@tangle-network/agent-eval 0.120.4 → 0.122.0

package/dist/fuzz.js

@@ -1,7 +1,7 @@
 import {
   CostLedger,
   CostReceiptCaptureError
-} from "./chunk-ZMXDQ4K7.js";
+} from "./chunk-JHOJHHU7.js";
 import "./chunk-VI2UW6B6.js";
 import {
   varianceBasedCurriculum

package/dist/index.d.ts

@@ -7830,11 +7830,15 @@ interface GenerationCandidate {
      *  reflective proposer grounds its next edit on. Keep `notes` GENERALIZABLE
      *  (which checks/lines/dimensions failed and how), NOT case-specific ground
      *  truth: leaking expected answers into the prompt is memorization, and the
-     *  held-out gate would reject it anyway. */
+     *  held-out gate would reject it anyway. `emitted` is a bounded excerpt of
+     *  the candidate's raw output for the scenario (worst rep when reps > 1) —
+     *  the "what it actually did" evidence; optional so trajectory capture is
+     *  never required of a dispatch. */
     scenarios: Array<{
         scenarioId: string;
         composite: number;
         notes?: string;
+        emitted?: string;
     }>;
     /** Proposer-supplied short label for the change. Present when the proposer
      *  returned a `ProposedCandidate`; absent for bare-surface mutators. */
@@ -13223,6 +13227,7 @@ interface IntentMatchOptions {
     llm?: LlmClientOptions;
     costLedger?: CostLedgerHandle;
     costPhase?: string;
+    costTags?: Record<string, string>;
     signal?: AbortSignal;
 }
 /**
@@ -13855,6 +13860,7 @@ interface SemanticConceptJudgeOptions {
     llm?: LlmClientOptions;
     costLedger?: CostLedgerHandle;
     costPhase?: string;
+    costTags?: Record<string, string>;
     signal?: AbortSignal;
     /**
      * Score aggregation strategy. Default `mean` — uniform average across
@@ -15253,24 +15259,6 @@ interface ReferenceReplaySteeringRowsOptions<Input = unknown> {
 declare function referenceReplayRunsToSteeringRows<Input = unknown>(runs: ReferenceReplayRun<Input>[], options?: ReferenceReplaySteeringRowsOptions<Input>): SteeringOptimizationRow[];
 declare function referenceReplayScenarioToRunScore(scenarioScore: ReferenceReplayScenarioScore, durationMs?: number): RunScore;
 
-/**
- * Reflective mutation — primitives for trace-conditioned prompt rewriting.
- *
- * Used by `prompt-evolution.ts` (and any consumer running iterative
- * improvement). Given a parent prompt + concrete trace evidence (top trials,
- * bottom trials, missed expectations), produce an LLM-ready prompt that
- * proposes targeted mutations — not blind rephrasings.
- *
- * Why this lives outside `prompt-evolution.ts`: any consumer that wants to
- * run reflective rewriting WITHOUT the population/Pareto machinery can
- * import these primitives directly.
- *
- * Quality bar (vs. naive "mutate this prompt"):
- *   - Show parent ↔ children diff, not just one variant
- *   - Quote specific missed goldens with their match phrases
- *   - Surface the model's actual emitted output side-by-side with what was expected
- *   - Quote concrete mutation primitives so the model has a vocabulary
- */
 interface TrialTrace {
     /** Stable id for the trial — surfaces in the prompt for grounding. */
     id: string;

package/dist/index.js

@@ -9,12 +9,12 @@ import {
   checkBehavioralCanary,
   checkCanaries,
   runBehavioralCanaries
-} from "./chunk-SQQED7ZH.js";
+} from "./chunk-3A246TSA.js";
 import {
   BENCHMARK_SPLIT_SEED,
   benchmarks_exports,
   deterministicSplit
-} from "./chunk-CVJP5TMD.js";
+} from "./chunk-PICTDURQ.js";
 import {
   DEFAULT_RULES,
   buildTrajectory,
@@ -43,15 +43,17 @@ import {
   agentProfileHash,
   agentProfileId,
   agentProfileModelId,
+  comparePairedArms,
   completionVerdict,
   createLlmCorrectnessChecker,
   createTokenRecallChecker,
   expandProfileAxes,
   extractProducedState,
   harnessAxisOf,
+  pairArms,
   parseCorrectnessResponse,
   verifyCompletion
-} from "./chunk-ZTJPIIVI.js";
+} from "./chunk-TDHX3BVA.js";
 import {
   DEFAULT_MUTATION_PRIMITIVES,
   DEFAULT_RED_TEAM_CORPUS,
@@ -65,7 +67,6 @@ import {
   campaignMeanComposite,
   codeExecutionJudge,
   coherenceJudge,
-  comparePairedArms,
   costReceiptFromTCloud,
   createCustomJudge,
   createDomainExpertJudge,
@@ -78,7 +79,6 @@ import {
   heldOutGate,
   llmJudge,
   maximumChargeForTCloudRequest,
-  pairArms,
   paretoFrontier,
   paretoFrontierWithCrowding,
   parseReflectionResponse,
@@ -91,7 +91,7 @@ import {
   scoreRedTeamOutput,
   surfaceContentHash,
   toolNamesForRun
-} from "./chunk-7QBFOJ3E.js";
+} from "./chunk-VZ6VKOJT.js";
 import {
   BackendIntegrityError,
   assertRealAgentReceipts,
@@ -103,7 +103,7 @@ import {
   inMemoryVerdictCache,
   summarizeAgentReceiptIntegrity,
   summarizeBackendIntegrity
-} from "./chunk-SYV364BL.js";
+} from "./chunk-XDIRG3TO.js";
 import {
   DEFAULT_COMPLEXITY_WEIGHTS,
   FindingsStore,
@@ -116,12 +116,12 @@ import {
   defaultIsMaterial,
   diffFindings,
   runSemanticConceptJudge
-} from "./chunk-3XH4Y2SS.js";
+} from "./chunk-P62I523A.js";
 import {
   buildDefaultAnalystRegistry,
   computeTraceMetrics,
   createChatClient
-} from "./chunk-5BYTIDZ7.js";
+} from "./chunk-QWMPPZ3X.js";
 import "./chunk-HHWE3POT.js";
 import {
   Mutex
@@ -157,7 +157,7 @@ import {
   scorePolicyEditReadiness,
   validatePolicyEdit,
   validatePolicyEditCandidateRecord
-} from "./chunk-5CVUPHJ4.js";
+} from "./chunk-S5TT5R3L.js";
 import {
   allCriticalPassed,
   controlFailureClassFromVerification,
@@ -267,7 +267,7 @@ import {
   CostReservationExceededError,
   costForUsage,
   modelPriceKey
-} from "./chunk-ZMXDQ4K7.js";
+} from "./chunk-JHOJHHU7.js";
 import {
   MODEL_PRICING,
   MetricsCollector,
@@ -8119,6 +8119,7 @@ async function runIntentMatchJudge(input, options = {}) {
     llm: options.llm ?? {},
     costLedger: options.costLedger ?? new CostLedger(),
     costPhase: options.costPhase ?? "judge.intent-match",
+    costTags: options.costTags ?? {},
     signal: options.signal ?? new AbortController().signal
   };
   if (input.sourceFiles.length === 0 && !input.servedHtml) {
@@ -8154,6 +8155,7 @@ async function runIntentMatchJudge(input, options = {}) {
       phase: opts.costPhase,
       actor: "intent-match",
       model: opts.model,
+      ...Object.keys(opts.costTags).length > 0 ? { tags: opts.costTags } : {},
       maximumCharge: maximumChargeForLlmRequest(request, opts.llm),
       signal: opts.signal,
       execute: (signal, callId) => callLlmJson(request, {