@tangle-network/agent-eval 0.109.1 → 0.110.1

package/dist/{insight-report-B4xrdwEK.d.ts → insight-report-D4cXFsLt.d.ts}

@@ -1,4 +1,4 @@
-import { G as GainDistributionBin, P as ParetoFigureSpec } from './summary-report-QMZVe3P-.js';
+import { G as GainDistributionBin, P as ParetoFigureSpec } from './summary-report-Bz-0-t8v.js';
 import { C as ContinuousAgreement } from './judge-calibration-7C-IDmKr.js';
 
 /**

package/dist/{integrity-DqGZg3st.d.ts → integrity-qemeBAyx.d.ts}

@@ -1,6 +1,6 @@
 import { C as CaptureIntegrityError } from './errors-oeQrLqXC.js';
 import { R as RawProviderSink } from './raw-provider-sink-C46HDghv.js';
-import { T as TraceStore } from './store-BcFXE6LG.js';
+import { T as TraceStore } from './store-BsVi7ncX.js';
 
 /**
  * Run-completion integrity check — at end of run, verify the expected event

package/dist/{types-D1ytG0Yg.d.ts → kind-factory-20hcaYpf.d.ts}

@@ -1,5 +1,7 @@
-import { R as RunRecord } from './run-record-I-Z3JNvO.js';
+import { AxAIService, AxFunction } from '@ax-llm/ax';
 import { T as TraceAnalysisStore } from './store-C1YxJDEK.js';
+import { z } from 'zod';
+import { R as RunRecord } from './run-record-DksGsfgv.js';
 import { a as JudgeInput } from './types-C7DGg5ex.js';
 import { b as LlmCallRequest, c as LlmCallResult } from './llm-client-DyqEH4jH.js';
 
@@ -93,6 +95,54 @@ interface MockTransportOpts extends BaseTransportOpts {
  */
 declare function createChatClient(opts: CreateChatClientOpts): ChatClient;
 
+/**
+ * Typed Ax output for analyst findings.
+ *
+ * Replaces the legacy `findings:string[]` pattern (where every bullet
+ * became a flat-severity `AnalystFinding`) with a structured object
+ * array. Ax binds the field as `findings:json[]` so the provider emits
+ * native structured output; at the kind-factory boundary we Zod-validate
+ * each emitted finding so malformed rows fail loud instead of being
+ * silently lifted with default severity.
+ *
+ * Why not `f.object().array()` directly in the signature? The Ax
+ * signature string `question:string -> findings:json[]` already lets
+ * the provider emit JSON arrays. A Zod boundary is required either
+ * way (the provider can return any JSON), and Zod gives us a single
+ * validation surface independent of which Ax version is installed.
+ */
+
+declare const ANALYST_SEVERITIES: readonly ["critical", "high", "medium", "low", "info"];
+declare const RawAnalystFindingSchema: z.ZodObject<{
+    severity: z.ZodEnum<{
+        info: "info";
+        critical: "critical";
+        medium: "medium";
+        low: "low";
+        high: "high";
+    }>;
+    claim: z.ZodString;
+    subject: z.ZodOptional<z.ZodString>;
+    evidence_uri: z.ZodString;
+    evidence_excerpt: z.ZodOptional<z.ZodString>;
+    confidence: z.ZodNumber;
+    rationale: z.ZodOptional<z.ZodString>;
+    recommended_action: z.ZodOptional<z.ZodString>;
+}, z.core.$strict>;
+type RawAnalystFinding = z.infer<typeof RawAnalystFindingSchema>;
+/**
+ * Description embedded into the actor prompt so the LLM knows what
+ * shape to emit. Kept here so kinds share one source of truth rather
+ * than restating the schema in every prompt.
+ */
+declare const RAW_FINDING_SCHEMA_PROMPT = "Each finding MUST be a JSON object with these fields:\n  - severity: one of \"critical\" | \"high\" | \"medium\" | \"low\" | \"info\"\n  - claim: one-sentence statement (max 2000 chars)\n  - subject?: the routing locus this finding is about. It MUST be one of the exact subject forms listed in this kind's instructions above (e.g. `system-prompt:<section>`, `agent-knowledge:wiki:<slug>`, `tool-doc:<tool>`). A free phrase, a bare noun, or any form not in that list is REJECTED at parse time and the finding is discarded \u2014 omit subject entirely rather than guess a form.\n  - evidence_uri: REQUIRED, never blank. Exactly one of \"span://<trace_id>/<span_id>\" (trace evidence), \"artifact://<relative-path>\" (files), \"metric://<name>\" (named scalars) \u2014 ALWAYS cite a real id surfaced by the tools. If you have no citable id, do not emit the finding.\n  - evidence_excerpt?: short quote (<=2000 chars) from the cited span/artifact\n  - confidence: number 0..1 \u2014 0.9+ when backed by exact quotes, 0.6-0.8 for inferred patterns, <0.5 for speculative\n  - rationale?: one or two sentences explaining the reasoning\n  - recommended_action?: concrete change phrased as an imperative (\"Add ...\", \"Replace ...\", \"Stop ...\") \u2014 omit when the finding is purely descriptive\n\nEmit an empty array when the question has no findings to report. Do not fabricate evidence.";
+/**
+ * Validate one row emitted by the LLM. Returns the typed finding on
+ * success; returns `null` and logs the reason on failure so the kind
+ * factory can skip-and-count rather than abort the whole analyst run.
+ */
+declare function parseRawFinding(row: unknown, log?: (msg: string, fields?: Record<string, unknown>) => void): RawAnalystFinding | null;
+
 /**
  * Analyst contract — the missing orchestration layer over agent-eval's
  * existing analyzers (analyzeTraces, MultiLayerVerifier, RunCritic,
@@ -338,4 +388,121 @@ type AnalystRunEvent = {
     result: AnalystRunResult;
 };
 
-export { type AnalystFinding as A, type ChatClient as C, type DirectProviderTransportOpts as D, type EvidenceRef as E, type MockTransportOpts as M, type RouterTransportOpts as R, type SandboxSdkTransportOpts as S, type Analyst as a, type AnalystContext as b, type AnalystRunSummary as c, type AnalystRunResult as d, type AnalystRunInputs as e, type AnalystRunEvent as f, type AnalystCost as g, type AnalystSeverity as h, type AnalystInputKind as i, type AnalystRequirements as j, type ChatCallOpts as k, type ChatRequest as l, type ChatResponse as m, type ChatTransport as n, type CliBridgeTransportOpts as o, type CreateChatClientOpts as p, computeFindingId as q, createChatClient as r, makeFinding as s };
+/**
+ * Analyst-kind factory — the typed way to define trace analysts.
+ *
+ * A "kind" is a specialized analyst whose actor prompt, tool subset,
+ * and Ax recursion config target one failure-mode lens (failure-mode
+ * classification, knowledge gap discovery, knowledge poisoning, recursive
+ * self-improvement, ...). Kinds emit findings in the typed `RawAnalystFinding`
+ * shape via a JSON-array Ax output; the factory validates each row with
+ * Zod and lifts it into `AnalystFinding[]` with no shape guessing.
+ *
+ * Composition rules:
+ *   - Each kind owns its actor description. No generic "answer this
+ *     question" prompt — the prompt names the failure lens.
+ *   - Each kind picks a narrow tool subset from `ANALYST_TOOL_GROUPS`.
+ *     A kind that never needs full-trace dumps can drop `viewTrace` /
+ *     `viewSpans` and stay cheap.
+ *   - Each kind declares its recursion + parallelism budget. Discovery-
+ *     heavy kinds (failure-mode) get higher `maxDepth`; lens kinds
+ *     (poisoning) usually stay at 0 since they have a tighter brief.
+ *
+ * Optimizer hook: kinds may declare `goldens` — labeled examples used
+ * by `AxMiPRO` / `AxBootstrapFewShot` / `AxGEPA` to fit the actor
+ * description programmatically. Stored on the kind, not the registry,
+ * because the right metric is kind-specific.
+ */
+
+/**
+ * Per-kind specification. The factory turns this into a regular
+ * `Analyst<TraceAnalysisStore>` ready for `AnalystRegistry.register()`.
+ */
+interface TraceAnalystKindSpec {
+    /** Stable id. Appears in finding_id, telemetry, and registry exclusions. */
+    id: string;
+    /** One-sentence description shown in `registry.list()`. */
+    description: string;
+    /** Coarse classification stamped on every emitted finding (`failure-mode`, `knowledge-gap`, ...). */
+    area: string;
+    /** Bump on any breaking change to the actor prompt or output schema. */
+    version: string;
+    /** Actor system prompt. Must instruct the LLM to emit `findings` per the schema. */
+    actorDescription: string;
+    /** Responder system prompt; falls back to a minimal "format the findings" instruction. */
+    responderDescription?: string;
+    /** Tool functions the actor may call. Pick narrow subsets via `ANALYST_TOOL_GROUPS`. */
+    buildTools: (store: TraceAnalysisStore) => AxFunction[];
+    /** Recursion budget. `maxDepth: 0` disables subagents. */
+    recursion?: {
+        maxDepth: number;
+        maxParallelSubagents?: number;
+    };
+    /** Actor turn cap. Default 12. */
+    maxTurns?: number;
+    /** Runtime char cap. Default 6000. */
+    maxRuntimeChars?: number;
+    /** Cost classification surfaced in `registry.list()` and budget enforcement. */
+    cost: AnalystCost;
+    /** Per-finding-row hook — kinds may reject / rewrite before lifting. */
+    postProcess?: (row: RawAnalystFinding, ctx: AnalystContext) => RawAnalystFinding | null;
+    /** Optional optimizer hook — populated when a kind wants to fit its prompt against labeled examples. */
+    goldens?: TraceAnalystGolden[];
+}
+/**
+ * One labeled example consumed by Ax optimizers (MIPRO / GEPA / Bootstrap).
+ * Each input is the same `{question}` an analyst would receive; `expected`
+ * is the ground-truth finding set a fitted prompt should produce on this
+ * input. Metric: kind-specific (default: F1 on `finding_id` overlap).
+ */
+interface TraceAnalystGolden {
+    question: string;
+    expected: ReadonlyArray<Omit<RawAnalystFinding, 'confidence'>>;
+}
+interface CreateTraceAnalystKindOpts {
+    /** AxAIService bound at registration time. */
+    ai: AxAIService;
+    /** Optional model override; falls back to the AI service's default. */
+    model?: string;
+    /** Override the spec's `version` (e.g. when an optimizer has fitted a new prompt). */
+    versionSuffix?: string;
+    /**
+     * Optional two-phase recovery: when the agentic harvest is empty but the
+     * actor produced a substantive free-form `report`, extract findings from that
+     * prose via a tolerant chat-completions pass (`structureFindings`) — no
+     * strict-emission contract, so it works on weak models. Omit to leave the
+     * actor's harvest as-is (the report is still surfaced fail-loud either way).
+     */
+    recovery?: {
+        baseUrl: string;
+        apiKey?: string;
+        model?: string;
+        fetchImpl?: typeof fetch;
+    };
+}
+/**
+ * Build an `Analyst<TraceAnalysisStore>` from a kind spec.
+ *
+ * Lifts the Ax pipeline once at registration time so the registry
+ * gets a stateless analyst. The Ax agent is freshly constructed per
+ * `analyze()` call (the agent carries chat-log + usage state we don't
+ * want shared across analyst runs).
+ */
+declare function createTraceAnalystKind(spec: TraceAnalystKindSpec, opts: CreateTraceAnalystKindOpts): Analyst<TraceAnalysisStore>;
+/**
+ * Render a compact prior-findings block the actor reads alongside its
+ * brief. Each row is one line so the actor can scan dozens cheaply.
+ * The kind's prompt instructs the actor to (a) check whether a new
+ * cluster matches a prior `finding_id` (carry the id forward via
+ * `id_basis` to keep diffs stable) and (b) raise severity / confidence
+ * when a prior finding has reappeared without remediation.
+ *
+ * Returns the empty string when there are no prior findings — most
+ * runs are "first-of-its-kind" and the prompt stays unchanged.
+ *
+ * Exported for tests + for consumers that build their own actor
+ * prompts (e.g. specialized analysts living outside the default kinds).
+ */
+declare function renderPriorFindings(prior: AnalystContext['priorFindings']): string;
+
+export { type AnalystFinding as A, parseRawFinding as B, type ChatClient as C, type DirectProviderTransportOpts as D, type EvidenceRef as E, renderPriorFindings as F, type MockTransportOpts as M, RAW_FINDING_SCHEMA_PROMPT as R, type SandboxSdkTransportOpts as S, type TraceAnalystKindSpec as T, type Analyst as a, type AnalystContext as b, type AnalystRunSummary as c, type AnalystRunResult as d, type AnalystRunInputs as e, type AnalystRunEvent as f, type AnalystSeverity as g, ANALYST_SEVERITIES as h, type AnalystCost as i, type AnalystInputKind as j, type AnalystRequirements as k, type ChatCallOpts as l, type ChatRequest as m, type ChatResponse as n, type ChatTransport as o, type CliBridgeTransportOpts as p, type CreateChatClientOpts as q, type CreateTraceAnalystKindOpts as r, type RawAnalystFinding as s, RawAnalystFindingSchema as t, type RouterTransportOpts as u, type TraceAnalystGolden as v, computeFindingId as w, createChatClient as x, createTraceAnalystKind as y, makeFinding as z };

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 { 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-Dz8TQV4y.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-DPnyG-CE.js';
+export { R as RubricOutcomePair, a as RubricPredictiveValidityInput, b as RubricPredictiveValidityReport, c as RubricRanking, r as rubricPredictiveValidity } from '../rubric-predictive-validity-C-fMteAW.js';
 import { C as ContinuousAgreement, a as CalibrationResult, b as ContinuousCalibrationResult, c as CandidateScore, G as GoldenItem } from '../judge-calibration-7C-IDmKr.js';
 import { S as SeriesConvergenceOptions, a as SeriesConvergenceResult } from '../series-convergence-D5OWMBg6.js';
 import { C as CorpusAgreementReport } from '../statistics-D88peojY.js';
-import '../store-BcFXE6LG.js';
-import '../schema-m0gsnbt3.js';
-import '../run-record-I-Z3JNvO.js';
+import '../store-BsVi7ncX.js';
+import '../schema-SGWcK9wa.js';
+import '../run-record-DksGsfgv.js';
 import '@tangle-network/agent-interface';
 import '../errors-oeQrLqXC.js';
 import '../types-C7DGg5ex.js';

package/dist/meta-eval/index.js

@@ -19,8 +19,7 @@ import {
 import {
   aggregateLlm,
   llmSpans
-} from "../chunk-JZXGWLK5.js";
-import "../chunk-5BKGXME7.js";
+} from "../chunk-MHNQWM4I.js";
 import {
   ValidationError
 } from "../chunk-ONWEPEDO.js";

package/dist/meta-eval/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/meta-eval/correlation-study.ts","../../src/meta-eval/sentinel.ts"],"sourcesContent":["/**\n * Correlation study — \"does our eval score predict real-world outcomes?\"\n *\n * This is the load-bearing signal. Takes a TraceStore + OutcomeStore,\n * joins on runId, computes Pearson + Spearman + bootstrap CI for every\n * (evalMetric, outcomeMetric) pair the caller declares.\n *\n * Without this number the framework is ornamental. With it and r > 0.6\n * the framework is a moat — no other agent-eval tool publishes one.\n */\n\nimport { pearsonR, spearmanR } from '../statistics'\nimport { aggregateLlm, llmSpans } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\nimport type { DeploymentOutcome, OutcomeFilter, OutcomeStore } from './outcome-store'\n\nexport interface EvalMetricSpec {\n  id: string\n  /** Extract a scalar from a run (defaults cover score/pass/durationMs/costUsd/tokens). */\n  extract?: (run: Run, store: TraceStore) => Promise<number | null>\n}\n\nexport interface OutcomePair {\n  evalMetric: string\n  outcomeMetric: string\n}\n\nexport interface CorrelationResult {\n  evalMetric: string\n  outcomeMetric: string\n  n: number\n  pearson: number\n  spearman: number\n  /** 95% bootstrap CI for Pearson. */\n  pearsonCi95: { lower: number; upper: number }\n  /** Rough verdict: 'strong' ≥ 0.7, 'moderate' ≥ 0.4, else 'weak'. */\n  verdict: 'strong' | 'moderate' | 'weak'\n}\n\nexport interface CorrelationStudyResult {\n  pairs: CorrelationResult[]\n  joinedSamples: number\n  skippedRuns: number\n}\n\nexport interface CorrelationStudyOptions {\n  /** Only join outcomes captured within this window after run.startedAt. */\n  maxCaptureLagMs?: number\n  /** Restrict to a subset of outcomes (cohort, region, source). */\n  outcomeFilter?: OutcomeFilter\n  /** Which outcome per run to use when multiple exist. Default 'latest'. */\n  reduction?: 'latest' | 'mean' | 'max'\n  /** Bootstrap iterations for the CI. Default 500. */\n  bootstrapIterations?: number\n}\n\nexport async function correlationStudy(\n  traceStore: TraceStore,\n  outcomeStore: OutcomeStore,\n  evalMetrics: EvalMetricSpec[],\n  outcomeMetricNames: string[],\n  options: CorrelationStudyOptions = {},\n): Promise<CorrelationStudyResult> {\n  const runs = await traceStore.listRuns()\n  const outcomes = await outcomeStore.list(options.outcomeFilter)\n  const outcomesByRun = new Map<string, DeploymentOutcome[]>()\n  for (const o of outcomes) {\n    const arr = outcomesByRun.get(o.runId) ?? []\n    arr.push(o)\n    outcomesByRun.set(o.runId, arr)\n  }\n\n  const reduction = options.reduction ?? 'latest'\n  const maxLag = options.maxCaptureLagMs ?? Infinity\n\n  const pairs: Array<{ evalMetric: string; outcomeMetric: string; xs: number[]; ys: number[] }> = []\n  for (const em of evalMetrics) {\n    for (const om of outcomeMetricNames) {\n      pairs.push({ evalMetric: em.id, outcomeMetric: om, xs: [], ys: [] })\n    }\n  }\n\n  let joined = 0\n  let skipped = 0\n  for (const run of runs) {\n    const os = outcomesByRun.get(run.runId)\n    if (!os || os.length === 0) {\n      skipped++\n      continue\n    }\n    const eligible = os.filter((o) => o.capturedAt - run.startedAt <= maxLag)\n    if (eligible.length === 0) {\n      skipped++\n      continue\n    }\n\n    for (const em of evalMetrics) {\n      const extract = em.extract ?? defaultExtract(em.id)\n      const x = await extract(run, traceStore)\n      if (x === null || !Number.isFinite(x)) continue\n\n      for (const om of outcomeMetricNames) {\n        const values = eligible\n          .map((o) => o.metrics[om])\n          .filter((v): v is number => typeof v === 'number' && Number.isFinite(v))\n        if (values.length === 0) continue\n        const y = reduce(values, reduction, eligible)\n        if (y === null) continue\n        const pair = pairs.find((p) => p.evalMetric === em.id && p.outcomeMetric === om)!\n        pair.xs.push(x)\n        pair.ys.push(y)\n      }\n    }\n    joined++\n  }\n\n  const results: CorrelationResult[] = pairs\n    .filter((p) => p.xs.length >= 3)\n    .map((p) => {\n      const pearson = pearsonR(p.xs, p.ys)\n      const spearman = spearmanR(p.xs, p.ys)\n      const pearsonCi95 = bootstrapPearsonCi(p.xs, p.ys, options.bootstrapIterations ?? 500)\n      const verdict: CorrelationResult['verdict'] =\n        Math.abs(pearson) >= 0.7 ? 'strong' : Math.abs(pearson) >= 0.4 ? 'moderate' : 'weak'\n      return {\n        evalMetric: p.evalMetric,\n        outcomeMetric: p.outcomeMetric,\n        n: p.xs.length,\n        pearson,\n        spearman,\n        pearsonCi95,\n        verdict,\n      }\n    })\n\n  return { pairs: results, joinedSamples: joined, skippedRuns: skipped }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────\n\nfunction reduce(\n  values: number[],\n  kind: 'latest' | 'mean' | 'max',\n  outcomes: DeploymentOutcome[],\n): number | null {\n  if (values.length === 0) return null\n  if (kind === 'mean') return values.reduce((a, b) => a + b, 0) / values.length\n  if (kind === 'max') return Math.max(...values)\n  // 'latest': pick the outcome captured last, then lookup its metric\n  const latest = [...outcomes].sort((a, b) => b.capturedAt - a.capturedAt)[0]\n  if (!latest) return null\n  const latestKey = Object.keys(latest.metrics)[0]\n  const v = latestKey !== undefined ? latest.metrics[latestKey] : undefined\n  // For 'latest' we already have `values` aligned; use the last-captured one\n  const paired = outcomes\n    .map((o) => {\n      const k = Object.keys(o.metrics)[0]\n      return {\n        at: o.capturedAt,\n        v: k !== undefined ? values.find((x) => o.metrics[k] === x) : undefined,\n      }\n    })\n    .filter((p) => p.v !== undefined)\n  if (paired.length === 0) return v ?? null\n  return paired.sort((a, b) => b.at - a.at)[0]?.v ?? null\n}\n\nfunction bootstrapPearsonCi(\n  xs: number[],\n  ys: number[],\n  iterations: number,\n): { lower: number; upper: number } {\n  const n = xs.length\n  if (n < 3) return { lower: NaN, upper: NaN }\n  const rs: number[] = []\n  for (let b = 0; b < iterations; b++) {\n    const rx: number[] = new Array(n)\n    const ry: number[] = new Array(n)\n    for (let i = 0; i < n; i++) {\n      const idx = Math.floor(Math.random() * n)\n      rx[i] = xs[idx]!\n      ry[i] = ys[idx]!\n    }\n    const r = pearsonR(rx, ry)\n    if (Number.isFinite(r)) rs.push(r)\n  }\n  rs.sort((a, b) => a - b)\n  if (rs.length === 0) return { lower: NaN, upper: NaN }\n  return {\n    lower: rs[Math.floor(0.025 * rs.length)]!,\n    upper: rs[Math.min(rs.length - 1, Math.floor(0.975 * rs.length))]!,\n  }\n}\n\nfunction defaultExtract(metric: string): (run: Run, store: TraceStore) => Promise<number | null> {\n  return async (run, store) => {\n    switch (metric) {\n      case 'score':\n      case 'overallScore':\n        return run.outcome?.score ?? null\n      case 'pass':\n        return run.outcome?.pass === true ? 1 : 0\n      case 'durationMs':\n        return run.endedAt && run.startedAt ? run.endedAt - run.startedAt : null\n      case 'costUsd': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).costUsd\n      }\n      case 'inputTokens': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).inputTokens\n      }\n      default:\n        return null\n    }\n  }\n}\n","/**\n * Judge sentinel — eval trustworthiness as a continuously measured,\n * alarmed trend.\n *\n * Judges are models; models change underneath us; calibration decays.\n * This module composes three existing instruments into one loop:\n *\n *   snapshot — adapters turn real instrument outputs (`calibrateJudge` /\n *              `calibrateJudgeContinuous`, `corpusInterRaterAgreement` /\n *              `continuousAgreement`, a rerun sentinel golden set) into\n *              `SentinelSnapshot`s\n *   series   — `analyzeSeries` (src/series-convergence) classifies each\n *              judge×metric history as stabilized / drifting / noisy\n *   alarm    — `judgeSentinelReport` turns trends + thresholds into named\n *              alarms; `evalHealthStamp` is the tiny object a campaign\n *              attaches to its verdicts\n *\n * Alarm conditions:\n *   - convergence state `drifting-down` on any tracked metric\n *   - irr below `minIrr`\n *   - calibrationKappa / sentinelPassRate dropped vs the series baseline\n *     beyond `maxKappaDrop` / `maxSentinelDrop`\n *   - judge model changed with no post-change golden-grounded snapshot\n *     (the silent-upgrade trap: judges agreeing with each other after a\n *     model swap proves nothing — only re-measuring against gold does)\n *   - newest snapshot older than `staleAfterDays` relative to `asOf`\n *\n * Wire-in: run the snapshot adapters from a post-campaign hook or a\n * nightly job, append to a `SentinelStore`, then compute\n * `judgeSentinelReport(await store.history(), { asOf })` with `asOf` =\n * the campaign timestamp. Attach `evalHealthStamp(report)` to campaign\n * verdicts: an alarmed sentinel means every downstream verdict carries\n * an integrity warning until the judge is recalibrated.\n *\n * The module is clock-free — every timestamp (`at`, `asOf`) is\n * caller-supplied ISO-8601, so reports are reproducible.\n */\n\nimport { ValidationError } from '../errors'\nimport type {\n  CalibrationResult,\n  CandidateScore,\n  ContinuousAgreement,\n  ContinuousCalibrationResult,\n  GoldenItem,\n} from '../judge-calibration'\nimport {\n  analyzeSeries,\n  type SeriesConvergenceOptions,\n  type SeriesConvergenceResult,\n} from '../series-convergence'\nimport type { CorpusAgreementReport } from '../statistics'\n\nexport const SENTINEL_METRIC_NAMES = ['irr', 'calibrationKappa', 'sentinelPassRate'] as const\n\nexport type SentinelMetricName = (typeof SENTINEL_METRIC_NAMES)[number]\n\nexport interface SentinelMetrics {\n  /** Inter-rater reliability — ICC(2,1) from agreement instruments. */\n  irr?: number\n  /** Weighted κ vs the human golden set (`calibrateJudge*`). */\n  calibrationKappa?: number\n  /** Fraction of a fixed sentinel golden set the judge still scores within tolerance. */\n  sentinelPassRate?: number\n}\n\nexport interface SentinelSnapshot {\n  /** Caller-supplied ISO-8601 timestamp — the module never reads a clock. */\n  at: string\n  judgeId: string\n  /** Exact model identity behind the judge (pin the full version string). */\n  judgeModel: string\n  /**\n   * Metrics measured at `at`. May be empty: an empty-metrics snapshot is a\n   * model-change marker — it records the new `judgeModel` immediately and\n   * the silent-upgrade alarm stays raised until a golden-grounded snapshot\n   * (calibrationKappa or sentinelPassRate) follows.\n   */\n  metrics: SentinelMetrics\n}\n\n/** Identity + timestamp the caller supplies alongside an instrument output. */\nexport interface SnapshotMeta {\n  at: string\n  judgeId: string\n  judgeModel: string\n}\n\nfunction parseIso(value: string, label: string): number {\n  const ms = typeof value === 'string' ? Date.parse(value) : NaN\n  if (!Number.isFinite(ms)) {\n    throw new ValidationError(\n      `judge-sentinel: ${label} is not a parseable ISO timestamp: ${JSON.stringify(value)}`,\n    )\n  }\n  return ms\n}\n\n/** Shape guard for snapshots crossing a boundary (store append, JSONL load, report input). */\nexport function validateSentinelSnapshot(snapshot: SentinelSnapshot, source = 'snapshot'): void {\n  parseIso(snapshot.at, `${source}.at`)\n  if (typeof snapshot.judgeId !== 'string' || snapshot.judgeId.length === 0) {\n    throw new ValidationError(`judge-sentinel: ${source}.judgeId must be a non-empty string`)\n  }\n  if (typeof snapshot.judgeModel !== 'string' || snapshot.judgeModel.length === 0) {\n    throw new ValidationError(`judge-sentinel: ${source}.judgeModel must be a non-empty string`)\n  }\n  if (snapshot.metrics === null || typeof snapshot.metrics !== 'object') {\n    throw new ValidationError(`judge-sentinel: ${source}.metrics must be an object (may be empty)`)\n  }\n  for (const [key, value] of Object.entries(snapshot.metrics)) {\n    if (!(SENTINEL_METRIC_NAMES as readonly string[]).includes(key)) {\n      // A typo'd key would silently never trend — reject instead.\n      throw new ValidationError(\n        `judge-sentinel: ${source}.metrics carries unknown key \"${key}\" — known: ${SENTINEL_METRIC_NAMES.join(', ')}`,\n      )\n    }\n    if (value !== undefined && !Number.isFinite(value)) {\n      throw new ValidationError(\n        `judge-sentinel: ${source}.metrics.${key} is not a finite number: ${String(value)}`,\n      )\n    }\n  }\n}\n\n// ── Snapshot adapters — real instrument output → SentinelSnapshot ───\n\n/**\n * From `calibrateJudge` / `calibrateJudgeContinuous` output. Prefers the\n * un-rounded continuous κ_w when the report carries one (integer κ\n * discards information for [0,1] judges).\n */\nexport function snapshotFromCalibration(\n  report: CalibrationResult | ContinuousCalibrationResult,\n  meta: SnapshotMeta,\n): SentinelSnapshot {\n  const kappa = 'weightedKappaContinuous' in report ? report.weightedKappaContinuous : report.kappa\n  if (!Number.isFinite(kappa)) {\n    throw new ValidationError(\n      `snapshotFromCalibration: κ is not finite (n=${report.n}) — calibrate against ≥2 joined golden items before snapshotting`,\n    )\n  }\n  const snapshot: SentinelSnapshot = { ...meta, metrics: { calibrationKappa: kappa } }\n  validateSentinelSnapshot(snapshot)\n  return snapshot\n}\n\n/**\n * From `corpusInterRaterAgreement` (statistics.ts) or `continuousAgreement`\n * (judge-calibration.ts) output. IRR = ICC(2,1) — the reliability\n * coefficient both instruments compute. For ensemble-level agreement,\n * `meta.judgeId` names the ensemble and `meta.judgeModel` pins its\n * composition (e.g. a joined list of member model strings).\n */\nexport function snapshotFromAgreement(\n  report: CorpusAgreementReport | ContinuousAgreement,\n  meta: SnapshotMeta,\n): SentinelSnapshot {\n  const irr = 'overallIcc' in report ? report.overallIcc : report.icc\n  if (!Number.isFinite(irr)) {\n    throw new ValidationError(\n      'snapshotFromAgreement: ICC is not finite — agreement needs ≥2 raters and ≥2 complete items',\n    )\n  }\n  const snapshot: SentinelSnapshot = { ...meta, metrics: { irr } }\n  validateSentinelSnapshot(snapshot)\n  return snapshot\n}\n\nexport interface SentinelSetOptions {\n  /** Max |judge − human| for an item to count as passed. Default 0.1 (tuned for [0,1] scores). */\n  tolerance?: number\n}\n\n/**\n * From a rerun of a fixed sentinel golden set: the same `GoldenItem`s the\n * judge was originally calibrated on, re-scored periodically. Pass rate =\n * fraction of joined items where the judge stays within `tolerance` of the\n * human score. Items the judge didn't score are excluded from the\n * denominator; zero joined items is an error, not a 0% pass rate.\n */\nexport function snapshotFromSentinelSet(\n  scores: CandidateScore[],\n  golden: GoldenItem[],\n  meta: SnapshotMeta,\n  options: SentinelSetOptions = {},\n): SentinelSnapshot {\n  const tolerance = options.tolerance ?? 0.1\n  if (!Number.isFinite(tolerance) || tolerance < 0) {\n    throw new ValidationError(\n      `snapshotFromSentinelSet: tolerance must be a finite number ≥ 0, got ${tolerance}`,\n    )\n  }\n  const goldenById = new Map<string, number>()\n  for (const item of golden) {\n    if (!Number.isFinite(item.humanScore)) {\n      throw new ValidationError(\n        `snapshotFromSentinelSet: golden item \"${item.itemId}\" has a non-finite humanScore`,\n      )\n    }\n    if (goldenById.has(item.itemId)) {\n      throw new ValidationError(`snapshotFromSentinelSet: duplicate golden itemId \"${item.itemId}\"`)\n    }\n    goldenById.set(item.itemId, item.humanScore)\n  }\n  let joined = 0\n  let passed = 0\n  for (const s of scores) {\n    const human = goldenById.get(s.itemId)\n    if (human === undefined) continue\n    if (!Number.isFinite(s.score)) {\n      throw new ValidationError(\n        `snapshotFromSentinelSet: judge score for item \"${s.itemId}\" is not finite`,\n      )\n    }\n    joined++\n    if (Math.abs(s.score - human) <= tolerance) passed++\n  }\n  if (joined === 0) {\n    throw new ValidationError(\n      'snapshotFromSentinelSet: no judge scores joined the golden set — itemIds mismatch or empty sentinel run',\n    )\n  }\n  const snapshot: SentinelSnapshot = { ...meta, metrics: { sentinelPassRate: passed / joined } }\n  validateSentinelSnapshot(snapshot)\n  return snapshot\n}\n\n// ── Persistence ──────────────────────────────────────────────────────\n\nexport interface SentinelStore {\n  append(snapshot: SentinelSnapshot): Promise<void>\n  /** All snapshots (optionally for one judge), in append order. */\n  history(judgeId?: string): Promise<SentinelSnapshot[]>\n}\n\nexport function inMemorySentinelStore(initial: SentinelSnapshot[] = []): SentinelStore {\n  for (const s of initial) validateSentinelSnapshot(s)\n  const items = initial.map((s) => ({ ...s, metrics: { ...s.metrics } }))\n  return {\n    async append(snapshot) {\n      validateSentinelSnapshot(snapshot)\n      items.push({ ...snapshot, metrics: { ...snapshot.metrics } })\n    },\n    async history(judgeId) {\n      const all = items.map((s) => ({ ...s, metrics: { ...s.metrics } }))\n      return judgeId === undefined ? all : all.filter((s) => s.judgeId === judgeId)\n    },\n  }\n}\n\n/**\n * JSONL store — one snapshot per line, appended atomically per call.\n * A corrupt or shape-invalid line is a loud error naming the file and\n * line number: a sentinel history that silently drops records would\n * defeat the drift detection it exists to provide.\n */\nexport function fileSentinelStore(path: string): SentinelStore {\n  return {\n    async append(snapshot) {\n      validateSentinelSnapshot(snapshot)\n      const fs = await import('node:fs/promises')\n      const pathMod = await import('node:path')\n      await fs.mkdir(pathMod.dirname(path), { recursive: true })\n      await fs.appendFile(path, `${JSON.stringify(snapshot)}\\n`, 'utf8')\n    },\n    async history(judgeId) {\n      const fs = await import('node:fs/promises')\n      let raw: string\n      try {\n        raw = await fs.readFile(path, 'utf8')\n      } catch (err) {\n        if ((err as NodeJS.ErrnoException).code === 'ENOENT') return []\n        throw err\n      }\n      const lines = raw.split('\\n')\n      const snapshots: SentinelSnapshot[] = []\n      for (let i = 0; i < lines.length; i++) {\n        const line = lines[i]!\n        if (line.trim().length === 0) continue\n        let parsed: unknown\n        try {\n          parsed = JSON.parse(line)\n        } catch (err) {\n          throw new ValidationError(\n            `judge-sentinel: store at ${path} has a corrupt JSONL line ${i + 1}: ${(err as Error).message}`,\n          )\n        }\n        const snapshot = parsed as SentinelSnapshot\n        validateSentinelSnapshot(snapshot, `${path}:${i + 1}`)\n        snapshots.push(snapshot)\n      }\n      return judgeId === undefined ? snapshots : snapshots.filter((s) => s.judgeId === judgeId)\n    },\n  }\n}\n\n// ── Trend report ─────────────────────────────────────────────────────\n\nexport interface SentinelThresholds {\n  /** Floor for irr — below it the judge ensemble is unreliable regardless of trend. Default 0.6. */\n  minIrr?: number\n  /** Max tolerated calibrationKappa drop vs the series baseline. Default 0.15. */\n  maxKappaDrop?: number\n  /** Max tolerated sentinelPassRate drop vs the series baseline. Default 0.1. */\n  maxSentinelDrop?: number\n  /** Newest snapshot older than this (days, vs asOf) = the sentinel itself went blind. Default 30. */\n  staleAfterDays?: number\n}\n\nexport interface JudgeSentinelOptions {\n  /** Caller-supplied ISO timestamp staleness is measured against. */\n  asOf: string\n  thresholds?: SentinelThresholds\n  /** Forwarded to `analyzeSeries` (window, stableCv, driftRun). */\n  convergence?: SeriesConvergenceOptions\n}\n\nexport interface SentinelTrend {\n  judgeId: string\n  metric: SentinelMetricName\n  /** Verbatim `analyzeSeries` state for this judge×metric history. */\n  state: SeriesConvergenceResult['state']\n  /** Newest value in the series. */\n  current: number\n  /** Oldest value in the series — the reference the drop thresholds compare against. */\n  baseline: number\n  /** current − baseline (signed; negative = decayed). */\n  drift: number\n  alarmed: boolean\n  reason?: string\n}\n\nexport interface SentinelReport {\n  perJudge: SentinelTrend[]\n  alarms: string[]\n  healthy: boolean\n  /**\n   * `judgeId:metric` pairs with too few snapshots for the convergence\n   * machine. Named, not counted — a blind spot you can't see is a blind\n   * spot you won't fix. Floor/drop checks still apply to thin series, so\n   * insufficient history alone never masks an alarm.\n   */\n  insufficientHistory: string[]\n}\n\nconst MS_PER_DAY = 86_400_000\n\nexport function judgeSentinelReport(\n  history: SentinelSnapshot[],\n  opts: JudgeSentinelOptions,\n): SentinelReport {\n  const asOfMs = parseIso(opts.asOf, 'asOf')\n  const minIrr = opts.thresholds?.minIrr ?? 0.6\n  const maxKappaDrop = opts.thresholds?.maxKappaDrop ?? 0.15\n  const maxSentinelDrop = opts.thresholds?.maxSentinelDrop ?? 0.1\n  const staleAfterDays = opts.thresholds?.staleAfterDays ?? 30\n\n  const byJudge = new Map<string, Array<SentinelSnapshot & { atMs: number }>>()\n  for (const snapshot of history) {\n    validateSentinelSnapshot(snapshot)\n    const atMs = parseIso(snapshot.at, `snapshot.at (judge \"${snapshot.judgeId}\")`)\n    const arr = byJudge.get(snapshot.judgeId) ?? []\n    arr.push({ ...snapshot, atMs })\n    byJudge.set(snapshot.judgeId, arr)\n  }\n\n  const perJudge: SentinelTrend[] = []\n  const alarms: string[] = []\n  const insufficientHistory: string[] = []\n\n  for (const judgeId of [...byJudge.keys()].sort()) {\n    const snaps = byJudge.get(judgeId)!.sort((a, b) => a.atMs - b.atMs)\n    const newest = snaps[snaps.length - 1]!\n\n    const ageDays = (asOfMs - newest.atMs) / MS_PER_DAY\n    if (ageDays > staleAfterDays) {\n      alarms.push(\n        `judge \"${judgeId}\": stale — newest snapshot ${newest.at} is ${ageDays.toFixed(1)}d old vs asOf (limit ${staleAfterDays}d)`,\n      )\n    }\n\n    // Silent-upgrade trap: only the latest model change matters for current\n    // trust, and only golden-grounded metrics clear it — irr alone proves\n    // judges still agree with each other, not that they're still right.\n    let changeIdx = -1\n    for (let i = 1; i < snaps.length; i++) {\n      if (snaps[i]!.judgeModel !== snaps[i - 1]!.judgeModel) changeIdx = i\n    }\n    if (changeIdx >= 0) {\n      const recalibrated = snaps\n        .slice(changeIdx)\n        .some(\n          (s) =>\n            Number.isFinite(s.metrics.calibrationKappa) ||\n            Number.isFinite(s.metrics.sentinelPassRate),\n        )\n      if (!recalibrated) {\n        alarms.push(\n          `judge \"${judgeId}\": model changed \"${snaps[changeIdx - 1]!.judgeModel}\" → \"${snaps[changeIdx]!.judgeModel}\" at ${snaps[changeIdx]!.at} with no post-change calibration snapshot — silent-upgrade trap; rerun calibrateJudge or the sentinel set against the new model`,\n        )\n      }\n    }\n\n    for (const metric of SENTINEL_METRIC_NAMES) {\n      const values = snaps\n        .filter((s) => typeof s.metrics[metric] === 'number')\n        .map((s) => s.metrics[metric]!)\n      if (values.length === 0) continue\n\n      const convergence = analyzeSeries(values, opts.convergence)\n      if (convergence.state === 'insufficient-data')\n        insufficientHistory.push(`${judgeId}:${metric}`)\n\n      const current = values[values.length - 1]!\n      const baseline = values[0]!\n      const drift = current - baseline\n      const reasons: string[] = []\n      if (convergence.state === 'drifting-down') {\n        reasons.push(\n          `convergence state drifting-down (tail run ${convergence.tailRun}, window mean ${convergence.windowMean.toFixed(3)})`,\n        )\n      }\n      if (metric === 'irr' && current < minIrr) {\n        reasons.push(`irr ${current.toFixed(3)} below floor ${minIrr}`)\n      }\n      if (metric === 'calibrationKappa' && baseline - current > maxKappaDrop) {\n        reasons.push(\n          `calibrationKappa dropped ${(baseline - current).toFixed(3)} from baseline ${baseline.toFixed(3)} (limit ${maxKappaDrop})`,\n        )\n      }\n      if (metric === 'sentinelPassRate' && baseline - current > maxSentinelDrop) {\n        reasons.push(\n          `sentinelPassRate dropped ${(baseline - current).toFixed(3)} from baseline ${baseline.toFixed(3)} (limit ${maxSentinelDrop})`,\n        )\n      }\n\n      const alarmed = reasons.length > 0\n      const trend: SentinelTrend = {\n        judgeId,\n        metric,\n        state: convergence.state,\n        current,\n        baseline,\n        drift,\n        alarmed,\n      }\n      if (alarmed) {\n        trend.reason = reasons.join('; ')\n        alarms.push(`judge \"${judgeId}\" ${metric}: ${trend.reason}`)\n      }\n      perJudge.push(trend)\n    }\n  }\n\n  return { perJudge, alarms, healthy: alarms.length === 0, insufficientHistory }\n}\n\n// ── Verdict stamp ────────────────────────────────────────────────────\n\nexport interface EvalHealthStamp {\n  healthy: boolean\n  alarms: string[]\n}\n\n/**\n * The object a campaign attaches to its verdicts. Compute it from the\n * sentinel report in a post-campaign hook (or a nightly job feeding the\n * next day's campaigns) and store it alongside the verdict payload.\n * `healthy: false` means the judges that produced those verdicts have an\n * unresolved drift / decay / silent-upgrade / staleness alarm — treat the\n * verdicts as carrying an integrity warning until recalibration clears it.\n */\nexport function evalHealthStamp(report: SentinelReport): EvalHealthStamp {\n  return { healthy: report.healthy, alarms: [...report.alarms] }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAyDA,eAAsB,iBACpB,YACA,cACA,aACA,oBACA,UAAmC,CAAC,GACH;AACjC,QAAM,OAAO,MAAM,WAAW,SAAS;AACvC,QAAM,WAAW,MAAM,aAAa,KAAK,QAAQ,aAAa;AAC9D,QAAM,gBAAgB,oBAAI,IAAiC;AAC3D,aAAW,KAAK,UAAU;AACxB,UAAM,MAAM,cAAc,IAAI,EAAE,KAAK,KAAK,CAAC;AAC3C,QAAI,KAAK,CAAC;AACV,kBAAc,IAAI,EAAE,OAAO,GAAG;AAAA,EAChC;AAEA,QAAM,YAAY,QAAQ,aAAa;AACvC,QAAM,SAAS,QAAQ,mBAAmB;AAE1C,QAAM,QAA0F,CAAC;AACjG,aAAW,MAAM,aAAa;AAC5B,eAAW,MAAM,oBAAoB;AACnC,YAAM,KAAK,EAAE,YAAY,GAAG,IAAI,eAAe,IAAI,IAAI,CAAC,GAAG,IAAI,CAAC,EAAE,CAAC;AAAA,IACrE;AAAA,EACF;AAEA,MAAI,SAAS;AACb,MAAI,UAAU;AACd,aAAW,OAAO,MAAM;AACtB,UAAM,KAAK,cAAc,IAAI,IAAI,KAAK;AACtC,QAAI,CAAC,MAAM,GAAG,WAAW,GAAG;AAC1B;AACA;AAAA,IACF;AACA,UAAM,WAAW,GAAG,OAAO,CAAC,MAAM,EAAE,aAAa,IAAI,aAAa,MAAM;AACxE,QAAI,SAAS,WAAW,GAAG;AACzB;AACA;AAAA,IACF;AAEA,eAAW,MAAM,aAAa;AAC5B,YAAM,UAAU,GAAG,WAAW,eAAe,GAAG,EAAE;AAClD,YAAM,IAAI,MAAM,QAAQ,KAAK,UAAU;AACvC,UAAI,MAAM,QAAQ,CAAC,OAAO,SAAS,CAAC,EAAG;AAEvC,iBAAW,MAAM,oBAAoB;AACnC,cAAM,SAAS,SACZ,IAAI,CAAC,MAAM,EAAE,QAAQ,EAAE,CAAC,EACxB,OAAO,CAAC,MAAmB,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,CAAC;AACzE,YAAI,OAAO,WAAW,EAAG;AACzB,cAAM,IAAI,OAAO,QAAQ,WAAW,QAAQ;AAC5C,YAAI,MAAM,KAAM;AAChB,cAAM,OAAO,MAAM,KAAK,CAAC,MAAM,EAAE,eAAe,GAAG,MAAM,EAAE,kBAAkB,EAAE;AAC/E,aAAK,GAAG,KAAK,CAAC;AACd,aAAK,GAAG,KAAK,CAAC;AAAA,MAChB;AAAA,IACF;AACA;AAAA,EACF;AAEA,QAAM,UAA+B,MAClC,OAAO,CAAC,MAAM,EAAE,GAAG,UAAU,CAAC,EAC9B,IAAI,CAAC,MAAM;AACV,UAAM,UAAU,SAAS,EAAE,IAAI,EAAE,EAAE;AACnC,UAAM,WAAW,UAAU,EAAE,IAAI,EAAE,EAAE;AACrC,UAAM,cAAc,mBAAmB,EAAE,IAAI,EAAE,IAAI,QAAQ,uBAAuB,GAAG;AACrF,UAAM,UACJ,KAAK,IAAI,OAAO,KAAK,MAAM,WAAW,KAAK,IAAI,OAAO,KAAK,MAAM,aAAa;AAChF,WAAO;AAAA,MACL,YAAY,EAAE;AAAA,MACd,eAAe,EAAE;AAAA,MACjB,GAAG,EAAE,GAAG;AAAA,MACR;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF,CAAC;AAEH,SAAO,EAAE,OAAO,SAAS,eAAe,QAAQ,aAAa,QAAQ;AACvE;AAIA,SAAS,OACP,QACA,MACA,UACe;AACf,MAAI,OAAO,WAAW,EAAG,QAAO;AAChC,MAAI,SAAS,OAAQ,QAAO,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;AACvE,MAAI,SAAS,MAAO,QAAO,KAAK,IAAI,GAAG,MAAM;AAE7C,QAAM,SAAS,CAAC,GAAG,QAAQ,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,aAAa,EAAE,UAAU,EAAE,CAAC;AAC1E,MAAI,CAAC,OAAQ,QAAO;AACpB,QAAM,YAAY,OAAO,KAAK,OAAO,OAAO,EAAE,CAAC;AAC/C,QAAM,IAAI,cAAc,SAAY,OAAO,QAAQ,SAAS,IAAI;AAEhE,QAAM,SAAS,SACZ,IAAI,CAAC,MAAM;AACV,UAAM,IAAI,OAAO,KAAK,EAAE,OAAO,EAAE,CAAC;AAClC,WAAO;AAAA,MACL,IAAI,EAAE;AAAA,MACN,GAAG,MAAM,SAAY,OAAO,KAAK,CAAC,MAAM,EAAE,QAAQ,CAAC,MAAM,CAAC,IAAI;AAAA,IAChE;AAAA,EACF,CAAC,EACA,OAAO,CAAC,MAAM,EAAE,MAAM,MAAS;AAClC,MAAI,OAAO,WAAW,EAAG,QAAO,KAAK;AACrC,SAAO,OAAO,KAAK,CAAC,GAAG,MAAM,EAAE,KAAK,EAAE,EAAE,EAAE,CAAC,GAAG,KAAK;AACrD;AAEA,SAAS,mBACP,IACA,IACA,YACkC;AAClC,QAAM,IAAI,GAAG;AACb,MAAI,IAAI,EAAG,QAAO,EAAE,OAAO,KAAK,OAAO,IAAI;AAC3C,QAAM,KAAe,CAAC;AACtB,WAAS,IAAI,GAAG,IAAI,YAAY,KAAK;AACnC,UAAM,KAAe,IAAI,MAAM,CAAC;AAChC,UAAM,KAAe,IAAI,MAAM,CAAC;AAChC,aAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAM,MAAM,KAAK,MAAM,KAAK,OAAO,IAAI,CAAC;AACxC,SAAG,CAAC,IAAI,GAAG,GAAG;AACd,SAAG,CAAC,IAAI,GAAG,GAAG;AAAA,IAChB;AACA,UAAM,IAAI,SAAS,IAAI,EAAE;AACzB,QAAI,OAAO,SAAS,CAAC,EAAG,IAAG,KAAK,CAAC;AAAA,EACnC;AACA,KAAG,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AACvB,MAAI,GAAG,WAAW,EAAG,QAAO,EAAE,OAAO,KAAK,OAAO,IAAI;AACrD,SAAO;AAAA,IACL,OAAO,GAAG,KAAK,MAAM,QAAQ,GAAG,MAAM,CAAC;AAAA,IACvC,OAAO,GAAG,KAAK,IAAI,GAAG,SAAS,GAAG,KAAK,MAAM,QAAQ,GAAG,MAAM,CAAC,CAAC;AAAA,EAClE;AACF;AAEA,SAAS,eAAe,QAAyE;AAC/F,SAAO,OAAO,KAAK,UAAU;AAC3B,YAAQ,QAAQ;AAAA,MACd,KAAK;AAAA,MACL,KAAK;AACH,eAAO,IAAI,SAAS,SAAS;AAAA,MAC/B,KAAK;AACH,eAAO,IAAI,SAAS,SAAS,OAAO,IAAI;AAAA,MAC1C,KAAK;AACH,eAAO,IAAI,WAAW,IAAI,YAAY,IAAI,UAAU,IAAI,YAAY;AAAA,MACtE,KAAK,WAAW;AACd,cAAM,MAAM,MAAM,SAAS,OAAO,IAAI,KAAK;AAC3C,eAAO,aAAa,GAAG,EAAE;AAAA,MAC3B;AAAA,MACA,KAAK,eAAe;AAClB,cAAM,MAAM,MAAM,SAAS,OAAO,IAAI,KAAK;AAC3C,eAAO,aAAa,GAAG,EAAE;AAAA,MAC3B;AAAA,MACA;AACE,eAAO;AAAA,IACX;AAAA,EACF;AACF;;;ACpKO,IAAM,wBAAwB,CAAC,OAAO,oBAAoB,kBAAkB;AAmCnF,SAAS,SAAS,OAAe,OAAuB;AACtD,QAAM,KAAK,OAAO,UAAU,WAAW,KAAK,MAAM,KAAK,IAAI;AAC3D,MAAI,CAAC,OAAO,SAAS,EAAE,GAAG;AACxB,UAAM,IAAI;AAAA,MACR,mBAAmB,KAAK,sCAAsC,KAAK,UAAU,KAAK,CAAC;AAAA,IACrF;AAAA,EACF;AACA,SAAO;AACT;AAGO,SAAS,yBAAyB,UAA4B,SAAS,YAAkB;AAC9F,WAAS,SAAS,IAAI,GAAG,MAAM,KAAK;AACpC,MAAI,OAAO,SAAS,YAAY,YAAY,SAAS,QAAQ,WAAW,GAAG;AACzE,UAAM,IAAI,gBAAgB,mBAAmB,MAAM,qCAAqC;AAAA,EAC1F;AACA,MAAI,OAAO,SAAS,eAAe,YAAY,SAAS,WAAW,WAAW,GAAG;AAC/E,UAAM,IAAI,gBAAgB,mBAAmB,MAAM,wCAAwC;AAAA,EAC7F;AACA,MAAI,SAAS,YAAY,QAAQ,OAAO,SAAS,YAAY,UAAU;AACrE,UAAM,IAAI,gBAAgB,mBAAmB,MAAM,2CAA2C;AAAA,EAChG;AACA,aAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,SAAS,OAAO,GAAG;AAC3D,QAAI,CAAE,sBAA4C,SAAS,GAAG,GAAG;AAE/D,YAAM,IAAI;AAAA,QACR,mBAAmB,MAAM,iCAAiC,GAAG,mBAAc,sBAAsB,KAAK,IAAI,CAAC;AAAA,MAC7G;AAAA,IACF;AACA,QAAI,UAAU,UAAa,CAAC,OAAO,SAAS,KAAK,GAAG;AAClD,YAAM,IAAI;AAAA,QACR,mBAAmB,MAAM,YAAY,GAAG,4BAA4B,OAAO,KAAK,CAAC;AAAA,MACnF;AAAA,IACF;AAAA,EACF;AACF;AASO,SAAS,wBACd,QACA,MACkB;AAClB,QAAM,QAAQ,6BAA6B,SAAS,OAAO,0BAA0B,OAAO;AAC5F,MAAI,CAAC,OAAO,SAAS,KAAK,GAAG;AAC3B,UAAM,IAAI;AAAA,MACR,oDAA+C,OAAO,CAAC;AAAA,IACzD;AAAA,EACF;AACA,QAAM,WAA6B,EAAE,GAAG,MAAM,SAAS,EAAE,kBAAkB,MAAM,EAAE;AACnF,2BAAyB,QAAQ;AACjC,SAAO;AACT;AASO,SAAS,sBACd,QACA,MACkB;AAClB,QAAM,MAAM,gBAAgB,SAAS,OAAO,aAAa,OAAO;AAChE,MAAI,CAAC,OAAO,SAAS,GAAG,GAAG;AACzB,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,QAAM,WAA6B,EAAE,GAAG,MAAM,SAAS,EAAE,IAAI,EAAE;AAC/D,2BAAyB,QAAQ;AACjC,SAAO;AACT;AAcO,SAAS,wBACd,QACA,QACA,MACA,UAA8B,CAAC,GACb;AAClB,QAAM,YAAY,QAAQ,aAAa;AACvC,MAAI,CAAC,OAAO,SAAS,SAAS,KAAK,YAAY,GAAG;AAChD,UAAM,IAAI;AAAA,MACR,4EAAuE,SAAS;AAAA,IAClF;AAAA,EACF;AACA,QAAM,aAAa,oBAAI,IAAoB;AAC3C,aAAW,QAAQ,QAAQ;AACzB,QAAI,CAAC,OAAO,SAAS,KAAK,UAAU,GAAG;AACrC,YAAM,IAAI;AAAA,QACR,yCAAyC,KAAK,MAAM;AAAA,MACtD;AAAA,IACF;AACA,QAAI,WAAW,IAAI,KAAK,MAAM,GAAG;AAC/B,YAAM,IAAI,gBAAgB,qDAAqD,KAAK,MAAM,GAAG;AAAA,IAC/F;AACA,eAAW,IAAI,KAAK,QAAQ,KAAK,UAAU;AAAA,EAC7C;AACA,MAAI,SAAS;AACb,MAAI,SAAS;AACb,aAAW,KAAK,QAAQ;AACtB,UAAM,QAAQ,WAAW,IAAI,EAAE,MAAM;AACrC,QAAI,UAAU,OAAW;AACzB,QAAI,CAAC,OAAO,SAAS,EAAE,KAAK,GAAG;AAC7B,YAAM,IAAI;AAAA,QACR,kDAAkD,EAAE,MAAM;AAAA,MAC5D;AAAA,IACF;AACA;AACA,QAAI,KAAK,IAAI,EAAE,QAAQ,KAAK,KAAK,UAAW;AAAA,EAC9C;AACA,MAAI,WAAW,GAAG;AAChB,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,QAAM,WAA6B,EAAE,GAAG,MAAM,SAAS,EAAE,kBAAkB,SAAS,OAAO,EAAE;AAC7F,2BAAyB,QAAQ;AACjC,SAAO;AACT;AAUO,SAAS,sBAAsB,UAA8B,CAAC,GAAkB;AACrF,aAAW,KAAK,QAAS,0BAAyB,CAAC;AACnD,QAAM,QAAQ,QAAQ,IAAI,CAAC,OAAO,EAAE,GAAG,GAAG,SAAS,EAAE,GAAG,EAAE,QAAQ,EAAE,EAAE;AACtE,SAAO;AAAA,IACL,MAAM,OAAO,UAAU;AACrB,+BAAyB,QAAQ;AACjC,YAAM,KAAK,EAAE,GAAG,UAAU,SAAS,EAAE,GAAG,SAAS,QAAQ,EAAE,CAAC;AAAA,IAC9D;AAAA,IACA,MAAM,QAAQ,SAAS;AACrB,YAAM,MAAM,MAAM,IAAI,CAAC,OAAO,EAAE,GAAG,GAAG,SAAS,EAAE,GAAG,EAAE,QAAQ,EAAE,EAAE;AAClE,aAAO,YAAY,SAAY,MAAM,IAAI,OAAO,CAAC,MAAM,EAAE,YAAY,OAAO;AAAA,IAC9E;AAAA,EACF;AACF;AAQO,SAAS,kBAAkB,MAA6B;AAC7D,SAAO;AAAA,IACL,MAAM,OAAO,UAAU;AACrB,+BAAyB,QAAQ;AACjC,YAAM,KAAK,MAAM,OAAO,aAAkB;AAC1C,YAAM,UAAU,MAAM,OAAO,MAAW;AACxC,YAAM,GAAG,MAAM,QAAQ,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;AACzD,YAAM,GAAG,WAAW,MAAM,GAAG,KAAK,UAAU,QAAQ,CAAC;AAAA,GAAM,MAAM;AAAA,IACnE;AAAA,IACA,MAAM,QAAQ,SAAS;AACrB,YAAM,KAAK,MAAM,OAAO,aAAkB;AAC1C,UAAI;AACJ,UAAI;AACF,cAAM,MAAM,GAAG,SAAS,MAAM,MAAM;AAAA,MACtC,SAAS,KAAK;AACZ,YAAK,IAA8B,SAAS,SAAU,QAAO,CAAC;AAC9D,cAAM;AAAA,MACR;AACA,YAAM,QAAQ,IAAI,MAAM,IAAI;AAC5B,YAAM,YAAgC,CAAC;AACvC,eAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,cAAM,OAAO,MAAM,CAAC;AACpB,YAAI,KAAK,KAAK,EAAE,WAAW,EAAG;AAC9B,YAAI;AACJ,YAAI;AACF,mBAAS,KAAK,MAAM,IAAI;AAAA,QAC1B,SAAS,KAAK;AACZ,gBAAM,IAAI;AAAA,YACR,4BAA4B,IAAI,6BAA6B,IAAI,CAAC,KAAM,IAAc,OAAO;AAAA,UAC/F;AAAA,QACF;AACA,cAAM,WAAW;AACjB,iCAAyB,UAAU,GAAG,IAAI,IAAI,IAAI,CAAC,EAAE;AACrD,kBAAU,KAAK,QAAQ;AAAA,MACzB;AACA,aAAO,YAAY,SAAY,YAAY,UAAU,OAAO,CAAC,MAAM,EAAE,YAAY,OAAO;AAAA,IAC1F;AAAA,EACF;AACF;AAmDA,IAAM,aAAa;AAEZ,SAAS,oBACd,SACA,MACgB;AAChB,QAAM,SAAS,SAAS,KAAK,MAAM,MAAM;AACzC,QAAM,SAAS,KAAK,YAAY,UAAU;AAC1C,QAAM,eAAe,KAAK,YAAY,gBAAgB;AACtD,QAAM,kBAAkB,KAAK,YAAY,mBAAmB;AAC5D,QAAM,iBAAiB,KAAK,YAAY,kBAAkB;AAE1D,QAAM,UAAU,oBAAI,IAAwD;AAC5E,aAAW,YAAY,SAAS;AAC9B,6BAAyB,QAAQ;AACjC,UAAM,OAAO,SAAS,SAAS,IAAI,uBAAuB,SAAS,OAAO,IAAI;AAC9E,UAAM,MAAM,QAAQ,IAAI,SAAS,OAAO,KAAK,CAAC;AAC9C,QAAI,KAAK,EAAE,GAAG,UAAU,KAAK,CAAC;AAC9B,YAAQ,IAAI,SAAS,SAAS,GAAG;AAAA,EACnC;AAEA,QAAM,WAA4B,CAAC;AACnC,QAAM,SAAmB,CAAC;AAC1B,QAAM,sBAAgC,CAAC;AAEvC,aAAW,WAAW,CAAC,GAAG,QAAQ,KAAK,CAAC,EAAE,KAAK,GAAG;AAChD,UAAM,QAAQ,QAAQ,IAAI,OAAO,EAAG,KAAK,CAAC,GAAG,MAAM,EAAE,OAAO,EAAE,IAAI;AAClE,UAAM,SAAS,MAAM,MAAM,SAAS,CAAC;AAErC,UAAM,WAAW,SAAS,OAAO,QAAQ;AACzC,QAAI,UAAU,gBAAgB;AAC5B,aAAO;AAAA,QACL,UAAU,OAAO,mCAA8B,OAAO,EAAE,OAAO,QAAQ,QAAQ,CAAC,CAAC,wBAAwB,cAAc;AAAA,MACzH;AAAA,IACF;AAKA,QAAI,YAAY;AAChB,aAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,UAAI,MAAM,CAAC,EAAG,eAAe,MAAM,IAAI,CAAC,EAAG,WAAY,aAAY;AAAA,IACrE;AACA,QAAI,aAAa,GAAG;AAClB,YAAM,eAAe,MAClB,MAAM,SAAS,EACf;AAAA,QACC,CAAC,MACC,OAAO,SAAS,EAAE,QAAQ,gBAAgB,KAC1C,OAAO,SAAS,EAAE,QAAQ,gBAAgB;AAAA,MAC9C;AACF,UAAI,CAAC,cAAc;AACjB,eAAO;AAAA,UACL,UAAU,OAAO,qBAAqB,MAAM,YAAY,CAAC,EAAG,UAAU,aAAQ,MAAM,SAAS,EAAG,UAAU,QAAQ,MAAM,SAAS,EAAG,EAAE;AAAA,QACxI;AAAA,MACF;AAAA,IACF;AAEA,eAAW,UAAU,uBAAuB;AAC1C,YAAM,SAAS,MACZ,OAAO,CAAC,MAAM,OAAO,EAAE,QAAQ,MAAM,MAAM,QAAQ,EACnD,IAAI,CAAC,MAAM,EAAE,QAAQ,MAAM,CAAE;AAChC,UAAI,OAAO,WAAW,EAAG;AAEzB,YAAM,cAAc,cAAc,QAAQ,KAAK,WAAW;AAC1D,UAAI,YAAY,UAAU;AACxB,4BAAoB,KAAK,GAAG,OAAO,IAAI,MAAM,EAAE;AAEjD,YAAM,UAAU,OAAO,OAAO,SAAS,CAAC;AACxC,YAAM,WAAW,OAAO,CAAC;AACzB,YAAM,QAAQ,UAAU;AACxB,YAAM,UAAoB,CAAC;AAC3B,UAAI,YAAY,UAAU,iBAAiB;AACzC,gBAAQ;AAAA,UACN,6CAA6C,YAAY,OAAO,iBAAiB,YAAY,WAAW,QAAQ,CAAC,CAAC;AAAA,QACpH;AAAA,MACF;AACA,UAAI,WAAW,SAAS,UAAU,QAAQ;AACxC,gBAAQ,KAAK,OAAO,QAAQ,QAAQ,CAAC,CAAC,gBAAgB,MAAM,EAAE;AAAA,MAChE;AACA,UAAI,WAAW,sBAAsB,WAAW,UAAU,cAAc;AACtE,gBAAQ;AAAA,UACN,6BAA6B,WAAW,SAAS,QAAQ,CAAC,CAAC,kBAAkB,SAAS,QAAQ,CAAC,CAAC,WAAW,YAAY;AAAA,QACzH;AAAA,MACF;AACA,UAAI,WAAW,sBAAsB,WAAW,UAAU,iBAAiB;AACzE,gBAAQ;AAAA,UACN,6BAA6B,WAAW,SAAS,QAAQ,CAAC,CAAC,kBAAkB,SAAS,QAAQ,CAAC,CAAC,WAAW,eAAe;AAAA,QAC5H;AAAA,MACF;AAEA,YAAM,UAAU,QAAQ,SAAS;AACjC,YAAM,QAAuB;AAAA,QAC3B;AAAA,QACA;AAAA,QACA,OAAO,YAAY;AAAA,QACnB;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,MACF;AACA,UAAI,SAAS;AACX,cAAM,SAAS,QAAQ,KAAK,IAAI;AAChC,eAAO,KAAK,UAAU,OAAO,KAAK,MAAM,KAAK,MAAM,MAAM,EAAE;AAAA,MAC7D;AACA,eAAS,KAAK,KAAK;AAAA,IACrB;AAAA,EACF;AAEA,SAAO,EAAE,UAAU,QAAQ,SAAS,OAAO,WAAW,GAAG,oBAAoB;AAC/E;AAiBO,SAAS,gBAAgB,QAAyC;AACvE,SAAO,EAAE,SAAS,OAAO,SAAS,QAAQ,CAAC,GAAG,OAAO,MAAM,EAAE;AAC/D;","names":[]}
+{"version":3,"sources":["../../src/meta-eval/correlation-study.ts","../../src/meta-eval/sentinel.ts"],"sourcesContent":["/**\n * Correlation study — \"does our eval score predict real-world outcomes?\"\n *\n * This is the load-bearing signal. Takes a TraceStore + OutcomeStore,\n * joins on runId, computes Pearson + Spearman + bootstrap CI for every\n * (evalMetric, outcomeMetric) pair the caller declares.\n *\n * Without this number the framework is ornamental. With it and r > 0.6\n * the framework is a moat — no other agent-eval tool publishes one.\n */\n\nimport { pearsonR, spearmanR } from '../statistics'\nimport { aggregateLlm, llmSpans } from '../trace/query'\nimport type { Run } from '../trace/schema'\nimport type { TraceStore } from '../trace/store'\nimport type { DeploymentOutcome, OutcomeFilter, OutcomeStore } from './outcome-store'\n\nexport interface EvalMetricSpec {\n  id: string\n  /** Extract a scalar from a run (defaults cover score/pass/durationMs/costUsd/tokens). */\n  extract?: (run: Run, store: TraceStore) => Promise<number | null>\n}\n\nexport interface OutcomePair {\n  evalMetric: string\n  outcomeMetric: string\n}\n\nexport interface CorrelationResult {\n  evalMetric: string\n  outcomeMetric: string\n  n: number\n  pearson: number\n  spearman: number\n  /** 95% bootstrap CI for Pearson. */\n  pearsonCi95: { lower: number; upper: number }\n  /** Rough verdict: 'strong' ≥ 0.7, 'moderate' ≥ 0.4, else 'weak'. */\n  verdict: 'strong' | 'moderate' | 'weak'\n}\n\nexport interface CorrelationStudyResult {\n  pairs: CorrelationResult[]\n  joinedSamples: number\n  skippedRuns: number\n}\n\nexport interface CorrelationStudyOptions {\n  /** Only join outcomes captured within this window after run.startedAt. */\n  maxCaptureLagMs?: number\n  /** Restrict to a subset of outcomes (cohort, region, source). */\n  outcomeFilter?: OutcomeFilter\n  /** Which outcome per run to use when multiple exist. Default 'latest'. */\n  reduction?: 'latest' | 'mean' | 'max'\n  /** Bootstrap iterations for the CI. Default 500. */\n  bootstrapIterations?: number\n}\n\nexport async function correlationStudy(\n  traceStore: TraceStore,\n  outcomeStore: OutcomeStore,\n  evalMetrics: EvalMetricSpec[],\n  outcomeMetricNames: string[],\n  options: CorrelationStudyOptions = {},\n): Promise<CorrelationStudyResult> {\n  const runs = await traceStore.listRuns()\n  const outcomes = await outcomeStore.list(options.outcomeFilter)\n  const outcomesByRun = new Map<string, DeploymentOutcome[]>()\n  for (const o of outcomes) {\n    const arr = outcomesByRun.get(o.runId) ?? []\n    arr.push(o)\n    outcomesByRun.set(o.runId, arr)\n  }\n\n  const reduction = options.reduction ?? 'latest'\n  const maxLag = options.maxCaptureLagMs ?? Infinity\n\n  const pairs: Array<{ evalMetric: string; outcomeMetric: string; xs: number[]; ys: number[] }> = []\n  for (const em of evalMetrics) {\n    for (const om of outcomeMetricNames) {\n      pairs.push({ evalMetric: em.id, outcomeMetric: om, xs: [], ys: [] })\n    }\n  }\n\n  let joined = 0\n  let skipped = 0\n  for (const run of runs) {\n    const os = outcomesByRun.get(run.runId)\n    if (!os || os.length === 0) {\n      skipped++\n      continue\n    }\n    const eligible = os.filter((o) => o.capturedAt - run.startedAt <= maxLag)\n    if (eligible.length === 0) {\n      skipped++\n      continue\n    }\n\n    for (const em of evalMetrics) {\n      const extract = em.extract ?? defaultExtract(em.id)\n      const x = await extract(run, traceStore)\n      if (x === null || !Number.isFinite(x)) continue\n\n      for (const om of outcomeMetricNames) {\n        const values = eligible\n          .map((o) => o.metrics[om])\n          .filter((v): v is number => typeof v === 'number' && Number.isFinite(v))\n        if (values.length === 0) continue\n        const y = reduce(values, reduction, eligible)\n        if (y === null) continue\n        const pair = pairs.find((p) => p.evalMetric === em.id && p.outcomeMetric === om)!\n        pair.xs.push(x)\n        pair.ys.push(y)\n      }\n    }\n    joined++\n  }\n\n  const results: CorrelationResult[] = pairs\n    .filter((p) => p.xs.length >= 3)\n    .map((p) => {\n      const pearson = pearsonR(p.xs, p.ys)\n      const spearman = spearmanR(p.xs, p.ys)\n      const pearsonCi95 = bootstrapPearsonCi(p.xs, p.ys, options.bootstrapIterations ?? 500)\n      const verdict: CorrelationResult['verdict'] =\n        Math.abs(pearson) >= 0.7 ? 'strong' : Math.abs(pearson) >= 0.4 ? 'moderate' : 'weak'\n      return {\n        evalMetric: p.evalMetric,\n        outcomeMetric: p.outcomeMetric,\n        n: p.xs.length,\n        pearson,\n        spearman,\n        pearsonCi95,\n        verdict,\n      }\n    })\n\n  return { pairs: results, joinedSamples: joined, skippedRuns: skipped }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────\n\nfunction reduce(\n  values: number[],\n  kind: 'latest' | 'mean' | 'max',\n  outcomes: DeploymentOutcome[],\n): number | null {\n  if (values.length === 0) return null\n  if (kind === 'mean') return values.reduce((a, b) => a + b, 0) / values.length\n  if (kind === 'max') return Math.max(...values)\n  // 'latest': pick the outcome captured last, then lookup its metric\n  const latest = [...outcomes].sort((a, b) => b.capturedAt - a.capturedAt)[0]\n  if (!latest) return null\n  const latestKey = Object.keys(latest.metrics)[0]\n  const v = latestKey !== undefined ? latest.metrics[latestKey] : undefined\n  // For 'latest' we already have `values` aligned; use the last-captured one\n  const paired = outcomes\n    .map((o) => {\n      const k = Object.keys(o.metrics)[0]\n      return {\n        at: o.capturedAt,\n        v: k !== undefined ? values.find((x) => o.metrics[k] === x) : undefined,\n      }\n    })\n    .filter((p) => p.v !== undefined)\n  if (paired.length === 0) return v ?? null\n  return paired.sort((a, b) => b.at - a.at)[0]?.v ?? null\n}\n\nfunction bootstrapPearsonCi(\n  xs: number[],\n  ys: number[],\n  iterations: number,\n): { lower: number; upper: number } {\n  const n = xs.length\n  if (n < 3) return { lower: NaN, upper: NaN }\n  const rs: number[] = []\n  for (let b = 0; b < iterations; b++) {\n    const rx: number[] = new Array(n)\n    const ry: number[] = new Array(n)\n    for (let i = 0; i < n; i++) {\n      const idx = Math.floor(Math.random() * n)\n      rx[i] = xs[idx]!\n      ry[i] = ys[idx]!\n    }\n    const r = pearsonR(rx, ry)\n    if (Number.isFinite(r)) rs.push(r)\n  }\n  rs.sort((a, b) => a - b)\n  if (rs.length === 0) return { lower: NaN, upper: NaN }\n  return {\n    lower: rs[Math.floor(0.025 * rs.length)]!,\n    upper: rs[Math.min(rs.length - 1, Math.floor(0.975 * rs.length))]!,\n  }\n}\n\nfunction defaultExtract(metric: string): (run: Run, store: TraceStore) => Promise<number | null> {\n  return async (run, store) => {\n    switch (metric) {\n      case 'score':\n      case 'overallScore':\n        return run.outcome?.score ?? null\n      case 'pass':\n        return run.outcome?.pass === true ? 1 : 0\n      case 'durationMs':\n        return run.endedAt && run.startedAt ? run.endedAt - run.startedAt : null\n      case 'costUsd': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).costUsd\n      }\n      case 'inputTokens': {\n        const llm = await llmSpans(store, run.runId)\n        return aggregateLlm(llm).inputTokens\n      }\n      default:\n        return null\n    }\n  }\n}\n","/**\n * Judge sentinel — eval trustworthiness as a continuously measured,\n * alarmed trend.\n *\n * Judges are models; models change underneath us; calibration decays.\n * This module composes three existing instruments into one loop:\n *\n *   snapshot — adapters turn real instrument outputs (`calibrateJudge` /\n *              `calibrateJudgeContinuous`, `corpusInterRaterAgreement` /\n *              `continuousAgreement`, a rerun sentinel golden set) into\n *              `SentinelSnapshot`s\n *   series   — `analyzeSeries` (src/series-convergence) classifies each\n *              judge×metric history as stabilized / drifting / noisy\n *   alarm    — `judgeSentinelReport` turns trends + thresholds into named\n *              alarms; `evalHealthStamp` is the tiny object a campaign\n *              attaches to its verdicts\n *\n * Alarm conditions:\n *   - convergence state `drifting-down` on any tracked metric\n *   - irr below `minIrr`\n *   - calibrationKappa / sentinelPassRate dropped vs the series baseline\n *     beyond `maxKappaDrop` / `maxSentinelDrop`\n *   - judge model changed with no post-change golden-grounded snapshot\n *     (the silent-upgrade trap: judges agreeing with each other after a\n *     model swap proves nothing — only re-measuring against gold does)\n *   - newest snapshot older than `staleAfterDays` relative to `asOf`\n *\n * Wire-in: run the snapshot adapters from a post-campaign hook or a\n * nightly job, append to a `SentinelStore`, then compute\n * `judgeSentinelReport(await store.history(), { asOf })` with `asOf` =\n * the campaign timestamp. Attach `evalHealthStamp(report)` to campaign\n * verdicts: an alarmed sentinel means every downstream verdict carries\n * an integrity warning until the judge is recalibrated.\n *\n * The module is clock-free — every timestamp (`at`, `asOf`) is\n * caller-supplied ISO-8601, so reports are reproducible.\n */\n\nimport { ValidationError } from '../errors'\nimport type {\n  CalibrationResult,\n  CandidateScore,\n  ContinuousAgreement,\n  ContinuousCalibrationResult,\n  GoldenItem,\n} from '../judge-calibration'\nimport {\n  analyzeSeries,\n  type SeriesConvergenceOptions,\n  type SeriesConvergenceResult,\n} from '../series-convergence'\nimport type { CorpusAgreementReport } from '../statistics'\n\nexport const SENTINEL_METRIC_NAMES = ['irr', 'calibrationKappa', 'sentinelPassRate'] as const\n\nexport type SentinelMetricName = (typeof SENTINEL_METRIC_NAMES)[number]\n\nexport interface SentinelMetrics {\n  /** Inter-rater reliability — ICC(2,1) from agreement instruments. */\n  irr?: number\n  /** Weighted κ vs the human golden set (`calibrateJudge*`). */\n  calibrationKappa?: number\n  /** Fraction of a fixed sentinel golden set the judge still scores within tolerance. */\n  sentinelPassRate?: number\n}\n\nexport interface SentinelSnapshot {\n  /** Caller-supplied ISO-8601 timestamp — the module never reads a clock. */\n  at: string\n  judgeId: string\n  /** Exact model identity behind the judge (pin the full version string). */\n  judgeModel: string\n  /**\n   * Metrics measured at `at`. May be empty: an empty-metrics snapshot is a\n   * model-change marker — it records the new `judgeModel` immediately and\n   * the silent-upgrade alarm stays raised until a golden-grounded snapshot\n   * (calibrationKappa or sentinelPassRate) follows.\n   */\n  metrics: SentinelMetrics\n}\n\n/** Identity + timestamp the caller supplies alongside an instrument output. */\nexport interface SnapshotMeta {\n  at: string\n  judgeId: string\n  judgeModel: string\n}\n\nfunction parseIso(value: string, label: string): number {\n  const ms = typeof value === 'string' ? Date.parse(value) : NaN\n  if (!Number.isFinite(ms)) {\n    throw new ValidationError(\n      `judge-sentinel: ${label} is not a parseable ISO timestamp: ${JSON.stringify(value)}`,\n    )\n  }\n  return ms\n}\n\n/** Shape guard for snapshots crossing a boundary (store append, JSONL load, report input). */\nexport function validateSentinelSnapshot(snapshot: SentinelSnapshot, source = 'snapshot'): void {\n  parseIso(snapshot.at, `${source}.at`)\n  if (typeof snapshot.judgeId !== 'string' || snapshot.judgeId.length === 0) {\n    throw new ValidationError(`judge-sentinel: ${source}.judgeId must be a non-empty string`)\n  }\n  if (typeof snapshot.judgeModel !== 'string' || snapshot.judgeModel.length === 0) {\n    throw new ValidationError(`judge-sentinel: ${source}.judgeModel must be a non-empty string`)\n  }\n  if (snapshot.metrics === null || typeof snapshot.metrics !== 'object') {\n    throw new ValidationError(`judge-sentinel: ${source}.metrics must be an object (may be empty)`)\n  }\n  for (const [key, value] of Object.entries(snapshot.metrics)) {\n    if (!(SENTINEL_METRIC_NAMES as readonly string[]).includes(key)) {\n      // A typo'd key would silently never trend — reject instead.\n      throw new ValidationError(\n        `judge-sentinel: ${source}.metrics carries unknown key \"${key}\" — known: ${SENTINEL_METRIC_NAMES.join(', ')}`,\n      )\n    }\n    if (value !== undefined && !Number.isFinite(value)) {\n      throw new ValidationError(\n        `judge-sentinel: ${source}.metrics.${key} is not a finite number: ${String(value)}`,\n      )\n    }\n  }\n}\n\n// ── Snapshot adapters — real instrument output → SentinelSnapshot ───\n\n/**\n * From `calibrateJudge` / `calibrateJudgeContinuous` output. Prefers the\n * un-rounded continuous κ_w when the report carries one (integer κ\n * discards information for [0,1] judges).\n */\nexport function snapshotFromCalibration(\n  report: CalibrationResult | ContinuousCalibrationResult,\n  meta: SnapshotMeta,\n): SentinelSnapshot {\n  const kappa = 'weightedKappaContinuous' in report ? report.weightedKappaContinuous : report.kappa\n  if (!Number.isFinite(kappa)) {\n    throw new ValidationError(\n      `snapshotFromCalibration: κ is not finite (n=${report.n}) — calibrate against ≥2 joined golden items before snapshotting`,\n    )\n  }\n  const snapshot: SentinelSnapshot = { ...meta, metrics: { calibrationKappa: kappa } }\n  validateSentinelSnapshot(snapshot)\n  return snapshot\n}\n\n/**\n * From `corpusInterRaterAgreement` (statistics.ts) or `continuousAgreement`\n * (judge-calibration.ts) output. IRR = ICC(2,1) — the reliability\n * coefficient both instruments compute. For ensemble-level agreement,\n * `meta.judgeId` names the ensemble and `meta.judgeModel` pins its\n * composition (e.g. a joined list of member model strings).\n */\nexport function snapshotFromAgreement(\n  report: CorpusAgreementReport | ContinuousAgreement,\n  meta: SnapshotMeta,\n): SentinelSnapshot {\n  const irr = 'overallIcc' in report ? report.overallIcc : report.icc\n  if (!Number.isFinite(irr)) {\n    throw new ValidationError(\n      'snapshotFromAgreement: ICC is not finite — agreement needs ≥2 raters and ≥2 complete items',\n    )\n  }\n  const snapshot: SentinelSnapshot = { ...meta, metrics: { irr } }\n  validateSentinelSnapshot(snapshot)\n  return snapshot\n}\n\nexport interface SentinelSetOptions {\n  /** Max |judge − human| for an item to count as passed. Default 0.1 (tuned for [0,1] scores). */\n  tolerance?: number\n}\n\n/**\n * From a rerun of a fixed sentinel golden set: the same `GoldenItem`s the\n * judge was originally calibrated on, re-scored periodically. Pass rate =\n * fraction of joined items where the judge stays within `tolerance` of the\n * human score. Items the judge didn't score are excluded from the\n * denominator; zero joined items is an error, not a 0% pass rate.\n */\nexport function snapshotFromSentinelSet(\n  scores: CandidateScore[],\n  golden: GoldenItem[],\n  meta: SnapshotMeta,\n  options: SentinelSetOptions = {},\n): SentinelSnapshot {\n  const tolerance = options.tolerance ?? 0.1\n  if (!Number.isFinite(tolerance) || tolerance < 0) {\n    throw new ValidationError(\n      `snapshotFromSentinelSet: tolerance must be a finite number ≥ 0, got ${tolerance}`,\n    )\n  }\n  const goldenById = new Map<string, number>()\n  for (const item of golden) {\n    if (!Number.isFinite(item.humanScore)) {\n      throw new ValidationError(\n        `snapshotFromSentinelSet: golden item \"${item.itemId}\" has a non-finite humanScore`,\n      )\n    }\n    if (goldenById.has(item.itemId)) {\n      throw new ValidationError(`snapshotFromSentinelSet: duplicate golden itemId \"${item.itemId}\"`)\n    }\n    goldenById.set(item.itemId, item.humanScore)\n  }\n  let joined = 0\n  let passed = 0\n  for (const s of scores) {\n    const human = goldenById.get(s.itemId)\n    if (human === undefined) continue\n    if (!Number.isFinite(s.score)) {\n      throw new ValidationError(\n        `snapshotFromSentinelSet: judge score for item \"${s.itemId}\" is not finite`,\n      )\n    }\n    joined++\n    if (Math.abs(s.score - human) <= tolerance) passed++\n  }\n  if (joined === 0) {\n    throw new ValidationError(\n      'snapshotFromSentinelSet: no judge scores joined the golden set — itemIds mismatch or empty sentinel run',\n    )\n  }\n  const snapshot: SentinelSnapshot = { ...meta, metrics: { sentinelPassRate: passed / joined } }\n  validateSentinelSnapshot(snapshot)\n  return snapshot\n}\n\n// ── Persistence ──────────────────────────────────────────────────────\n\nexport interface SentinelStore {\n  append(snapshot: SentinelSnapshot): Promise<void>\n  /** All snapshots (optionally for one judge), in append order. */\n  history(judgeId?: string): Promise<SentinelSnapshot[]>\n}\n\nexport function inMemorySentinelStore(initial: SentinelSnapshot[] = []): SentinelStore {\n  for (const s of initial) validateSentinelSnapshot(s)\n  const items = initial.map((s) => ({ ...s, metrics: { ...s.metrics } }))\n  return {\n    async append(snapshot) {\n      validateSentinelSnapshot(snapshot)\n      items.push({ ...snapshot, metrics: { ...snapshot.metrics } })\n    },\n    async history(judgeId) {\n      const all = items.map((s) => ({ ...s, metrics: { ...s.metrics } }))\n      return judgeId === undefined ? all : all.filter((s) => s.judgeId === judgeId)\n    },\n  }\n}\n\n/**\n * JSONL store — one snapshot per line, appended atomically per call.\n * A corrupt or shape-invalid line is a loud error naming the file and\n * line number: a sentinel history that silently drops records would\n * defeat the drift detection it exists to provide.\n */\nexport function fileSentinelStore(path: string): SentinelStore {\n  return {\n    async append(snapshot) {\n      validateSentinelSnapshot(snapshot)\n      const fs = await import('node:fs/promises')\n      const pathMod = await import('node:path')\n      await fs.mkdir(pathMod.dirname(path), { recursive: true })\n      await fs.appendFile(path, `${JSON.stringify(snapshot)}\\n`, 'utf8')\n    },\n    async history(judgeId) {\n      const fs = await import('node:fs/promises')\n      let raw: string\n      try {\n        raw = await fs.readFile(path, 'utf8')\n      } catch (err) {\n        if ((err as NodeJS.ErrnoException).code === 'ENOENT') return []\n        throw err\n      }\n      const lines = raw.split('\\n')\n      const snapshots: SentinelSnapshot[] = []\n      for (let i = 0; i < lines.length; i++) {\n        const line = lines[i]!\n        if (line.trim().length === 0) continue\n        let parsed: unknown\n        try {\n          parsed = JSON.parse(line)\n        } catch (err) {\n          throw new ValidationError(\n            `judge-sentinel: store at ${path} has a corrupt JSONL line ${i + 1}: ${(err as Error).message}`,\n          )\n        }\n        const snapshot = parsed as SentinelSnapshot\n        validateSentinelSnapshot(snapshot, `${path}:${i + 1}`)\n        snapshots.push(snapshot)\n      }\n      return judgeId === undefined ? snapshots : snapshots.filter((s) => s.judgeId === judgeId)\n    },\n  }\n}\n\n// ── Trend report ─────────────────────────────────────────────────────\n\nexport interface SentinelThresholds {\n  /** Floor for irr — below it the judge ensemble is unreliable regardless of trend. Default 0.6. */\n  minIrr?: number\n  /** Max tolerated calibrationKappa drop vs the series baseline. Default 0.15. */\n  maxKappaDrop?: number\n  /** Max tolerated sentinelPassRate drop vs the series baseline. Default 0.1. */\n  maxSentinelDrop?: number\n  /** Newest snapshot older than this (days, vs asOf) = the sentinel itself went blind. Default 30. */\n  staleAfterDays?: number\n}\n\nexport interface JudgeSentinelOptions {\n  /** Caller-supplied ISO timestamp staleness is measured against. */\n  asOf: string\n  thresholds?: SentinelThresholds\n  /** Forwarded to `analyzeSeries` (window, stableCv, driftRun). */\n  convergence?: SeriesConvergenceOptions\n}\n\nexport interface SentinelTrend {\n  judgeId: string\n  metric: SentinelMetricName\n  /** Verbatim `analyzeSeries` state for this judge×metric history. */\n  state: SeriesConvergenceResult['state']\n  /** Newest value in the series. */\n  current: number\n  /** Oldest value in the series — the reference the drop thresholds compare against. */\n  baseline: number\n  /** current − baseline (signed; negative = decayed). */\n  drift: number\n  alarmed: boolean\n  reason?: string\n}\n\nexport interface SentinelReport {\n  perJudge: SentinelTrend[]\n  alarms: string[]\n  healthy: boolean\n  /**\n   * `judgeId:metric` pairs with too few snapshots for the convergence\n   * machine. Named, not counted — a blind spot you can't see is a blind\n   * spot you won't fix. Floor/drop checks still apply to thin series, so\n   * insufficient history alone never masks an alarm.\n   */\n  insufficientHistory: string[]\n}\n\nconst MS_PER_DAY = 86_400_000\n\nexport function judgeSentinelReport(\n  history: SentinelSnapshot[],\n  opts: JudgeSentinelOptions,\n): SentinelReport {\n  const asOfMs = parseIso(opts.asOf, 'asOf')\n  const minIrr = opts.thresholds?.minIrr ?? 0.6\n  const maxKappaDrop = opts.thresholds?.maxKappaDrop ?? 0.15\n  const maxSentinelDrop = opts.thresholds?.maxSentinelDrop ?? 0.1\n  const staleAfterDays = opts.thresholds?.staleAfterDays ?? 30\n\n  const byJudge = new Map<string, Array<SentinelSnapshot & { atMs: number }>>()\n  for (const snapshot of history) {\n    validateSentinelSnapshot(snapshot)\n    const atMs = parseIso(snapshot.at, `snapshot.at (judge \"${snapshot.judgeId}\")`)\n    const arr = byJudge.get(snapshot.judgeId) ?? []\n    arr.push({ ...snapshot, atMs })\n    byJudge.set(snapshot.judgeId, arr)\n  }\n\n  const perJudge: SentinelTrend[] = []\n  const alarms: string[] = []\n  const insufficientHistory: string[] = []\n\n  for (const judgeId of [...byJudge.keys()].sort()) {\n    const snaps = byJudge.get(judgeId)!.sort((a, b) => a.atMs - b.atMs)\n    const newest = snaps[snaps.length - 1]!\n\n    const ageDays = (asOfMs - newest.atMs) / MS_PER_DAY\n    if (ageDays > staleAfterDays) {\n      alarms.push(\n        `judge \"${judgeId}\": stale — newest snapshot ${newest.at} is ${ageDays.toFixed(1)}d old vs asOf (limit ${staleAfterDays}d)`,\n      )\n    }\n\n    // Silent-upgrade trap: only the latest model change matters for current\n    // trust, and only golden-grounded metrics clear it — irr alone proves\n    // judges still agree with each other, not that they're still right.\n    let changeIdx = -1\n    for (let i = 1; i < snaps.length; i++) {\n      if (snaps[i]!.judgeModel !== snaps[i - 1]!.judgeModel) changeIdx = i\n    }\n    if (changeIdx >= 0) {\n      const recalibrated = snaps\n        .slice(changeIdx)\n        .some(\n          (s) =>\n            Number.isFinite(s.metrics.calibrationKappa) ||\n            Number.isFinite(s.metrics.sentinelPassRate),\n        )\n      if (!recalibrated) {\n        alarms.push(\n          `judge \"${judgeId}\": model changed \"${snaps[changeIdx - 1]!.judgeModel}\" → \"${snaps[changeIdx]!.judgeModel}\" at ${snaps[changeIdx]!.at} with no post-change calibration snapshot — silent-upgrade trap; rerun calibrateJudge or the sentinel set against the new model`,\n        )\n      }\n    }\n\n    for (const metric of SENTINEL_METRIC_NAMES) {\n      const values = snaps\n        .filter((s) => typeof s.metrics[metric] === 'number')\n        .map((s) => s.metrics[metric]!)\n      if (values.length === 0) continue\n\n      const convergence = analyzeSeries(values, opts.convergence)\n      if (convergence.state === 'insufficient-data')\n        insufficientHistory.push(`${judgeId}:${metric}`)\n\n      const current = values[values.length - 1]!\n      const baseline = values[0]!\n      const drift = current - baseline\n      const reasons: string[] = []\n      if (convergence.state === 'drifting-down') {\n        reasons.push(\n          `convergence state drifting-down (tail run ${convergence.tailRun}, window mean ${convergence.windowMean.toFixed(3)})`,\n        )\n      }\n      if (metric === 'irr' && current < minIrr) {\n        reasons.push(`irr ${current.toFixed(3)} below floor ${minIrr}`)\n      }\n      if (metric === 'calibrationKappa' && baseline - current > maxKappaDrop) {\n        reasons.push(\n          `calibrationKappa dropped ${(baseline - current).toFixed(3)} from baseline ${baseline.toFixed(3)} (limit ${maxKappaDrop})`,\n        )\n      }\n      if (metric === 'sentinelPassRate' && baseline - current > maxSentinelDrop) {\n        reasons.push(\n          `sentinelPassRate dropped ${(baseline - current).toFixed(3)} from baseline ${baseline.toFixed(3)} (limit ${maxSentinelDrop})`,\n        )\n      }\n\n      const alarmed = reasons.length > 0\n      const trend: SentinelTrend = {\n        judgeId,\n        metric,\n        state: convergence.state,\n        current,\n        baseline,\n        drift,\n        alarmed,\n      }\n      if (alarmed) {\n        trend.reason = reasons.join('; ')\n        alarms.push(`judge \"${judgeId}\" ${metric}: ${trend.reason}`)\n      }\n      perJudge.push(trend)\n    }\n  }\n\n  return { perJudge, alarms, healthy: alarms.length === 0, insufficientHistory }\n}\n\n// ── Verdict stamp ────────────────────────────────────────────────────\n\nexport interface EvalHealthStamp {\n  healthy: boolean\n  alarms: string[]\n}\n\n/**\n * The object a campaign attaches to its verdicts. Compute it from the\n * sentinel report in a post-campaign hook (or a nightly job feeding the\n * next day's campaigns) and store it alongside the verdict payload.\n * `healthy: false` means the judges that produced those verdicts have an\n * unresolved drift / decay / silent-upgrade / staleness alarm — treat the\n * verdicts as carrying an integrity warning until recalibration clears it.\n */\nexport function evalHealthStamp(report: SentinelReport): EvalHealthStamp {\n  return { healthy: report.healthy, alarms: [...report.alarms] }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;AAyDA,eAAsB,iBACpB,YACA,cACA,aACA,oBACA,UAAmC,CAAC,GACH;AACjC,QAAM,OAAO,MAAM,WAAW,SAAS;AACvC,QAAM,WAAW,MAAM,aAAa,KAAK,QAAQ,aAAa;AAC9D,QAAM,gBAAgB,oBAAI,IAAiC;AAC3D,aAAW,KAAK,UAAU;AACxB,UAAM,MAAM,cAAc,IAAI,EAAE,KAAK,KAAK,CAAC;AAC3C,QAAI,KAAK,CAAC;AACV,kBAAc,IAAI,EAAE,OAAO,GAAG;AAAA,EAChC;AAEA,QAAM,YAAY,QAAQ,aAAa;AACvC,QAAM,SAAS,QAAQ,mBAAmB;AAE1C,QAAM,QAA0F,CAAC;AACjG,aAAW,MAAM,aAAa;AAC5B,eAAW,MAAM,oBAAoB;AACnC,YAAM,KAAK,EAAE,YAAY,GAAG,IAAI,eAAe,IAAI,IAAI,CAAC,GAAG,IAAI,CAAC,EAAE,CAAC;AAAA,IACrE;AAAA,EACF;AAEA,MAAI,SAAS;AACb,MAAI,UAAU;AACd,aAAW,OAAO,MAAM;AACtB,UAAM,KAAK,cAAc,IAAI,IAAI,KAAK;AACtC,QAAI,CAAC,MAAM,GAAG,WAAW,GAAG;AAC1B;AACA;AAAA,IACF;AACA,UAAM,WAAW,GAAG,OAAO,CAAC,MAAM,EAAE,aAAa,IAAI,aAAa,MAAM;AACxE,QAAI,SAAS,WAAW,GAAG;AACzB;AACA;AAAA,IACF;AAEA,eAAW,MAAM,aAAa;AAC5B,YAAM,UAAU,GAAG,WAAW,eAAe,GAAG,EAAE;AAClD,YAAM,IAAI,MAAM,QAAQ,KAAK,UAAU;AACvC,UAAI,MAAM,QAAQ,CAAC,OAAO,SAAS,CAAC,EAAG;AAEvC,iBAAW,MAAM,oBAAoB;AACnC,cAAM,SAAS,SACZ,IAAI,CAAC,MAAM,EAAE,QAAQ,EAAE,CAAC,EACxB,OAAO,CAAC,MAAmB,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,CAAC;AACzE,YAAI,OAAO,WAAW,EAAG;AACzB,cAAM,IAAI,OAAO,QAAQ,WAAW,QAAQ;AAC5C,YAAI,MAAM,KAAM;AAChB,cAAM,OAAO,MAAM,KAAK,CAAC,MAAM,EAAE,eAAe,GAAG,MAAM,EAAE,kBAAkB,EAAE;AAC/E,aAAK,GAAG,KAAK,CAAC;AACd,aAAK,GAAG,KAAK,CAAC;AAAA,MAChB;AAAA,IACF;AACA;AAAA,EACF;AAEA,QAAM,UAA+B,MAClC,OAAO,CAAC,MAAM,EAAE,GAAG,UAAU,CAAC,EAC9B,IAAI,CAAC,MAAM;AACV,UAAM,UAAU,SAAS,EAAE,IAAI,EAAE,EAAE;AACnC,UAAM,WAAW,UAAU,EAAE,IAAI,EAAE,EAAE;AACrC,UAAM,cAAc,mBAAmB,EAAE,IAAI,EAAE,IAAI,QAAQ,uBAAuB,GAAG;AACrF,UAAM,UACJ,KAAK,IAAI,OAAO,KAAK,MAAM,WAAW,KAAK,IAAI,OAAO,KAAK,MAAM,aAAa;AAChF,WAAO;AAAA,MACL,YAAY,EAAE;AAAA,MACd,eAAe,EAAE;AAAA,MACjB,GAAG,EAAE,GAAG;AAAA,MACR;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF,CAAC;AAEH,SAAO,EAAE,OAAO,SAAS,eAAe,QAAQ,aAAa,QAAQ;AACvE;AAIA,SAAS,OACP,QACA,MACA,UACe;AACf,MAAI,OAAO,WAAW,EAAG,QAAO;AAChC,MAAI,SAAS,OAAQ,QAAO,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;AACvE,MAAI,SAAS,MAAO,QAAO,KAAK,IAAI,GAAG,MAAM;AAE7C,QAAM,SAAS,CAAC,GAAG,QAAQ,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,aAAa,EAAE,UAAU,EAAE,CAAC;AAC1E,MAAI,CAAC,OAAQ,QAAO;AACpB,QAAM,YAAY,OAAO,KAAK,OAAO,OAAO,EAAE,CAAC;AAC/C,QAAM,IAAI,cAAc,SAAY,OAAO,QAAQ,SAAS,IAAI;AAEhE,QAAM,SAAS,SACZ,IAAI,CAAC,MAAM;AACV,UAAM,IAAI,OAAO,KAAK,EAAE,OAAO,EAAE,CAAC;AAClC,WAAO;AAAA,MACL,IAAI,EAAE;AAAA,MACN,GAAG,MAAM,SAAY,OAAO,KAAK,CAAC,MAAM,EAAE,QAAQ,CAAC,MAAM,CAAC,IAAI;AAAA,IAChE;AAAA,EACF,CAAC,EACA,OAAO,CAAC,MAAM,EAAE,MAAM,MAAS;AAClC,MAAI,OAAO,WAAW,EAAG,QAAO,KAAK;AACrC,SAAO,OAAO,KAAK,CAAC,GAAG,MAAM,EAAE,KAAK,EAAE,EAAE,EAAE,CAAC,GAAG,KAAK;AACrD;AAEA,SAAS,mBACP,IACA,IACA,YACkC;AAClC,QAAM,IAAI,GAAG;AACb,MAAI,IAAI,EAAG,QAAO,EAAE,OAAO,KAAK,OAAO,IAAI;AAC3C,QAAM,KAAe,CAAC;AACtB,WAAS,IAAI,GAAG,IAAI,YAAY,KAAK;AACnC,UAAM,KAAe,IAAI,MAAM,CAAC;AAChC,UAAM,KAAe,IAAI,MAAM,CAAC;AAChC,aAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAM,MAAM,KAAK,MAAM,KAAK,OAAO,IAAI,CAAC;AACxC,SAAG,CAAC,IAAI,GAAG,GAAG;AACd,SAAG,CAAC,IAAI,GAAG,GAAG;AAAA,IAChB;AACA,UAAM,IAAI,SAAS,IAAI,EAAE;AACzB,QAAI,OAAO,SAAS,CAAC,EAAG,IAAG,KAAK,CAAC;AAAA,EACnC;AACA,KAAG,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AACvB,MAAI,GAAG,WAAW,EAAG,QAAO,EAAE,OAAO,KAAK,OAAO,IAAI;AACrD,SAAO;AAAA,IACL,OAAO,GAAG,KAAK,MAAM,QAAQ,GAAG,MAAM,CAAC;AAAA,IACvC,OAAO,GAAG,KAAK,IAAI,GAAG,SAAS,GAAG,KAAK,MAAM,QAAQ,GAAG,MAAM,CAAC,CAAC;AAAA,EAClE;AACF;AAEA,SAAS,eAAe,QAAyE;AAC/F,SAAO,OAAO,KAAK,UAAU;AAC3B,YAAQ,QAAQ;AAAA,MACd,KAAK;AAAA,MACL,KAAK;AACH,eAAO,IAAI,SAAS,SAAS;AAAA,MAC/B,KAAK;AACH,eAAO,IAAI,SAAS,SAAS,OAAO,IAAI;AAAA,MAC1C,KAAK;AACH,eAAO,IAAI,WAAW,IAAI,YAAY,IAAI,UAAU,IAAI,YAAY;AAAA,MACtE,KAAK,WAAW;AACd,cAAM,MAAM,MAAM,SAAS,OAAO,IAAI,KAAK;AAC3C,eAAO,aAAa,GAAG,EAAE;AAAA,MAC3B;AAAA,MACA,KAAK,eAAe;AAClB,cAAM,MAAM,MAAM,SAAS,OAAO,IAAI,KAAK;AAC3C,eAAO,aAAa,GAAG,EAAE;AAAA,MAC3B;AAAA,MACA;AACE,eAAO;AAAA,IACX;AAAA,EACF;AACF;;;ACpKO,IAAM,wBAAwB,CAAC,OAAO,oBAAoB,kBAAkB;AAmCnF,SAAS,SAAS,OAAe,OAAuB;AACtD,QAAM,KAAK,OAAO,UAAU,WAAW,KAAK,MAAM,KAAK,IAAI;AAC3D,MAAI,CAAC,OAAO,SAAS,EAAE,GAAG;AACxB,UAAM,IAAI;AAAA,MACR,mBAAmB,KAAK,sCAAsC,KAAK,UAAU,KAAK,CAAC;AAAA,IACrF;AAAA,EACF;AACA,SAAO;AACT;AAGO,SAAS,yBAAyB,UAA4B,SAAS,YAAkB;AAC9F,WAAS,SAAS,IAAI,GAAG,MAAM,KAAK;AACpC,MAAI,OAAO,SAAS,YAAY,YAAY,SAAS,QAAQ,WAAW,GAAG;AACzE,UAAM,IAAI,gBAAgB,mBAAmB,MAAM,qCAAqC;AAAA,EAC1F;AACA,MAAI,OAAO,SAAS,eAAe,YAAY,SAAS,WAAW,WAAW,GAAG;AAC/E,UAAM,IAAI,gBAAgB,mBAAmB,MAAM,wCAAwC;AAAA,EAC7F;AACA,MAAI,SAAS,YAAY,QAAQ,OAAO,SAAS,YAAY,UAAU;AACrE,UAAM,IAAI,gBAAgB,mBAAmB,MAAM,2CAA2C;AAAA,EAChG;AACA,aAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,SAAS,OAAO,GAAG;AAC3D,QAAI,CAAE,sBAA4C,SAAS,GAAG,GAAG;AAE/D,YAAM,IAAI;AAAA,QACR,mBAAmB,MAAM,iCAAiC,GAAG,mBAAc,sBAAsB,KAAK,IAAI,CAAC;AAAA,MAC7G;AAAA,IACF;AACA,QAAI,UAAU,UAAa,CAAC,OAAO,SAAS,KAAK,GAAG;AAClD,YAAM,IAAI;AAAA,QACR,mBAAmB,MAAM,YAAY,GAAG,4BAA4B,OAAO,KAAK,CAAC;AAAA,MACnF;AAAA,IACF;AAAA,EACF;AACF;AASO,SAAS,wBACd,QACA,MACkB;AAClB,QAAM,QAAQ,6BAA6B,SAAS,OAAO,0BAA0B,OAAO;AAC5F,MAAI,CAAC,OAAO,SAAS,KAAK,GAAG;AAC3B,UAAM,IAAI;AAAA,MACR,oDAA+C,OAAO,CAAC;AAAA,IACzD;AAAA,EACF;AACA,QAAM,WAA6B,EAAE,GAAG,MAAM,SAAS,EAAE,kBAAkB,MAAM,EAAE;AACnF,2BAAyB,QAAQ;AACjC,SAAO;AACT;AASO,SAAS,sBACd,QACA,MACkB;AAClB,QAAM,MAAM,gBAAgB,SAAS,OAAO,aAAa,OAAO;AAChE,MAAI,CAAC,OAAO,SAAS,GAAG,GAAG;AACzB,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,QAAM,WAA6B,EAAE,GAAG,MAAM,SAAS,EAAE,IAAI,EAAE;AAC/D,2BAAyB,QAAQ;AACjC,SAAO;AACT;AAcO,SAAS,wBACd,QACA,QACA,MACA,UAA8B,CAAC,GACb;AAClB,QAAM,YAAY,QAAQ,aAAa;AACvC,MAAI,CAAC,OAAO,SAAS,SAAS,KAAK,YAAY,GAAG;AAChD,UAAM,IAAI;AAAA,MACR,4EAAuE,SAAS;AAAA,IAClF;AAAA,EACF;AACA,QAAM,aAAa,oBAAI,IAAoB;AAC3C,aAAW,QAAQ,QAAQ;AACzB,QAAI,CAAC,OAAO,SAAS,KAAK,UAAU,GAAG;AACrC,YAAM,IAAI;AAAA,QACR,yCAAyC,KAAK,MAAM;AAAA,MACtD;AAAA,IACF;AACA,QAAI,WAAW,IAAI,KAAK,MAAM,GAAG;AAC/B,YAAM,IAAI,gBAAgB,qDAAqD,KAAK,MAAM,GAAG;AAAA,IAC/F;AACA,eAAW,IAAI,KAAK,QAAQ,KAAK,UAAU;AAAA,EAC7C;AACA,MAAI,SAAS;AACb,MAAI,SAAS;AACb,aAAW,KAAK,QAAQ;AACtB,UAAM,QAAQ,WAAW,IAAI,EAAE,MAAM;AACrC,QAAI,UAAU,OAAW;AACzB,QAAI,CAAC,OAAO,SAAS,EAAE,KAAK,GAAG;AAC7B,YAAM,IAAI;AAAA,QACR,kDAAkD,EAAE,MAAM;AAAA,MAC5D;AAAA,IACF;AACA;AACA,QAAI,KAAK,IAAI,EAAE,QAAQ,KAAK,KAAK,UAAW;AAAA,EAC9C;AACA,MAAI,WAAW,GAAG;AAChB,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,QAAM,WAA6B,EAAE,GAAG,MAAM,SAAS,EAAE,kBAAkB,SAAS,OAAO,EAAE;AAC7F,2BAAyB,QAAQ;AACjC,SAAO;AACT;AAUO,SAAS,sBAAsB,UAA8B,CAAC,GAAkB;AACrF,aAAW,KAAK,QAAS,0BAAyB,CAAC;AACnD,QAAM,QAAQ,QAAQ,IAAI,CAAC,OAAO,EAAE,GAAG,GAAG,SAAS,EAAE,GAAG,EAAE,QAAQ,EAAE,EAAE;AACtE,SAAO;AAAA,IACL,MAAM,OAAO,UAAU;AACrB,+BAAyB,QAAQ;AACjC,YAAM,KAAK,EAAE,GAAG,UAAU,SAAS,EAAE,GAAG,SAAS,QAAQ,EAAE,CAAC;AAAA,IAC9D;AAAA,IACA,MAAM,QAAQ,SAAS;AACrB,YAAM,MAAM,MAAM,IAAI,CAAC,OAAO,EAAE,GAAG,GAAG,SAAS,EAAE,GAAG,EAAE,QAAQ,EAAE,EAAE;AAClE,aAAO,YAAY,SAAY,MAAM,IAAI,OAAO,CAAC,MAAM,EAAE,YAAY,OAAO;AAAA,IAC9E;AAAA,EACF;AACF;AAQO,SAAS,kBAAkB,MAA6B;AAC7D,SAAO;AAAA,IACL,MAAM,OAAO,UAAU;AACrB,+BAAyB,QAAQ;AACjC,YAAM,KAAK,MAAM,OAAO,aAAkB;AAC1C,YAAM,UAAU,MAAM,OAAO,MAAW;AACxC,YAAM,GAAG,MAAM,QAAQ,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;AACzD,YAAM,GAAG,WAAW,MAAM,GAAG,KAAK,UAAU,QAAQ,CAAC;AAAA,GAAM,MAAM;AAAA,IACnE;AAAA,IACA,MAAM,QAAQ,SAAS;AACrB,YAAM,KAAK,MAAM,OAAO,aAAkB;AAC1C,UAAI;AACJ,UAAI;AACF,cAAM,MAAM,GAAG,SAAS,MAAM,MAAM;AAAA,MACtC,SAAS,KAAK;AACZ,YAAK,IAA8B,SAAS,SAAU,QAAO,CAAC;AAC9D,cAAM;AAAA,MACR;AACA,YAAM,QAAQ,IAAI,MAAM,IAAI;AAC5B,YAAM,YAAgC,CAAC;AACvC,eAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,cAAM,OAAO,MAAM,CAAC;AACpB,YAAI,KAAK,KAAK,EAAE,WAAW,EAAG;AAC9B,YAAI;AACJ,YAAI;AACF,mBAAS,KAAK,MAAM,IAAI;AAAA,QAC1B,SAAS,KAAK;AACZ,gBAAM,IAAI;AAAA,YACR,4BAA4B,IAAI,6BAA6B,IAAI,CAAC,KAAM,IAAc,OAAO;AAAA,UAC/F;AAAA,QACF;AACA,cAAM,WAAW;AACjB,iCAAyB,UAAU,GAAG,IAAI,IAAI,IAAI,CAAC,EAAE;AACrD,kBAAU,KAAK,QAAQ;AAAA,MACzB;AACA,aAAO,YAAY,SAAY,YAAY,UAAU,OAAO,CAAC,MAAM,EAAE,YAAY,OAAO;AAAA,IAC1F;AAAA,EACF;AACF;AAmDA,IAAM,aAAa;AAEZ,SAAS,oBACd,SACA,MACgB;AAChB,QAAM,SAAS,SAAS,KAAK,MAAM,MAAM;AACzC,QAAM,SAAS,KAAK,YAAY,UAAU;AAC1C,QAAM,eAAe,KAAK,YAAY,gBAAgB;AACtD,QAAM,kBAAkB,KAAK,YAAY,mBAAmB;AAC5D,QAAM,iBAAiB,KAAK,YAAY,kBAAkB;AAE1D,QAAM,UAAU,oBAAI,IAAwD;AAC5E,aAAW,YAAY,SAAS;AAC9B,6BAAyB,QAAQ;AACjC,UAAM,OAAO,SAAS,SAAS,IAAI,uBAAuB,SAAS,OAAO,IAAI;AAC9E,UAAM,MAAM,QAAQ,IAAI,SAAS,OAAO,KAAK,CAAC;AAC9C,QAAI,KAAK,EAAE,GAAG,UAAU,KAAK,CAAC;AAC9B,YAAQ,IAAI,SAAS,SAAS,GAAG;AAAA,EACnC;AAEA,QAAM,WAA4B,CAAC;AACnC,QAAM,SAAmB,CAAC;AAC1B,QAAM,sBAAgC,CAAC;AAEvC,aAAW,WAAW,CAAC,GAAG,QAAQ,KAAK,CAAC,EAAE,KAAK,GAAG;AAChD,UAAM,QAAQ,QAAQ,IAAI,OAAO,EAAG,KAAK,CAAC,GAAG,MAAM,EAAE,OAAO,EAAE,IAAI;AAClE,UAAM,SAAS,MAAM,MAAM,SAAS,CAAC;AAErC,UAAM,WAAW,SAAS,OAAO,QAAQ;AACzC,QAAI,UAAU,gBAAgB;AAC5B,aAAO;AAAA,QACL,UAAU,OAAO,mCAA8B,OAAO,EAAE,OAAO,QAAQ,QAAQ,CAAC,CAAC,wBAAwB,cAAc;AAAA,MACzH;AAAA,IACF;AAKA,QAAI,YAAY;AAChB,aAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,UAAI,MAAM,CAAC,EAAG,eAAe,MAAM,IAAI,CAAC,EAAG,WAAY,aAAY;AAAA,IACrE;AACA,QAAI,aAAa,GAAG;AAClB,YAAM,eAAe,MAClB,MAAM,SAAS,EACf;AAAA,QACC,CAAC,MACC,OAAO,SAAS,EAAE,QAAQ,gBAAgB,KAC1C,OAAO,SAAS,EAAE,QAAQ,gBAAgB;AAAA,MAC9C;AACF,UAAI,CAAC,cAAc;AACjB,eAAO;AAAA,UACL,UAAU,OAAO,qBAAqB,MAAM,YAAY,CAAC,EAAG,UAAU,aAAQ,MAAM,SAAS,EAAG,UAAU,QAAQ,MAAM,SAAS,EAAG,EAAE;AAAA,QACxI;AAAA,MACF;AAAA,IACF;AAEA,eAAW,UAAU,uBAAuB;AAC1C,YAAM,SAAS,MACZ,OAAO,CAAC,MAAM,OAAO,EAAE,QAAQ,MAAM,MAAM,QAAQ,EACnD,IAAI,CAAC,MAAM,EAAE,QAAQ,MAAM,CAAE;AAChC,UAAI,OAAO,WAAW,EAAG;AAEzB,YAAM,cAAc,cAAc,QAAQ,KAAK,WAAW;AAC1D,UAAI,YAAY,UAAU;AACxB,4BAAoB,KAAK,GAAG,OAAO,IAAI,MAAM,EAAE;AAEjD,YAAM,UAAU,OAAO,OAAO,SAAS,CAAC;AACxC,YAAM,WAAW,OAAO,CAAC;AACzB,YAAM,QAAQ,UAAU;AACxB,YAAM,UAAoB,CAAC;AAC3B,UAAI,YAAY,UAAU,iBAAiB;AACzC,gBAAQ;AAAA,UACN,6CAA6C,YAAY,OAAO,iBAAiB,YAAY,WAAW,QAAQ,CAAC,CAAC;AAAA,QACpH;AAAA,MACF;AACA,UAAI,WAAW,SAAS,UAAU,QAAQ;AACxC,gBAAQ,KAAK,OAAO,QAAQ,QAAQ,CAAC,CAAC,gBAAgB,MAAM,EAAE;AAAA,MAChE;AACA,UAAI,WAAW,sBAAsB,WAAW,UAAU,cAAc;AACtE,gBAAQ;AAAA,UACN,6BAA6B,WAAW,SAAS,QAAQ,CAAC,CAAC,kBAAkB,SAAS,QAAQ,CAAC,CAAC,WAAW,YAAY;AAAA,QACzH;AAAA,MACF;AACA,UAAI,WAAW,sBAAsB,WAAW,UAAU,iBAAiB;AACzE,gBAAQ;AAAA,UACN,6BAA6B,WAAW,SAAS,QAAQ,CAAC,CAAC,kBAAkB,SAAS,QAAQ,CAAC,CAAC,WAAW,eAAe;AAAA,QAC5H;AAAA,MACF;AAEA,YAAM,UAAU,QAAQ,SAAS;AACjC,YAAM,QAAuB;AAAA,QAC3B;AAAA,QACA;AAAA,QACA,OAAO,YAAY;AAAA,QACnB;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,MACF;AACA,UAAI,SAAS;AACX,cAAM,SAAS,QAAQ,KAAK,IAAI;AAChC,eAAO,KAAK,UAAU,OAAO,KAAK,MAAM,KAAK,MAAM,MAAM,EAAE;AAAA,MAC7D;AACA,eAAS,KAAK,KAAK;AAAA,IACrB;AAAA,EACF;AAEA,SAAO,EAAE,UAAU,QAAQ,SAAS,OAAO,WAAW,GAAG,oBAAoB;AAC/E;AAiBO,SAAS,gBAAgB,QAAyC;AACvE,SAAO,EAAE,SAAS,OAAO,SAAS,QAAQ,CAAC,GAAG,OAAO,MAAM,EAAE;AAC/D;","names":[]}

package/dist/{multi-layer-verifier-CI4jdX-q.d.ts → multi-layer-verifier-BsqKuLyN.d.ts}

@@ -147,4 +147,4 @@ declare class MultiLayerVerifier<Env = unknown> {
     run(opts: VerifyOptions<Env>): Promise<VerificationReport>;
 }
 
-export { type Finding as F, type LayerResult as L, MultiLayerVerifier as M, type Severity as S, type VerifyOptions as V, type VerificationReport as a, type Layer as b, type VerifyContext as c, type LayerStatus as d, gradeSemanticStatus as g };
+export { type Finding as F, type Layer as L, MultiLayerVerifier as M, type Severity as S, type VerifyOptions as V, type VerificationReport as a, type LayerResult as b, type VerifyContext as c, type LayerStatus as d, gradeSemanticStatus as g };

package/dist/multishot/index.d.ts

@@ -1,9 +1,9 @@
-import { J as JudgeScore } from '../types-BdIv5dvA.js';
+import { J as JudgeScore } from '../types-v--ctu-b.js';
 import { AgentProfile } from '@tangle-network/agent-interface';
 import { M as MatrixResult } from '../types-BUxNaJ8c.js';
-import '../run-record-I-Z3JNvO.js';
+import '../run-record-DksGsfgv.js';
 import '../errors-oeQrLqXC.js';
-import '../schema-m0gsnbt3.js';
+import '../schema-SGWcK9wa.js';
 import '../verdict-C9MlYujm.js';
 
 interface MultishotMessage {

package/dist/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "@tangle-network/agent-eval — wire protocol",
-    "version": "0.109.1",
+    "version": "0.110.1",
     "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/pipelines/index.d.ts

@@ -1,10 +1,9 @@
-import { h as BudgetSpec, R as Run, T as ToolSpan } from '../schema-m0gsnbt3.js';
-import { T as TraceStore, R as RunFilter } from '../store-BcFXE6LG.js';
-export { a as FailureCluster, F as FailureClusterReport, f as failureClusterView } from '../failure-cluster-DH9Flgcf.js';
-import { a as TrajectoryStep } from '../trajectory-2TkpSEVh.js';
-import { B as BaselineOptions, a as BaselineReport } from '../baseline-Bbid3WoO.js';
-export { c as computeToolUseMetrics } from '../baseline-Bbid3WoO.js';
-import { l as llmSpans } from '../query-0aTmbmQe.js';
+import { h as BudgetSpec, R as Run, T as ToolSpan } from '../schema-SGWcK9wa.js';
+import { T as TraceStore, R as RunFilter } from '../store-BsVi7ncX.js';
+export { a as FailureCluster, F as FailureClusterReport, f as failureClusterView } from '../failure-cluster-C48PiReX.js';
+import { T as TrajectoryStep, B as BaselineOptions, a as BaselineReport } from '../baseline-DsNteOgR.js';
+export { c as computeToolUseMetrics } from '../baseline-DsNteOgR.js';
+import { l as llmSpans } from '../query-Ck190MOd.js';
 
 /**
  * BudgetBreachView — aggregates breach events across the corpus.

package/dist/pipelines/index.js

@@ -1,12 +1,10 @@
 import {
   DEFAULT_RULES,
+  buildTrajectory,
   classifyFailure,
   compareToBaseline,
   computeToolUseMetrics
-} from "../chunk-DBDRR6GF.js";
-import {
-  buildTrajectory
-} from "../chunk-RZTMDUO7.js";
+} from "../chunk-PLOMR3HP.js";
 import {
   interRaterReliability,
   pearsonR
@@ -17,8 +15,7 @@ import {
   llmSpans,
   runFailureClass,
   toolSpans
-} from "../chunk-JZXGWLK5.js";
-import "../chunk-5BKGXME7.js";
+} from "../chunk-MHNQWM4I.js";
 import "../chunk-ONWEPEDO.js";
 import "../chunk-PZ5AY32C.js";