@tangle-network/agent-eval 0.102.0 → 0.103.0

package/dist/{insight-report-C02J3q4T.d.ts → insight-report-k0sRTzKg.d.ts}

@@ -1,4 +1,4 @@
-import { G as GainDistributionBin, P as ParetoFigureSpec } from './summary-report-C4uzRWh8.js';
+import { G as GainDistributionBin, P as ParetoFigureSpec } from './summary-report-C0nnxOD8.js';
 import { C as ContinuousAgreement } from './judge-calibration-0p2QcWNE.js';
 
 /**

package/dist/{kind-factory-OgqQSvLi.d.ts → kind-factory-D0nk7AKV.d.ts}

@@ -1,7 +1,7 @@
 import { AxAIService, AxFunction } from '@ax-llm/ax';
 import { T as TraceAnalysisStore } from './store-C1YxJDEK.js';
 import { z } from 'zod';
-import { g as AnalystCost, b as AnalystContext, a as Analyst } from './types-BEzCBMQD.js';
+import { g as AnalystCost, b as AnalystContext, a as Analyst } from './types-Dz9cKF0g.js';
 
 /**
  * Typed Ax output for analyst findings.

package/dist/meta-eval/index.d.ts

@@ -1,12 +1,12 @@
 export { C as CalibrationBin, a as CalibrationOptions, b as CalibrationPair, c as CalibrationReport, d as CorrelationResult, e as CorrelationStudyOptions, f as CorrelationStudyResult, E as EvalMetricSpec, O as OutcomePair, g as calibrationCurve, h as calibrationFromPairs, i as correlationStudy } from '../calibration-BPmzuVPk.js';
 export { D as DeploymentOutcome, F as FileSystemOutcomeStore, a as FileSystemOutcomeStoreOptions, I as InMemoryOutcomeStore, O as OutcomeFilter, b as OutcomeStore } from '../outcome-store-rnXLEqSn.js';
-export { R as RubricOutcomePair, a as RubricPredictiveValidityInput, b as RubricPredictiveValidityReport, c as RubricRanking, r as rubricPredictiveValidity } from '../rubric-predictive-validity-w7tun-q3.js';
+export { R as RubricOutcomePair, a as RubricPredictiveValidityInput, b as RubricPredictiveValidityReport, c as RubricRanking, r as rubricPredictiveValidity } from '../rubric-predictive-validity-0MdjTt8R.js';
 import { C as ContinuousAgreement, a as CalibrationResult, b as ContinuousCalibrationResult, c as CandidateScore, G as GoldenItem } from '../judge-calibration-0p2QcWNE.js';
 import { S as SeriesConvergenceOptions, a as SeriesConvergenceResult } from '../series-convergence-D5OWMBg6.js';
 import { C as CorpusAgreementReport } from '../statistics-xP-cWc5k.js';
 import '../store-BcFXE6LG.js';
 import '../schema-m0gsnbt3.js';
-import '../run-record-DEwidcqn.js';
+import '../run-record-MRdJ-Kq2.js';
 import '@tangle-network/agent-interface';
 import '../errors-CzMUYo7b.js';
 import '../types-C7DGg5ex.js';

package/dist/multishot/index.d.ts

@@ -1,7 +1,7 @@
-import { J as JudgeScore } from '../types-fWqEJm7h.js';
+import { J as JudgeScore } from '../types-DFI_Z-ZL.js';
 import { AgentProfile } from '@tangle-network/agent-interface';
 import { M as MatrixResult } from '../types-BUxNaJ8c.js';
-import '../run-record-DEwidcqn.js';
+import '../run-record-MRdJ-Kq2.js';
 import '../errors-CzMUYo7b.js';
 import '../schema-m0gsnbt3.js';
 import '../verdict-C9MlYujm.js';

package/dist/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "@tangle-network/agent-eval — wire protocol",
-    "version": "0.102.0",
+    "version": "0.103.0",
     "description": "HTTP and stdio RPC interface to agent-eval. The TypeScript runtime is the source of truth; this spec is the contract that cross-language clients (Python, Rust, Go) generate from.\n\nWire-protocol version: 1.0.0. Bumps on breaking changes to request/response schemas.",
     "contact": {
       "name": "Tangle Network",

package/dist/{policy-edit-Dccm9tyA.d.ts → policy-edit-BDQzzsBU.d.ts}

@@ -1,6 +1,6 @@
-import { A as AgentProfileCell, a as AgentProfileJson } from './run-record-DEwidcqn.js';
+import { A as AgentProfileCell, a as AgentProfileJson } from './run-record-MRdJ-Kq2.js';
 import { V as ValidationError } from './errors-CzMUYo7b.js';
-import { A as AnalystFinding, E as EvidenceRef } from './types-BEzCBMQD.js';
+import { A as AnalystFinding, E as EvidenceRef } from './types-Dz9cKF0g.js';
 
 type PolicyEditSchemaVersion = 'policy-edit/v1';
 declare const POLICY_EDIT_AXES: readonly ["carrier", "representation", "budget", "sampling", "output_contract", "tool_contract", "routing", "memory", "agent_profile", "deployment_target"];

package/dist/{pre-registration-BjGZf9YA.d.ts → pre-registration-mWG2w8d-.d.ts}

@@ -1,8 +1,9 @@
 import { A as AgentEvalError } from './errors-CzMUYo7b.js';
-import { R as RunRecord } from './run-record-DEwidcqn.js';
-import { C as ChatClient } from './types-BEzCBMQD.js';
-import { c as JudgeDimension, S as Scenario, a as JudgeConfig } from './types-fWqEJm7h.js';
+import { R as RunRecord } from './run-record-MRdJ-Kq2.js';
+import { C as ChatClient } from './types-Dz9cKF0g.js';
+import { c as JudgeDimension, S as Scenario, a as JudgeConfig } from './types-DFI_Z-ZL.js';
 import { TCloud } from '@tangle-network/tcloud';
+import { R as RawProviderSink } from './raw-provider-sink-C46HDghv.js';
 import { D as DefaultVerdict } from './verdict-C9MlYujm.js';
 
 /**
@@ -166,9 +167,10 @@ declare function containsAll(name: string, required: string[], options?: {
  *      actually fulfils the requirement. A hallucinated artifact fails here;
  *      an absent one already failed stage 1.
  *
- * `completionRate` is satisfied / total. Quality dimensions are meaningless
- * on an incomplete task — callers gate on `fullyComplete` / `completionRate`
- * before scoring quality.
+ * `completionRate` is satisfied / MEASURABLE requirements (unmeasured rows —
+ * checker failures — are excluded from the denominator, never scored as
+ * zeros). Quality dimensions are meaningless on an incomplete task — callers
+ * gate on `fullyComplete` / `completionRate` before scoring quality.
  */
 
 /** What kind of produced state can satisfy a requirement structurally. */
@@ -210,12 +212,23 @@ interface RequirementCheck {
     structurallyPresent: boolean;
     /**
      * Whether the matched item actually fulfils the requirement. `null` when
-     * not structurally present, or when the matched item carries no content
-     * to assess.
+     * not structurally present, when the matched item carries no content
+     * to assess, or when the correctness check itself failed (`unmeasured`).
      */
     correct: boolean | null;
-    /** structurallyPresent && correct !== false. */
+    /** structurallyPresent && !unmeasured && correct !== false. */
     satisfied: boolean;
+    /**
+     * Set when the correctness check itself errored (LLM call failure or an
+     * unparseable response after retry). The requirement's fulfilment is
+     * UNKNOWN — `correct` stays null, `satisfied` is false, and
+     * `completionVerdict` excludes the row from `completionRate`'s
+     * denominator. Never folded into a zero: a synthetic zero is
+     * indistinguishable from a real failure (see `JudgeParseError`).
+     */
+    unmeasured?: true;
+    /** Why the correctness check could not be measured (present iff `unmeasured`). */
+    unmeasuredReason?: string;
     /** Human-readable evidence for the verdict. */
     evidence: string[];
 }
@@ -225,10 +238,12 @@ interface RequirementCheck {
 interface CompletionVerdict extends DefaultVerdict {
     taskId: string;
     requirements: RequirementCheck[];
-    /** satisfied / total requirements. */
+    /** satisfied / MEASURABLE requirements (unmeasured rows leave the denominator). */
     completionRate: number;
-    /** Every requirement satisfied. */
+    /** Every measurable requirement satisfied (false when anything is unmeasured). */
     fullyComplete: boolean;
+    /** Requirements whose correctness check errored — reported, never scored as zero. */
+    unmeasuredCount: number;
 }
 /**
  * Construct a `CompletionVerdict` from the per-requirement checks, deriving
@@ -262,8 +277,26 @@ interface LlmCorrectnessCheckerOpts {
     model?: string;
     /** Max chars of artifact content sent to the checker. */
     maxContentChars?: number;
+    /**
+     * Checker LLM calls per requirement before giving up (parse failures and
+     * call errors both consume attempts). The failure then surfaces as an
+     * `unmeasured` requirement, never a zero.
+     */
+    maxAttempts?: number;
+    /**
+     * Forensic capture of every checker request/response/error — without it a
+     * checker failure is unauditable (the agent-turn raws never contain the
+     * checker's own calls). Same sink contract as `LlmClient`.
+     */
+    rawSink?: RawProviderSink;
 }
-/** Parse the correctness checker's model response. Fails loud on a bad shape. */
+/**
+ * Parse the correctness checker's model response. Tolerates a response
+ * truncated mid-JSON (max_tokens cap) by auto-closing the prefix — the
+ * verdict boolean usually lands in the first few tokens, so a recovered
+ * prefix with a boolean `correct` is a real measurement, not a guess.
+ * Fails loud (JudgeParseError) when no boolean verdict is recoverable.
+ */
 declare function parseCorrectnessResponse(raw: string): {
     correct: boolean;
     reason: string;

package/dist/product-benchmark/index.d.ts

@@ -1,3 +1,95 @@
+import { R as RunRecord } from '../run-record-MRdJ-Kq2.js';
+import '@tangle-network/agent-interface';
+import '../errors-CzMUYo7b.js';
+import '../schema-m0gsnbt3.js';
+
+/**
+ * Export side of the product benchmark bundle contract: convert product
+ * eval run directories (`records.jsonl` of `RunRecord` rows + trace/raw
+ * artifacts) into a portable `product-benchmark-manifest.json` +
+ * `product-benchmark-records.jsonl` bundle that
+ * `validateProductBenchmarkRun` accepts.
+ *
+ * Product-specific policy (safety-split detection, tool-call recovery,
+ * profile id fallback, artifact materialization) enters through explicit
+ * options; everything else is the shared union of the tax/legal/creative
+ * exporters. Scenario catalogs, smoke runners, and CLIs stay in the
+ * products.
+ *
+ * Input rows are checked structurally, not with `validateRunRecord`:
+ * product harnesses record bare model aliases and partial provenance, and
+ * the bundle contract's own validators re-check every field that matters
+ * on the way out.
+ */
+
+/** Full mutable-surface superset a product arm may declare. */
+declare const productBenchmarkMutableSurfaces: readonly ["prompt", "resources.files", "tools", "mcp", "hooks", "subagents"];
+interface ProductBenchmarkExportOptions {
+    /** Source eval run directories, each containing a `records.jsonl` of RunRecord rows. */
+    readonly runDirs: readonly string[];
+    /** Destination directory for the bundle (manifest + records + materialized source runs). */
+    readonly outDir: string;
+    readonly projectId: string;
+    readonly benchmarkId: string;
+    /** Repo-relative path of the product's canonical agent profile source. */
+    readonly agentProfilePath: string;
+    /** Pass threshold applied when a row carries no explicit `outcome.raw.pass`. Default 0.7. */
+    readonly passThreshold?: number;
+    /**
+     * First scenario tag. Defaults to `projectId` with a trailing `-agent`
+     * stripped (`tax-agent` → `tax`), matching the product exporters.
+     */
+    readonly scenarioTagPrefix?: string;
+    /** Profile id used when a row has no `agentProfile.profileId`. Defaults to the row's arm id. */
+    readonly fallbackProfileId?: string;
+    /** Arm mutable surfaces recorded in the manifest. Defaults to the full superset. */
+    readonly mutableSurfaces?: readonly string[];
+    /**
+     * Copy each run dir into `<outDir>/source-runs/` and record
+     * bundle-relative artifact paths (portable, self-contained). When false,
+     * artifacts keep absolute paths into the original run dirs. Default true.
+     */
+    readonly materializeSourceRuns?: boolean;
+    /**
+     * Override split classification for a row. Return undefined to fall back
+     * to the default (`outcome.raw.safety === 1` → safety, then splitTag).
+     */
+    readonly classifySplit?: (record: RunRecord) => ProductBenchmarkSplit | undefined;
+    /** Recovers a tool-call count when the row's raw bag carries none (e.g. from turn artifacts). */
+    readonly toolCallFallback?: (record: RunRecord, runDir: string) => number;
+    /** Backend version recorded per row. Defaults to the cwd package.json's `@tangle-network/sandbox` range. */
+    readonly backendVersion?: string;
+    /**
+     * Explicit substrate versions for the manifest, merged over what the cwd
+     * package.json / node_modules resolve. Use when a substrate package is not
+     * installed where the export runs — the validator refuses an `'unknown'`
+     * version, so provide the real one rather than shipping the sentinel.
+     */
+    readonly substrate?: Partial<ProductBenchmarkManifest['substrate']>;
+}
+interface ProductBenchmarkSingleRunExportOptions extends Omit<ProductBenchmarkExportOptions, 'runDirs'> {
+    readonly runDir: string;
+}
+interface ProductBenchmarkExportResult {
+    readonly manifestPath: string;
+    readonly recordsPath: string;
+    readonly records: number;
+}
+/** Repo identity from the exporting process's cwd. `'unknown'` values are flagged by `validateProductBenchmarkRun`. */
+declare function productBenchmarkRepoIdentity(): ProductBenchmarkManifest['repo'];
+/** Map one RunRecord row to a validated product benchmark record. */
+declare function runRecordToProductBenchmarkRecord(record: RunRecord, runDir: string, artifactRoot: string, artifacts: ProductBenchmarkRecord['artifacts'], options: ProductBenchmarkExportOptions | ProductBenchmarkSingleRunExportOptions): ProductBenchmarkRecord;
+/** Derive the bundle manifest from already-normalized records. */
+declare function buildProductBenchmarkManifest(records: readonly ProductBenchmarkRecord[], options: Pick<ProductBenchmarkExportOptions, 'outDir' | 'projectId' | 'benchmarkId' | 'scenarioTagPrefix' | 'mutableSurfaces' | 'substrate'>): ProductBenchmarkManifest;
+/** Single-run convenience wrapper over `exportProductBenchmarkRuns`. */
+declare function exportProductBenchmark(options: ProductBenchmarkSingleRunExportOptions): ProductBenchmarkExportResult;
+/**
+ * Export one or more product eval run dirs into a validated product
+ * benchmark bundle at `outDir`. Both the manifest and every record are
+ * run through the contract validators before anything is written.
+ */
+declare function exportProductBenchmarkRuns(options: ProductBenchmarkExportOptions): ProductBenchmarkExportResult;
+
 declare const productBenchmarkSplits: readonly ["practice", "dev", "holdout", "safety", "sentinel"];
 type ProductBenchmarkSplit = (typeof productBenchmarkSplits)[number];
 interface ProductBenchmarkRepoRef {
@@ -116,6 +208,11 @@ interface ProductBenchmarkValidationReport {
     readonly manifestPath: string;
     readonly recordsPath: string;
     readonly records: number;
+    /** Manifest repo fields that are empty or the `'unknown'` export sentinel. */
+    readonly repoFailures: readonly string[];
+    /** Manifest substrate versions that are empty or the `'unknown'` export
+     *  sentinel — a bundle without substrate identity is not reproducible. */
+    readonly substrateFailures: readonly string[];
     readonly projects: readonly string[];
     readonly benchmarks: readonly string[];
     readonly arms: readonly string[];
@@ -140,5 +237,11 @@ declare function readProductBenchmarkRecords(path: string): ProductBenchmarkReco
 declare function readProductBenchmarkManifest(path: string): ProductBenchmarkManifest;
 declare function validateProductBenchmarkRun(input: ProductBenchmarkRunInput): ProductBenchmarkValidationReport;
 declare function findProductBenchmarkArtifacts(runDir: string): ProductBenchmarkArtifactPaths | null;
+/**
+ * Fail-loud gate over a bundle directory: locates the manifest + records,
+ * runs `validateProductBenchmarkRun`, and throws with every repo,
+ * integrity, and artifact failure listed. Returns the report when clean.
+ */
+declare function assertProductBenchmarkRun(runDir: string): ProductBenchmarkValidationReport;
 
-export { type AgentProfileRuntimeReceipt, type ProductBenchmarkArm, type ProductBenchmarkArtifactPaths, type ProductBenchmarkBudgets, type ProductBenchmarkManifest, type ProductBenchmarkProfileRef, type ProductBenchmarkRecord, type ProductBenchmarkRepoRef, type ProductBenchmarkRunInput, type ProductBenchmarkScenario, type ProductBenchmarkSplit, type ProductBenchmarkSubstrateVersions, type ProductBenchmarkValidationReport, type RuntimeResolution, findProductBenchmarkArtifacts, productBenchmarkIntegrityFailures, productBenchmarkSplits, readProductBenchmarkManifest, readProductBenchmarkRecords, validateProductBenchmarkManifest, validateProductBenchmarkRecord, validateProductBenchmarkRun };
+export { type AgentProfileRuntimeReceipt, type ProductBenchmarkArm, type ProductBenchmarkArtifactPaths, type ProductBenchmarkBudgets, type ProductBenchmarkExportOptions, type ProductBenchmarkExportResult, type ProductBenchmarkManifest, type ProductBenchmarkProfileRef, type ProductBenchmarkRecord, type ProductBenchmarkRepoRef, type ProductBenchmarkRunInput, type ProductBenchmarkScenario, type ProductBenchmarkSingleRunExportOptions, type ProductBenchmarkSplit, type ProductBenchmarkSubstrateVersions, type ProductBenchmarkValidationReport, type RuntimeResolution, assertProductBenchmarkRun, buildProductBenchmarkManifest, exportProductBenchmark, exportProductBenchmarkRuns, findProductBenchmarkArtifacts, productBenchmarkIntegrityFailures, productBenchmarkMutableSurfaces, productBenchmarkRepoIdentity, productBenchmarkSplits, readProductBenchmarkManifest, readProductBenchmarkRecords, runRecordToProductBenchmarkRecord, validateProductBenchmarkManifest, validateProductBenchmarkRecord, validateProductBenchmarkRun };

package/dist/product-benchmark/index.js

@@ -1,21 +1,35 @@
 import {
+  assertProductBenchmarkRun,
+  buildProductBenchmarkManifest,
+  exportProductBenchmark,
+  exportProductBenchmarkRuns,
   findProductBenchmarkArtifacts,
   productBenchmarkIntegrityFailures,
+  productBenchmarkMutableSurfaces,
+  productBenchmarkRepoIdentity,
   productBenchmarkSplits,
   readProductBenchmarkManifest,
   readProductBenchmarkRecords,
+  runRecordToProductBenchmarkRecord,
   validateProductBenchmarkManifest,
   validateProductBenchmarkRecord,
   validateProductBenchmarkRun
-} from "../chunk-63MBSQTX.js";
+} from "../chunk-RQNOLV3I.js";
 import "../chunk-3BFEG2F6.js";
 import "../chunk-PZ5AY32C.js";
 export {
+  assertProductBenchmarkRun,
+  buildProductBenchmarkManifest,
+  exportProductBenchmark,
+  exportProductBenchmarkRuns,
   findProductBenchmarkArtifacts,
   productBenchmarkIntegrityFailures,
+  productBenchmarkMutableSurfaces,
+  productBenchmarkRepoIdentity,
   productBenchmarkSplits,
   readProductBenchmarkManifest,
   readProductBenchmarkRecords,
+  runRecordToProductBenchmarkRecord,
   validateProductBenchmarkManifest,
   validateProductBenchmarkRecord,
   validateProductBenchmarkRun

package/dist/{provenance-BEITkFII.d.ts → provenance-BhJm32vN.d.ts}

@@ -1,9 +1,9 @@
-import { S as Scenario, g as Gate, G as GateResult, h as GateContext, C as CampaignResult, i as Mutator, f as SurfaceProposer, M as MutableSurface, j as GateDecision } from './types-fWqEJm7h.js';
+import { S as Scenario, g as Gate, G as GateResult, h as GateContext, C as CampaignResult, i as Mutator, f as SurfaceProposer, M as MutableSurface, j as GateDecision } from './types-DFI_Z-ZL.js';
 import { R as RedTeamCase } from './red-team-BWdoyleI.js';
-import { R as RunRecord } from './run-record-DEwidcqn.js';
+import { R as RunRecord } from './run-record-MRdJ-Kq2.js';
 import { D as Direction } from './pareto-E-pembql.js';
 import { a as PairedBootstrapResult } from './statistics-xP-cWc5k.js';
-import { R as RunCampaignOptions, C as CampaignStorage } from './gepa-bxuDoaO9.js';
+import { R as RunCampaignOptions, C as CampaignStorage } from './gepa-DeyPTlvx.js';
 import { HostedClient, TraceSpanEvent } from './hosted/index.js';
 
 /**

package/dist/{release-report-B1tA6pKu.d.ts → release-report-BQ1Ziyu-.d.ts}

@@ -1,6 +1,6 @@
 import { D as DatasetSplit, c as DatasetManifest, a as DatasetScenario } from './dataset-BbGkaN2I.js';
-import { m as GateDecision } from './summary-report-C4uzRWh8.js';
-import { R as RunRecord, b as RunSplitTag } from './run-record-DEwidcqn.js';
+import { m as GateDecision } from './summary-report-C0nnxOD8.js';
+import { R as RunRecord, b as RunSplitTag } from './run-record-MRdJ-Kq2.js';
 
 /**
  * Release confidence gate.

package/dist/reporting.d.ts

@@ -1,9 +1,9 @@
-export { R as RubricOutcomePair, a as RubricPredictiveValidityInput, b as RubricPredictiveValidityReport, c as RubricRanking, r as rubricPredictiveValidity } from './rubric-predictive-validity-w7tun-q3.js';
-export { B as BootstrapOptions, a as BootstrapResult, J as JudgeReplayGateArgs, R as ReleaseConfidenceAxis, b as ReleaseConfidenceAxisName, c as ReleaseConfidenceInput, d as ReleaseConfidenceIssue, e as ReleaseConfidenceMetrics, f as ReleaseConfidenceScorecard, g as ReleaseConfidenceStatus, h as ReleaseConfidenceThresholds, i as ReleaseTraceEvidence, j as RenderReleaseReportOptions, V as Verdict, k as assertReleaseConfidence, l as bootstrapCi, m as evaluateReleaseConfidence, n as judgeReplayGate, r as renderReleaseReport } from './release-report-B1tA6pKu.js';
+export { R as RubricOutcomePair, a as RubricPredictiveValidityInput, b as RubricPredictiveValidityReport, c as RubricRanking, r as rubricPredictiveValidity } from './rubric-predictive-validity-0MdjTt8R.js';
+export { B as BootstrapOptions, a as BootstrapResult, J as JudgeReplayGateArgs, R as ReleaseConfidenceAxis, b as ReleaseConfidenceAxisName, c as ReleaseConfidenceInput, d as ReleaseConfidenceIssue, e as ReleaseConfidenceMetrics, f as ReleaseConfidenceScorecard, g as ReleaseConfidenceStatus, h as ReleaseConfidenceThresholds, i as ReleaseTraceEvidence, j as RenderReleaseReportOptions, V as Verdict, k as assertReleaseConfidence, l as bootstrapCi, m as evaluateReleaseConfidence, n as judgeReplayGate, r as renderReleaseReport } from './release-report-BQ1Ziyu-.js';
 export { I as InterimReleaseConfidence, a as InterimReleaseConfidenceInput, P as PairedEvalueOptions, b as PairedEvalueSequence, c as PairedEvalueStep, S as SequentialDecision, e as evaluateInterimReleaseConfidence, p as pairedEvalueSequence } from './sequential-5iSVfzl2.js';
 export { P as PairedBootstrapOptions, a as PairedBootstrapResult, b as benjaminiHochberg, p as pairedBootstrap, w as wilcoxonSignedRank } from './statistics-xP-cWc5k.js';
-export { G as GainDistributionBin, a as GainDistributionFigureSpec, b as GainDistributionOptions, P as ParetoFigureSpec, c as ParetoPoint, R as RESEARCH_REPORT_HARD_PAIR_FLOOR, d as ResearchReport, e as ResearchReportCandidate, f as ResearchReportDecision, g as ResearchReportMethodology, h as ResearchReportOptions, i as ResearchReportRecommendation, S as SummaryTable, j as SummaryTableOptions, k as SummaryTableRow, l as gainHistogram, p as paretoChart, r as researchReport, s as summaryTable } from './summary-report-C4uzRWh8.js';
-import './run-record-DEwidcqn.js';
+export { G as GainDistributionBin, a as GainDistributionFigureSpec, b as GainDistributionOptions, P as ParetoFigureSpec, c as ParetoPoint, R as RESEARCH_REPORT_HARD_PAIR_FLOOR, d as ResearchReport, e as ResearchReportCandidate, f as ResearchReportDecision, g as ResearchReportMethodology, h as ResearchReportOptions, i as ResearchReportRecommendation, S as SummaryTable, j as SummaryTableOptions, k as SummaryTableRow, l as gainHistogram, p as paretoChart, r as researchReport, s as summaryTable } from './summary-report-C0nnxOD8.js';
+import './run-record-MRdJ-Kq2.js';
 import '@tangle-network/agent-interface';
 import './errors-CzMUYo7b.js';
 import './schema-m0gsnbt3.js';

package/dist/{researcher-Ba2y1Foi.d.ts → researcher-B_ODTAJs.d.ts}

@@ -1,6 +1,6 @@
-import { b as RunSplitTag, c as RunTokenUsage, d as RunJudgeMetadata, J as JudgeScoresRecord, A as AgentProfileCell, e as AgentProfileCellInput, R as RunRecord } from './run-record-DEwidcqn.js';
+import { b as RunSplitTag, c as RunTokenUsage, d as RunJudgeMetadata, J as JudgeScoresRecord, A as AgentProfileCell, e as AgentProfileCellInput, R as RunRecord } from './run-record-MRdJ-Kq2.js';
 import { L as LlmClientOptions, a as LlmRouteRequirements } from './llm-client-Bj7g0rqu.js';
-import { h as ResearchReportOptions, d as ResearchReport, m as GateDecision } from './summary-report-C4uzRWh8.js';
+import { h as ResearchReportOptions, d as ResearchReport, m as GateDecision } from './summary-report-C0nnxOD8.js';
 import { T as TraceEmitter, R as RunCompleteHook } from './emitter-C2rqGH_l.js';
 import { R as RunIntegrityExpectations, a as RunIntegrityReport } from './integrity-D2t12mMw.js';
 import { R as RawProviderSink } from './raw-provider-sink-C46HDghv.js';

package/dist/rl.d.ts

@@ -1,23 +1,23 @@
-import { R as RunRecord, b as RunSplitTag } from './run-record-DEwidcqn.js';
+import { R as RunRecord, b as RunSplitTag } from './run-record-MRdJ-Kq2.js';
 export { A as AdversarialMutation } from './adversarial-B7loGVVX.js';
-import { P as PreferenceExtractionReport, D as DpoExportRow, G as GrpoExportRow, S as SftExportRow, E as ExtractPreferencesOptions, a as DpoLookups, b as GrpoLookups, c as SftLookups } from './corpus-ONOzGFmG.js';
-export { d as CorpusAppendResult, C as CorpusRecord, e as DatasetFormat, f as ExtractStepRewardsOptions, H as HarvestOptions, g as PreferenceStrategy, h as PreferenceTriple, i as PrmExportRow, j as PrmLookups, k as PrmTrainingTriple, R as RewardKind, l as RewardStats, m as RlDatasetBundle, n as RlDatasetConfig, o as RlDatasetManifest, p as RlDatasetStats, q as RunwiseStepSummary, r as StepReward, s as StepRewardJsonlRow, t as StepScorer, u as appendToCorpus, v as buildDatasetFromCorpus, w as buildRlDataset, x as datasheetToMarkdown, y as extractPreferences, z as extractStepRewards, A as prmTrainingPairs, B as readCorpus, F as runwiseStepRewardSummary, I as stepRewardsToJsonl, J as toAnthropicFormat, K as toDpoJsonl, L as toDpoRows, M as toGrpoJsonl, N as toGrpoRows, O as toPrmJsonl, Q as toPrmRows, T as toSftJsonl, U as toSftRows, V as toTRLFormat } from './corpus-ONOzGFmG.js';
+import { P as PreferenceExtractionReport, D as DpoExportRow, G as GrpoExportRow, S as SftExportRow, E as ExtractPreferencesOptions, a as DpoLookups, b as GrpoLookups, c as SftLookups } from './corpus-CiSzzLa5.js';
+export { d as CorpusAppendResult, C as CorpusRecord, e as DatasetFormat, f as ExtractStepRewardsOptions, H as HarvestOptions, g as PreferenceStrategy, h as PreferenceTriple, i as PrmExportRow, j as PrmLookups, k as PrmTrainingTriple, R as RewardKind, l as RewardStats, m as RlDatasetBundle, n as RlDatasetConfig, o as RlDatasetManifest, p as RlDatasetStats, q as RunwiseStepSummary, r as StepReward, s as StepRewardJsonlRow, t as StepScorer, u as appendToCorpus, v as buildDatasetFromCorpus, w as buildRlDataset, x as datasheetToMarkdown, y as extractPreferences, z as extractStepRewards, A as prmTrainingPairs, B as readCorpus, F as runwiseStepRewardSummary, I as stepRewardsToJsonl, J as toAnthropicFormat, K as toDpoJsonl, L as toDpoRows, M as toGrpoJsonl, N as toGrpoRows, O as toPrmJsonl, Q as toPrmRows, T as toSftJsonl, U as toSftRows, V as toTRLFormat } from './corpus-CiSzzLa5.js';
 export { O as OffPolicyEstimate, a as OffPolicyOptions, b as OffPolicyTrajectory, d as doublyRobust, i as inverseProbabilityWeighting, o as offPolicyEstimateAll, s as selfNormalizedImportanceWeighting } from './off-policy-DiwuKKg7.js';
 import { b as OutcomeStore } from './outcome-store-rnXLEqSn.js';
 export { D as DeploymentOutcome, F as FileSystemOutcomeStore, a as FileSystemOutcomeStoreOptions, I as InMemoryOutcomeStore } from './outcome-store-rnXLEqSn.js';
-import { b as RubricPredictiveValidityReport } from './rubric-predictive-validity-w7tun-q3.js';
-import { R as Researcher, F as FailureMode, S as SteeringChange, E as ExperimentPlan, a as ExperimentResult, b as EvalCampaignResult, c as EvalCampaignOptions } from './researcher-Ba2y1Foi.js';
-export { r as runEvalCampaign } from './researcher-Ba2y1Foi.js';
+import { b as RubricPredictiveValidityReport } from './rubric-predictive-validity-0MdjTt8R.js';
+import { R as Researcher, F as FailureMode, S as SteeringChange, E as ExperimentPlan, a as ExperimentResult, b as EvalCampaignResult, c as EvalCampaignOptions } from './researcher-B_ODTAJs.js';
+export { r as runEvalCampaign } from './researcher-B_ODTAJs.js';
 import { a as VerificationReport } from './multi-layer-verifier-DUZXrPDA.js';
 import { I as InterimReleaseConfidence } from './sequential-5iSVfzl2.js';
-import { C as CampaignResult } from './types-fWqEJm7h.js';
+import { C as CampaignResult } from './types-DFI_Z-ZL.js';
 import '@tangle-network/agent-interface';
 import './errors-CzMUYo7b.js';
 import './schema-m0gsnbt3.js';
 import './store-BcFXE6LG.js';
 import './llm-client-Bj7g0rqu.js';
 import './raw-provider-sink-C46HDghv.js';
-import './summary-report-C4uzRWh8.js';
+import './summary-report-C0nnxOD8.js';
 import './failure-cluster-DH9Flgcf.js';
 import './emitter-C2rqGH_l.js';
 import './integrity-D2t12mMw.js';

package/dist/rl.js

@@ -16,7 +16,7 @@ import {
 } from "./chunk-3RF76KTD.js";
 import {
   runEvalCampaign
-} from "./chunk-4LWD6GC7.js";
+} from "./chunk-6FIAJHCU.js";
 import "./chunk-CWNP4DV4.js";
 import {
   rubricPredictiveValidity
@@ -36,7 +36,7 @@ import {
 } from "./chunk-VZSRQ272.js";
 import "./chunk-SBCB6VZY.js";
 import "./chunk-PC4UYEBM.js";
-import "./chunk-G6S73VA7.js";
+import "./chunk-2NSLDY4B.js";
 import "./chunk-TVVP3ZZQ.js";
 import "./chunk-ABOIVNXL.js";
 import "./chunk-VSMTAMNK.js";

package/dist/{rubric-predictive-validity-w7tun-q3.d.ts → rubric-predictive-validity-0MdjTt8R.d.ts}

@@ -1,4 +1,4 @@
-import { R as RunRecord } from './run-record-DEwidcqn.js';
+import { R as RunRecord } from './run-record-MRdJ-Kq2.js';
 import { b as OutcomeStore } from './outcome-store-rnXLEqSn.js';
 
 /**

package/dist/{run-campaign-RF3H6D4U.js → run-campaign-2L4WCJHR.js}

@@ -1,7 +1,7 @@
 import {
   planCampaignRun,
   runCampaign
-} from "./chunk-LSCBODPQ.js";
+} from "./chunk-XKA6ZGEY.js";
 import "./chunk-HRGUJTER.js";
 import "./chunk-3BFEG2F6.js";
 import "./chunk-PZ5AY32C.js";
@@ -9,4 +9,4 @@ export {
   planCampaignRun,
   runCampaign
 };
-//# sourceMappingURL=run-campaign-RF3H6D4U.js.map
+//# sourceMappingURL=run-campaign-2L4WCJHR.js.map

package/dist/{run-record-DEwidcqn.d.ts → run-record-MRdJ-Kq2.d.ts}

@@ -310,5 +310,16 @@ declare function parseRunRecordSafe(input: unknown): {
 };
 /** Round-trip helper — `JSON.parse(JSON.stringify(record))` then validate. */
 declare function roundTripRunRecord(record: RunRecord): RunRecord;
+/**
+ * Heuristic snapshot check. Accepts:
+ *   - `name@YYYY-MM-DD` (Anthropic style: `claude-sonnet-4-6@2025-04-15`)
+ *   - `name-YYYYMMDD`   (OpenAI style: `gpt-4o-2024-11-20`)
+ *   - `name@<arbitrary-token>` (allow opaque snapshots like `@v3`)
+ *   - explicit `:date-...` Vertex-style tags
+ *
+ * Rejects bare aliases like `claude-sonnet-4` or `gpt-4o` that remap
+ * silently as providers ship new snapshots.
+ */
+declare function modelHasSnapshot(model: string): boolean;
 
-export { type AgentProfileCell as A, toAgentProfileJson as B, validateAgentProfileCell as C, validateRunRecord as D, verifyAgentProfileCell as E, type JudgeScoresRecord as J, type RunRecord as R, type AgentProfileJson as a, type RunSplitTag as b, type RunTokenUsage as c, type RunJudgeMetadata as d, type AgentProfileCellInput as e, AGENT_PROFILE_KINDS as f, type AgentInterfaceProfileLike as g, type AgentProfileCellSchemaVersion as h, AgentProfileCellValidationError as i, type AgentProfileDimensionValue as j, type AgentProfileHarness as k, type AgentProfileKind as l, type AgentProfileSource as m, type AgentProfileSourceInput as n, type RunOutcome as o, RunRecordValidationError as p, agentProfileCellHashMaterial as q, agentProfileCellKey as r, assertRunAgentProfileCell as s, buildAgentInterfaceProfileCell as t, buildAgentProfileCell as u, groupRunsByAgentProfileCell as v, isRunRecord as w, parseRunRecordSafe as x, requireAgentProfileCell as y, roundTripRunRecord as z };
+export { type AgentProfileCell as A, roundTripRunRecord as B, toAgentProfileJson as C, validateAgentProfileCell as D, validateRunRecord as E, verifyAgentProfileCell as F, type JudgeScoresRecord as J, type RunRecord as R, type AgentProfileJson as a, type RunSplitTag as b, type RunTokenUsage as c, type RunJudgeMetadata as d, type AgentProfileCellInput as e, AGENT_PROFILE_KINDS as f, type AgentInterfaceProfileLike as g, type AgentProfileCellSchemaVersion as h, AgentProfileCellValidationError as i, type AgentProfileDimensionValue as j, type AgentProfileHarness as k, type AgentProfileKind as l, type AgentProfileSource as m, type AgentProfileSourceInput as n, type RunOutcome as o, RunRecordValidationError as p, agentProfileCellHashMaterial as q, agentProfileCellKey as r, assertRunAgentProfileCell as s, buildAgentInterfaceProfileCell as t, buildAgentProfileCell as u, groupRunsByAgentProfileCell as v, isRunRecord as w, modelHasSnapshot as x, parseRunRecordSafe as y, requireAgentProfileCell as z };

package/dist/{runtime-trajectory-OJDaTYHN.d.ts → runtime-trajectory-8w0_jmtR.d.ts}

@@ -1,4 +1,4 @@
-import { b as RunSplitTag } from './run-record-DEwidcqn.js';
+import { b as RunSplitTag } from './run-record-MRdJ-Kq2.js';
 
 interface RuntimeTrajectoryHookEvent {
     id: string;

package/dist/{semantic-concept-judge-J8xvjdc3.d.ts → semantic-concept-judge-D-IlH5v1.d.ts}

@@ -1,7 +1,7 @@
 import { AxAIService } from '@ax-llm/ax';
 import { z } from 'zod';
-import { A as AnalystFinding, a as Analyst, b as AnalystContext } from './types-BEzCBMQD.js';
-import { T as TraceAnalystKindSpec } from './kind-factory-OgqQSvLi.js';
+import { A as AnalystFinding, a as Analyst, b as AnalystContext } from './types-Dz9cKF0g.js';
+import { T as TraceAnalystKindSpec } from './kind-factory-D0nk7AKV.js';
 import { a as TraceAnalystSpan } from './store-C1YxJDEK.js';
 import { L as LlmClientOptions } from './llm-client-Bj7g0rqu.js';
 import { S as Severity } from './multi-layer-verifier-DUZXrPDA.js';

package/dist/{summary-report-C4uzRWh8.d.ts → summary-report-C0nnxOD8.d.ts}

@@ -1,4 +1,4 @@
-import { R as RunRecord } from './run-record-DEwidcqn.js';
+import { R as RunRecord } from './run-record-MRdJ-Kq2.js';
 import { F as FailureClusterReport } from './failure-cluster-DH9Flgcf.js';
 
 /**

package/dist/traces.d.ts

@@ -14,7 +14,7 @@ import { A as AnalyzeTracesOptions, b as AnalyzeTracesResult } from './analyst-C
 export { a as AnalyzeTracesInput, c as AnalyzeTracesTurnSnapshot, d as analyzeTraces } from './analyst-C8HHvfJp.js';
 import { h as TraceAnalystSpanKind, i as TraceAnalystSpanStatus, T as TraceAnalysisStore, g as TraceAnalystFilters, b as DatasetOverview, Q as QueryTracesPage, l as ViewTraceResult, V as ViewSpansResult, c as SearchTraceResult, S as SearchSpanResult } from './store-C1YxJDEK.js';
 export { D as DEFAULT_TRACE_ANALYST_BUDGETS, E as ErrorCluster, d as SpanMatchRecord, e as TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, f as TraceAnalystByteBudgets, a as TraceAnalystSpan, j as TraceAnalystTraceSummary, k as ViewTraceOversized } from './store-C1YxJDEK.js';
-import { b as RunSplitTag, c as RunTokenUsage, R as RunRecord } from './run-record-DEwidcqn.js';
+import { b as RunSplitTag, c as RunTokenUsage, R as RunRecord } from './run-record-MRdJ-Kq2.js';
 import { AxFunction } from '@ax-llm/ax';
 import '@tangle-network/agent-interface';
 

package/dist/traces.js

@@ -28,7 +28,7 @@ import {
   scoreTraceInsightReadiness,
   tokenizeDomainWords,
   traceAnalystOnRunComplete
-} from "./chunk-PMF5WIBX.js";
+} from "./chunk-7RBJANJD.js";
 import {
   TRACE_ANALYST_ACTOR_DESCRIPTION,
   TRACE_ANALYST_ACTOR_DESCRIPTION_VERSION,
@@ -106,7 +106,7 @@ import {
   defaultProviderRedactor,
   providerFromBaseUrl
 } from "./chunk-PC4UYEBM.js";
-import "./chunk-G6S73VA7.js";
+import "./chunk-2NSLDY4B.js";
 import {
   TraceEmitter,
   llmSpanFromProvider

package/dist/{types-fWqEJm7h.d.ts → types-DFI_Z-ZL.d.ts}

@@ -1,4 +1,4 @@
-import { c as RunTokenUsage } from './run-record-DEwidcqn.js';
+import { c as RunTokenUsage } from './run-record-MRdJ-Kq2.js';
 
 /**
  * Pass A substrate types — `runCampaign` is the one primitive every
@@ -347,9 +347,20 @@ interface CampaignCostMeter {
      *  `tokenUsage`, so a cell that never reports tokens reads as a stub. Any
      *  dispatch that calls an LLM MUST report its usage. */
     observeTokens(usage: CampaignTokenUsage): void;
+    /** Record the concrete model the backend RESOLVED this cell to at runtime.
+     *  The substrate cannot see the LLM call, so it cannot know which model a
+     *  vendor-locked harness actually served — only the dispatch, reading the
+     *  backend's usage/terminal events, can. A dispatch whose profile declares a
+     *  runtime-resolved model (the `HARNESS_NATIVE_MODEL` sentinel) MUST report
+     *  the resolved, snapshot-bearing id here so the RunRecord pins a real model
+     *  instead of the sentinel. Last write wins (a cell issues one logical run);
+     *  optional because most dispatches declare a concrete model up front. */
+    observeModel?(model: string): void;
     current(): number;
     /** Accumulated token usage for this cell (zeros if never observed). */
     tokens(): CampaignTokenUsage;
+    /** The runtime-resolved model reported via `observeModel`, if any. */
+    resolvedModel?(): string | undefined;
 }
 /** Source tag — required on every store write. Used by the
  *  default training-source filter (production-trace samples NOT used as
@@ -446,6 +457,12 @@ interface CampaignCellResult<TArtifact> {
      *  `{ input: 0, output: 0 }` when the dispatch reported none — which the
      *  backend-integrity guard reads as a stub. */
     tokenUsage: CampaignTokenUsage;
+    /** The concrete model the backend resolved this cell to at runtime, reported
+     *  by the dispatch via `ctx.cost.observeModel`. Set only when the dispatch
+     *  reported it — a profile that declares a concrete model up front has no
+     *  need to. Consumed by `buildRunRecord` to pin the real model when the
+     *  declared model is the `HARNESS_NATIVE_MODEL` sentinel. */
+    resolvedModel?: string;
     durationMs: number;
     seed: number;
     cached: boolean;

package/dist/{types-BEzCBMQD.d.ts → types-Dz9cKF0g.d.ts}

@@ -1,4 +1,4 @@
-import { R as RunRecord } from './run-record-DEwidcqn.js';
+import { R as RunRecord } from './run-record-MRdJ-Kq2.js';
 import { T as TraceAnalysisStore } from './store-C1YxJDEK.js';
 import { a as JudgeInput } from './types-C7DGg5ex.js';
 import { b as LlmCallRequest, c as LlmCallResult } from './llm-client-Bj7g0rqu.js';

package/dist/workflow/index.d.ts

@@ -1,7 +1,7 @@
 import { W as WorkflowTopology } from '../harness-optimizer-mOl9XX_O.js';
-import { b as RunSplitTag, c as RunTokenUsage, R as RunRecord } from '../run-record-DEwidcqn.js';
-import { A as AnalystFinding, h as AnalystSeverity, E as EvidenceRef } from '../types-BEzCBMQD.js';
-import { F as FailureClusterInsight } from '../insight-report-C02J3q4T.js';
+import { b as RunSplitTag, c as RunTokenUsage, R as RunRecord } from '../run-record-MRdJ-Kq2.js';
+import { A as AnalystFinding, h as AnalystSeverity, E as EvidenceRef } from '../types-Dz9cKF0g.js';
+import { F as FailureClusterInsight } from '../insight-report-k0sRTzKg.js';
 import { a as VerificationReport, L as LayerResult } from '../multi-layer-verifier-DUZXrPDA.js';
 import { F as FailureClusterReport } from '../failure-cluster-DH9Flgcf.js';
 import { R as RedactionRule, a as RedactionReport } from '../redact-B40YG2M_.js';
@@ -19,7 +19,7 @@ import '../types-C7DGg5ex.js';
 import '@tangle-network/tcloud';
 import '../llm-client-Bj7g0rqu.js';
 import '../raw-provider-sink-C46HDghv.js';
-import '../summary-report-C4uzRWh8.js';
+import '../summary-report-C0nnxOD8.js';
 import '../judge-calibration-0p2QcWNE.js';
 import '../verdict-C9MlYujm.js';
 import '../control-runtime-Acf9CGhw.js';

package/dist/workflow/index.js

@@ -7,7 +7,7 @@ import {
 } from "../chunk-GGE4NNQT.js";
 import {
   validateRunRecord
-} from "../chunk-G6S73VA7.js";
+} from "../chunk-2NSLDY4B.js";
 import "../chunk-ABOIVNXL.js";
 import "../chunk-VSMTAMNK.js";
 import {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-eval",
-  "version": "0.102.0",
+  "version": "0.103.0",
   "description": "Evaluate and improve AI agents from runs, traces, judges, and feedback. Compare candidates, cluster failures, measure lift, and gate releases.",
   "homepage": "https://github.com/tangle-network/agent-eval#readme",
   "repository": {