@tangle-network/agent-eval 0.128.2 → 0.129.0

package/dist/index.js

@@ -9,11 +9,15 @@ import {
   checkBehavioralCanary,
   checkCanaries,
   runBehavioralCanaries
-} from "./chunk-NACAGYSY.js";
+} from "./chunk-M4YBQKIJ.js";
 import {
-  mintRolloutRows,
-  rolloutReward
-} from "./chunk-IHQDPH7D.js";
+  ATIF_SCHEMA_VERSION,
+  HARBOR_IMPORT_GAP,
+  fromHarborTrajectory,
+  relabelImportedSplit,
+  toHarborTrajectories,
+  toHarborTrajectory
+} from "./chunk-RXHCETDZ.js";
 import {
   SUPERVISOR_RUN_SCHEMA,
   analyzeSupervisorRun,
@@ -27,25 +31,14 @@ import {
   showMeasured,
   supervisorRunRolloutLines,
   writeSupervisorRunReport
-} from "./chunk-TSN7JT6D.js";
-import "./chunk-VBQ3CRKH.js";
-import {
-  toJsonl,
-  toRewardRows,
-  toSftRows
-} from "./chunk-EJGRPCO3.js";
-import {
-  ROLLOUT_SCHEMA,
-  assertRolloutLine,
-  isRolloutLine,
-  isTrainableSplit,
-  validateRolloutLine
-} from "./chunk-UWZZKKU7.js";
+} from "./chunk-X4YIBDER.js";
+import "./chunk-HPWUNB47.js";
+import "./chunk-3OCR4R5I.js";
 import {
   BENCHMARK_SPLIT_SEED,
   benchmarks_exports,
   deterministicSplit
-} from "./chunk-XPRT64IE.js";
+} from "./chunk-IYCLP2N2.js";
 import {
   DEFAULT_RULES,
   classifyFailure,
@@ -53,10 +46,7 @@ import {
   computeToolUseMetrics,
   iqr,
   welchsTTest
-} from "./chunk-P5W7RQKK.js";
-import {
-  buildTrajectory
-} from "./chunk-RZTMDUO7.js";
+} from "./chunk-FXTVJPYD.js";
 import {
   analyzeSeries
 } from "./chunk-BOD4O7OF.js";
@@ -84,7 +74,7 @@ import {
   harnessAxisOf,
   parseCorrectnessResponse,
   verifyCompletion
-} from "./chunk-UB2LOJ6Q.js";
+} from "./chunk-QB6BDBP2.js";
 import {
   DEFAULT_MUTATION_PRIMITIVES,
   DEFAULT_RED_TEAM_CORPUS,
@@ -93,20 +83,12 @@ import {
   JudgeParseError,
   REFERENCE_EQUIVALENCE_INPUT_LIMITS,
   REFERENCE_EQUIVALENCE_JUDGE_VERSION,
-  adversarialJudge,
   buildReflectionPrompt,
-  codeExecutionJudge,
-  coherenceJudge,
-  costReceiptFromTCloud,
-  createCustomJudge,
-  createDomainExpertJudge,
   createReferenceEquivalenceJudge,
   crowdingDistance,
-  defaultJudges,
   dominates,
   hashScenarios,
   llmJudge,
-  maximumChargeForTCloudRequest,
   paretoFrontier,
   paretoFrontierWithCrowding,
   parseReflectionResponse,
@@ -118,7 +100,7 @@ import {
   scoreRedTeamOutput,
   surfaceContentHash,
   toolNamesForRun
-} from "./chunk-NKAGIDE2.js";
+} from "./chunk-2QU3YOPR.js";
 import {
   BackendIntegrityError,
   assertRealAgentReceipts,
@@ -130,7 +112,7 @@ import {
   inMemoryVerdictCache,
   summarizeAgentReceiptIntegrity,
   summarizeBackendIntegrity
-} from "./chunk-EZJEIH2R.js";
+} from "./chunk-C6LXANRU.js";
 import {
   DEFAULT_COMPLEXITY_WEIGHTS,
   FindingsStore,
@@ -143,7 +125,7 @@ import {
   defaultIsMaterial,
   diffFindings,
   runSemanticConceptJudge
-} from "./chunk-ZUUWPZCV.js";
+} from "./chunk-DODXQREJ.js";
 import {
   AnalystRegistry,
   DEFAULT_TRACE_ANALYST_KINDS,
@@ -160,7 +142,7 @@ import {
   makeFinding,
   renderPriorFindings,
   renderUpstreamFindings
-} from "./chunk-DJKY2TSY.js";
+} from "./chunk-BSO5JDQH.js";
 import "./chunk-HHWE3POT.js";
 import {
   DEFAULT_RUN_SCORE_WEIGHTS,
@@ -189,19 +171,39 @@ import {
   stopOnNoProgress,
   stopOnRepeatedAction,
   subjectiveEval
-} from "./chunk-BYT7ELPS.js";
+} from "./chunk-WVATSFCP.js";
 import {
   assertReleaseConfidence,
   bootstrapCi,
   evaluateReleaseConfidence,
   judgeReplayGate,
   renderReleaseReport
-} from "./chunk-XDWDC2MP.js";
+} from "./chunk-JQSF5DQT.js";
 import {
   runEvalCampaign
-} from "./chunk-TBL77AUT.js";
-import "./chunk-NYLOYM6N.js";
-import "./chunk-2MKQIFS4.js";
+} from "./chunk-YQN4ICPP.js";
+import {
+  mintRolloutRows
+} from "./chunk-H23X7XKK.js";
+import {
+  toJsonl,
+  toRewardRows,
+  toSftRows
+} from "./chunk-OWN5NPMC.js";
+import {
+  ROLLOUT_SCHEMA,
+  assertMinted,
+  assertMintedLines,
+  assertRolloutLine,
+  isRolloutLine,
+  isTrainableSplit,
+  validateRolloutLine
+} from "./chunk-PC5DOSM7.js";
+import {
+  buildTrajectory
+} from "./chunk-RZTMDUO7.js";
+import "./chunk-EG66UGL4.js";
+import "./chunk-E7QXT7SX.js";
 import {
   LlmCallError,
   LlmClient,
@@ -217,7 +219,7 @@ import {
   maximumChargeForLlmRequest,
   probeLlm,
   stripFencedJson
-} from "./chunk-PBE2LOSS.js";
+} from "./chunk-SFLLL76A.js";
 import {
   evaluateInterimReleaseConfidence,
   pairedEvalueSequence
@@ -228,12 +230,12 @@ import {
   paretoChart,
   researchReport,
   summaryTable
-} from "./chunk-VLOATJQ2.js";
+} from "./chunk-TJVT4QFF.js";
 import {
   comparePairedArms,
   pairArms,
   pairRunRecords
-} from "./chunk-DPUHNQLN.js";
+} from "./chunk-7FO3TNPI.js";
 import {
   benjaminiHochberg,
   bonferroni,
@@ -275,7 +277,7 @@ import {
   weightedMean,
   wilcoxonSignedRank,
   wilson
-} from "./chunk-MHELPNRP.js";
+} from "./chunk-ZHTZ4EYI.js";
 import {
   CostAccountingIncompleteError,
   CostCallConflictError,
@@ -287,7 +289,7 @@ import {
   costForTokenPricing,
   costForUsage,
   modelPriceKey
-} from "./chunk-WS3NZZQQ.js";
+} from "./chunk-VCZ5FQYW.js";
 import {
   MODEL_PRICING,
   MetricsCollector,
@@ -298,8 +300,6 @@ import {
   resolveModelPricing
 } from "./chunk-VI2UW6B6.js";
 import {
-  FileSystemTraceStore,
-  InMemoryTraceStore,
   OTEL_AGENT_EVAL_SCOPE,
   ReplayCache,
   ReplayCacheMissError,
@@ -326,7 +326,7 @@ import {
   scoreTraceInsightReadiness,
   tokenizeDomainWords,
   traceAnalystOnRunComplete
-} from "./chunk-EOSZT7PL.js";
+} from "./chunk-U4L7JRPZ.js";
 import "./chunk-7ZZMD7UK.js";
 import {
   extractUsage,
@@ -377,10 +377,12 @@ import {
   traceSpanKindToOpenInferenceKind
 } from "./chunk-P6FYH6K4.js";
 import {
+  FileSystemTraceStore,
+  InMemoryTraceStore,
   RunIntegrityError,
   assertRunCaptured,
   throwIfRunIncomplete
-} from "./chunk-TT4KNT67.js";
+} from "./chunk-U4PHLT2N.js";
 import {
   FileSystemRawProviderSink,
   InMemoryRawProviderSink,
@@ -417,7 +419,7 @@ import {
   validateRunRecord,
   verifyAgentProfileCell,
   verifyManifest
-} from "./chunk-2JX3CFMB.js";
+} from "./chunk-56TAVBOK.js";
 import {
   FAILURE_CLASSES,
   TRACE_SCHEMA_VERSION,
@@ -427,6 +429,14 @@ import {
   isSandboxSpan,
   isToolSpan
 } from "./chunk-MA6HLL3S.js";
+import {
+  isRealnessGated,
+  observedScore,
+  observedSplitScore,
+  scoreOrigin,
+  trainingReward,
+  trainingScore
+} from "./chunk-OIUOT4QD.js";
 import {
   AgentEvalError,
   CaptureIntegrityError,
@@ -764,15 +774,11 @@ function renderCommitMessage(input) {
 }
 
 // src/executor.ts
-function describeShape(content, message) {
-  if (message === void 0 || message === null) return "no choices[0].message";
-  return `message.content of type ${content === null ? "null" : typeof content}`;
-}
 function errMessage(err) {
   if (err instanceof Error) return `${err.name}: ${err.message}`;
   return String(err);
 }
-async function executeScenario(tc, scenario, config) {
+async function executeScenario(chat, scenario, config) {
   const startTime = Date.now();
   const model = config.model ?? "gpt-4o";
   const costLedger = config.costLedger ?? new CostLedger();
@@ -803,19 +809,18 @@ async function executeScenario(tc, scenario, config) {
       phase: config.costPhase ?? "benchmark.agent",
       actor: "scenario-agent",
       model,
-      maximumCharge: maximumChargeForTCloudRequest(request, config.tcloudMaximumAttempts),
+      maximumCharge: chat.maximumAttempts === void 0 ? void 0 : maximumChargeForLlmRequest(request, { maximumAttempts: chat.maximumAttempts }),
       tags: costTags,
       signal: config.signal,
-      execute: () => tc.chat(request),
-      receipt: (response) => costReceiptFromTCloud(response, model)
+      execute: (signal, callId) => chat.chat(request, { signal, idempotencyKey: callId }),
+      receipt: costReceiptFromLlm,
+      receiptFromError: costReceiptFromLlmError
     });
     if (!paid.succeeded) throw paid.error;
-    const resp = paid.value;
-    const message = resp.choices?.[0]?.message;
-    const rawContent = message?.content;
-    if (message === void 0 || message === null || typeof rawContent !== "string") {
+    const rawContent = paid.value.content;
+    if (typeof rawContent !== "string") {
       throw new CaptureIntegrityError(
-        `chat response for scenario "${scenario.id}" turn ${i} is malformed: expected choices[0].message.content to be a string, got ${describeShape(rawContent, message)}`
+        `chat response for scenario "${scenario.id}" turn ${i} is malformed: expected content to be a string, got ${rawContent === null ? "null" : typeof rawContent}`
       );
     }
     const content = rawContent;
@@ -903,8 +908,7 @@ async function executeScenario(tc, scenario, config) {
     costLedger,
     costPhase: config.costPhase ?? "benchmark.judge",
     costTags,
-    signal: config.signal,
-    tcloudMaximumAttempts: config.tcloudMaximumAttempts
+    signal: config.signal
   };
   const judgeResults = [];
   let failedJudges = 0;
@@ -923,7 +927,7 @@ async function executeScenario(tc, scenario, config) {
           console.log(`    ${judgeName} retry ${attempt}/2 (waiting ${wait / 1e3}s)`);
           await sleep2(wait);
         }
-        const scores = await judge(tc, judgeInput);
+        const scores = await judge(chat, judgeInput);
         judgeResults.push(scores);
         await sleep2(3e3);
         lastError = void 0;
@@ -983,10 +987,10 @@ async function executeScenario(tc, scenario, config) {
 
 // src/benchmark.ts
 var BenchmarkRunner = class {
-  tc;
+  chat;
   config;
-  constructor(tc, config) {
-    this.tc = tc;
+  constructor(chat, config) {
+    this.chat = chat;
     this.config = config;
   }
   async run(scenarios) {
@@ -1008,13 +1012,12 @@ var BenchmarkRunner = class {
       console.log(`[${i + 1}/${toRun.length}] ${scenario.id} (${scenario.persona})`);
       console.log(`  thesis: ${scenario.thesis}`);
       console.log(`  turns: ${scenario.turns.length}`);
-      const result = await executeScenario(this.tc, scenario, {
+      const result = await executeScenario(this.chat, scenario, {
         systemPrompt: this.config.systemPrompt,
         model: this.config.model,
         judges: this.config.judges,
         costLedger,
-        costTags,
-        tcloudMaximumAttempts: this.config.tcloudMaximumAttempts
+        costTags
       });
       results.push(result);
       for (const turn of result.turns) {
@@ -1717,19 +1720,17 @@ var RIGOR_STANCE = {
   relentless: "Your stance: a senior partner reviewing this work for a client who will litigate if it is wrong. You interrogate every claim. You accept nothing undefended. You find the single weakest point in every answer and attack it. Courteous, never satisfied."
 };
 var AgentDriver = class {
-  tc;
+  chat;
   client;
   driverModel;
   productContext;
   costLedger;
-  tcloudMaximumAttempts;
-  constructor(tc, config) {
-    this.tc = tc;
+  constructor(chat, config) {
+    this.chat = chat;
     this.client = config.client;
     this.driverModel = config.driverModel ?? "claude-sonnet-4-6";
     this.productContext = config.productContext ?? "";
     this.costLedger = config.costLedger ?? new CostLedger();
-    this.tcloudMaximumAttempts = config.tcloudMaximumAttempts;
   }
   /**
    * Run a persona through the product.
@@ -1812,15 +1813,14 @@ var AgentDriver = class {
   }
   /** Use the driver LLM to decide what the "user" says next */
   async decideNextMessage(persona, state, history, costTags) {
-    return decideNextUserTurn(this.tc, {
+    return decideNextUserTurn(this.chat, {
       persona,
       state,
       history,
       productContext: this.productContext,
       model: this.driverModel,
       costLedger: this.costLedger,
-      costTags,
-      tcloudMaximumAttempts: this.tcloudMaximumAttempts
+      costTags
     });
   }
   /** Handle pending approvals based on persona feedback patterns */
@@ -1925,7 +1925,7 @@ COMPLETION: drive toward the goal's real acceptance check. Do not declare done \
 
 Output ONLY your next instruction to the worker \u2014 direct, detailed, actionable, in the first person as the driver. No meta-commentary, no preamble.`;
 }
-async function decideNextUserTurn(tc, opts) {
+async function decideNextUserTurn(chat, opts) {
   const { persona, state, history, productContext = "", model = "claude-sonnet-4-6" } = opts;
   const lastResponse = history.length > 0 ? history[history.length - 1].content.slice(0, 2e3) : "(no conversation yet \u2014 this is the first message)";
   const recentHistory = history.slice(-6).map((h) => `${h.role}: ${h.content.slice(0, 500)}`).join("\n\n");
@@ -1951,14 +1951,13 @@ ${lastResponse}` : "No conversation yet. Send your opening message \u2014 in cha
     actor: "decideNextUserTurn",
     model,
     tags: opts.costTags,
-    maximumCharge: maximumChargeForTCloudRequest(request, opts.tcloudMaximumAttempts),
-    execute: () => tc.chat(request),
-    receipt: (response) => costReceiptFromTCloud(response, model)
+    maximumCharge: chat.maximumAttempts === void 0 ? void 0 : maximumChargeForLlmRequest(request, { maximumAttempts: chat.maximumAttempts }),
+    execute: (signal, callId) => chat.chat(request, { signal, idempotencyKey: callId }),
+    receipt: costReceiptFromLlm,
+    receiptFromError: costReceiptFromLlmError
   });
   if (!paid.succeeded) throw paid.error;
-  const resp = paid.value;
-  const content = resp.choices?.[0]?.message?.content ?? "";
-  return content.trim();
+  return paid.value.content.trim();
 }
 
 // src/feedback-trajectory.ts
@@ -4842,7 +4841,7 @@ function assertNonNegative(n, name) {
 
 // src/eval-trace-store.ts
 function runScore(record) {
-  return runTaskScore(record);
+  return trainingScore(record);
 }
 function matches(record, f) {
   if (f.experimentId && record.experimentId !== f.experimentId) return false;
@@ -4938,13 +4937,19 @@ var EvalTraceStore = class {
    * Highest-scoring run for a scenario (optionally restricted to a candidate).
    * Returns null when no run matches. Ties resolve to the earliest-appended run
    * so the result is stable.
+   *
+   * Runs flagged as gamed are DROPPED, not zeroed. The caller's use for this is
+   * few-shot seeding — the returned trajectory becomes an example to copy — so
+   * the same rule as SFT applies: a faked success must not be in the candidate
+   * set at all. When every run for the scenario is gated the honest answer is
+   * `null` (no exemplar), never the least-bad fake.
    */
   async getBest(scenarioId, opts = {}) {
-    const rows = await this.query({
+    const rows = (await this.query({
       scenarioId,
       candidateId: opts.candidateId,
       splitTag: opts.splitTag
-    });
+    })).filter((r) => !isRealnessGated(r));
     const scored = rows.flatMap((record) => {
       const score = runScore(record);
       return score === void 0 ? [] : [{ record, score }];
@@ -4966,6 +4971,9 @@ var EvalTraceStore = class {
    * ran a scenario more than once, its best `runScore` for that scenario is
    * used. Throws when there is no paired scenario — an unpaired "comparison" is
    * not one.
+   *
+   * Realness-gated runs are excluded and counted in `realnessGatedRuns`, never
+   * folded in as a zero.
    */
   async compareRuns(candidateA, candidateB) {
     if (candidateA === candidateB) {
@@ -4974,12 +4982,17 @@ var EvalTraceStore = class {
       );
     }
     const rows = await this.backend.load();
+    let realnessGatedRuns = 0;
     const bestByScenario = (candidate) => {
       const m = /* @__PURE__ */ new Map();
       for (const r of rows) {
         if (r.candidateId !== candidate) continue;
         const sid = r.scenarioId;
         if (!sid) continue;
+        if (isRealnessGated(r)) {
+          realnessGatedRuns++;
+          continue;
+        }
         const s = runScore(r);
         if (s === void 0) continue;
         const prev = m.get(sid);
@@ -4992,7 +5005,7 @@ var EvalTraceStore = class {
     const paired = [...aScores.keys()].filter((sid) => bScores.has(sid)).sort();
     if (paired.length === 0) {
       throw new ValidationError(
-        `EvalTraceStore.compareRuns: "${candidateA}" and "${candidateB}" share no scenario (need scenarioId on records)`
+        realnessGatedRuns > 0 ? `EvalTraceStore.compareRuns: "${candidateA}" and "${candidateB}" share no scenario with honest runs on both sides (${realnessGatedRuns} run(s) were realness-gated)` : `EvalTraceStore.compareRuns: "${candidateA}" and "${candidateB}" share no scenario (need scenarioId on records)`
       );
     }
     let sumA = 0;
@@ -5020,7 +5033,8 @@ var EvalTraceStore = class {
       meanDelta: meanB - meanA,
       bWins,
       ties,
-      aWins
+      aWins,
+      realnessGatedRuns
     };
   }
 };
@@ -9610,16 +9624,20 @@ var HeldOutGate = class {
     const candidateId = inferCandidateId2(candidate, this.baselineKey);
     const baselineId = this.baselineKey;
     assertScenarioIdentities([...candidate, ...baseline]);
-    const candidateSearch = scoredRuns(candidate, "searchScore", "search");
-    const baselineSearch = scoredRuns(baseline, "searchScore", "search");
-    const candidateHoldout = scoredRuns(candidate, "holdoutScore", "holdout");
-    const baselineHoldout = scoredRuns(baseline, "holdoutScore", "holdout");
+    const realnessGatedRuns = [...candidate, ...baseline].filter(isRealnessGated).length;
+    const honestCandidate = candidate.filter((run) => !isRealnessGated(run));
+    const honestBaseline = baseline.filter((run) => !isRealnessGated(run));
+    const candidateSearch = scoredRuns(honestCandidate, "search");
+    const baselineSearch = scoredRuns(honestBaseline, "search");
+    const candidateHoldout = scoredRuns(honestCandidate, "holdout");
+    const baselineHoldout = scoredRuns(honestBaseline, "holdout");
     const searchPairing = pairRunRecords(baselineSearch, candidateSearch);
     const holdoutPairing = pairRunRecords(baselineHoldout, candidateHoldout);
-    const beforeSearch = searchPairing.pairs.map((pair) => pair.baseline.outcome.searchScore);
-    const afterSearch = searchPairing.pairs.map((pair) => pair.treatment.outcome.searchScore);
-    const beforeHoldout = holdoutPairing.pairs.map((pair) => pair.baseline.outcome.holdoutScore);
-    const afterHoldout = holdoutPairing.pairs.map((pair) => pair.treatment.outcome.holdoutScore);
+    const splitScoreOf = (run, split) => observedSplitScore(run, split);
+    const beforeSearch = searchPairing.pairs.map((pair) => splitScoreOf(pair.baseline, "search"));
+    const afterSearch = searchPairing.pairs.map((pair) => splitScoreOf(pair.treatment, "search"));
+    const beforeHoldout = holdoutPairing.pairs.map((pair) => splitScoreOf(pair.baseline, "holdout"));
+    const afterHoldout = holdoutPairing.pairs.map((pair) => splitScoreOf(pair.treatment, "holdout"));
     const productiveRuns = beforeHoldout.length;
     const candidateSearchMean = meanOrNull(afterSearch);
     const candidateHoldoutMean = meanOrNull(afterHoldout);
@@ -9638,7 +9656,8 @@ var HeldOutGate = class {
       overfitGap,
       baselineOverfitGap,
       medianCandidateCost,
-      medianBaselineCost
+      medianBaselineCost,
+      realnessGatedRuns
     };
     const missingSplitScores = [
       candidateSearch.length === 0 ? "candidate search" : null,
@@ -9755,10 +9774,12 @@ function assertScenarioIdentities(runs) {
     }
   }
 }
-function scoredRuns(runs, field, splitFilter) {
-  return runs.filter(
-    (run) => run.splitTag === splitFilter && typeof run.outcome[field] === "number" && Number.isFinite(run.outcome[field])
-  );
+function scoredRuns(runs, split) {
+  return runs.filter((run) => {
+    if (run.splitTag !== split) return false;
+    const v = observedSplitScore(run, split);
+    return typeof v === "number" && Number.isFinite(v);
+  });
 }
 function meanOrNull(xs) {
   if (xs.length === 0) return null;
@@ -10315,7 +10336,7 @@ async function tracedAnalyzeTraces(input, options, traceOpts) {
 
 // src/traced-judges.ts
 function traceJudge(judge, judgeName, opts) {
-  return async (tc, input) => {
+  return async (chat, input) => {
     const span = await opts.emitter.span({
       kind: "llm",
       name: `judge:${judgeName}`,
@@ -10326,7 +10347,7 @@ function traceJudge(judge, judgeName, opts) {
       }
     });
     try {
-      const scores = await judge(tc, input);
+      const scores = await judge(chat, input);
       const composite = scores.length > 0 ? scores.reduce((sum4, s) => sum4 + s.score, 0) / scores.length : 0;
       await span.end({
         attributes: {
@@ -10344,7 +10365,7 @@ function traceJudge(judge, judgeName, opts) {
   };
 }
 function traceJudgeEnsemble(judges, judgeNames, opts) {
-  return async (tc, input) => {
+  return async (chat, input) => {
     const ensembleSpan = await opts.emitter.span({
       kind: "custom",
       name: "judge:ensemble",
@@ -10365,7 +10386,7 @@ function traceJudgeEnsemble(judges, judgeNames, opts) {
           parentSpanId: ensembleSpan.span.spanId
         });
         try {
-          const scores = await tracedFn(tc, input);
+          const scores = await tracedFn(chat, input);
           allScores.push(...scores);
         } catch (err) {
           if (!(err instanceof JudgeParseError)) throw err;
@@ -11387,6 +11408,7 @@ ${failures.join("\n")}`);
 }
 export {
   AGENT_PROFILE_KINDS,
+  ATIF_SCHEMA_VERSION,
   ATTESTATION_ALGORITHM,
   AgentDriver,
   AgentEvalError,
@@ -11439,6 +11461,7 @@ export {
   FileSystemRawProviderSink,
   FileSystemTraceStore,
   FindingsStore,
+  HARBOR_IMPORT_GAP,
   HARNESS_NATIVE_MODEL,
   HeldOutGate,
   HoldoutAuditor,
@@ -11532,7 +11555,6 @@ export {
   ValidationError,
   VerificationError,
   acquisitionPlansForKnowledgeGaps,
-  adversarialJudge,
   agentProfileCellHashMaterial,
   agentProfileCellKey,
   agentProfileHash,
@@ -11558,6 +11580,8 @@ export {
   assertCapabilityHeadroom,
   assertCrossFamily,
   assertLlmRoute,
+  assertMinted,
+  assertMintedLines,
   assertModelsServed,
   assertNoHiddenLeak,
   assertProductBenchmarkRun,
@@ -11617,9 +11641,7 @@ export {
   claudeCodeSupervisorRunReader,
   cliffsDelta,
   clusteredPairedBinary,
-  codeExecutionJudge,
   cohensD,
-  coherenceJudge,
   collectionPreserved,
   commentsForSource,
   commitBisect,
@@ -11654,9 +11676,7 @@ export {
   createAnalystAi,
   createAntiSlopJudge,
   createChatClient,
-  createCustomJudge,
   createDefaultReviewer,
-  createDomainExpertJudge,
   createFeedbackTrajectory,
   createIntentMatchJudge,
   createLlmCorrectnessChecker,
@@ -11677,7 +11697,6 @@ export {
   decideReferenceReplayRunPromotion,
   defaultBlendWeights,
   defaultIsMaterial,
-  defaultJudges,
   defaultProviderRedactor,
   defaultReferenceReplayMatcher,
   defaultTraceInsightPanel,
@@ -11738,6 +11757,7 @@ export {
   formatDriverReport,
   formatFindings,
   formatScorecardDiff,
+  fromHarborTrajectory,
   gainHistogram,
   gateTreatmentApplied,
   gateTreatmentFromMetrics,
@@ -11777,6 +11797,7 @@ export {
   isModelPriced,
   isOtelConfigured,
   isOtlpModelCall,
+  isRealnessGated,
   isRetrievalSpan,
   isRolloutLine,
   isRunRecord,
@@ -11829,6 +11850,8 @@ export {
   notBlocked,
   objectiveEval,
   observeAll,
+  observedScore,
+  observedSplitScore,
   otelRunCompleteHook,
   otlpRowsToRunRecords,
   otlpRowsToTraceRunRecords,
@@ -11891,6 +11914,7 @@ export {
   referenceReplayScenarioToRunScore,
   regexMatch,
   regexMatches,
+  relabelImportedSplit,
   renderMarkdownReport,
   renderPlaybookMarkdown,
   renderPreferenceMemoryMarkdown,
@@ -11911,7 +11935,6 @@ export {
   researchReport,
   resolveModelPricing,
   resolveSeat,
-  rolloutReward,
   rollupSupervisorRuns,
   roundTripRunRecord,
   routeFields,
@@ -11948,6 +11971,7 @@ export {
   scoreContinuity,
   scoreFromEvals,
   scoreKnowledgeReadiness,
+  scoreOrigin,
   scorePrReviewComments,
   scorePrReviewSource,
   scoreRedTeamOutput,
@@ -11979,6 +12003,8 @@ export {
   textInSnapshot,
   throwIfRunIncomplete,
   toAgentProfileJson,
+  toHarborTrajectories,
+  toHarborTrajectory,
   toJsonl,
   toLangfuseEnvelope,
   toOpenAiTool,
@@ -11995,6 +12021,8 @@ export {
   traceJudgeEnsemble,
   traceSpanKindToOpenInferenceKind,
   tracedAnalyzeTraces,
+  trainingReward,
+  trainingScore,
   typoMutator,
   urlContains,
   userQuestionsForKnowledgeGaps,