@tangle-network/agent-eval 0.115.2 → 0.116.0

package/dist/{pre-registration-D8h7ZxNL.d.ts → pre-registration--vU0mMtD.d.ts}

@@ -1,10 +1,10 @@
 import { A as AgentEvalError } from './errors-oeQrLqXC.js';
-import { R as RunRecord } from './run-record-B7RTi_ix.js';
+import { R as RunRecord } from './run-record-CZmcpWPo.js';
 import { P as PairedBootstrapOptions, M as McNemarResult, R as RiskDifferenceResult, a as PairedBootstrapResult } from './statistics-oUbOJe-S.js';
-import { C as ChatClient } from './kind-factory-DcNg13sZ.js';
-import { a as JudgeDimension, S as Scenario, b as JudgeConfig } from './types-C5gJrOVT.js';
+import { C as ChatClient } from './policy-edit-Clb2v6Oa.js';
+import { a as JudgeDimension, S as Scenario, b as JudgeConfig } from './types-Ca_63YSD.js';
 import { TCloud } from '@tangle-network/tcloud';
-import { R as RawProviderSink } from './raw-provider-sink-C46HDghv.js';
+import { R as RawProviderSink } from './store-9cAScOcb.js';
 import { D as DefaultVerdict } from './verdict-C9MlYujm.js';
 
 /**

package/dist/{provenance-Bibyg1U9.d.ts → provenance-BbVagC68.d.ts}

@@ -1,7 +1,8 @@
-import { S as Scenario, g as Gate, G as GateResult, n as GateContext, C as CampaignResult, p as Mutator, f as SurfaceProposer, M as MutableSurface, h as GateDecision } from './types-C5gJrOVT.js';
-import { e as RedTeamCase, D as Direction, b as RunCampaignOptions } from './gepa-dne9JDPL.js';
-import { R as RunRecord } from './run-record-B7RTi_ix.js';
+import { S as Scenario, g as Gate, G as GateResult, n as GateContext, C as CampaignResult, p as Mutator, f as SurfaceProposer, M as MutableSurface, o as GenerationCandidate, h as GateDecision } from './types-Ca_63YSD.js';
+import { e as RedTeamCase, D as Direction, b as RunCampaignOptions } from './gepa-CQelRtuC.js';
+import { R as RunRecord } from './run-record-CZmcpWPo.js';
 import { a as PairedBootstrapResult } from './statistics-oUbOJe-S.js';
+import { P as PolicyEditCandidateRecord } from './policy-edit-Clb2v6Oa.js';
 import { HostedClient, TraceSpanEvent } from './hosted/index.js';
 import { C as CampaignStorage } from './storage-Dw_f7WMt.js';
 
@@ -356,7 +357,7 @@ declare function evolutionaryProposer<TFindings = unknown>(opts: EvolutionaryPro
  * Two artifacts, one source of truth:
  *
  *   1. `LoopProvenanceRecord` — a structured JSON record capturing every
- *      candidate (surfaceHash + label + rationale), its measured composite,
+ *      candidate (surfaceHash + label + rationale + structured cause), its measured composite,
  *      the gate decision + reasons + delta, the held-out lift, the explicit
  *      baseline→candidate diff, and BACKEND PROVENANCE (the
  *      `assertRealBackend` verdict + worker call count + model). This is the
@@ -387,6 +388,18 @@ interface LoopProvenanceCandidate {
     /** Proposer rationale — the "because Z". When the proposer returned a bare
      *  surface (blind mutator) this is absent. */
     rationale?: string;
+    /** Exact validated cause when the proposer emitted a structured record. */
+    candidateRecord?: PolicyEditCandidateRecord;
+    /** Exact complete incumbent this candidate mutated. */
+    parentSurfaceHash: string;
+    /** Search-split composite of the exact parent. */
+    parentComposite: number;
+    /** Search-split composite change relative to the exact parent. */
+    observedDeltaFromParent?: number;
+    /** Whether the candidate completed every designed cell and could be selected. */
+    eligibleForPromotion: boolean;
+    /** Designed-denominator receipt retained even for incomplete candidates. */
+    coverage: NonNullable<GenerationCandidate['coverage']>;
     /** Mean composite this candidate scored on the search split. */
     composite: number;
     /** Whether this candidate was promoted out of its generation. */
@@ -409,7 +422,7 @@ interface LoopProvenanceBackend {
  * the bare hosted event) + backend provenance.
  */
 interface LoopProvenanceRecord {
-    schema: 'tangle.loop-provenance.v2';
+    schema: 'tangle.loop-provenance.v3';
     runId: string;
     runDir: string;
     timestamp: string;
@@ -421,8 +434,10 @@ interface LoopProvenanceRecord {
     winnerRationale?: string;
     /** The explicit baseline→winner unified diff the gate decided on. */
     diff: string;
-    /** Every candidate across every generation, each carrying its rationale. */
+    /** Every candidate across every generation, with its rationale and structured cause. */
     candidates: LoopProvenanceCandidate[];
+    /** Baseline composite on the search split that generated the candidates. */
+    baselineSearchComposite: number;
     /** The gate verdict — decision + reasons + contributing gates + delta. */
     gate: {
         decision: GateDecision;
@@ -453,18 +468,15 @@ interface BuildLoopProvenanceArgs<TArtifact, TScenario extends Scenario> {
     winnerLabel?: string;
     winnerRationale?: string;
     diff: string;
+    /** Baseline composite on the search split, distinct from holdout scoring. */
+    baselineSearchComposite: number;
     /** Per-generation candidate records straight off the loop result. */
     generations: Array<{
         generationIndex: number;
-        candidates: Array<{
-            surfaceHash: string;
-            composite: number;
-            label?: string;
-            rationale?: string;
-        }>;
+        candidates: GenerationCandidate[];
         promoted: string[];
-        /** Surfaces measured this generation, keyed positionally to candidates so
-         *  the content hash can be computed from the real surface text. */
+        /** Surfaces measured this generation, keyed by surface hash so the content
+         *  hash can be computed and the loop identity rechecked from real bytes. */
         surfaces: Array<{
             surfaceHash: string;
             surface: MutableSurface;

package/dist/{release-report-CCtzajxP.d.ts → release-report-CamNDe90.d.ts}

@@ -1,6 +1,6 @@
 import { a as DatasetSplit, b as DatasetManifest, D as DatasetScenario } from './dataset-NENEzRgk.js';
-import { m as GateDecision } from './summary-report-BJ5aNwZ1.js';
-import { R as RunRecord, b as RunSplitTag } from './run-record-B7RTi_ix.js';
+import { m as GateDecision } from './summary-report-DTNgQycC.js';
+import { R as RunRecord, a as RunSplitTag } from './run-record-CZmcpWPo.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-DYTLjGWu.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-CCtzajxP.js';
+export { R as RubricOutcomePair, a as RubricPredictiveValidityInput, b as RubricPredictiveValidityReport, c as RubricRanking, r as rubricPredictiveValidity } from './rubric-predictive-validity-BIdf9h4R.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-CamNDe90.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-oUbOJe-S.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-BJ5aNwZ1.js';
-import './run-record-B7RTi_ix.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-DTNgQycC.js';
+import './run-record-CZmcpWPo.js';
 import '@tangle-network/agent-interface';
 import './errors-oeQrLqXC.js';
 import './schema-SGWcK9wa.js';

package/dist/{researcher-Dq-EtpbE.d.ts → researcher-Dwbo_Fxx.d.ts}

@@ -1,9 +1,9 @@
-import { b as RunSplitTag, d as RunTokenUsage, e as RunJudgeMetadata, J as JudgeScoresRecord, A as AgentProfileCell, f as AgentProfileCellInput, R as RunRecord } from './run-record-B7RTi_ix.js';
-import { L as LlmClientOptions, a as LlmRouteRequirements } from './llm-client-DyqEH4jH.js';
-import { h as ResearchReportOptions, d as ResearchReport, m as GateDecision } from './summary-report-BJ5aNwZ1.js';
+import { a as RunSplitTag, c as RunTokenUsage, d as RunJudgeMetadata, J as JudgeScoresRecord, A as AgentProfileCell, e as AgentProfileCellInput, R as RunRecord } from './run-record-CZmcpWPo.js';
+import { L as LlmClientOptions, a as LlmRouteRequirements } from './policy-edit-Clb2v6Oa.js';
+import { h as ResearchReportOptions, d as ResearchReport, m as GateDecision } from './summary-report-DTNgQycC.js';
 import { T as TraceEmitter, R as RunCompleteHook } from './emitter-BRchAAAx.js';
-import { R as RunIntegrityExpectations, a as RunIntegrityReport } from './integrity-qemeBAyx.js';
-import { R as RawProviderSink } from './raw-provider-sink-C46HDghv.js';
+import { R as RunIntegrityExpectations, a as RunIntegrityReport } from './integrity-C6PZ73iC.js';
+import { R as RawProviderSink } from './store-9cAScOcb.js';
 import { F as FailureClass } from './schema-SGWcK9wa.js';
 import { T as TraceStore } from './store-BsVi7ncX.js';
 

package/dist/rl.d.ts

@@ -1,24 +1,26 @@
-import { R as RunRecord, b as RunSplitTag } from './run-record-B7RTi_ix.js';
+import { R as RunRecord, a as RunSplitTag } from './run-record-CZmcpWPo.js';
 export { A as AdversarialMutation } from './adversarial-B7loGVVX.js';
 import { S as Span } from './schema-SGWcK9wa.js';
 import { T as TraceStore } from './store-BsVi7ncX.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-DYTLjGWu.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-Dq-EtpbE.js';
-export { r as runEvalCampaign } from './researcher-Dq-EtpbE.js';
+import { b as RubricPredictiveValidityReport } from './rubric-predictive-validity-BIdf9h4R.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-Dwbo_Fxx.js';
+export { r as runEvalCampaign } from './researcher-Dwbo_Fxx.js';
 import { a as VerificationReport } from './multi-layer-verifier-BsqKuLyN.js';
 import { I as InterimReleaseConfidence } from './sequential-5iSVfzl2.js';
-import { C as CampaignResult } from './types-C5gJrOVT.js';
+import { C as CampaignResult } from './types-Ca_63YSD.js';
 import '@tangle-network/agent-interface';
 import './errors-oeQrLqXC.js';
-import './llm-client-DyqEH4jH.js';
-import './raw-provider-sink-C46HDghv.js';
-import './summary-report-BJ5aNwZ1.js';
+import './policy-edit-Clb2v6Oa.js';
+import './store-9cAScOcb.js';
+import './types-C7DGg5ex.js';
+import '@tangle-network/tcloud';
+import './summary-report-DTNgQycC.js';
 import './failure-cluster-C48PiReX.js';
 import './emitter-BRchAAAx.js';
-import './integrity-qemeBAyx.js';
+import './integrity-C6PZ73iC.js';
 import './verdict-C9MlYujm.js';
 
 /**

package/dist/rl.js

@@ -10,7 +10,7 @@ import {
 } from "./chunk-3RF76KTD.js";
 import {
   runEvalCampaign
-} from "./chunk-QG5F6463.js";
+} from "./chunk-ONM6PEAE.js";
 import {
   detectRewardHacking,
   extractVerifiableReward,
@@ -38,7 +38,7 @@ import "./chunk-TVVP3ZZQ.js";
 import "./chunk-5UF54T55.js";
 import "./chunk-XJYR7XFV.js";
 import "./chunk-VSMTAMNK.js";
-import "./chunk-FUCQVFMU.js";
+import "./chunk-GY4SYVPJ.js";
 import "./chunk-PC4UYEBM.js";
 import {
   ValidationError

package/dist/{rubric-predictive-validity-DYTLjGWu.d.ts → rubric-predictive-validity-BIdf9h4R.d.ts}

@@ -1,4 +1,4 @@
-import { R as RunRecord } from './run-record-B7RTi_ix.js';
+import { R as RunRecord } from './run-record-CZmcpWPo.js';
 import { b as OutcomeStore } from './outcome-store-rnXLEqSn.js';
 
 /**

package/dist/{run-record-B7RTi_ix.d.ts → run-record-CZmcpWPo.d.ts}

@@ -357,4 +357,4 @@ declare function roundTripRunRecord(record: RunRecord): RunRecord;
  */
 declare function modelHasSnapshot(model: string): boolean;
 
-export { type AgentProfileCell as A, requireAgentProfileCell as B, resolveRunCostProvenance as C, roundTripRunRecord as D, toAgentProfileJson as E, validateAgentProfileCell as F, validateRunRecord as G, verifyAgentProfileCell as H, type JudgeScoresRecord as J, type RunRecord as R, type AgentProfileJson as a, type RunSplitTag as b, type RunCostProvenance as c, type RunTokenUsage as d, type RunJudgeMetadata as e, type AgentProfileCellInput as f, AGENT_PROFILE_KINDS as g, type AgentInterfaceProfileLike as h, type AgentProfileCellSchemaVersion as i, AgentProfileCellValidationError as j, type AgentProfileDimensionValue as k, type AgentProfileHarness as l, type AgentProfileKind as m, type AgentProfileSource as n, type AgentProfileSourceInput as o, type RunOutcome as p, RunRecordValidationError as q, agentProfileCellHashMaterial as r, agentProfileCellKey as s, assertRunAgentProfileCell as t, buildAgentInterfaceProfileCell as u, buildAgentProfileCell as v, groupRunsByAgentProfileCell as w, isRunRecord as x, modelHasSnapshot as y, parseRunRecordSafe as z };
+export { type AgentProfileCell as A, requireAgentProfileCell as B, resolveRunCostProvenance as C, roundTripRunRecord as D, toAgentProfileJson as E, validateAgentProfileCell as F, validateRunRecord as G, verifyAgentProfileCell as H, type JudgeScoresRecord as J, type RunRecord as R, type RunSplitTag as a, type RunCostProvenance as b, type RunTokenUsage as c, type RunJudgeMetadata as d, type AgentProfileCellInput as e, type AgentProfileJson as f, AGENT_PROFILE_KINDS as g, type AgentInterfaceProfileLike as h, type AgentProfileCellSchemaVersion as i, AgentProfileCellValidationError as j, type AgentProfileDimensionValue as k, type AgentProfileHarness as l, type AgentProfileKind as m, type AgentProfileSource as n, type AgentProfileSourceInput as o, type RunOutcome as p, RunRecordValidationError as q, agentProfileCellHashMaterial as r, agentProfileCellKey as s, assertRunAgentProfileCell as t, buildAgentInterfaceProfileCell as u, buildAgentProfileCell as v, groupRunsByAgentProfileCell as w, isRunRecord as x, modelHasSnapshot as y, parseRunRecordSafe as z };

package/dist/{runtime-trajectory-Dws7Kpgi.d.ts → runtime-trajectory-CC0jx9ql.d.ts}

@@ -1,4 +1,4 @@
-import { b as RunSplitTag } from './run-record-B7RTi_ix.js';
+import { a as RunSplitTag } from './run-record-CZmcpWPo.js';
 
 interface RuntimeTrajectoryHookEvent {
     id: string;

package/dist/{semantic-concept-judge-DxJmRkyJ.d.ts → semantic-concept-judge-CKjePUMh.d.ts}

@@ -1,10 +1,10 @@
 import { AxAIService } from '@ax-llm/ax';
 import { z } from 'zod';
-import { A as AnalystFinding, T as TraceAnalystKindSpec, a as Analyst, b as AnalystContext } from './kind-factory-DcNg13sZ.js';
-import { a as TraceAnalystSpan } from './store-C1YxJDEK.js';
+import { d as AnalystFinding, A as Analyst, b as AnalystContext, L as LlmClientOptions } from './policy-edit-Clb2v6Oa.js';
+import { T as TraceAnalystKindSpec } from './kind-factory-DWOvXjR_.js';
+import { a as TraceAnalystSpan } from './store-9cAScOcb.js';
 import { R as Run, S as Span, e as TraceEvent, A as Artifact, B as BudgetLedgerEntry } from './schema-SGWcK9wa.js';
 import { T as TraceStore } from './store-BsVi7ncX.js';
-import { L as LlmClientOptions } from './llm-client-DyqEH4jH.js';
 import { S as Severity } from './multi-layer-verifier-BsqKuLyN.js';
 
 interface CreateAnalystAiConfig {

package/dist/{store-C1YxJDEK.d.ts → store-9cAScOcb.d.ts}

@@ -1,3 +1,134 @@
+/**
+ * RawProviderSink — first-class persistence for the actual HTTP-level
+ * request/response bodies of every LLM provider call.
+ *
+ * Why this is a separate sink from the structured `LlmSpan`:
+ *
+ *   - `LlmSpan` records the *intent* — model name, messages, output text,
+ *     usage. It's what dashboards read; it's NOT enough for forensics.
+ *   - When a downstream consumer reports "the verifier used the wrong route"
+ *     or "tokens look right but reasoning was missing," the only way to
+ *     answer is the raw HTTP body. Span fields can lie (a proxy can echo
+ *     a different `model` value than what actually answered); the raw
+ *     response is ground truth.
+ *
+ * Default behaviour: opt-in. Pass `rawSink` to `LlmClientOptions` (or the
+ * matrix runner / BuilderSession sets it up automatically) and every
+ * request, response, and error is recorded — including retries, with the
+ * attempt index attached so a flaky call's full event chain is recoverable.
+ *
+ * Redaction is enforced at sink time. The default redactor strips
+ * `Authorization`, `X-Api-Key`, `X-Auth-Token`, `Cookie` headers and any
+ * payload field whose key matches `apiKey | api_key | bearer | password |
+ * secret | token` (case-insensitive). Override via the sink constructor or
+ * the per-call `redactor`. The `redactedFields` array on the persisted
+ * event lets a reviewer see what was stripped without exposing the values.
+ */
+type RawProviderDirection = 'request' | 'response' | 'error';
+interface RawProviderEvent {
+    /** Stable id. Generated by the sink if omitted. */
+    eventId: string;
+    /** Trace context populated by `LlmClient` when the call is wrapped in a span. */
+    runId?: string;
+    spanId?: string;
+    /**
+     * Logical provider name. Free-form so callers can use whatever id matches
+     * their topology (`'openai'`, `'anthropic'`, `'tangle-router'`, …). When
+     * omitted, derived from `baseUrl` in `LlmClientOptions`.
+     */
+    provider: string;
+    model: string;
+    /** Endpoint path, e.g. `'/v1/chat/completions'`. */
+    endpoint: string;
+    /** Base URL used for the call (already-normalised — no trailing slash). */
+    baseUrl: string;
+    /** 0-indexed retry attempt. The first attempt is 0; a retried call gets 1, 2, … */
+    attemptIndex: number;
+    direction: RawProviderDirection;
+    /** Unix ms. */
+    timestamp: number;
+    /** Wall-clock duration of the call leg. Set on `response` and `error` events; null on `request`. */
+    durationMs?: number;
+    statusCode?: number;
+    requestHeaders?: Record<string, string>;
+    requestBody?: unknown;
+    responseHeaders?: Record<string, string>;
+    responseBody?: unknown;
+    /** Set on `direction: 'error'` events. */
+    errorMessage?: string;
+    /** Field paths the redactor stripped from this event ('header:Authorization', 'body.apiKey', …). */
+    redactedFields: string[];
+}
+interface RawProviderSinkFilter {
+    runId?: string;
+    spanId?: string;
+    direction?: RawProviderDirection;
+    attemptIndex?: number;
+}
+interface RawProviderSink {
+    record(event: RawProviderEvent): Promise<void>;
+    /** Optional listing — implementations that durably persist (file, db) should support this. */
+    list?(filter?: RawProviderSinkFilter): Promise<RawProviderEvent[]>;
+    /** Optional teardown for backed implementations. */
+    close?(): Promise<void>;
+}
+type ProviderRedactor = (event: RawProviderEvent) => RawProviderEvent;
+/**
+ * Default redactor — strips well-known auth headers and any body field whose
+ * key matches the credential pattern. Records every redacted path on
+ * `event.redactedFields` so a downstream reviewer can see what was removed.
+ */
+declare function defaultProviderRedactor(event: RawProviderEvent): RawProviderEvent;
+interface InMemoryRawProviderSinkOptions {
+    redactor?: ProviderRedactor;
+}
+declare class InMemoryRawProviderSink implements RawProviderSink {
+    private events;
+    private redactor;
+    constructor(opts?: InMemoryRawProviderSinkOptions);
+    record(event: RawProviderEvent): Promise<void>;
+    list(filter?: RawProviderSinkFilter): Promise<RawProviderEvent[]>;
+    size(): number;
+}
+declare class NoopRawProviderSink implements RawProviderSink {
+    record(): Promise<void>;
+    /**
+     * Returns an empty array. Implemented so `assertRunCaptured` does not
+     * trip the `no_raw_sink` issue when a caller explicitly opts out of
+     * capture by passing this sink — opt-out is a deliberate choice, not a
+     * misconfiguration.
+     */
+    list(): Promise<RawProviderEvent[]>;
+}
+interface FileSystemRawProviderSinkOptions {
+    /** Directory the NDJSON file is written into. Created if missing. */
+    dir: string;
+    /** File name; default `'raw-provider-events.ndjson'`. */
+    fileName?: string;
+    /** Bytes after which the writer rolls over to a new file (default 32 MiB). */
+    rollAtBytes?: number;
+    redactor?: ProviderRedactor;
+}
+declare class FileSystemRawProviderSink implements RawProviderSink {
+    private dir;
+    private fileName;
+    private rollAtBytes;
+    private redactor;
+    private bytesWritten;
+    private rollIndex;
+    private initPromise;
+    constructor(opts: FileSystemRawProviderSinkOptions);
+    private ensureInit;
+    private currentPath;
+    record(event: RawProviderEvent): Promise<void>;
+    list(filter?: RawProviderSinkFilter): Promise<RawProviderEvent[]>;
+}
+/**
+ * Best-effort provider id from a base URL. Falls back to the URL host when
+ * none of the well-known patterns match.
+ */
+declare function providerFromBaseUrl(baseUrl: string): string;
+
 /**
  * Shared types for the trace-analyst module.
  *
@@ -245,4 +376,4 @@ interface TraceAnalysisStore {
     }): Promise<SearchSpanResult>;
 }
 
-export { DEFAULT_TRACE_ANALYST_BUDGETS as D, type ErrorCluster as E, type QueryTracesPage as Q, type SearchSpanResult as S, type TraceAnalysisStore as T, type ViewSpansResult as V, type TraceAnalystSpan as a, type DatasetOverview as b, type SearchTraceResult as c, type SpanMatchRecord as d, TRACE_ANALYST_TRUNCATION_MARKER_PREFIX as e, type TraceAnalystByteBudgets as f, type TraceAnalystFilters as g, type TraceAnalystSpanKind as h, type TraceAnalystSpanStatus as i, type TraceAnalystTraceSummary as j, type ViewTraceOversized as k, type ViewTraceResult as l };
+export { DEFAULT_TRACE_ANALYST_BUDGETS as D, type ErrorCluster as E, FileSystemRawProviderSink as F, InMemoryRawProviderSink as I, NoopRawProviderSink as N, type ProviderRedactor as P, type QueryTracesPage as Q, type RawProviderSink as R, type SearchSpanResult as S, type TraceAnalysisStore as T, type ViewSpansResult as V, type TraceAnalystSpan as a, type DatasetOverview as b, type FileSystemRawProviderSinkOptions as c, type InMemoryRawProviderSinkOptions as d, type RawProviderDirection as e, type RawProviderEvent as f, type RawProviderSinkFilter as g, type SearchTraceResult as h, type SpanMatchRecord as i, TRACE_ANALYST_TRUNCATION_MARKER_PREFIX as j, type TraceAnalystByteBudgets as k, type TraceAnalystFilters as l, type TraceAnalystSpanKind as m, type TraceAnalystSpanStatus as n, type TraceAnalystTraceSummary as o, type ViewTraceOversized as p, type ViewTraceResult as q, defaultProviderRedactor as r, providerFromBaseUrl as s };

package/dist/{summary-report-BJ5aNwZ1.d.ts → summary-report-DTNgQycC.d.ts}

@@ -1,4 +1,4 @@
-import { R as RunRecord } from './run-record-B7RTi_ix.js';
+import { R as RunRecord } from './run-record-CZmcpWPo.js';
 import { F as FailureClusterReport } from './failure-cluster-C48PiReX.js';
 
 /**

package/dist/traces.d.ts

@@ -1,19 +1,17 @@
 import { N as NotFoundError, R as ReplayError } from './errors-oeQrLqXC.js';
-import { P as ProviderRedactor, R as RawProviderSink, d as RawProviderEvent } from './raw-provider-sink-C46HDghv.js';
-export { F as FileSystemRawProviderSink, a as FileSystemRawProviderSinkOptions, I as InMemoryRawProviderSink, b as InMemoryRawProviderSinkOptions, N as NoopRawProviderSink, c as RawProviderDirection, e as RawProviderSinkFilter, f as defaultProviderRedactor, p as providerFromBaseUrl } from './raw-provider-sink-C46HDghv.js';
+import { m as TraceAnalystSpanKind, n as TraceAnalystSpanStatus, T as TraceAnalysisStore, l as TraceAnalystFilters, b as DatasetOverview, Q as QueryTracesPage, q as ViewTraceResult, V as ViewSpansResult, h as SearchTraceResult, S as SearchSpanResult, P as ProviderRedactor, R as RawProviderSink, f as RawProviderEvent } from './store-9cAScOcb.js';
+export { D as DEFAULT_TRACE_ANALYST_BUDGETS, E as ErrorCluster, F as FileSystemRawProviderSink, c as FileSystemRawProviderSinkOptions, I as InMemoryRawProviderSink, d as InMemoryRawProviderSinkOptions, N as NoopRawProviderSink, e as RawProviderDirection, g as RawProviderSinkFilter, i as SpanMatchRecord, j as TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, k as TraceAnalystByteBudgets, a as TraceAnalystSpan, o as TraceAnalystTraceSummary, p as ViewTraceOversized, r as defaultProviderRedactor, s as providerFromBaseUrl } from './store-9cAScOcb.js';
 import { a as RunCompleteHookContext, R as RunCompleteHook } from './emitter-BRchAAAx.js';
 export { S as SpanHandle, T as TraceEmitter, b as TraceEmitterOptions, l as llmSpanFromProvider } from './emitter-BRchAAAx.js';
-export { b as RunIntegrityError, R as RunIntegrityExpectations, c as RunIntegrityIssue, d as RunIntegrityIssueCode, a as RunIntegrityReport, e as assertRunCaptured, t as throwIfRunIncomplete } from './integrity-qemeBAyx.js';
+export { b as RunIntegrityError, R as RunIntegrityExpectations, c as RunIntegrityIssue, d as RunIntegrityIssueCode, a as RunIntegrityReport, e as assertRunCaptured, t as throwIfRunIncomplete } from './integrity-C6PZ73iC.js';
 import { T as TraceStore } from './store-BsVi7ncX.js';
 export { E as EventFilter, F as FileSystemTraceStore, a as FileSystemTraceStoreOptions, I as InMemoryTraceStore, R as RunFilter, S as SpanFilter } from './store-BsVi7ncX.js';
 export { a as aggregateLlm, b as argHash, g as groupBy, j as judgeSpans, l as llmSpans, r as runFailureClass, c as runsForScenario, t as toolSpans } from './query-Ck190MOd.js';
 import { R as Run } from './schema-SGWcK9wa.js';
 export { A as Artifact, B as BudgetLedgerEntry, h as BudgetSpec, E as EventKind, i as FAILURE_CLASSES, F as FailureClass, G as GenericSpan, J as JudgeSpan, L as LlmSpan, M as Message, c as RetrievalSpan, g as RunLayer, a as RunOutcome, f as RunStatus, d as SandboxSpan, S as Span, j as SpanBase, b as SpanKind, k as SpanStatus, l as TRACE_SCHEMA_VERSION, T as ToolSpan, e as TraceEvent, m as isJudgeSpan, n as isLlmSpan, o as isRetrievalSpan, p as isSandboxSpan, q as isToolSpan } from './schema-SGWcK9wa.js';
-import { A as AnalyzeTracesOptions, b as AnalyzeTracesResult } from './analyst-C8HHvfJp.js';
-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, d as RunTokenUsage, R as RunRecord } from './run-record-B7RTi_ix.js';
+import { A as AnalyzeTracesOptions, b as AnalyzeTracesResult } from './analyst-CFBc14Wc.js';
+export { a as AnalyzeTracesInput, c as AnalyzeTracesTurnSnapshot, d as analyzeTraces } from './analyst-CFBc14Wc.js';
+import { a as RunSplitTag, c as RunTokenUsage, R as RunRecord } from './run-record-CZmcpWPo.js';
 import { AxFunction } from '@ax-llm/ax';
 import '@tangle-network/agent-interface';
 

package/dist/{types-C5gJrOVT.d.ts → types-Ca_63YSD.d.ts}

@@ -1,4 +1,5 @@
-import { d as RunTokenUsage } from './run-record-B7RTi_ix.js';
+import { P as PolicyEditCandidateRecord } from './policy-edit-Clb2v6Oa.js';
+import { c as RunTokenUsage } from './run-record-CZmcpWPo.js';
 
 /**
  * Pass A substrate types — `runCampaign` is the one primitive every
@@ -168,6 +169,9 @@ interface ProposedCandidate {
      *  primitive it used. Survives to `GenerationCandidate.rationale` and the
      *  emitted provenance record. */
     rationale: string;
+    /** Structured, JSON-safe cause for this exact candidate when the proposer
+     *  can provide one. Policy edits retain the full validated edit here. */
+    candidateRecord?: PolicyEditCandidateRecord;
 }
 /** Type guard: a proposal carrying its rationale vs a bare
  *  surface. The loop branches on this to populate `GenerationCandidate`. */
@@ -194,6 +198,28 @@ interface ParetoParent {
     label?: string;
     rationale?: string;
 }
+/** Exact measured state for the surface an optimizer is learning from.
+ *  Unlike a model-authored expected gain, every value here comes from a
+ *  completed campaign over the designed denominator. */
+interface ScoredSurfaceOutcome {
+    /** Optimization/search evidence only. Held-out results must never flow back
+     *  into a proposer through this type. */
+    split: 'search';
+    /** Generation that actually measured this surface (`-1` for the baseline). */
+    generation: number;
+    surfaceHash: string;
+    composite: number;
+    dimensions: Record<string, number>;
+    scenarios: Array<{
+        scenarioId: string;
+        composite: number;
+        notes?: string;
+    }>;
+    coverage: {
+        expectedCells: number;
+        scorableCells: number;
+    };
+}
 /** Stateless surface mutation — given findings + current
  *  surface, return N candidate surfaces. Pure transform, no generation
  *  awareness. Reflective-mutation and `AxGEPA` mutators conform. Wrapped by
@@ -221,6 +247,12 @@ interface ProposeContext<TFindings = unknown> {
     populationSize: number;
     generation: number;
     signal: AbortSignal;
+    /** Measured baseline for this optimization run. `runOptimization` always
+     *  supplies it; optional for standalone proposer callers. */
+    baselineOutcome?: ScoredSurfaceOutcome;
+    /** Measured result for `currentSurface`, the complete global incumbent every
+     *  new candidate mutates. `runOptimization` always supplies it. */
+    incumbentOutcome?: ScoredSurfaceOutcome;
     /** Optional analysis report produced before proposal. Opaque to the substrate:
      *  the proposer that consumes it owns the shape. */
     report?: unknown;
@@ -517,6 +549,29 @@ interface GenerationCandidate {
     surfaceHash: string;
     composite: number;
     ci95: [number, number];
+    /** Exact surface this candidate mutated. */
+    parentSurfaceHash?: string;
+    /** Measured search-split composite of the exact parent surface. */
+    parentComposite?: number;
+    /** Candidate composite minus its parent's composite. Present only when the
+     *  candidate completed the designed denominator. */
+    observedDeltaFromParent?: number;
+    /** Whether this candidate had a scorable result for every designed campaign
+     *  cell and was therefore eligible for ranking, promotion, and Pareto
+     *  selection. Older externally-authored records may omit this field; loop
+     *  records always populate it. */
+    eligibleForPromotion?: boolean;
+    /** Exact denominator receipt for selection eligibility. Scores stay
+     *  descriptive: an incomplete candidate is retained with its observed score
+     *  and errors instead of receiving an invented penalty. */
+    coverage?: {
+        expectedCells: number;
+        scorableCells: number;
+        unscorableCells: Array<{
+            cellId: string;
+            reason: string;
+        }>;
+    };
     /** Mean score per judge dimension across all cells (scenarios × reps ×
      *  judges that reported the dimension). */
     dimensions: Record<string, number>;
@@ -538,6 +593,8 @@ interface GenerationCandidate {
      *  "because rationale Z" the audit requires to survive to the result.
      *  Present when the proposer returned a `ProposedCandidate`. */
     rationale?: string;
+    /** Exact structured cause threaded from the proposer, when available. */
+    candidateRecord?: PolicyEditCandidateRecord;
 }
 interface CampaignAggregates {
     byJudge: Record<string, JudgeAggregate>;
@@ -572,4 +629,4 @@ interface CampaignResult<TArtifact = unknown, TScenario extends Scenario = Scena
     scenarios: Array<Pick<TScenario, 'id' | 'kind'>>;
 }
 
-export { type JudgeAggregate as A, type ScenarioAggregate as B, type CampaignResult as C, type DispatchContext as D, isProposedCandidate as E, labelTrustRank as F, type GateResult as G, type JudgeScore as J, type LabeledScenarioStore as L, type MutableSurface as M, type OptimizationProposer as O, type ParetoParent as P, type RedactionStatus as R, type Scenario as S, type TraceSpan as T, type JudgeDimension as a, type JudgeConfig as b, type DispatchFn as c, type CampaignTraceWriter as d, type GenerationRecord as e, type SurfaceProposer as f, type Gate as g, type GateDecision as h, type CampaignAggregates as i, type CampaignArtifactWriter as j, type CampaignCellResult as k, type CampaignCostMeter as l, type CodeSurface as m, type GateContext as n, type GenerationCandidate as o, type Mutator as p, type OptimizerConfig as q, type SessionScript as r, type LabeledScenarioWrite as s, type LabeledScenarioSampleArgs as t, type LabeledScenarioRecord as u, type LabelTrust as v, type ProposedCandidate as w, type ProposeContext as x, type LabeledScenarioSource as y, type CampaignTokenUsage as z };
+export { type JudgeAggregate as A, type ScenarioAggregate as B, type CampaignResult as C, type DispatchContext as D, type ScoredSurfaceOutcome as E, isProposedCandidate as F, type GateResult as G, labelTrustRank as H, type JudgeScore as J, type LabeledScenarioStore as L, type MutableSurface as M, type OptimizationProposer as O, type ParetoParent as P, type RedactionStatus as R, type Scenario as S, type TraceSpan as T, type JudgeDimension as a, type JudgeConfig as b, type DispatchFn as c, type CampaignTraceWriter as d, type GenerationRecord as e, type SurfaceProposer as f, type Gate as g, type GateDecision as h, type CampaignAggregates as i, type CampaignArtifactWriter as j, type CampaignCellResult as k, type CampaignCostMeter as l, type CodeSurface as m, type GateContext as n, type GenerationCandidate as o, type Mutator as p, type OptimizerConfig as q, type SessionScript as r, type LabeledScenarioWrite as s, type LabeledScenarioSampleArgs as t, type LabeledScenarioRecord as u, type LabelTrust as v, type ProposedCandidate as w, type ProposeContext as x, type LabeledScenarioSource as y, type CampaignTokenUsage as z };

package/dist/wire/index.js

@@ -34,8 +34,8 @@ import {
   runRpcOnce,
   startServer,
   startServerAsync
-} from "../chunk-DRPIZQIT.js";
-import "../chunk-FUCQVFMU.js";
+} from "../chunk-4D5RVB3W.js";
+import "../chunk-GY4SYVPJ.js";
 import "../chunk-PC4UYEBM.js";
 import "../chunk-ONWEPEDO.js";
 import "../chunk-PZ5AY32C.js";

package/docs/design/loop-taxonomy.md

@@ -201,8 +201,7 @@ consume the **same dataset** the flywheel builds.
   agent-runtime.
 - **runCampaign** — a measurement: a surface scored over N scenarios × M reps.
   agent-eval. (A "campaign" = a coordinated batch of measurements.)
-- **runOptimization** — the improvement loop body: proposer suggests surfaces,
-  each measured by a campaign, top-K promoted per generation. agent-eval.
+- **runOptimization** — the improvement loop body: proposer suggests surfaces, each is measured, and only a candidate that beats the global incumbent is promoted. agent-eval.
 - **runImprovementLoop** — `runOptimization` + holdout re-score + release gate
   + optional PR. agent-eval.
 - **runAnalystLoop** — reflective autoresearch: findings + knowledge updates +

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-eval",
-  "version": "0.115.2",
+  "version": "0.116.0",
   "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": {