaking into the loop.\n    derived_from_judge: true,\n    metadata: { judge_name: s.judgeName, dimension: s.dimension, score_10: score10 },\n  })\n}\n\n// ── 4. SemanticConceptJudge → Analyst ──────────────────────────────\n\nexport interface SemanticConceptJudgeAdapterOpts {\n  id?: string\n  area?: string\n  /** Registry context owns cancellation and the per-analyst cost ledger. */\n  options?: Omit<SemanticConceptJudgeOptions, 'costLedger' | 'signal'>\n  /** Maximum post-cancellation wait for a provider receipt. Default 5 seconds. */\n  settlementTimeoutMs?: number\n}\n\nexport function createSemanticConceptJudgeAdapter(\n  opts: SemanticConceptJudgeAdapterOpts = {},\n): Analyst<SemanticConceptJudgeInput> {\n  const id = opts.id ?? 'semantic-concept-judge'\n  const area = opts.area ?? 'concept-coverage'\n  const settlementTimeoutMs = validateUsageSettlementTimeout(opts.settlementTimeoutMs)\n  return {\n    id,\n    description:\n      'Runs the semantic-concept judge and surfaces missing / weak concepts as findings.',\n    inputKind: 'custom',\n    cost: { kind: 'llm', models: opts.options?.model ? [opts.options.model] : undefined },\n    version: `${SEMANTIC_CONCEPT_JUDGE_VERSION}-adapter-${ADAPTER_REV}`,\n    async analyze(input, ctx) {\n      const costLedger = new CostLedger(ctx.budgetUsd)\n      let result: SemanticConceptJudgeResult\n      try {\n        result = await runSemanticConceptJudge(input, {\n          ...opts.options,\n          costLedger,\n          signal: ctx.signal,\n        })\n      } finally {\n        const usage = await settleUsageReceiptFromCostLedger(costLedger, {\n          channel: 'judge',\n          timeoutMs: settlementTimeoutMs,\n        })\n        if (!usage.settled) {\n          ctx.log?.('semantic-concept judge provider settlement timed out', {\n            pending_calls: usage.pendingCalls,\n            timeout_ms: settlementTimeoutMs,\n          })\n        }\n        ctx.recordUsage?.(usage.receipt)\n      }\n      if (!result.available) {\n        return [\n          makeFinding({\n            analyst_id: id,\n            area,\n            claim: 'semantic-concept judge unavailable',\n            rationale: result.error,\n            severity: 'info',\n            confidence: 1,\n            evidence_refs: [],\n            metadata: { reason: result.error },\n          }),\n        ]\n      }\n      const out: AnalystFinding[] = []\n      for (const f of result.findings) {\n        // Only surface gaps: missing concepts or low scores. Concepts at\n        // 7+/10 with present=true are not findings — they're successes.\n        if (f.present && f.score >= 7) continue\n        out.push(\n          makeFinding({\n            analyst_id: id,\n            area,\n            subject: f.concept,\n            claim: f.present\n              ? `concept \"${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            },\n          }),\n        )\n      }\n      return out\n    },\n  }\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;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;AAaO,SAAS,kCACd,OAAwC,CAAC,GACL;AACpC,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,sBAAsB,+BAA+B,KAAK,mBAAmB;AACnF,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,KAAK;AACxB,YAAM,aAAa,IAAI,WAAW,IAAI,SAAS;AAC/C,UAAI;AACJ,UAAI;AACF,iBAAS,MAAM,wBAAwB,OAAO;AAAA,UAC5C,GAAG,KAAK;AAAA,UACR;AAAA,UACA,QAAQ,IAAI;AAAA,QACd,CAAC;AAAA,MACH,UAAE;AACA,cAAM,QAAQ,MAAM,iCAAiC,YAAY;AAAA,UAC/D,SAAS;AAAA,UACT,WAAW;AAAA,QACb,CAAC;AACD,YAAI,CAAC,MAAM,SAAS;AAClB,cAAI,MAAM,wDAAwD;AAAA,YAChE,eAAe,MAAM;AAAA,YACrB,YAAY;AAAA,UACd,CAAC;AAAA,QACH;AACA,YAAI,cAAc,MAAM,OAAO;AAAA,MACjC;AACA,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,YACd;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;","names":[]}
+{"version":3,"sources":["../../src/analyst/adapters.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 { CostLedger } from '../cost-ledger'\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  type SemanticConceptJudgeResult,\n} from '../semantic-concept-judge'\nimport type { JudgeFn, JudgeInput, JudgeScore, TCloud } from '../types'\nimport type { Analyst, AnalystFinding, AnalystSeverity } from './types'\nimport { makeFinding } from './types'\nimport { settleUsageReceiptFromCostLedger, validateUsageSettlementTimeout } from './usage-receipt'\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. 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}: ${layer.reason ?? 'no reason given'}`,\n              severity:\n                layer.status === 'error' ? 'high' : layer.status === 'timeout' ? 'medium' : 'high',\n              confidence: 1,\n              evidence_refs: [],\n              metadata: {\n                layer_status: layer.status,\n                duration_ms: layer.durationMs,\n                score: layer.score,\n                diagnostics: layer.diagnostics,\n              },\n            }),\n          )\n        }\n      }\n      ctx.log?.('verifier complete', {\n        layers: report.layers.length,\n        blended: report.blendedScore,\n        all_pass: report.allPass,\n      })\n      return out\n    },\n  }\n}\n\nfunction liftLayerFinding(\n  analyst_id: string,\n  area: string,\n  layer: string,\n  f: LayerFinding,\n): AnalystFinding {\n  return makeFinding({\n    analyst_id,\n    area,\n    subject: f.layer ?? layer,\n    claim: f.message,\n    severity: liftSeverity(f.severity),\n    confidence: 0.85,\n    evidence_refs: f.evidence\n      ? [{ kind: 'artifact', uri: 'inline:evidence', excerpt: f.evidence }]\n      : [],\n    metadata: f.detail,\n  })\n}\n\n// ── 2. RunCritic → Analyst ──────────────────────────────────────────\n\nexport interface RunCriticAdapterOpts {\n  id?: string\n  area?: string\n  critic?: RunCritic\n  /** Optional threshold below which a dimension is reported as a finding. Default 0.5. */\n  threshold?: number\n}\n\nexport function createRunCriticAdapter(opts: RunCriticAdapterOpts = {}): Analyst<RunTrace> {\n  const id = opts.id ?? 'run-critic'\n  const area = opts.area ?? 'run-quality'\n  const critic = opts.critic ?? new RunCritic()\n  const threshold = opts.threshold ?? 0.5\n  return {\n    id,\n    description:\n      'Scores a single run across success / grounding / drift / tool-quality and surfaces below-threshold dimensions.',\n    inputKind: 'custom',\n    cost: { kind: 'deterministic' },\n    version: `run-critic-${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// ── 3. 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 ${score10.toFixed(1)}/10`,\n    rationale: s.reasoning,\n    severity,\n    confidence: 0.8,\n    evidence_refs: s.evidence\n      ? [{ kind: 'artifact', uri: 'inline:evidence', excerpt: s.evidence }]\n      : [],\n    // Provenance: this finding IS a judge verdict (an acceptance score), not an\n    // observation of behavior. The steer firewall (assertNoJudgeVerdict) rejects\n    // it from steering — even when it cites an artifact above — because letting a\n    // verdict steer the next attempt is the held-out judge leaking into the loop.\n    derived_from_judge: true,\n    metadata: { judge_name: s.judgeName, dimension: s.dimension, score_10: score10 },\n  })\n}\n\n// ── 4. SemanticConceptJudge → Analyst ──────────────────────────────\n\nexport interface SemanticConceptJudgeAdapterOpts {\n  id?: string\n  area?: string\n  /** Registry context owns cancellation and the per-analyst cost ledger. */\n  options?: Omit<SemanticConceptJudgeOptions, 'costLedger' | 'signal'>\n  /** Maximum post-cancellation wait for a provider receipt. Default 5 seconds. */\n  settlementTimeoutMs?: number\n}\n\nexport function createSemanticConceptJudgeAdapter(\n  opts: SemanticConceptJudgeAdapterOpts = {},\n): Analyst<SemanticConceptJudgeInput> {\n  const id = opts.id ?? 'semantic-concept-judge'\n  const area = opts.area ?? 'concept-coverage'\n  const settlementTimeoutMs = validateUsageSettlementTimeout(opts.settlementTimeoutMs)\n  return {\n    id,\n    description:\n      'Runs the semantic-concept judge and surfaces missing / weak concepts as findings.',\n    inputKind: 'custom',\n    cost: {\n      kind: 'llm',\n      models: opts.options?.model ? [opts.options.model] : undefined,\n      settlement_timeout_ms: settlementTimeoutMs,\n    },\n    version: `${SEMANTIC_CONCEPT_JUDGE_VERSION}-adapter-${ADAPTER_REV}`,\n    async analyze(input, ctx) {\n      const costLedger = new CostLedger(ctx.budgetUsd)\n      let result: SemanticConceptJudgeResult\n      try {\n        result = await runSemanticConceptJudge(input, {\n          ...opts.options,\n          costLedger,\n          signal: ctx.signal,\n        })\n      } finally {\n        const usage = await settleUsageReceiptFromCostLedger(costLedger, {\n          channel: 'judge',\n          timeoutMs: settlementTimeoutMs,\n        })\n        if (!usage.settled) {\n          ctx.log?.('semantic-concept judge provider settlement timed out', {\n            pending_calls: usage.pendingCalls,\n            timeout_ms: settlementTimeoutMs,\n          })\n        }\n        ctx.recordUsage?.(usage.receipt)\n      }\n      if (!result.available) {\n        return [\n          makeFinding({\n            analyst_id: id,\n            area,\n            claim: 'semantic-concept judge unavailable',\n            rationale: result.error,\n            severity: 'info',\n            confidence: 1,\n            evidence_refs: [],\n            metadata: { reason: result.error },\n          }),\n        ]\n      }\n      const out: AnalystFinding[] = []\n      for (const f of result.findings) {\n        // Only surface gaps: missing concepts or low scores. Concepts at\n        // 7+/10 with present=true are not findings — they're successes.\n        if (f.present && f.score >= 7) continue\n        out.push(\n          makeFinding({\n            analyst_id: id,\n            area,\n            subject: f.concept,\n            claim: f.present\n              ? `concept \"${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            },\n          }),\n        )\n      }\n      return out\n    },\n  }\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;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;AAaO,SAAS,kCACd,OAAwC,CAAC,GACL;AACpC,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,sBAAsB,+BAA+B,KAAK,mBAAmB;AACnF,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM;AAAA,MACJ,MAAM;AAAA,MACN,QAAQ,KAAK,SAAS,QAAQ,CAAC,KAAK,QAAQ,KAAK,IAAI;AAAA,MACrD,uBAAuB;AAAA,IACzB;AAAA,IACA,SAAS,GAAG,8BAA8B,YAAY,WAAW;AAAA,IACjE,MAAM,QAAQ,OAAO,KAAK;AACxB,YAAM,aAAa,IAAI,WAAW,IAAI,SAAS;AAC/C,UAAI;AACJ,UAAI;AACF,iBAAS,MAAM,wBAAwB,OAAO;AAAA,UAC5C,GAAG,KAAK;AAAA,UACR;AAAA,UACA,QAAQ,IAAI;AAAA,QACd,CAAC;AAAA,MACH,UAAE;AACA,cAAM,QAAQ,MAAM,iCAAiC,YAAY;AAAA,UAC/D,SAAS;AAAA,UACT,WAAW;AAAA,QACb,CAAC;AACD,YAAI,CAAC,MAAM,SAAS;AAClB,cAAI,MAAM,wDAAwD;AAAA,YAChE,eAAe,MAAM;AAAA,YACrB,YAAY;AAAA,UACd,CAAC;AAAA,QACH;AACA,YAAI,cAAc,MAAM,OAAO;AAAA,MACjC;AACA,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,YACd;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;","names":[]}

package/dist/benchmarks/index.d.ts

@@ -356,6 +356,10 @@ interface Scenario {
     kind: string;
     tags?: string[];
 }
+/** Redacted identity of a complete scenario payload retained in campaign results. */
+interface CampaignScenarioIdentity extends Pick<Scenario, 'id' | 'kind'> {
+    scenarioDigest: `sha256:${string}`;
+}
 /** Context handed to every dispatch invocation. Scoped — every
  *  trace/span carries the cellId, every artifact write lands under the cell's
  *  artifact root, the cost meter accumulates per cell. */
@@ -564,7 +568,11 @@ interface CampaignAggregates {
 interface CampaignResult<TArtifact = unknown, TScenario extends Scenario = Scenario> {
     /** sha256(scenarios, judges, dispatch source ref, optimizer config, seed). Stable identity for reruns. */
     manifestHash: string;
+    /** Canonical identity of the exact scenario payloads and replicate count. */
+    splitDigest: `sha256:${string}`;
     seed: number;
+    /** Replicates designed for every scenario in this campaign. */
+    reps: number;
     startedAt: string;
     endedAt: string;
     durationMs: number;
@@ -578,11 +586,9 @@ interface CampaignResult<TArtifact = unknown, TScenario extends Scenario = Scena
     prUrl?: string;
     runDir: string;
     artifactsByPath: Record<string, string>;
-    /** Substrate strips the input scenarios to id+kind for the result manifest;
-     *  consumers needing full payload look it up via the original input. The
-     *  type parameter `TScenario` is propagated for downstream consumers that
-     *  want narrowed types when extending `CampaignResult`. */
-    scenarios: Array<Pick<TScenario, 'id' | 'kind'>>;
+    /** Redacted identities that let consumers verify the exact scenario payloads
+     *  without retaining customer task content in the result. */
+    scenarios: Array<CampaignScenarioIdentity & Pick<TScenario, 'id' | 'kind'>>;
 }
 
 /**

package/dist/benchmarks/index.js

@@ -16,20 +16,20 @@ import {
   routing_exports,
   runBenchmarkAdapter,
   summarizeBenchmarkCampaign
-} from "../chunk-D7AEXSM5.js";
-import "../chunk-F6YUH3L4.js";
-import "../chunk-RLCJQ6VZ.js";
-import "../chunk-PAHNGS65.js";
+} from "../chunk-PICTDURQ.js";
+import "../chunk-JN2FCO5W.js";
+import "../chunk-32BZXMSO.js";
+import "../chunk-XDIRG3TO.js";
 import "../chunk-3YYRZDON.js";
-import "../chunk-7A4LIMMY.js";
+import "../chunk-S5TT5R3L.js";
 import "../chunk-ARU2PZFM.js";
 import "../chunk-NJC7U437.js";
 import "../chunk-PJQFMIOX.js";
 import "../chunk-JHOJHHU7.js";
 import "../chunk-VI2UW6B6.js";
-import "../chunk-FIUFRXP3.js";
+import "../chunk-WW2A73HW.js";
 import "../chunk-GGE4NNQT.js";
-import "../chunk-6QIM2EAP.js";
+import "../chunk-PXD6ZFNY.js";
 import "../chunk-PC4UYEBM.js";
 import "../chunk-S3UZOQ5Y.js";
 import "../chunk-MA6HLL3S.js";