@tangle-network/agent-eval 0.105.0 → 0.106.1

package/dist/adapters/http.d.ts

@@ -1,4 +1,4 @@
-import { S as Scenario, D as DispatchFn, b as DispatchContext } from '../types-Mh6I44ub.js';
+import { S as Scenario, D as DispatchFn, b as DispatchContext } from '../types-Ctf7XIAL.js';
 import '../run-record-I-Z3JNvO.js';
 import '@tangle-network/agent-interface';
 import '../errors-oeQrLqXC.js';

package/dist/adapters/langchain.d.ts

@@ -1,4 +1,4 @@
-import { S as Scenario, J as JudgeScore, D as DispatchFn, a as JudgeConfig } from '../types-Mh6I44ub.js';
+import { S as Scenario, J as JudgeScore, D as DispatchFn, a as JudgeConfig } from '../types-Ctf7XIAL.js';
 import '../run-record-I-Z3JNvO.js';
 import '@tangle-network/agent-interface';
 import '../errors-oeQrLqXC.js';

package/dist/adapters/otel.d.ts

@@ -1,5 +1,5 @@
 import { TraceSpanEvent, HostedClient } from '../hosted/index.js';
-import '../types-Mh6I44ub.js';
+import '../types-Ctf7XIAL.js';
 import '../run-record-I-Z3JNvO.js';
 import '@tangle-network/agent-interface';
 import '../errors-oeQrLqXC.js';

package/dist/campaign/index.d.ts

@@ -1,11 +1,11 @@
-import { S as SignedManifest, B as BackendIntegrityReport, C as CompletionRequirement, R as RuntimeEventLike, a as CompletionVerdict, P as ProducedState, b as CorrectnessChecker } from '../pre-registration-D8Zr8-rr.js';
-export { L as LlmJudgeDimension, c as LlmJudgeOptions, l as llmJudge } from '../pre-registration-D8Zr8-rr.js';
+import { S as SignedManifest, B as BackendIntegrityReport, C as CompletionRequirement, R as RuntimeEventLike, a as CompletionVerdict, P as ProducedState, b as CorrectnessChecker } from '../pre-registration-BkkTmtQG.js';
+export { L as LlmJudgeDimension, c as LlmJudgeOptions, l as llmJudge } from '../pre-registration-BkkTmtQG.js';
 import { A as AnalyzeTracesOptions, a as AnalyzeTracesInput, b as AnalyzeTracesResult } from '../analyst-C8HHvfJp.js';
-import { S as Scenario, M as MutableSurface, b as DispatchContext, a as JudgeConfig, e as GenerationRecord, g as Gate, J as JudgeScore, L as LabeledScenarioStore, s as LabeledScenarioWrite, t as LabeledScenarioSampleArgs, u as LabeledScenarioRecord, v as LabelTrust, f as SurfaceProposer, w as ProposedCandidate, x as ProposeContext, y as LabeledScenarioSource, C as CampaignResult, o as CodeSurface } from '../types-Mh6I44ub.js';
-export { k as CampaignAggregates, l as CampaignArtifactWriter, m as CampaignCellResult, n as CampaignCostMeter, z as CampaignTokenUsage, d as CampaignTraceWriter, D as DispatchFn, h as GateContext, j as GateDecision, G as GateResult, p as GenerationCandidate, A as JudgeAggregate, c as JudgeDimension, i as Mutator, O as OptimizationProposer, q as OptimizerConfig, P as ParetoParent, R as RedactionStatus, B as ScenarioAggregate, r as SessionScript, T as TraceSpan, E as isProposedCandidate, F as labelTrustRank } from '../types-Mh6I44ub.js';
-import { e as CampaignRunPlan, P as PlanCampaignRunOptions, C as CampaignStorage, R as RunCampaignOptions, c as RunImprovementLoopOptions } from '../gepa-Dn0gtj9n.js';
-export { h as CampaignRunPlanCell, j as GepaProposerConstraints, G as GepaProposerOptions, O as OpenAutoPrOptions, k as OpenAutoPrResult, b as RunImprovementLoopResult, a as RunOptimizationOptions, l as RunOptimizationResult, m as countSentenceEdits, n as defaultRenderDiff, o as extractH2Sections, f as fsCampaignStorage, g as gepaProposer, i as inMemoryCampaignStorage, p as openAutoPr, q as planCampaignRun, r as runCampaign, d as runImprovementLoop, s as runOptimization, t as surfaceHash } from '../gepa-Dn0gtj9n.js';
-export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, k as BuildLoopProvenanceArgs, D as DefaultProductionGateOptions, l as EmitLoopProvenanceArgs, m as EmitLoopProvenanceResult, E as EvidenceVector, b as EvolutionaryProposerOptions, H as HeldOutGateOptions, n as LoopProvenanceBackend, o as LoopProvenanceCandidate, L as LoopProvenanceRecord, O as ObjectiveSource, P as ParetoSignificanceGateOptions, c as PromotionObjective, d as PromotionPolicy, R as RunEvalOptions, e as buildEvidenceVector, q as buildLoopProvenanceRecord, f as composeGate, g as defaultProductionGate, s as emitLoopProvenance, h as evolutionaryProposer, i as heldOutGate, t as loopProvenanceSpans, p as paretoPolicy, j as paretoSignificanceGate, u as provenanceRecordPath, v as provenanceSpansPath, r as runEval, w as surfaceContentHash } from '../provenance-phkRBl4h.js';
+import { S as Scenario, M as MutableSurface, b as DispatchContext, a as JudgeConfig, e as GenerationRecord, g as Gate, J as JudgeScore, L as LabeledScenarioStore, s as LabeledScenarioWrite, t as LabeledScenarioSampleArgs, u as LabeledScenarioRecord, v as LabelTrust, f as SurfaceProposer, w as ProposedCandidate, x as ProposeContext, y as LabeledScenarioSource, C as CampaignResult, m as CodeSurface } from '../types-Ctf7XIAL.js';
+export { i as CampaignAggregates, j as CampaignArtifactWriter, k as CampaignCellResult, l as CampaignCostMeter, z as CampaignTokenUsage, d as CampaignTraceWriter, D as DispatchFn, n as GateContext, h as GateDecision, G as GateResult, o as GenerationCandidate, A as JudgeAggregate, c as JudgeDimension, p as Mutator, O as OptimizationProposer, q as OptimizerConfig, P as ParetoParent, R as RedactionStatus, B as ScenarioAggregate, r as SessionScript, T as TraceSpan, E as isProposedCandidate, F as labelTrustRank } from '../types-Ctf7XIAL.js';
+import { e as CampaignRunPlan, P as PlanCampaignRunOptions, C as CampaignStorage, b as RunCampaignOptions, c as RunImprovementLoopOptions } from '../gepa-DQ3ruj18.js';
+export { h as CampaignRunPlanCell, j as GepaProposerConstraints, G as GepaProposerOptions, O as OpenAutoPrOptions, k as OpenAutoPrResult, a as RunImprovementLoopResult, R as RunOptimizationOptions, l as RunOptimizationResult, m as countSentenceEdits, n as defaultRenderDiff, o as extractH2Sections, f as fsCampaignStorage, g as gepaProposer, i as inMemoryCampaignStorage, p as openAutoPr, q as planCampaignRun, r as runCampaign, d as runImprovementLoop, s as runOptimization, t as surfaceHash } from '../gepa-DQ3ruj18.js';
+export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, l as BuildLoopProvenanceArgs, D as DefaultProductionGateOptions, m as EmitLoopProvenanceArgs, n as EmitLoopProvenanceResult, E as EvidenceVector, b as EvolutionaryProposerOptions, H as HeldOutGateOptions, o as LoopProvenanceBackend, q as LoopProvenanceCandidate, L as LoopProvenanceRecord, O as ObjectiveSource, c as ParetoSignificanceGateOptions, P as PowerPreflight, s as PowerPreflightOptions, d as PromotionObjective, e as PromotionPolicy, R as RunEvalOptions, f as buildEvidenceVector, t as buildLoopProvenanceRecord, g as composeGate, h as defaultProductionGate, u as emitLoopProvenance, i as evolutionaryProposer, j as heldOutGate, v as loopProvenanceSpans, p as paretoPolicy, k as paretoSignificanceGate, w as powerPreflight, x as provenanceRecordPath, y as provenanceSpansPath, r as runEval, z as surfaceContentHash } from '../provenance-BGPBEFoV.js';
 import { E as EProcessState, a as PairedBootstrapResult } from '../statistics-D88peojY.js';
 import { L as LlmClientOptions } from '../llm-client-DyqEH4jH.js';
 import { AgentProfile } from '@tangle-network/agent-interface';
@@ -391,7 +391,17 @@ interface PairedHoldout {
 declare function pairHoldout(candidate: Map<string, Record<string, JudgeScore>>, baseline: Map<string, Record<string, JudgeScore>>, scenarioIds: Set<string>, select: (s: JudgeScore) => number | undefined): PairedHoldout;
 interface HeldoutSignificance {
     paired: PairedHoldout;
+    /** The bootstrap the ship decision keys on — of the MEAN paired delta by
+     *  default (see the tie note on `heldoutSignificance`). */
     bootstrap: PairedBootstrapResult;
+    /** The MEDIAN paired-delta bootstrap, reported as a diagnostic. When many
+     *  scenarios are tied (both sides solve them), the median is pinned near 0
+     *  regardless of the mean lift — comparing the two exposes tie-domination. */
+    medianBootstrap: PairedBootstrapResult;
+    /** Fraction of paired observations that are exact ties (|delta| < 1e-9). A
+     *  high tie fraction is WHY a median-based gate would have missed a real lift;
+     *  it is the observability the tie fix adds. */
+    tieFraction: number;
     /** n paired observations. */
     n: number;
     /** True iff n >= minProductiveRuns AND the CI lower bound clears the threshold. */

package/dist/campaign/index.js

@@ -2,15 +2,12 @@ import {
   buildEvidenceVector,
   composeGate,
   defaultProductionGate,
-  detectScale,
-  dimensionRegressions,
   evolutionaryProposer,
-  heldoutSignificance,
-  pairHoldout,
   paretoPolicy,
   paretoSignificanceGate,
+  powerPreflight,
   runEval
-} from "../chunk-Z7FK73R5.js";
+} from "../chunk-IA3X6CSF.js";
 import {
   HARNESS_NATIVE_MODEL,
   agentProfileHash,
@@ -20,21 +17,25 @@ import {
   harnessAxisOf,
   llmJudge,
   verifyCompletion
-} from "../chunk-BBQILVBH.js";
+} from "../chunk-2HLM4SJJ.js";
 import {
   buildLoopProvenanceRecord,
   campaignBreakdown,
   campaignMeanComposite,
   countSentenceEdits,
   defaultRenderDiff,
+  detectScale,
+  dimensionRegressions,
   emitLoopProvenance,
   extractH2Sections,
   gepaProposer,
   heldOutGate,
+  heldoutSignificance,
   isProposedCandidate,
   labelTrustRank,
   loopProvenanceSpans,
   openAutoPr,
+  pairHoldout,
   provenanceRecordPath,
   provenanceSpansPath,
   renderAnalystEvidence,
@@ -42,7 +43,7 @@ import {
   runOptimization,
   surfaceContentHash,
   surfaceHash
-} from "../chunk-ZKIL3VOD.js";
+} from "../chunk-3E5KUXYZ.js";
 import {
   assertRealBackend,
   contentHash,
@@ -2778,6 +2779,7 @@ export {
   planCampaignRun,
   planEvalFixtureRun,
   policyEditProposer,
+  powerPreflight,
   provenanceRecordPath,
   provenanceSpansPath,
   renderScoreboardMarkdown,