@tangle-network/agent-eval 0.124.0 → 0.126.0

package/dist/index.d.ts

@@ -1153,10 +1153,14 @@ interface CostReceipt extends CostCallBase, CostUsage {
     costUsd: number;
     costUnknown: boolean;
     usageUnknown?: boolean;
+    /** Rates used to estimate cost locally. Absent when cost is provider-reported or unknown. */
     pricing?: {
         inputUsdPerThousand: number;
+        cachedInputUsdPerThousand?: number;
+        cacheWriteUsdPerThousand?: number;
         outputUsdPerThousand: number;
     };
+    /** Cost reported by the provider, not a local token-price calculation. */
     actualCostUsd?: number;
     error?: string;
 }
@@ -1164,20 +1168,27 @@ interface CostReceipt extends CostCallBase, CostUsage {
 type CostLedgerEntry = Omit<CostReceipt, 'status' | 'callId' | 'phase' | 'actor' | 'maximumCostUsd' | 'usageUnknown' | 'pricing' | 'error'>;
 interface CostReceiptInput extends CostUsage {
     model: string;
+    /** Caller-supplied rates for a local estimate when the provider does not report billed cost. */
+    customTokenPricing?: CustomTokenPricing;
     actualCostUsd?: number;
     costUnknown?: boolean;
     usageUnknown?: boolean;
 }
 /** Per-million token rates for a model or endpoint not covered by package pricing. */
 interface CustomTokenPricing {
+    /** Non-cached input tokens. */
     inputUsdPerMillion: number;
+    /** Cache-read tokens. Falls back to the normal input rate when omitted. */
+    cachedInputUsdPerMillion?: number;
+    /** Cache-creation or cache-write tokens. Falls back to the normal input rate when omitted. */
+    cacheWriteUsdPerMillion?: number;
     outputUsdPerMillion: number;
 }
 type MaximumCharge = {
     externallyEnforcedMaximumUsd: number;
 } | ({
     customTokenPricing: CustomTokenPricing;
-} & Pick<CostUsage, 'inputTokens' | 'outputTokens'>) | ({
+} & Pick<CostUsage, 'inputTokens' | 'outputTokens' | 'cachedTokens' | 'cacheWriteTokens'>) | ({
     model: string;
 } & CostUsage);
 interface RunPaidCallInput<T> {
@@ -1245,6 +1256,8 @@ interface CostLedgerFilter {
 interface CostLedgerWaitOptions {
     /** Maximum time to wait for active provider calls. Default 5 seconds. */
     timeoutMs?: number;
+    /** Wait only for calls matching this attribution filter. */
+    filter?: CostLedgerFilter;
 }
 /** Append-only storage. `append` must atomically reject stale revisions. */
 interface CostLedgerPersistence {
@@ -1339,8 +1352,8 @@ interface CostResult {
     costUnknown: boolean;
 }
 declare function costForUsage(model: string, usage: CostUsage): CostResult;
-/** Price input and output token counts with caller-supplied per-million rates. */
-declare function costForTokenPricing(pricing: CustomTokenPricing, usage: Pick<CostUsage, 'inputTokens' | 'outputTokens'>): number;
+/** Price token counts with caller-supplied per-million rates. */
+declare function costForTokenPricing(pricing: CustomTokenPricing, usage: Pick<CostUsage, 'inputTokens' | 'outputTokens' | 'cachedTokens' | 'cacheWriteTokens'>): number;
 
 /**
  * RawProviderSink — first-class persistence for the actual HTTP-level
@@ -3680,112 +3693,6 @@ declare class SkillUsageAnalyst implements Analyst<SkillUsageReport> {
 }
 declare const SKILL_USAGE_ANALYST: SkillUsageAnalyst;
 
-type PolicyEditSchemaVersion = 'policy-edit/v1';
-declare const POLICY_EDIT_AXES: readonly ["carrier", "representation", "budget", "sampling", "output_contract", "tool_contract", "routing", "memory", "agent_profile", "deployment_target"];
-type PolicyEditAxis = (typeof POLICY_EDIT_AXES)[number];
-declare const POLICY_EDIT_TARGET_SURFACES: readonly ["prompt", "tool-contract", "runtime-config", "memory", "agent-profile", "code", "deployment"];
-type PolicyEditTargetSurface = (typeof POLICY_EDIT_TARGET_SURFACES)[number];
-type PolicyEditRisk = 'low' | 'medium' | 'high' | 'unknown';
-type PolicyEditGainDirection = 'increase' | 'decrease';
-type PolicyEditGainUnit = 'absolute' | 'relative' | 'percent' | 'score';
-interface PolicyEditTarget {
-    surface: PolicyEditTargetSurface;
-    /** Stable path inside the target surface, for example `system-prompt:tools`
-     * or `budget.maxTurns`. */
-    path?: string;
-    /** Optional canonical deployment identity. Store the existing cell, not a
-     * local profile shape. */
-    agentProfileCell?: AgentProfileCell;
-    /** Human label when the path is not enough for a readable audit trail. */
-    label?: string;
-}
-type PolicyEditChange = {
-    kind: 'text';
-    mode: 'append' | 'prepend' | 'replace';
-    value: string;
-    /** Required when `mode === 'replace'`; exact match only. */
-    find?: string;
-} | {
-    kind: 'json';
-    mode: 'set' | 'merge' | 'remove';
-    path: string;
-    value?: AgentProfileJson;
-};
-interface PolicyEditExpectedGain {
-    /** Metric this edit is expected to move, e.g. `holdout.composite`. */
-    metric: string;
-    direction: PolicyEditGainDirection;
-    /** Positive magnitude in the metric's native units. */
-    amount: number;
-    unit?: PolicyEditGainUnit;
-    rationale?: string;
-}
-interface PolicyEditSource {
-    findingIds: string[];
-    analystIds: string[];
-    evidenceRefs: EvidenceRef[];
-    /** Mirrors `AnalystFinding.derived_from_judge`; admission rejects it. */
-    derivedFromJudge?: boolean;
-}
-interface PolicyEdit {
-    schemaVersion: PolicyEditSchemaVersion;
-    editId: string;
-    axis: PolicyEditAxis;
-    target: PolicyEditTarget;
-    change: PolicyEditChange;
-    claim: string;
-    expectedGain: PolicyEditExpectedGain;
-    confidence: number;
-    risk: PolicyEditRisk;
-    source: PolicyEditSource;
-    rationale?: string;
-    validationPlan?: string;
-    metadata?: Record<string, unknown>;
-}
-declare const POLICY_EDIT_CANDIDATE_RECORD_SCHEMA: "tangle.policy-edit-candidate.v1";
-/** JSON-safe attribution carried with a measured candidate and its scores. */
-interface PolicyEditCandidateRecord {
-    schema: typeof POLICY_EDIT_CANDIDATE_RECORD_SCHEMA;
-    policyEdit: PolicyEdit;
-}
-type PolicyEditInit = Omit<PolicyEdit, 'schemaVersion' | 'editId'> & {
-    schemaVersion?: PolicyEditSchemaVersion;
-    editId?: string;
-};
-declare class PolicyEditValidationError extends ValidationError {
-    readonly path: string;
-    constructor(message: string, path?: string);
-}
-interface FindingToPolicyEditOptions {
-    expectedGain?: PolicyEditExpectedGain | ((finding: AnalystFinding) => PolicyEditExpectedGain | null | undefined);
-    risk?: PolicyEditRisk | ((finding: AnalystFinding) => PolicyEditRisk);
-    defaultAxis?: PolicyEditAxis;
-    defaultTargetSurface?: PolicyEditTargetSurface;
-}
-interface PolicyEditAdmissionOptions {
-    minScore?: number;
-    minExpectedGain?: number;
-    allowHighRisk?: boolean;
-    requireEvidence?: boolean;
-}
-interface PolicyEditAdmission {
-    edit: PolicyEdit;
-    decision: 'admit' | 'reject';
-    score: number;
-    reasons: string[];
-}
-declare function makePolicyEdit(init: PolicyEditInit): PolicyEdit;
-declare function computePolicyEditId(edit: Omit<PolicyEdit, 'editId'> | PolicyEdit): string;
-declare function validatePolicyEdit(input: unknown): PolicyEdit;
-declare function makePolicyEditCandidateRecord(edit: PolicyEdit): PolicyEditCandidateRecord;
-declare function validatePolicyEditCandidateRecord(input: unknown): PolicyEditCandidateRecord;
-declare function isPolicyEdit(input: unknown): input is PolicyEdit;
-declare function policyEditsFromFindings(findings: ReadonlyArray<AnalystFinding>, opts?: FindingToPolicyEditOptions): PolicyEdit[];
-declare function policyEditFromFinding(finding: AnalystFinding, opts?: FindingToPolicyEditOptions): PolicyEdit | null;
-declare function scorePolicyEditReadiness(edit: PolicyEdit, opts?: PolicyEditAdmissionOptions): number;
-declare function admitPolicyEdit(edit: PolicyEdit, opts?: PolicyEditAdmissionOptions): PolicyEditAdmission;
-declare function applyPolicyEditToSurface(surface: unknown, edit: PolicyEdit): unknown;
-
 /**
  * Automated pull-request transports for the production loop.
  *
@@ -6521,6 +6428,34 @@ interface WorkerLogSource {
     readonly inbox: string | null;
     /** Worker patch byte length, or null when absent. */
     readonly patchBytes: number | null;
+    /** Where this worker's transcript lives, for the rollout row. Null = no such artifact. */
+    readonly transcriptRef?: string | null;
+    /** Where this worker's delivered patch lives. Null = the store keeps no patch per worker. */
+    readonly patchPath?: string | null;
+    /** This worker's own inference tokens, when the store records them per worker. */
+    readonly tokensIn?: number | null;
+    readonly tokensOut?: number | null;
+    readonly cacheRead?: number | null;
+    readonly cacheWrite?: number | null;
+}
+/**
+ * Facts a SOURCE structurally cannot express, each with the reason.
+ *
+ * The difference between "the artifact is missing" and "this store never
+ * records that fact" is the difference between a run that spent $0 and a
+ * harness that does not price inference — and the second harness is where a
+ * loops-shaped assumption becomes a fabricated zero. A reader declares its
+ * limits once; the analyzer reports `unavailable` for everything downstream.
+ *
+ * `null` on a field means the source DOES carry that fact.
+ */
+interface SourceLimits {
+    /** Reason inference spend has no price in this store (null = the store prices it). */
+    readonly spendUsd: string | null;
+    /** Reason workers carry no pass/fail verdict (null = verdicts are recorded). */
+    readonly workerVerdicts: string | null;
+    /** Reason no delivered artifact (patch/diff) is retained per worker (null = retained). */
+    readonly deliverables: string | null;
 }
 /**
  * Everything the pure analyzer reads — already-read bytes, never paths. Each
@@ -6575,8 +6510,24 @@ interface SupervisorRunSources {
         sessions: number;
         input: number;
         output: number;
+        /** Cached prompt tokens, when the store counts them separately. */
+        cacheRead?: number;
+        cacheWrite?: number;
     } | null;
     readonly harnessMissingReason: string | null;
+    /** What this store structurally cannot record. See `SourceLimits`. */
+    readonly limits: SourceLimits;
+    /**
+     * Where the ROOT invocation's transcript lives. Undefined lets the rollout
+     * minter fall back to the loops layout (`<supRunDir>/journal.jsonl`); any
+     * other store must say, or the row points at a path that never existed.
+     */
+    readonly rootTranscriptRef?: string | null;
+    /**
+     * The `traces` CLI command that covers this run's harness-session layer.
+     * Null falls back to the analyzer's default (an opencode worker fleet).
+     */
+    readonly traceCommand: string | null;
 }
 /**
  * A source of supervisor-run bytes. Implementations own their storage layout;
@@ -6648,15 +6599,23 @@ interface DecisionMetrics {
 interface RoleSpend {
     readonly tokensIn: Measured<number>;
     readonly tokensOut: Measured<number>;
+    /**
+     * Cached prompt tokens read/written. On a harness that caches aggressively
+     * these dwarf `tokensIn`, so a report that omits them understates the context
+     * each invocation actually consumed. `unavailable` = the store has no such counter.
+     */
+    readonly cacheRead: Measured<number>;
+    readonly cacheWrite: Measured<number>;
     readonly usd: Measured<number>;
     readonly source: string;
 }
 interface PerWorkerRow {
     readonly worker: string;
     readonly wallMs: number | null;
-    readonly tokensIn: number;
-    readonly tokensOut: number;
-    readonly usd: number;
+    /** `null` = this store does not attribute tokens per worker (NOT "zero tokens"). */
+    readonly tokensIn: number | null;
+    readonly tokensOut: number | null;
+    readonly usd: number | null;
     readonly patchBytes: number | null;
     readonly passed: boolean | null;
 }
@@ -6787,6 +6746,88 @@ declare function analyzeSupervisorRunSources(src: SupervisorRunSources, now?: ()
  */
 declare function rollupSupervisorRuns(reports: readonly SupervisorRunReport[]): SupervisorRunRollup;
 
+/**
+ * Supervision-tree reader over a THIRD-PARTY harness: Claude Code.
+ *
+ * `loops-reader.ts` reads a supervisor we wrote, whose journal was designed
+ * for this analysis. This reader reads a harness we do not control, whose
+ * transcript was designed for replaying a chat — and recovers the same tree
+ * from it. If both produce a `SupervisorRunSources`, the tree model is a
+ * property of multi-agent runs, not of our journal format.
+ *
+ * ## Where the tree hides in a Claude Code transcript
+ *
+ * | Tree fact | Claude Code evidence |
+ * |---|---|
+ * | spawn      | assistant `tool_use` (`Agent` / `Task`), answered by a `tool_result` whose `toolUseResult.agentId` names the child |
+ * | settle     | a `<task-notification>` block in a later user line: `<task-id>` = agentId, `<status>` |
+ * | steer      | assistant `tool_use` (`SendMessage`) with `input.to` = agentId — mid-task, to a LIVE child |
+ * | delivered  | that steer's `tool_result` carrying `success` / `resumedAgentId` |
+ * | cancel     | assistant `tool_use` (`TaskStop`) targeting an agentId |
+ * | brain spend| `message.usage` on the main thread's assistant lines |
+ * | worker spend| `message.usage` inside `<session>/subagents/agent-<id>.jsonl` |
+ * | depth      | a child transcript that itself contains `Agent` tool_use lines |
+ *
+ * Every one of those is read through `parseClaudeEntries` — the SAME line
+ * parser `src/rollout/readers/claude-jsonl.ts` uses for solo rollouts. There
+ * is no second transcript parser.
+ *
+ * ## What Claude Code cannot say
+ *
+ * It records tokens but never a price, runs no per-worker verify, and keeps no
+ * per-worker patch. Those are declared once in `limits`, so the analyzer
+ * reports `unavailable — <reason>` instead of the $0 / 0-accepted that summing
+ * an empty field would produce. See `SourceLimits`.
+ *
+ * ## Metric coverage vs the loops journal
+ *
+ * Measured on a real 52-agent session (fixture:
+ * `tests/fixtures/supervisor-run/claude-code-session-*`).
+ *
+ * | Metric | loops | Claude Code | Why |
+ * |---|---|---|---|
+ * | workersSpawned / Settled / Cancelled | full | full | spawn tool_use + task-notification + TaskStop |
+ * | steers / steersDelivered / steersByWorker | full | full | `SendMessage`; delivery from its tool_result |
+ * | waves / waveSizes / maxConcurrency | full | full | derived from spawn/settle instants |
+ * | respawns / repeatedLabels | full | full | same derivation |
+ * | delegationDepth | full | full | a child transcript's own spawn calls |
+ * | timeToFirstSpawn / supervisorWall | full | full | transcript instants |
+ * | idleMs / idlePct / workerUtilization | full | PARTIAL | an agent that never notifies is counted live to the end of the transcript |
+ * | observeThenRespawn / respawnWithoutEvidence | full | full | ordering of spawn vs settle instants |
+ * | workerEvidenceBytes | full | PARTIAL | the child's closing message; 0 for pruned transcripts |
+ * | brain tokens in/out + cache | full | full | main-thread `message.usage` |
+ * | worker tokens in/out + cache | via harness join | PARTIAL | only for retained subagent transcripts |
+ * | perWorker wall | full | full | spawn → settle instants |
+ * | accepted / rejected / emptyPass / settledVerdicts | full | NONE | no per-worker verify step exists |
+ * | brain/worker/total usd, costPerAcceptedPatch | full | NONE | transcripts carry no price |
+ * | patch stats, delivered, verifyPass/Rc | full | NONE | no diff is handed back |
+ * | judgeResolved / Score / Passed / Total | full | NONE | no judge in the loop |
+ * | driverSteerCalls, brainTruncations | full | NONE | no outer driver log, no per-call finish_reason tap |
+ */
+
+interface ClaudeCodeReaderOptions {
+    /** The main session transcript: `~/.claude/projects/<slug>/<sessionId>.jsonl`. */
+    readonly transcriptPath: string;
+    /**
+     * Directory of child transcripts. Defaults to `<transcript-dir>/<sessionId>/subagents`.
+     * `null` skips the join, and every per-worker token count becomes unavailable.
+     */
+    readonly subagentsDir?: string | null;
+    readonly runRef?: string;
+    readonly instanceId?: string | null;
+    readonly arm?: string | null;
+    readonly spawnTools?: readonly string[];
+    readonly steerTools?: readonly string[];
+    readonly cancelTools?: readonly string[];
+}
+/**
+ * Read a Claude Code session (plus its subagent transcripts) as supervision-tree
+ * source bytes. Never throws on a missing artifact.
+ */
+declare function readClaudeCodeSupervisorRun(opts: ClaudeCodeReaderOptions): Promise<SupervisorRunSources>;
+/** A `SupervisorRunReader` over a Claude Code session — the same contract loops implements. */
+declare function claudeCodeSupervisorRunReader(opts: ClaudeCodeReaderOptions): SupervisorRunReader;
+
 /**
  * ONE implementation of `SupervisorRunReader`: the on-disk layout the loops
  * supervisor writes — `<runDir>/ws/.loops/supervisor/<id>/{journal.jsonl,
@@ -8382,98 +8423,8 @@ interface JudgeScore {
     /** Ensemble extras: each surviving judge's per-dimension scores. */
     perJudge?: Record<string, Record<string, number>>;
 }
-/** A tier-4 code surface — a finalized candidate change to the agent's
- *  IMPLEMENTATION, not its prompt. Produced by autoresearch (reads codebase +
- *  trace findings → opens a worktree). `worktreeRef` locates the candidate;
- *  the exact commits, tree, and binary-patch digest identify it. See the
- *  improvement-tier table in `docs/design/loop-taxonomy.md`. */
-interface CodeSurface {
-    readonly kind: 'code';
-    /** Worktree path or git ref holding the candidate code change. This is a
-     *  mutable locator and is deliberately excluded from content hashes. */
-    readonly worktreeRef: string;
-    /** Human-readable ref the worktree was forked from. Not identity-bearing. */
-    readonly baseRef: string;
-    /** Exact commit the candidate was forked from. */
-    readonly baseCommit: string;
-    /** Exact tree object for `baseCommit`. */
-    readonly baseTree: string;
-    /** Exact finalized candidate commit. */
-    readonly candidateCommit: string;
-    /** Exact tree object for `candidateCommit`. */
-    readonly candidateTree: string;
-    /** Identity of the exact patch artifact. The deployable candidate bundle
-     *  carries the same descriptor plus its base64-encoded content. */
-    readonly patch: {
-        readonly format: 'git-diff-binary';
-        readonly sha256: `sha256:${string}`;
-        readonly byteLength: number;
-    };
-    /** Human summary of what changed — rendered into the auto-PR body. */
-    readonly summary?: string;
-}
-/** The mutable surface a proposer changes. Tiers (see
- *  `docs/design/loop-taxonomy.md`):
- *   - `string`      — tiers 1-2: system-prompt addendum / serialized tool
- *                     config. Cheap, reversible, text-diffable.
- *   - `CodeSurface` — tier 4: an implementation change behind a worktree ref.
- *  Tier 3 (knowledge) is owned by agent-knowledge and rides its own adapter,
- *  not this type. */
-type MutableSurface = string | CodeSurface;
-/** A non-dominated parent on the GEPA Pareto frontier — a
- *  surface that, across the per-scenario objective vectors, no other tried
- *  surface beats on every scenario. A candidate worse on the mean composite
- *  but uniquely best on one hard scenario is non-dominated and survives here;
- *  the composite-best ranking would discard the lesson it carries. The loop
- *  computes the frontier across ALL generations and hands it to the proposer so
- *  a reflective proposer can combine complementary lessons (GEPA, Agrawal et
- *  al., arXiv:2507.19457). See `pareto.ts` (`paretoFrontier`). */
-interface ParetoParent {
-    surface: MutableSurface;
-    surfaceHash: string;
-    /** The objective vector: per-scenario composite (higher is better). The
-     *  axes the frontier is computed over. */
-    objectives: Record<string, number>;
-    /** Mean composite across the objective scenarios — the scalar summary used
-     *  for ordering + display, NOT for dominance. */
-    composite: number;
-    /** Generation that produced this surface (`-1` for the baseline). */
-    generation: number;
-    label?: string;
-    rationale?: string;
-}
 /** Five-valued verdict taxonomy (MOSS-paper alignment). */
 type GateDecision = 'ship' | 'hold' | 'need_more_work' | 'model_ceiling' | 'arch_ceiling';
-interface GateContext<TArtifact, TScenario extends Scenario> {
-    candidateArtifacts: Map<string, TArtifact>;
-    baselineArtifacts?: Map<string, TArtifact>;
-    /** Candidate (winner) judge scores, keyed by cellId. */
-    judgeScores: Map<string, Record<string, JudgeScore>>;
-    /** Baseline judge scores, keyed by cellId. SEPARATE from `judgeScores` —
-     *  baseline + candidate share cellIds (same scenarios), so a single map
-     *  cannot represent both. A gate computing a holdout delta MUST read
-     *  candidate from `judgeScores` and baseline from here. */
-    baselineJudgeScores?: Map<string, Record<string, JudgeScore>>;
-    /** Neutralized-arm judge scores, keyed by cellId — the winner surface with its
-     *  content footprint-matched-blanked (via a `neutralize` fn). Same scenarios as
-     *  `judgeScores`. Present ONLY when `runImprovementLoop` was given a `neutralize`
-     *  function. A placebo gate (`neutralizationGate`) compares this arm's lift
-     *  against the candidate's to reject decorative wins (lift from footprint, not
-     *  content). Undefined otherwise. */
-    neutralizedJudgeScores?: Map<string, Record<string, JudgeScore>>;
-    /** Neutralized-arm artifacts, keyed by cellId. Present alongside
-     *  `neutralizedJudgeScores`. */
-    neutralizedArtifacts?: Map<string, TArtifact>;
-    scenarios: TScenario[];
-    cost: {
-        candidate: number;
-        baseline: number;
-    };
-    /** Shared run spend account and receipt attribution phase. */
-    costLedger?: CostLedgerHandle;
-    costPhase?: string;
-    signal: AbortSignal;
-}
 interface GateResult {
     decision: GateDecision;
     reasons: string[];
@@ -8484,11 +8435,6 @@ interface GateResult {
     }>;
     delta?: number;
 }
-/** Composable promotion gate. */
-interface Gate<TArtifact = unknown, TScenario extends Scenario = Scenario> {
-    name: string;
-    decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult>;
-}
 /** Scoped trace writer handed to each dispatch — every span
  *  auto-tagged with the cellId so traces filter cleanly. */
 interface CampaignTraceWriter {
@@ -8619,8 +8565,6 @@ interface GenerationCandidate {
      *  "because rationale Z" the audit requires to survive to the result.
      *  Present when the proposer returned a `ProposedCandidate`. */
     rationale?: string;
-    /** Exact structured cause threaded from the proposer, when available. */
-    candidateRecord?: PolicyEditCandidateRecord;
 }
 interface CampaignAggregates {
     byJudge: Record<string, JudgeAggregate>;
@@ -10154,9 +10098,8 @@ declare function securityJudge(id: string, config: HarnessConfig): SandboxJudgeS
  * The `JudgeConfig` contract (src/campaign/types.ts) is deliberately a
  * function, not a fixed LLM-prompt shape: real consumers judge with
  * ensembles, deterministic checks, or one LLM call. `ensembleJudge`
- * (src/judge-panel.ts) covers the multi-model case; `buildAgreementJudge`
- * (src/campaign/distillation) covers the pure-comparator case. `llmJudge`
- * covers the common single-call case the `JudgeConfig` doc-comment names:
+ * (src/judge-panel.ts) covers the multi-model case. `llmJudge` covers the
+ * common single-call case the `JudgeConfig` doc-comment names:
  * one model call against `prompt`, parsed into the canonical `JudgeScore`
  * (`{ dimensions, composite, notes }`) on the campaign [0,1] scale.
  *
@@ -14882,195 +14825,6 @@ interface CampaignStorage {
     append?(path: string, content: string, expectedBytes: number): number | undefined;
 }
 
-/**
- * `openAutoPr` — thin shell-out helper for the `runImprovementLoop` preset's
- * `autoOnPromote: 'pr'` mode. Substitutes for the per-product PR-opening
- * code consumers duplicated 4 times. The PR body includes the campaign's
- * manifest hash, gate verdict, and scorecard summary so reviewers can see
- * exactly what was promoted + why.
- *
- * NOT a deploy mechanism — this only OPENS a PR. The human reviews + merges.
- * The Shape B (`autoOnPromote: 'config'`) live-runtime-mutation path is
- * deferred to Pass B with the full shadow / canary / rollback stack.
- */
-
-interface OpenAutoPrOptions<TArtifact, TScenario extends Scenario> {
-    /** Campaign result to attach to the PR. */
-    result: CampaignResult<TArtifact, TScenario>;
-    /** Gate verdict explaining the promotion. Substrate refuses to open a PR
-     *  when `gate.decision !== 'ship'` — fails loud. */
-    gate: GateResult;
-    /** Promoted surface diff — typically the new system prompt addendum or
-     *  full profile diff. Substrate writes it as the PR body. */
-    promotedDiff: string;
-    /** GH owner/repo target (e.g., `tangle-network/gtm-agent`). */
-    ghOwner: string;
-    ghRepo: string;
-    /** Branch name for the PR. Default `auto/<manifestHash[:12]>`. */
-    branch?: string;
-    /** PR title. Default includes manifest hash. */
-    title?: string;
-    /** Whether to actually open the PR or just dry-run. Default reads
-     *  `GH_AUTO_PR_TOKEN` env — present = open, absent = dry-run. */
-    dryRun?: boolean;
-    /** Test seam — substitute `gh pr create` invocation. */
-    ghExec?: (args: string[]) => {
-        stdout: string;
-        stderr: string;
-        status: number;
-    };
-}
-interface OpenAutoPrResult {
-    opened: boolean;
-    prUrl?: string;
-    dryRun: boolean;
-    reason: string;
-}
-/**
- * Open a GitHub PR for a gate-approved surface promotion, attaching the manifest hash, gate verdict, and diff as the PR body.
- */
-declare function openAutoPr<TArtifact, TScenario extends Scenario>(options: OpenAutoPrOptions<TArtifact, TScenario>): OpenAutoPrResult;
-
-/**
- * `runOptimization` — the improvement loop body. Runs N generations: the
- * `SurfaceProposer` proposes K candidate surfaces per generation, each
- * candidate runs a campaign (the measurement), and only a candidate that beats
- * the single global incumbent becomes the next generation's parent.
- * Proposer-agnostic — the same loop runs an evolutionary population mutator
- * (`evolutionaryProposer`) or any reflective / agentic proposer; they differ
- * only in how `propose()` picks candidates.
- *
- * This is `runLoop`'s shape (plan → measure → decide) specialized to surface
- * improvement: `proposer.propose` = plan, `runCampaign` = the measurement
- * (which runs the worker behind `dispatch`), the mean-composite ranking = the
- * validator, `proposer.decide` = the stop check.
- *
- * The gated-promotion shell (`runImprovementLoop`) wraps this with a holdout
- * re-score + release gate + optional PR.
- */
-
-interface RunOptimizationResult<TArtifact, TScenario extends Scenario> {
-    generations: Array<{
-        record: GenerationRecord;
-        surfaces: Array<{
-            surfaceHash: string;
-            surface: MutableSurface;
-            campaign: CampaignResult<TArtifact, TScenario>;
-        }>;
-    }>;
-    winnerSurface: MutableSurface;
-    winnerSurfaceHash: string;
-    /** Proposer label for the promoted surface. Present when the winning
-     *  candidate came from a `ProposedCandidate` (a reflective proposer);
-     *  absent when the winner is the baseline or a bare-surface mutator. */
-    winnerLabel?: string;
-    /** Proposer rationale for the promoted surface — the "because Z" that
-     *  motivated the winning change. Survives to `SelfImproveResult` and the
-     *  emitted provenance record. Absent when the winner is the baseline. */
-    winnerRationale?: string;
-    baselineCampaign: CampaignResult<TArtifact, TScenario>;
-    /** Run-wide spend, including agents, proposers, analysts, and judges. */
-    cost: CostLedgerSummary;
-    /** The GEPA Pareto frontier across every scored surface (baseline + all
-     *  generations) by per-scenario objective vector — the non-dominated set.
-     *  Each generation's `propose()` received the frontier-so-far as
-     *  `ctx.paretoParents`; this is the final frontier. A surface here that is
-     *  NOT the winner is uniquely best on some scenario the winner loses on. */
-    paretoFrontier: ParetoParent[];
-}
-
-/**
- * `runImprovementLoop` — the gated-promotion shell around the improvement
- * loop body (`runOptimization`). Proposes candidate surfaces via the
- * `SurfaceProposer`, re-scores the winner against the baseline on a
- * holdout set, runs the release gate, and optionally opens a PR.
- *
- * Role vocabulary (see docs/design/loop-taxonomy.md):
- *   - PROPOSER   = the `SurfaceProposer` (evolutionary GEPA mutator OR
- *                  reflective analyst). Proposes candidate SURFACES — the
- *                  worker's system prompt / tool config — NOT conversation
- *                  turns.
- *   - MEASUREMENT= `runCampaign`. Scores one surface by running the worker
- *                  (via `dispatch`) over scenarios and judging the output.
- *   - WORKER     = the agent harness in the sandbox, invoked behind the
- *                  topology-opaque `dispatch` seam — never referenced here.
- *
- * Distinct from `runLoop` in `@tangle-network/agent-runtime`, which is the
- * INNER conversation loop (execution driver ↔ workers in a sandbox). `runImprovementLoop`
- * is the OUTER loop: it improves the surface that those workers run.
- *
- * Hard-refuses unsafe configurations:
- *   - `tracing: 'off'` when a proposer is wired (improvement is unattributable)
- *   - `autoOnPromote: 'config'` — live mutation is unsupported without
- *     isolated deployment, rollback, and independent validation.
- */
-
-interface RunImprovementLoopResult<TArtifact, TScenario extends Scenario> extends RunOptimizationResult<TArtifact, TScenario> {
-    baselineOnHoldout: CampaignResult<TArtifact, TScenario>;
-    winnerOnHoldout: CampaignResult<TArtifact, TScenario>;
-    neutralizedOnHoldout?: CampaignResult<TArtifact, TScenario>;
-    neutralizedSurface?: MutableSurface;
-    gateResult: Awaited<ReturnType<Gate<TArtifact, TScenario>['decide']>>;
-    /** Present iff the loop ran with `holdout: 'deferred'`. When set,
-     *  `baselineOnHoldout`/`winnerOnHoldout` are the shared EMPTY campaign (zero
-     *  cells dispatched) and the gate verdict is the forced `'hold'`. */
-    holdout?: 'deferred';
-    /** Unified baseline→winner surface diff. Computed UNCONDITIONALLY (not only
-     *  when `autoOnPromote === 'pr'`) so the diff that the gate decided on is
-     *  always present on the result + in the emitted provenance record. Empty
-     *  string when winner == baseline (no change to diff). */
-    promotedDiff: string;
-    prResult?: ReturnType<typeof openAutoPr>;
-}
-
-/**
- * `gepaProposer` — a reflective `SurfaceProposer` for prompt-tier surfaces.
- * Each generation it reflects on the prior best candidate's per-scenario
- * scores + weakest dimensions, asks an LLM to propose targeted rewrites of
- * the current surface, and returns them as the next population.
- *
- * Maps onto the GEPA paper (Agrawal et al., arXiv:2507.19457):
- *   - *Reflection*: each generation reflects on the best parent's weakest
- *     dimensions + per-scenario top/bottom scores to propose targeted rewrites.
- *   - *Pareto frontier*: `runOptimization` maintains the non-dominated set of
- *     surfaces across generations (per-scenario objective vectors) and supplies
- *     it as `ctx.paretoParents`. A surface uniquely best on one hard scenario
- *     survives even when its mean composite is lower.
- *   - *Combine complementary lessons*: when the frontier has >1 member, the
- *     first population slot is a merge of those parents' strengths (one LLM
- *     call citing each parent's winning scenarios). Toggle via `combineParents`.
- * Dominance is computed by the package-canonical `paretoFrontier` (`pareto.ts`).
- *
- * Optional `constraints` move structured-doc guards into the proposer
- * (preserve H2 section headings, cap sentence-level edits) — useful when
- * the surface IS a structured procedure like a SKILL.md / runbook /
- * judge rubric. When `constraints` is omitted, behavior is unchanged.
- *
- * The proposer is surface-agnostic — any string surface in any consumer opts
- * in by selecting it. Reuses the generic reflection primitive
- * (`buildReflectionPrompt` / `parseReflectionResponse`) and the router client.
- *
- * Earns its keep where there is real per-instance signal (which the
- * dimensional + per-scenario evidence + the `LabeledScenarioStore` flywheel
- * now provide). For thin-signal surfaces it degrades to plain reflection.
- * On generation 0 (no history) it reflects on the current surface against
- * the mutation primitives alone.
- */
-
-interface GepaProposerConstraints {
-    /** H2 section headings that MUST appear unchanged in every candidate.
-     *  When set, the proposer auto-detects current H2s if this is empty AND
-     *  rejects any candidate that drops or renames a preserved heading.
-     *  Use when the surface is a structured doc (SKILL.md, runbook,
-     *  sectioned system prompt, judge rubric). */
-    preserveSections?: string[];
-    /** Maximum sentence-level edits per candidate vs the parent surface.
-     *  Rejection threshold = maxSentenceEdits × 2 (counts adds + removes).
-     *  Inspired by SkillOpt's edit-budget as a "textual learning rate."
-     *  Cap prevents an LLM rewrite from overwriting useful prior rules. */
-    maxSentenceEdits?: number;
-}
-
 interface BenchmarkRunOptions<TPayload = unknown, TArtifact = string> {
     adapter: BenchmarkAdapter<BenchmarkDatasetItem<TPayload>, TPayload, TArtifact>;
     respond: BenchmarkResponder<TPayload, TArtifact>;
@@ -16607,290 +16361,6 @@ declare function withOtelPipeline(opts?: OtelPipelineOptions): OtelPipelineHandl
  */
 declare function isOtelConfigured(): boolean;
 
-/**
- * Traced analyst wrapper — instruments `analyzeTraces` with spans so the
- * analyst's internal model turns appear in the trace tree. Also wraps each
- * actor turn callback with a span.
- *
- * The wrapper records the Ax turn loop at its public boundaries:
- *   1. A parent span for the entire analyst run.
- *   2. Per-turn child spans from the `onTurn` callback (captures code,
- *      output size, error status).
- *   3. Summary attributes on the parent (total turns, usage, findings).
- */
-
-interface TracedAnalystOptions {
-    /** TraceEmitter for span emission. */
-    emitter: TraceEmitter;
-    /** Parent span id. If omitted, uses emitter stack. */
-    parentSpanId?: string;
-}
-/**
- * Run `analyzeTraces` wrapped in a parent span with per-turn child spans.
- */
-declare function tracedAnalyzeTraces(input: AnalyzeTracesInput, options: AnalyzeTracesOptions, traceOpts: TracedAnalystOptions): Promise<AnalyzeTracesResult>;
-
-/**
- * Traced judge wrappers — instruments every LLM call inside the judge
- * ensemble with child spans so OTEL sinks see per-judge latency, model,
- * token counts, and score dimensions.
- *
- * The ensemble parent span groups all individual judge spans; each judge
- * gets its own child span with model + score as attributes.
- */
-
-interface TracedJudgeOptions {
-    /** TraceEmitter to emit spans into. */
-    emitter: TraceEmitter;
-    /** Parent span id for the ensemble. If omitted, uses the emitter stack. */
-    parentSpanId?: string;
-}
-/**
- * Wrap a single JudgeFn so its LLM call emits a traced span.
- */
-declare function traceJudge(judge: JudgeFn, judgeName: string, opts: TracedJudgeOptions): JudgeFn;
-/**
- * Wrap an array of JudgeFns with tracing, running them inside an ensemble
- * parent span. Returns a single function that calls all judges and merges
- * their scores.
- */
-declare function traceJudgeEnsemble(judges: JudgeFn[], judgeNames: string[], opts: TracedJudgeOptions): JudgeFn;
-
-/**
- * Gold scenarios for teacher→student distillation. The TEACHER is an
- * expensive workflow (e.g. the 70-agent skill audit) whose verdicts are
- * frozen as gold labels; the STUDENT is a cheap single-shot analyst whose
- * prompt GEPA optimizes toward reproducing those labels.
- *
- * A `GoldScenario` is a `Scenario` (the substrate's input contract) carrying
- * an OPAQUE `input` (what the student sees) and an OPAQUE `label` (the gold
- * verdict the student's output is scored against). Both are typed `unknown`
- * here: this module is domain-agnostic — it distills ANY analyst against ANY
- * gold JSONL. The agreement comparator (see `agreement-judge.ts`) is what
- * knows the label's shape.
- *
- * Loading + splitting are DETERMINISTIC and LLM-free: the gold set is the
- * fixed ground truth, never regenerated here.
- */
-
-/** A held gold record: opaque student-input + opaque gold-label, carried as a
- *  substrate `Scenario` so it flows through `runCampaign` unchanged. */
-interface GoldScenario<TInput = unknown, TLabel = unknown> extends Scenario {
-    kind: 'gold';
-    /** What the student analyst is shown (rendered into its user prompt). */
-    input: TInput;
-    /** The teacher's gold verdict — the target the student's output is scored
-     *  against by the agreement judge. NEVER shown to the student. */
-    label: TLabel;
-}
-/** Read a gold JSONL (one `{scenarioId|id, input, label, split?}` per line) into
- *  `GoldScenario[]`. Deterministic, no LLM. Blank lines are skipped; a line
- *  missing an id, `input`, or `label` throws (a silent skip would corrupt the
- *  split silently — fail loud on a malformed gold set). */
-declare function loadGoldScenarios<TInput = unknown, TLabel = unknown>(jsonlPath: string): GoldScenario<TInput, TLabel>[];
-/** Parse gold JSONL text directly (no fs). Exported so tests + in-memory
- *  callers exercise the same parse path as {@link loadGoldScenarios}. */
-declare function parseGoldJsonl<TInput = unknown, TLabel = unknown>(text: string, sourceLabel?: string): GoldScenario<TInput, TLabel>[];
-interface SplitGoldOptions {
-    /** Every Nth scenario (0-based index) goes to the TEST/holdout split; the
-     *  rest train. Default 4 ⇒ a 25% holdout. Ignored for any scenario that
-     *  carries an explicit `split:` tag (that is honored verbatim). */
-    testEveryNth?: number;
-}
-interface GoldSplit<TInput, TLabel> {
-    /** Training scenarios — the optimization pool the proposer searches over. */
-    train: GoldScenario<TInput, TLabel>[];
-    /** Held-out scenarios — kept OUT of training; scored only at the gate. */
-    test: GoldScenario<TInput, TLabel>[];
-}
-/** Deterministic train/test split. A scenario tagged `split:train|test` is
- *  routed by that tag; the rest fall to a modulo split (`index % testEveryNth
- *  === 0 ⇒ test`). Pure — same input always yields the same split, so a gold
- *  set's holdout is stable across runs (a shuffled split would let a lucky
- *  seed flatter the gate). */
-declare function splitGold<TInput, TLabel>(scenarios: GoldScenario<TInput, TLabel>[], options?: SplitGoldOptions): GoldSplit<TInput, TLabel>;
-
-/**
- * Agreement judge for teacher→student distillation. Scores a STUDENT artifact
- * (the cheap analyst's produced label) against the GoldScenario's gold label
- * (the teacher's verdict). The score IS the distillation objective: 1.0 means
- * the student reproduced the teacher exactly, 0.0 means total disagreement.
- *
- * The comparison function is INJECTED (`compareLabels`) so the judge is
- * domain-agnostic — distilling a skill-audit analyst, a triage analyst, or any
- * other student is a one-line comparator swap. A default `fieldAgreement`
- * comparator is provided for the common case: a flat verdict object with
- * categorical and array fields.
- *
- * Everything here is PURE + unit-testable — no LLM. (The student spends tokens
- * producing the artifact; scoring it against frozen gold does not.)
- */
-
-/** What an injected comparator returns: a [0,1] composite plus the per-field
- *  (per-dimension) agreement breakdown the GEPA proposer reflects on to learn
- *  WHICH part of the verdict the student is getting wrong. */
-interface AgreementResult {
-    /** Overall agreement in [0,1]. */
-    score: number;
-    /** Per-dimension agreement in [0,1] — keyed by field/aspect name. The
-     *  reflective proposer surfaces the weakest of these as the lever to fix. */
-    dimensions: Record<string, number>;
-}
-/** Compare a produced label against a gold label → agreement. Injected so the
- *  judge is domain-agnostic. */
-type CompareLabels<TProduced = unknown, TLabel = unknown> = (produced: TProduced, gold: TLabel) => AgreementResult;
-interface BuildAgreementJudgeOptions<TProduced = unknown, TLabel = unknown> {
-    /** Judge name surfaced in `CampaignResult.aggregates.byJudge` + the gate. */
-    name?: string;
-    /** The agreement function — produced student label vs gold teacher label. */
-    compareLabels: CompareLabels<TProduced, TLabel>;
-    /** Dimension keys the judge declares up-front (for `JudgeConfig.dimensions`).
-     *  When omitted, the dimensions present on the first scored result are used
-     *  for display only; the composite is unaffected. */
-    dimensionKeys?: string[];
-    /** Only score `gold`-kind scenarios. Default true — a mixed campaign won't
-     *  mis-apply the agreement judge to non-gold scenarios. */
-    goldOnly?: boolean;
-}
-/** Build a `JudgeConfig` that scores a produced student artifact against the
- *  scenario's gold label. Conforms to the substrate `JudgeConfig` contract:
- *  `score({artifact, scenario, signal}) => JudgeScore`. The `composite` is the
- *  comparator's `score`; `dimensions` carries its per-field breakdown plus the
- *  scalar `agreement` so a single-dimension consumer still sees the number. */
-declare function buildAgreementJudge<TProduced, TInput, TLabel>(options: BuildAgreementJudgeOptions<TProduced, TLabel>): JudgeConfig<TProduced, GoldScenario<TInput, TLabel>>;
-interface FieldAgreementSpec {
-    /** Categorical fields — scored exact-match (1 if equal, else 0). Compared
-     *  with `===` after `JSON`-normalizing so `true`/`'high'`/`3` all work. */
-    categorical?: string[];
-    /** Array fields — scored by Jaccard overlap (|A∩B| / |A∪B|). Two empty
-     *  arrays agree perfectly (1.0). Order-insensitive; elements compared by
-     *  their `JSON.stringify`. */
-    array?: string[];
-}
-/** Default comparator: average per-field agreement over a flat verdict object.
- *  Categorical fields score exact-match; array fields score set-overlap
- *  (Jaccard). The composite is the unweighted mean across all declared fields,
- *  so missing a single boolean (e.g. `public_leak_risk`) costs `1/nFields` of
- *  the score — the leak-detection lever the audit cares about is a real,
- *  non-trivial fraction of the objective, not rounding noise.
- *
- *  Pure. A field absent from BOTH produced + gold is treated as agreeing
- *  (both undefined ⇒ 1.0); a field present in only one side disagrees. */
-declare function fieldAgreement<TProduced extends Record<string, unknown>, TLabel>(spec: FieldAgreementSpec): CompareLabels<TProduced, TLabel>;
-
-/**
- * `runDistillation` — the teacher→student distillation loop. COMPOSES existing
- * substrate primitives; reimplements none of them:
- *
- *   - DRIVER       = `gepaProposer` (reflective prompt optimizer)
- *   - LOOP         = `runImprovementLoop` (outer: optimize → holdout re-score → gate)
- *   - MEASUREMENT  = `runCampaign` (inside the loop) scoring the student
- *   - JUDGE        = `buildAgreementJudge` — student label vs gold teacher label
- *   - GATE         = caller-supplied (`heldOutGate` / `defaultProductionGate`)
- *   - STUDENT      = a cheap single-shot analyst whose system prompt is the
- *                    `MutableSurface` GEPA mutates; it calls the LLM through
- *                    `createChatClient` and emits a JSON label.
- *
- * The surface IS the student's system prompt. Each generation GEPA rewrites it;
- * `dispatchWithSurface` renders {surface + scenario.input} into a chat request,
- * calls the (cheap) model, parses the produced JSON label, and returns it as
- * the artifact. The agreement judge scores that label against the gold label.
- *
- * `autoOnPromote: 'none'` is FORCED — the loop never opens a PR; the caller
- * (the `distill` CLI) decides what to do with the winning prompt.
- */
-
-/** Render the student's prompt from {current surface, scenario input}. The
- *  surface is the system prompt; the scenario input is the user turn. Override
- *  to inject few-shot framing or a JSON-schema reminder. */
-type RenderStudentPrompt<TInput> = (args: {
-    surface: string;
-    input: TInput;
-    scenarioId: string;
-}) => ChatRequest['messages'];
-/** Parse the model's raw text into a typed produced label. Throws on
- *  unparseable output — a thrown dispatch is recorded as a failed cell (never
- *  silently scored 0), which is the honest signal that the prompt isn't
- *  emitting valid JSON yet. */
-type ParseStudentLabel<TProduced> = (rawContent: string, scenarioId: string) => TProduced;
-interface RunDistillationOptions<TProduced, TInput, TLabel> {
-    /** The student analyst's INITIAL system prompt — the baseline surface GEPA
-     *  searches from. */
-    baselinePrompt: string;
-    /** Training scenarios (the optimization pool). */
-    train: GoldScenario<TInput, TLabel>[];
-    /** Held-out scenarios — kept OUT of training; scored only at the gate. */
-    holdout: GoldScenario<TInput, TLabel>[];
-    /** Transport for BOTH the student (cheap model) and the GEPA reflection
-     *  (the optimizer model). The student calls it via `createChatClient`. */
-    llm: CreateChatClientOpts;
-    /** Router transport the GEPA proposer reflects through. `gepaProposer` uses the
-     *  package `LlmClient` directly (`LlmClientOptions`), not the ChatClient —
-     *  pass the router creds here. A test may inject `fetch` to stub the
-     *  reflection HTTP and exercise the wiring without real tokens. */
-    reflectionLlm: LlmClientOptions;
-    /** Cheap model the student runs (e.g. a small/fast model). */
-    studentModel: string;
-    /** Model GEPA uses to propose prompt rewrites (typically a stronger model). */
-    optimizerModel: string;
-    /** Agreement judge — produced student label vs gold teacher label. */
-    judge: JudgeConfig<TProduced, GoldScenario<TInput, TLabel>>;
-    /** Promotion gate. Default: `heldOutGate` over the holdout. Pass
-     *  `defaultProductionGate({ holdoutScenarios: holdout, ... })` for the full
-     *  red-team / reward-hacking / canary stack. */
-    gate?: Gate<TProduced, GoldScenario<TInput, TLabel>>;
-    /** GEPA population size (candidates per generation). Default 4. */
-    populationSize?: number;
-    /** GEPA generations. Default 3. */
-    maxGenerations?: number;
-    /** Campaign reps per scenario. Default 1 — raise for CI bands on a flaky
-     *  student. */
-    reps?: number;
-    /** Where campaign artifacts + traces land. Default a temp dir under cwd. */
-    runDir?: string;
-    /** Levers offered to the GEPA reflection prompt. */
-    mutationPrimitives?: string[];
-    /** GEPA structured-doc constraints (preserve sections, edit budget). */
-    constraints?: GepaProposerConstraints;
-    /** Gate's minimum holdout-agreement delta to ship. Default 0.0 — a
-     *  distillation run reports the lift; the caller decides the bar. Only used
-     *  when `gate` is omitted (the default `heldOutGate`). */
-    deltaThreshold?: number;
-    /** Render the student prompt. Default: surface as system, JSON-stringified
-     *  input as the user turn with a JSON-only instruction. */
-    renderStudentPrompt?: RenderStudentPrompt<TInput>;
-    /** Parse the model's text into a produced label. Default: strict JSON parse
-     *  with fenced-block stripping. */
-    parseStudentLabel?: ParseStudentLabel<TProduced>;
-    /** Per-student-call sampling temperature. Default 0 (deterministic student;
-     *  the optimization signal must come from the PROMPT, not sampling noise). */
-    studentTemperature?: number;
-    /** Per-student-call max tokens. Default 1024. */
-    studentMaxTokens?: number;
-}
-interface RunDistillationResult<TProduced, TInput, TLabel> extends RunImprovementLoopResult<TProduced, GoldScenario<TInput, TLabel>> {
-    /** The winning student prompt (a string surface). */
-    winnerPrompt: string;
-    /** Mean agreement on the HOLDOUT — baseline vs winner. The headline number:
-     *  did distillation move the student closer to the teacher on UNSEEN gold? */
-    holdoutAgreement: {
-        baseline: number;
-        winner: number;
-        delta: number;
-    };
-}
-declare function runDistillation<TProduced, TInput, TLabel>(opts: RunDistillationOptions<TProduced, TInput, TLabel>): Promise<RunDistillationResult<TProduced, TInput, TLabel>>;
-/** Default student prompt render: surface as system, JSON input as the user
- *  turn, with a JSON-only output instruction. */
-declare function defaultRenderStudentPrompt<TInput>(args: {
-    surface: string;
-    input: TInput;
-    scenarioId: string;
-}): ChatRequest['messages'];
-/** Default label parse: strip a ```json fence if present, then `JSON.parse`.
- *  Throws on failure so the cell is recorded as failed, not silently zeroed. */
-declare function defaultParseStudentLabel<TProduced>(rawContent: string, scenarioId: string): TProduced;
-
 /**
  * Strong, generically-useful baseline ROLES — the top zone of an `AgentProfile`
  * before any domain layer. A product composes one of these with its own
@@ -17051,6 +16521,55 @@ declare namespace index {
   export { type index_AgentProfileSection as AgentProfileSection, index_BASELINE_ROLES as BASELINE_ROLES, type index_BaselineRoleKey as BaselineRoleKey, type index_ProfileSkill as ProfileSkill, type index_PromptProfile as PromptProfile, index_applyDomainPatch as applyDomainPatch, index_baselineProfile as baselineProfile, index_baselineProfileFromRole as baselineProfileFromRole, index_engineerRole as engineerRole, index_generalistRole as generalistRole, index_prodProfile as prodProfile, index_profileToSurface as profileToSurface, index_renderProfile as renderProfile, index_researcherRole as researcherRole, index_sectionHash as sectionHash };
 }
 
+/**
+ * Traced analyst wrapper — instruments `analyzeTraces` with spans so the
+ * analyst's internal model turns appear in the trace tree. Also wraps each
+ * actor turn callback with a span.
+ *
+ * The wrapper records the Ax turn loop at its public boundaries:
+ *   1. A parent span for the entire analyst run.
+ *   2. Per-turn child spans from the `onTurn` callback (captures code,
+ *      output size, error status).
+ *   3. Summary attributes on the parent (total turns, usage, findings).
+ */
+
+interface TracedAnalystOptions {
+    /** TraceEmitter for span emission. */
+    emitter: TraceEmitter;
+    /** Parent span id. If omitted, uses emitter stack. */
+    parentSpanId?: string;
+}
+/**
+ * Run `analyzeTraces` wrapped in a parent span with per-turn child spans.
+ */
+declare function tracedAnalyzeTraces(input: AnalyzeTracesInput, options: AnalyzeTracesOptions, traceOpts: TracedAnalystOptions): Promise<AnalyzeTracesResult>;
+
+/**
+ * Traced judge wrappers — instruments every LLM call inside the judge
+ * ensemble with child spans so OTEL sinks see per-judge latency, model,
+ * token counts, and score dimensions.
+ *
+ * The ensemble parent span groups all individual judge spans; each judge
+ * gets its own child span with model + score as attributes.
+ */
+
+interface TracedJudgeOptions {
+    /** TraceEmitter to emit spans into. */
+    emitter: TraceEmitter;
+    /** Parent span id for the ensemble. If omitted, uses the emitter stack. */
+    parentSpanId?: string;
+}
+/**
+ * Wrap a single JudgeFn so its LLM call emits a traced span.
+ */
+declare function traceJudge(judge: JudgeFn, judgeName: string, opts: TracedJudgeOptions): JudgeFn;
+/**
+ * Wrap an array of JudgeFns with tracing, running them inside an ensemble
+ * parent span. Returns a single function that calls all judges and merges
+ * their scores.
+ */
+declare function traceJudgeEnsemble(judges: JudgeFn[], judgeNames: string[], opts: TracedJudgeOptions): JudgeFn;
+
 /**
  * Program cost report — a thin projection over `CostLedger.summary()` that
  * adds the per-model rollup the summary lacks, plus `attachCostToReport`, the
@@ -17107,9 +16626,8 @@ declare function attachCostToReport<R extends object>(report: R, ledger: CostLed
  *  - `judges`     → `ensembleJudge({ models: seats.judges, … })` (src/judge-panel.ts)
  *                   and the `JudgeConfig`s handed to `makeEvalTools({ judges })`
  *                   (src/eval-tools.ts).
- *  - `reflection` → `selfImprove({ llm: { model: seats.reflection } })` — the
- *                   `gepaProposer` reflection model (src/contract/self-improve.ts);
- *                   same seat for any custom `SurfaceProposer`'s LLM.
+ *  - `reflection` → the model configured by a custom `SurfaceProposer` or
+ *                   external optimization engine.
  *  - `worker`     → the dispatch model the agent itself calls — the model an
  *                   `AgentProfile` declares.
  *  - `analyst`    → the LLM behind `analyzeRuns` / analyst-registry kinds.
@@ -17128,7 +16646,7 @@ interface ModelSeats {
     judges?: string[];
     /** Analyst model — `analyzeRuns` / analyst-registry LLM calls. */
     analyst?: string;
-    /** Reflection/proposer model — `gepaProposer` mutation proposals. */
+    /** Reflection or candidate-generation model. */
     reflection?: string;
     /** Verifier model — completion/objective checking. */
     verifier?: string;
@@ -17547,4 +17065,4 @@ type CachedJudge<TArtifact, TScenario extends Scenario = Scenario> = JudgeConfig
  */
 declare function cachedJudge<TArtifact, TScenario extends Scenario = Scenario>(judge: JudgeConfig<TArtifact, TScenario>, store: VerdictCacheStore, options: CachedJudgeOptions): CachedJudge<TArtifact, TScenario>;
 
-export { AGENT_PROFILE_KINDS, ATTESTATION_ALGORITHM, type ActionExecutionPolicy, type ActionPolicyDecision, type ActionableSideInfo, type ActiveLearningOptions, type AdapterRun, AgentDriver, type AgentDriverConfig, AgentEvalError, type AgentEvalErrorCode, type AgentInterfaceProfileLike, type AgentProfileCell, type AgentProfileCellInput, type AgentProfileCellSchemaVersion, AgentProfileCellValidationError, type AgentProfileDimensionValue, type AgentProfileHarness, type AgentProfileJson, type AgentProfileJsonObject, type AgentProfileKind, type AgentProfileRuntimeReceipt, type AgentProfileSource, type AgentProfileSourceInput, type AgreementResult, type AlignmentOp, type Analyst, type AnalystContext, type AnalystCost, type AnalystFinding, type AnalystHooks, type AnalystInputKind, AnalystRegistry, type AnalystRegistryOptions, type AnalystRequirements, type AnalystRunEvent, type AnalystRunInputs, type AnalystRunResult, type AnalystRunSummary, type AnalystSeverity, type AnalystUsageReceipt, type AnalyzeTracesInput, type AnalyzeTracesOptions, type AnalyzeTracesResult, type AnalyzeTracesTurnSnapshot, type AntiSlopConfig, type AntiSlopIssue, type AntiSlopReport, type Artifact$1 as Artifact, type ArtifactCheck, type Artifact as ArtifactCheckArtifact, type ArtifactEventLike, type ArtifactResult, type ArtifactValidator, type AsiSeverity, type AssertCapabilityHeadroomOptions, type AssertCrossFamilyOptions, type AssertSingleBackendOptions, type AttestationProvenance, type AttestationVerification, type AttestedReport, type AutoPrClient, AxGepaSteeringOptimizer, type AxSteeringOptimizerConfig, BENCHMARK_SPLIT_SEED, type BackendDescriptor, BackendIntegrityError, type BackendIntegrityReport, type BaselineOptions, type BaselineReport, BehaviorAssertion, type BehavioralMetrics, type BehavioralTokenSequence, type BenchmarkAdapter, type BenchmarkDatasetItem, type BenchmarkEvaluation, type BenchmarkFamily, type BenchmarkReport$1 as BenchmarkReport, type BenchmarkResponder, BenchmarkRunner, type BenchmarkRunnerConfig, type BenchmarkScenario, type BenchmarkSource, type BenchmarkTaskKind, type BisectOptions, type BisectResult, type BisectStep, type BlendWeights, type BootstrapOptions, type BootstrapResult, BudgetBreachError, BudgetGuard, type BudgetLedgerEntry, type BudgetPolicy, type BudgetSpec, type BuildAgreementJudgeOptions, CODING_HARNESSES, type CachedJudge, type CachedJudgeOptions, type CalibrationResult, CallExpectation, CallbackResearcher, type CallbackResearcherOptions, type CampaignFactoryParams, type CampaignIntegrityPolicy, type CampaignRunContext, type CampaignRunOutcome, type CampaignRunner, type CampaignScenario, type CampaignVariant, type CanaryAlert, type CanaryKind, type CanaryLeak, type CanaryOptions, type CanaryReport, type CanarySeverity, type CandidateComparison, type CandidateScenario, type CandidateScore, type CanonicalRawAnalystFinding, type CapabilityHeadroomOptions, type CapabilityHeadroomResult, type CaptureFetchContext, type CaptureFetchOptions, CaptureIntegrityError, type CausalAttributionReport, type CellVerdict, type ChannelRollup, type ChatCallOpts, type ChatClient, type ChatMessage, type ChatRequest, type ChatResponse, type ChatToolCall, type ChatTransport, type CheckResult, type CliBridgeTransportOpts, type CliffsMagnitude, type ClusterBootstrapInterval, type ClusterSignFlipAlternative, type ClusterSignFlipResult, type ClusteredBinaryCluster, type ClusteredMatchedPair, type ClusteredPairedBinaryOptions, type ClusteredPairedBinaryResult, type ClusteredPairedBinaryStatistics, type CollectedArtifacts, type CommandRunner, type CompareLabels, type ComparePairedArmsOptions, type CompletionCriterion, type CompletionRequirement, type CompletionVerdict, type ConceptComplexity, type ConceptFinding, type ConceptSpec, type ConceptWeightStrategy, ConfigError, type ContinuityCheck, type ContinuityCheckResult, type ContinuityReport, type ContinuitySnapshotPair, type ContinuousAgreement, type ContinuousAgreementOptions, type ContinuousCalibrationResult, type ContractCheckResult, type ContractJudgeOptions, type ContractMetric, type ContractReport, type ContractRule, type ContractRuleKind, type ContractSpan, type ContractVerdict, type ContractViolation, type ControlActionFailureMode, type ControlActionOutcome, type ControlBudget, type ControlContext, type ControlDecision, type ControlEvalResult, type ControlRunResult, type ControlRunToRunRecordOptions, type ControlRuntimeConfig, type ControlRuntimeError, type ControlSeverity, type ControlStep, type ControlStopPolicies, ConvergenceTracker, type CorpusAgreementOptions, type CorpusAgreementPerDimension, type CorpusAgreementReport, type CorpusScoreRecord, type CorrectnessChecker, CostAccountingIncompleteError, CostCallConflictError, CostCeilingReachedError, type CostChannel, type CostEntry, CostLedger, type CostLedgerEntry, type CostLedgerFilter, type CostLedgerHandle, type CostLedgerOptions, type CostLedgerPersistence, CostLedgerPersistenceError, type CostLedgerSummary, type CostReceipt, CostReceiptCaptureError, type CostReceiptInput, type CostReport, CostReservationExceededError, type CostResult, type CostSummary, CostTracker, type CostUsage, type CounterfactualContext, type CounterfactualMutation, type CounterfactualResult, type CounterfactualRunner, type CreateAnalystAiConfig, type CreateChatClientOpts, type CreateDefaultReviewerOptions, type CreateExperimentInput, type CreateSandboxPoolOpts, type CreateTraceAnalystKindOpts, CrossFamilyError, type CrossTraceDiff, type CrossTraceDiffOptions, type CustomTokenPricing, DEFAULT_AGENT_SLOS, DEFAULT_COMPLEXITY_WEIGHTS, DEFAULT_RULES as DEFAULT_FAILURE_RULES, DEFAULT_FINDERS, DEFAULT_HARNESS_OBJECTIVES, DEFAULT_MUTATION_PRIMITIVES, DEFAULT_MUTATORS, DEFAULT_PR_REVIEW_SCORE_WEIGHTS, DEFAULT_REDACTION_RULES, DEFAULT_RED_TEAM_CORPUS, DEFAULT_RUN_SCORE_WEIGHTS, DEFAULT_SEVERITY_WEIGHTS, DEFAULT_TRACE_ANALYST_BUDGETS, DEFAULT_TRACE_ANALYST_KINDS, type DataAcquisitionPlan, Dataset, type DatasetDifficulty, type DatasetManifest, type DatasetOverview, type DatasetProvenance, type DatasetScenario, type DatasetSplit, type DecideNextUserTurnOpts, type DefaultAnalystRegistryOptions, type DefaultVerdict, type DeployFamily, type DeployGateLayerInput, type DeployRunResult, type DeployRunner, type DescriptionLengthCandidate, type DescriptionLengthConfig, type DescriptionLengthDecision, type DescriptionLengthEvidence, DescriptionLengthGate, type DescriptionLengthRejectionCode, type DetectorEvent, type DetectorSeverity, type DetectorSignal, type DiffPolicy, type DiffScorecardOptions, type DirEntry, type DirectProviderTransportOpts, type Direction, type DiscoverPersonasOptions, type DiscoveredPersona, DockerSandboxDriver, type DriverResult, type DriverState, DualAgentBench, type DualAgentBenchConfig, type DualAgentReport, type DualAgentRound, type DualAgentScenario, type DualAgentScenarioResult, type EProcess, type EProcessOptions, type EProcessState, type EProcessStep, ERROR_COUNT_PATTERNS, type EnsembleAggregate, type EnsembleJudgeOptions, type ErrorCluster, type ErrorCountPattern, type ErrorStreakOptions, type EvalCampaignOptions, type EvalCampaignResult, type EvalResult, type EvalToolDef, EvalTraceStore, type EventFilter, type EventKind, type EvidenceRef, type EvolutionRound, type ExecutorConfig, type Expectation, type Experiment, type ExperimentPlan, type ExperimentProvenance, type ExperimentRep, type ExperimentResult, type ExperimentStats, type ExperimentStore, ExperimentTracker, type ExperimentTrackerOptions, type ExperimentVerdict, type ExportableSpan, type ExportedRewardModel, type ExtractOptions, type ExtractResult, type ExtractUsageFromSseOptions, type ExtractedUsage, FAILURE_CLASSES, FAILURE_MODE_KIND_SPEC, type FactorContribution, type FactorialCell, type FailedRun, type FailureClass, type FailureClassification, type FailureContext, type FailureMode, type FailureRule, type FeedbackArtifactType, type FeedbackAttempt, type FeedbackLabel, type FeedbackLabelKind, type FeedbackLabelSource, type FeedbackOptimizerRow, type FeedbackOutcome, type FeedbackPattern, type FeedbackReplayAdapter, type FeedbackReplayResult, type FeedbackSeverity, type FeedbackSplitPolicy, type FeedbackTask, type FeedbackTrajectory, type FeedbackTrajectoryFilter, type FeedbackTrajectoryStore, type FieldAgreementSpec, type FieldDestination, type FileChange, FileSystemFeedbackTrajectoryStore, FileSystemRawProviderSink, type FileSystemRawProviderSinkOptions, FileSystemTraceStore, type FileSystemTraceStoreOptions, type Finding, type FindingSubject, type FindingSubjectKind, type FindingToPolicyEditOptions, type FindingsDiff, FindingsStore, type FlattenOtlpOptions, type FlowAction, type FlowLayerEnv, type FlowLayerFactoryInput, type FlowRunner, type FlowRunnerStepResult, type FlowSpec, type FlowStep, type GainDistributionBin, type GainDistributionFigureSpec, type GainDistributionOptions, type GateDecision$1 as GateDecision, type GateEvidence, type GenericSpan, type GhCliClientOptions, type GoldScenario, type GoldSplit, type GoldenItem, type GoldenSeverity, type GoldenSpec, HARNESS_NATIVE_MODEL, type HarnessAdapter, type HarnessConfig, type HarnessExperimentConfig, type HarnessExperimentResult, type HarnessIntervention, type HarnessRunRequest, type HarnessRunResult, type HarnessScenario, type HarnessSelection, type HarnessVariant, type HarnessVariantReport, type HeadroomClass, type HeadroomInput, HeldOutGate, type HeldOutGateConfig, type HeldOutGateRejectionCode, type HeldOutPartition, type HiddenCriteriaGrader, type HiddenGradeResult, type HiddenLeak, HoldoutAuditor, HoldoutLockedError, type HttpGithubClientOptions, type HypothesisManifest, type HypothesisResult, IMPROVEMENT_KIND_SPEC, INPUT_VALUE, INTENT_MATCH_JUDGE_VERSION, type ImageData, type ImprovementThresholds, type ImprovementVerdictResult, InMemoryFeedbackTrajectoryStore, InMemoryRawProviderSink, type InMemoryRawProviderSinkOptions, InMemoryTraceStore, InMemoryWorkspaceInspector, type InferenceScorer, type InspectorContext, type IntentMatchInput, type IntentMatchOptions, type IntentMatchResult, type InteractionContribution, type InterimReleaseConfidence, type InterimReleaseConfidenceInput, type JudgeConfig$1 as JudgeConfig, JudgeError, type JudgeFamily, type JudgeFleetOptions, type JudgeFn, type JudgeInput, JudgeParseError, type JudgeReplayGateArgs, type JudgeReplayResult, type JudgeRetryOutcome, type JudgeRetryPolicy, type JudgeRubric, JudgeRunner, type JudgeScore$1 as JudgeScore, type JudgeScoreInput, type JudgeScoresRecord, type JudgeSpan, type JudgeVerdict, KNOWLEDGE_GAP_KIND_SPEC, KNOWLEDGE_POISONING_KIND_SPEC, type KeywordConceptSpec, type KeywordCoverageFinding, type KeywordCoverageOptions, type KeywordCoverageResult, type KnowledgeAcquisitionMode, type KnowledgeBundle, type KnowledgeFallbackPolicy, type KnowledgeFreshness, type KnowledgeImportance, type KnowledgeReadinessReport, type KnowledgeRecommendedAction, type KnowledgeRequirement, type KnowledgeRequirementCategory, type KnowledgeResponsibleSurface, type KnowledgeSensitivity, LLM_CACHED_TOKENS, LLM_CACHED_TOKEN_ATTR_KEYS, LLM_CACHE_WRITE_TOKENS, LLM_CACHE_WRITE_TOKEN_ATTR_KEYS, LLM_CONTEXT_TOKENS, LLM_COST_ATTR_KEYS, LLM_COST_USD, LLM_INPUT_TOKENS, LLM_INPUT_TOKEN_ATTR_KEYS, LLM_MODEL_ATTR_KEYS, LLM_MODEL_NAME, LLM_OUTPUT_TOKENS, LLM_OUTPUT_TOKEN_ATTR_KEYS, LLM_REASONING_TOKENS, LLM_REASONING_TOKEN_ATTR_KEYS, type LangfuseEnvelope, type LangfuseGeneration, type LangfuseScore, type Layer, type LayerResult, type LayerStatus, type LeaderboardOptions, type LeaderboardRow, type LiveProofArtifact, type LiveProofConfig, type LiveProofContext, type LiveProofResult, LlmCallError, type LlmCallMetadata, type LlmCallRequest, type LlmCallResult, LlmClient, type LlmClientOptions, type LlmCorrectnessCheckerOpts, type LlmJsonCall, type LlmJudgeDimension, type LlmJudgeOptions, type LlmMessage, LlmResponseError, type LlmReviewerConfig, LlmRouteAssertionError, type LlmRouteRequirements, type LlmSpan, type LlmSpanOtlpInput, type LlmUsage, LockedJsonlAppender, MODEL_PRICING, type MakeEvalToolsConfig, type MatchResult, type MatchedPair, type MatcherResult, type MaximumCharge, type McNemarResult, type Measured, type MeasurementPolicy, type MergeOptions, type Message, type MetricSamples, type MetricVerdict, MetricsCollector, type MintRolloutOptions, type MintRolloutResult, type MockTransportOpts, type ModelCostRollup, type ModelPreflight, type ModelSeats, ModelsUnreachableError, type MuffledFinder, type MuffledFinding, MultiLayerVerifier, type MultiToolchainLayerConfig, type Mutator, Mutex, type NoLeakOptions, type NoProgressOptions, NoopRawProviderSink, NoopResearcher, NotFoundError, OPENINFERENCE_SPAN_KIND, OTEL_AGENT_EVAL_SCOPE, OUTPUT_VALUE, type Objective, type Oracle, type OracleObservation, type OracleReport, type OracleResult, type OrthogonalityInput, type OrthogonalityResult, type OtelExportConfig, type OtelExporter, type OtelPipelineHandle, type OtelPipelineOptions, type OtlpExport, OtlpFileTraceStore, type OtlpFileTraceStoreOptions, type OtlpFlatLine, type OtlpResourceSpans, type OtlpSpan, type OtlpToRunRecordsOptions, type OtlpTraceRunRecord, POLICY_EDIT_AXES, POLICY_EDIT_CANDIDATE_RECORD_SCHEMA, POLICY_EDIT_TARGET_SURFACES, type PaidCallResult, type PairArmsOptions, type PairArmsResult, type PairedArmRow, type PairedArmsComparison, type PairedBootstrapOptions, type PairedBootstrapResult, type PairedCorrectness, type PairedEvalueOptions, type PairedEvalueSequence, type PairedEvalueStep, type PairedMetricDelta, type PairedSignTestResult, PairwiseSteeringOptimizer, type ParaphraseRobustnessScenarioInput, type ParaphraseRobustnessScenarioResult, type ParetoFigureSpec, type ParetoPoint, type ParetoResult, type ParseStudentLabel, type PartitionHeldOutOptions, type PendingCostCall, type PendingCostCallView, type PersistedFinding, type PersonaConfig, type PersonaRigor, type Playbook, type PlaybookEntry, type PolicyEdit, type PolicyEditAdmission, type PolicyEditAdmissionOptions, type PolicyEditAxis, type PolicyEditCandidateRecord, type PolicyEditChange, type PolicyEditExpectedGain, type PolicyEditGainDirection, type PolicyEditGainUnit, type PolicyEditInit, type PolicyEditRisk, type PolicyEditSchemaVersion, type PolicyEditSource, type PolicyEditTarget, type PolicyEditTargetSurface, PolicyEditValidationError, type PoolSlot, type PositionalBiasResult, type PrReviewAuditCase, type PrReviewBenchmarkSummary, type PrReviewComment, type PrReviewMatchedFinding, type PrReviewOutcome, type PrReviewReferenceFinding, type PrReviewScore, type PrReviewScoreWeights, type PrReviewSeverity, type PrReviewSource, type PreferenceMemoryEntry, type PreflightModelsOptions, type PreflightOutcome, type ProducedProposal, type ProducedState, type ProductBenchmarkArm, type ProductBenchmarkArtifactPaths, type ProductBenchmarkBudgets, type ProductBenchmarkExportOptions, type ProductBenchmarkExportResult, type ProductBenchmarkManifest, type ProductBenchmarkProfileRef, type ProductBenchmarkRecord, type ProductBenchmarkRepoRef, type ProductBenchmarkRunInput, type ProductBenchmarkScenario, type ProductBenchmarkSingleRunExportOptions, type ProductBenchmarkSplit, type ProductBenchmarkSubstrateVersions, type ProductBenchmarkValidationReport, ProductClient, type ProductClientConfig, type ProfileAxisSpec, type ProjectRuntimeTrajectoryEvidenceOptions, type ProjectedOtlpSpan, type PromptHandle, PromptRegistry, type ProportionInterval, type ProposalEventLike, type ProposeAutomatedPullRequestInput, type ProposeAutomatedPullRequestResult, type ProposeFn, type ProposeInput, type ProposeOutput, type ProposeReviewConfig, type ProposeReviewControlAction, type ProposeReviewControlConfig, type ProposeReviewControlResult, type ProposeReviewControlState, type ProposeReviewReport, type ProposeReviewShot, type ProposedSideEffect, type ProvenanceReader, type ProviderRedactor, type QueryTracesPage, REDACTION_VERSION, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, RESEARCH_REPORT_HARD_PAIR_FLOOR, ROLLOUT_FORMAT, ROLLOUT_SCHEMA, RUN_COST_ATTR_KEYS, type RawAnalystEvidence, type RawAnalystFinding, type RawProviderDirection, type RawProviderEvent, type RawProviderSink, type RawProviderSinkFilter, type RecordRunsOptions, type RedTeamCase, type RedTeamCategory, type RedTeamFinding, type RedTeamPayload, type RedTeamReport, type RedactionReport, type RedactionRule, type ReferenceEquivalenceJudgeInput, type ReferenceEquivalenceJudgeOptions, type ReferenceEquivalenceJudgeResult, type ReferenceEquivalenceScenario, type ReferenceMatchResult, type ReferenceReplayAdapter, type ReferenceReplayAdapterFn, type ReferenceReplayAdapterLike, type ReferenceReplayAggregate, type ReferenceReplayCandidate, type ReferenceReplayCase, type ReferenceReplayCaseRun, type ReferenceReplayExecutionScenario, type ReferenceReplayItem, type ReferenceReplayMatch, type ReferenceReplayMatchStrategy, type ReferenceReplayMatcher, type ReferenceReplayPromotionDecision, type ReferenceReplayPromotionPolicy, type ReferenceReplayRun, type ReferenceReplayRunContext, type ReferenceReplayRunOptions, type ReferenceReplayRunStore, type ReferenceReplayScenario, type ReferenceReplayScenarioScore, type ReferenceReplayScore, type ReferenceReplayScoreOptions, type ReferenceReplaySplit, type ReferenceReplaySplitComparison, type ReferenceReplaySteeringRowsOptions, type ReflectionContext, type ReflectionProposal, type RegistryRunOpts, type ReleaseConfidenceAxis, type ReleaseConfidenceAxisName, type ReleaseConfidenceInput, type ReleaseConfidenceIssue, type ReleaseConfidenceMetrics, type ReleaseConfidenceScorecard, type ReleaseConfidenceStatus, type ReleaseConfidenceThresholds, type ReleaseTraceEvidence, type RenderReleaseReportOptions, type RenderStudentPrompt, type RepeatedActionOptions, ReplayCache, type ReplayCacheEntry, ReplayCacheMissError, type ReplayCacheStats, ReplayError, type ReplayFetchOptions, type RepoRef, type RequirementCheck, type ResearchReport, type ResearchReportCandidate, type ResearchReportDecision, type ResearchReportMethodology, type ResearchReportOptions, type ResearchReportRecommendation, type Researcher, type RetrievalSpan, type Review, type ReviewFn, type ReviewInput, type ReviewMemoryEntry, type ReviewMemoryStore, type ReviewerMemoryEntry, type ReviewerOutput, type ReviewerPromptInput, type ReviewerSoftFailDefaults, type ReviewerVerificationSummary, type RewardRow, type RiskDifferenceResult, type RobustnessResult, type RolloutCapture, type RolloutLine, type RolloutRole, type RolloutScrubber, type RolloutSplit, type RolloutStep, type RouteMap, type RoutedField, type RouterTransportOpts, type RubricDimension, type Run, type RunCommandInput, type RunCommandResult, type RunCompleteHook, type RunCompleteHookContext, type RunCostProvenance, RunCritic, type RunCriticOptions, type RunDistillationOptions, type RunDistillationResult, type RunEvidenceMetadata, type RunFilter, RunIntegrityError, type RunIntegrityExpectations, type RunIntegrityIssue, type RunIntegrityIssueCode, type RunIntegrityReport, type RunJudgeMetadata, type RunLayer, type RunOutcome, type RunPaidCallInput, type RunRecord, type RunRecordBackend, type RunRecordFilter, RunRecordValidationError, type RunScore, type RunScoreWeights, type RunSplitTag, type RunStatus, type RunTokenUsage, type RunTrace, type RuntimeEventLike, type RuntimeResolution, type RuntimeTrajectoryEvidenceProjection, type RuntimeTrajectoryEvidenceSummary, type RuntimeTrajectoryHookEvent, type RuntimeTrajectoryRecord, type RuntimeTrajectoryRunRecord, SEMANTIC_CONCEPT_JUDGE_VERSION, SKILL_USAGE_ANALYST, SPAN_KIND_ATTR_KEYS, SUPERVISOR_RUN_SCHEMA, type SandboxDriver, SandboxHarness, type SandboxHarnessResult, type SandboxJudgeKind, type SandboxJudgeResult, type SandboxJudgeSpec, type SandboxPool, type SandboxResult, type SandboxSdkTransportOpts, type SandboxSpan, type SatisfiedBy, type ScanOptions, type Scenario$1 as Scenario, type ScenarioCost, type ScenarioFile, ScenarioRegistry, type ScenarioResult, type ScoreKnowledgeReadinessOptions, type Scorecard, type ScorecardCell, type ScorecardCellDiff, type ScorecardDiff, type ScorecardEntry, type ScorecardLogLine, type ScoredTarget, type SearchSpanResult, type SearchTraceResult, type SeatName, type SeatPresetName, SeatUnsetError, type SelfPlayOptions, type SelfPlayProposer, type SelfPlayScorer, type SelfPreferenceResult, type SemanticConceptJudgeInput, type SemanticConceptJudgeOptions, type SemanticConceptJudgeResult, type SequentialDecision, type SerializedRegex, type SeriesConvergenceOptions, type SeriesConvergenceResult, type Severity, type SftExportOptions, type SftRow, type SignTestAlternative, type SignedManifest, type SignedManifestAlgo, type SingleBackendDivergence, SingleBackendError, type SingleBackendField, type SingleBackendReport, SkillUsageAnalyst, type SliceOptions, type Slo, type SloCheckResult, type SloComparator, type SloReport, type SloSeverity, type SlopCategory, type SlotFactory, type Span, type SpanBase, type SpanFilter, type SpanHandle, type SpanKind, type SpanMatchRecord, SpanNotFoundError, type SpanPredicate, type SpanStatus, type SplitGoldOptions, type SseUsageMode, type SteeringBundle, type SteeringChange, type SteeringDelta, type SteeringOptimizationResult, type SteeringOptimizationRow, type SteeringOptimizationSelector, type SteeringOptimizerBackend, type SteeringOptimizerConfig, type SteeringRolePrompt, type StepAttribution, type StopDecision, type StreamingDetector, type SuboptimalCode, type SuboptimalSignal, SubprocessSandboxDriver, type SubprocessSandboxDriverOptions, type SummaryTable, type SummaryTableOptions, type SummaryTableRow, type SupervisorRunReader, type SupervisorRunReport, type SupervisorRunRollup, type SupervisorRunSources, type SupervisorRunTree, type SynthesisReason, type SynthesisTarget, TOOL_ARGS_CAPTURED, TOOL_LATENCY_MS, TOOL_NAME, TOOL_NAME_ATTR_KEYS, TRACE_ANALYST_ACTOR_DESCRIPTION, TRACE_ANALYST_ACTOR_DESCRIPTION_VERSION, TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, TRACE_SCHEMA_VERSION, type TaskGold, type TaskHeadroom, type TestGradedRunOptions, type TestGradedRunResult, type TestGradedScenario, type TestOutputParser, type TestResult, type TextMatcher, type ThresholdContract, TokenCounter, type TokenSpec, type ToolCallEventLike, type ToolDef, type ToolMatcher, type ToolSpan, type ToolSpanOtlpInput, type ToolStats, type ToolUseMetrics, type ToolUseOptions, type TraceAggregate, type TraceAnalysisStore, type TraceAnalystByteBudgets, type TraceAnalystFilters, type TraceAnalystGolden, type TraceAnalystHookOptions, type TraceAnalystKindSpec, type TraceAnalystSpan, type TraceAnalystSpanKind, type TraceAnalystSpanStatus, type TraceAnalystTraceSummary, type TraceContract, TraceContractBuilder, TraceEmitter, type TraceEmitterOptions, type TraceEvent, TraceFileMissingError, type TraceInsightContext, type TraceInsightFinding, type TraceInsightPanelRole, type TraceInsightPromptInput, type TraceInsightQualityGate, type TraceInsightQuestion, type TraceInsightReadiness, type TraceInsightSuite, type TraceInsightTask, TraceNotFoundError, type TraceStore, type TraceStoreSource, type TraceStoreToOtlpOptions, type TracedAnalystOptions, type TracedJudgeOptions, type TracesToOtlpResult, type Trajectory, type TrajectoryStep, type TreatmentClass, type TreatmentGate, type TreatmentGateInput, type TreatmentGateOptions, type TrialTrace, type Turn, type TurnMetrics, type TurnResult, UI_FINDING_SEVERITIES, UI_LENSES, UNIVERSAL_FINDERS, type UiFinding, type UiFindingScreenshot, type UiFindingSeverity, type UiLens, type Unavailable, type UserQuestion, type ValidationContext, ValidationError, type ValidationIssue, type ValidationResult, type VerbosityBiasResult, type Verdict, type VerdictCacheStats, type VerdictCacheStore, type Verification, VerificationError, type VerificationReport, type VerifyContext, type VerifyFn, type VerifyOptions, type ViewSpansResult, type ViewTraceOversized, type ViewTraceResult, type VisualDiffOptions, type VisualDiffResult, type ViteDeployRunnerInput, type WeightedCompositeInput, type WeightedCompositeResult, type WorkerDriverContext, type WorkflowTopology, type WorkspaceAssertion, type WorkspaceAssertionResult, type WorkspaceInspector, type WorkspaceSnapshot, type WranglerDeployRunnerInput, acquisitionPlansForKnowledgeGaps, admitPolicyEdit, adversarialJudge, agentProfileCellHashMaterial, agentProfileCellKey, agentProfileHash, agentProfileId, agentProfileModelId, agentVisibleFields, aggregateJudgeVerdicts, aggregateLlm, aggregatePrReviewScore, aggregateRunScore, allCriticalPassed, analyzeAntiSlop, analyzeSeries, analyzeSupervisorRun, analyzeSupervisorRunSources, analyzeTraces, appendScorecard, applyLlmSpanOtlpAttributes, applyPolicyEditToSurface, applyToolSpanOtlpAttributes, argHash, asNumber, asString, assertCapabilityHeadroom, assertCrossFamily, assertLlmRoute, assertModelsServed, assertNoHiddenLeak, assertProductBenchmarkRun, assertRealAgentReceipts, assertRealBackend, assertReleaseConfidence, assertRolloutLine, assertRunAgentProfileCell, assertRunCaptured, assertSingleBackend, assignFeedbackSplit, assignHeldOutTag, attachCostToReport, attest, attributeCounterfactuals, backoffMs, deterministicSplit as benchmarkDeterministicSplit, index$1 as benchmarks, benjaminiHochberg, bisect, blendHeldout, blockingKnowledgeEval, bonferroni, bootstrapCi, buildAgentInterfaceProfileCell, buildAgentProfileCell, buildAgreementJudge, buildDefaultAnalystRegistry, buildDriverSystemPrompt, buildProductBenchmarkManifest, buildReflectionPrompt, buildReviewerPrompt, buildTraceAnalystTools, buildTraceInsightContext, buildTraceInsightPrompt, buildTrajectory, buildWorkerDriverSystemPrompt, byteLengthRange, cachedJudge, calibrateJudge, calibrateJudgeContinuous, callLlm, callLlmJson, canaryLeakView, canonicalJson, canonicalize, capabilityHeadroom, captureFetchToRawSink, causalAttribution, checkBehavioralCanary, checkCanaries, checkSlos, checkTraceContracts, clamp01, classifyFailure, classifyTreatment, cliffsDelta, clusteredPairedBinary, codeExecutionJudge, cohensD, coherenceJudge, collectionPreserved, commentsForSource, commitBisect, comparePairedArms, compareReferenceReplay, compareToBaseline, compilerJudge, completionVerdict, composeParsers, composeValidators, computeExperimentStats, computeFindingId, computePolicyEditId, computeToolUseMetrics, computeTraceMetrics, confidenceInterval, containsAll, contentHash, contextInputTokens, continuousAgreement, contractJudge, controlFailureClassFromVerification, controlRunToFeedbackTrajectory, controlRunToRunRecord, convertTraceStoresToOtlp, corpusInterRaterAgreement, corpusInterRaterAgreementFromJudgeScores, costForTokenPricing, costForUsage, costReceiptFromLlm, costReceiptFromLlmError, costReport, createAnalystAi, createAntiSlopJudge, createChatClient, createCustomJudge, createDefaultReviewer, createDomainExpertJudge, createFeedbackTrajectory, createIntentMatchJudge, createLlmCorrectnessChecker, createLlmReviewer, createOtelExporter, createOtelTracingStore, createReferenceEquivalenceJudge, createReplayFetch, createSandboxPool, createSemanticConceptJudge, createTokenRecallChecker, createTraceAnalystKind, crossTraceDiff, crowdingDistance, dataDescriptionBits, decideNextUserTurn, decideReferenceReplayPromotion, decideReferenceReplayRunPromotion, defaultBlendWeights, defaultIsMaterial, defaultJudges, defaultParseStudentLabel, defaultProviderRedactor, defaultReferenceReplayMatcher, defaultRenderStudentPrompt, defaultTraceInsightPanel, deployGateLayer, describeTraceInsightScope, diffFindings, diffScorecard, discoverPersonas, distillPlaybook, domainEvidencePattern, dominates, eProcess, ensembleJudge, errorStreakDetector, estimateCost, estimateTokens, evaluateActionPolicy, evaluateContract, evaluateHypothesis, evaluateInterimReleaseConfidence, evaluateOracles, evaluateReleaseConfidence, evaluateTraceContract, executeScenario, expandProfileAxes, expectAgent, exportProductBenchmark, exportProductBenchmarkRuns, exportRewardModel, exportRunAsOtlp, extractAssetUrls, extractErrorCount, extractOtlpAttributes, extractProducedState, extractUsage, extractUsageFromResponse, extractUsageFromSse, feedbackTrajectoriesToDatasetScenarios, feedbackTrajectoriesToOptimizerRows, feedbackTrajectoryToDatasetScenario, feedbackTrajectoryToOptimizerRow, fieldAgreement, fileContains, fileExists, fileExperimentStore, fileVerdictCache, findAutoMatchNoExpectation, findConstructorCwdDropped, findFallbackToPass, findLiteralTruePass, findProductBenchmarkArtifacts, findSkipCountsAsPass, firstNumberAttr, firstStringAttr, flattenOtlpExportToNdjson, flowLayer, fnv1a32, formatBenchmarkReport, formatDriverReport, formatFindings, formatScorecardDiff, gainHistogram, gateTreatmentApplied, gateTreatmentFromMetrics, gateTreatmentFromSpans, gateTreatmentFromToolSpans, ghCliClient, gitProvenanceReader, precision as goldenPrecision, gradeOnHidden, gradeSemanticStatus, groupBy, groupRunsByAgentProfileCell, harnessAxisOf, hasCapturedToolArgs, hashContent, hashJson, hashScenarios, hashToUnit, hiddenGrade, holm, htmlContainsElement, httpGithubClient, improvementVerdict, inMemoryExperimentStore, inMemoryReferenceReplayStore, inMemoryReviewStore, inMemoryRunRecordBackend, inMemoryVerdictCache, inferDomainKeywords, inferOtlpKind, interRaterReliability, interpretCliffs, iqr, isHiddenDestination, isJudgeSpan, isLlmSpan, isModelPriced, isOtelConfigured, isPolicyEdit, isRetrievalSpan, isRolloutLine, isRunRecord, isSandboxSpan, isToolSpan, isTrainableSplit, isTransientLlmError, isUnavailable, iterateRawCalls, jestTestParser, jsonHasKeys, jsonShape, jsonlReferenceReplayStore, jsonlReviewStore, jsonlRunRecordBackend, judgeFamily, judgeReplayGate, judgeSpans, keyPreserved, knowledgeReadinessTracePayload, leaderboard, linterJudge, llmJudge, llmSpanFromProvider, llmSpans, loadGoldScenarios, loadScorecard, loadScorerFromGrader, localCommandRunner, lowercaseMutator, makeEvalTools, makeFinding, makePolicyEdit, makePolicyEditCandidateRecord, mannWhitneyU, mapConcurrent, matchGoldens, matchSpan, maximumChargeForLlmRequest, mcnemar, mcnemarPower, mcnemarRequiredN, mergeLayerResults, mergeSteeringBundle, mintRolloutRows, modelDescriptionBits, modelHasSnapshot, modelPriceKey, mulberry32, multiToolchainLayer, noProgressDetector, normalizeScores, notBlocked, objectiveEval, observeAll, otelRunCompleteHook, otlpRowsToRunRecords, otlpRowsToTraceRunRecords, otlpToRunRecords, otlpToTraceRunRecords, pairArms, pairedBootstrap, pairedEvalueSequence, pairedMde, pairedRiskDifference, pairedSignTest, pairedTTest, paraphraseRobustness, paraphraseRobustnessScenarios, paretoChart, paretoFrontier, paretoFrontierWithCrowding, parseCorrectnessResponse, parseFeedbackTrajectoriesJsonl, parseGoldJsonl, parseReflectionResponse, parseRunRecordSafe, parseRuntimeTrajectoryHookEvent, partialCredit, partitionHeldOut, passAtK, passOrthogonality, pearsonR, pixelDeltaRatio, planTraceInsightQuestions, policyEditFromFinding, policyEditsFromFindings, politenessPrefixMutator, positionalBias, preflightModels, printDriverSummary, probeLlm, productBenchmarkIntegrityFailures, productBenchmarkMutableSurfaces, productBenchmarkRepoIdentity, productBenchmarkSplits, index as profile, projectOtlpFlatLine, projectRuntimeTrajectoryEvidence, promptBisect, proposeSynthesisTargets, providerFromBaseUrl, pytestTestParser, ranks, readOtlpStatus, readProductBenchmarkManifest, readProductBenchmarkRecords, recordRuns, recordRunsToScorecard, redTeamDataset, redTeamReport, redactString, redactValue, referenceReplayRunsToSteeringRows, referenceReplayScenarioToRunScore, regexMatch, regexMatches, renderMarkdownReport, renderPlaybookMarkdown, renderPreferenceMemoryMarkdown, renderPriorFindings, renderReleaseReport, renderSteeringText, renderSupervisorRunHeadline, renderSupervisorRunMarkdown, renderUpstreamFindings, repeatedActionDetector, replayFeedbackTrajectories, replayFeedbackTrajectory, replayScorerOverCorpus, replayTraceThroughJudge, requireAgentProfileCell, requiredSampleSize, researchReport, resolveModelPricing, resolveRunCostProvenance, resolveSeat, rolloutReward, rollupSupervisorRuns, roundTripRunRecord, routeFields, rowCount, rowWhere, runAgentControlLoop, runAssertions, runBehavioralCanaries, runCanaries, runCounterfactual, runDistillation, runE2EWorkflow, runEvalCampaign, runExpectations, runFailureClass, runHarnessExperiment, runIntentMatchJudge, runJudgeFleet, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runLiveProof, runProposeReview, runProposeReviewAsControlLoop, runRecordToProductBenchmarkRecord, runReferenceEquivalenceJudge, runReferenceReplay, runScore, runSelfPlay, runSemanticConceptJudge, runTestGradedScenario, runsForScenario, scalarScore, scanForMuffledGates, scoreContinuity, scoreFromEvals, scoreKnowledgeReadiness, scorePolicyEditReadiness, scorePrReviewComments, scorePrReviewSource, scoreRedTeamOutput, scoreReferenceReplay, scoreTraceInsightReadiness, seatPresets, securityJudge, selectHarnessVariant, selfPreference, sentenceReorderMutator, serializeFeedbackTrajectoriesJsonl, showMeasured, signManifest, spearmanR, splitGold, statusAdvanced, stopOnNoProgress, stopOnRepeatedAction, stringField, stripFencedJson, subjectiveEval, summarizeAgentReceiptIntegrity, summarizeBackendIntegrity, summarizeHarnessResults, summarizePrReviewBenchmark, summarizePreferenceMemory, summaryTable, supervisorRunRolloutLines, testJudge, textInSnapshot, throwIfRunIncomplete, toAgentProfileJson, toJsonl, toLangfuseEnvelope, toOpenAiTool, toPrometheusText, toRewardRows, toSftRows, tokenizeDomainWords, toolNamesForRun, toolSpans, traceAnalystFunctionGroup, traceAnalystOnRunComplete, traceContract, traceJudge, traceJudgeEnsemble, traceSpanKindToOpenInferenceKind, tracedAnalyzeTraces, typoMutator, urlContains, userQuestionsForKnowledgeGaps, validateAgentProfileCell, validatePolicyEdit, validatePolicyEditCandidateRecord, validateProductBenchmarkManifest, validateProductBenchmarkRecord, validateProductBenchmarkRun, validateRolloutLine, validateRunRecord, verbosityBias, verifyAgentProfileCell, verifyAttestation, verifyCompletion, verifyManifest, visualDiff, viteDeployRunner, vitestTestParser, weightedComposite, weightedMean, weightedRecall, welchsTTest, whitespaceCollapseMutator, wilcoxonSignedRank, wilson, withAssignedFeedbackSplit, withHeldoutBlend, withJudgeRetry, withOtelPipeline, wranglerDeployRunner, writeSupervisorRunReport };
+export { AGENT_PROFILE_KINDS, ATTESTATION_ALGORITHM, type ActionExecutionPolicy, type ActionPolicyDecision, type ActionableSideInfo, type ActiveLearningOptions, type AdapterRun, AgentDriver, type AgentDriverConfig, AgentEvalError, type AgentEvalErrorCode, type AgentInterfaceProfileLike, type AgentProfileCell, type AgentProfileCellInput, type AgentProfileCellSchemaVersion, AgentProfileCellValidationError, type AgentProfileDimensionValue, type AgentProfileHarness, type AgentProfileJson, type AgentProfileJsonObject, type AgentProfileKind, type AgentProfileRuntimeReceipt, type AgentProfileSource, type AgentProfileSourceInput, type AlignmentOp, type Analyst, type AnalystContext, type AnalystCost, type AnalystFinding, type AnalystHooks, type AnalystInputKind, AnalystRegistry, type AnalystRegistryOptions, type AnalystRequirements, type AnalystRunEvent, type AnalystRunInputs, type AnalystRunResult, type AnalystRunSummary, type AnalystSeverity, type AnalystUsageReceipt, type AnalyzeTracesInput, type AnalyzeTracesOptions, type AnalyzeTracesResult, type AnalyzeTracesTurnSnapshot, type AntiSlopConfig, type AntiSlopIssue, type AntiSlopReport, type Artifact$1 as Artifact, type ArtifactCheck, type Artifact as ArtifactCheckArtifact, type ArtifactEventLike, type ArtifactResult, type ArtifactValidator, type AsiSeverity, type AssertCapabilityHeadroomOptions, type AssertCrossFamilyOptions, type AssertSingleBackendOptions, type AttestationProvenance, type AttestationVerification, type AttestedReport, type AutoPrClient, AxGepaSteeringOptimizer, type AxSteeringOptimizerConfig, BENCHMARK_SPLIT_SEED, type BackendDescriptor, BackendIntegrityError, type BackendIntegrityReport, type BaselineOptions, type BaselineReport, BehaviorAssertion, type BehavioralMetrics, type BehavioralTokenSequence, type BenchmarkAdapter, type BenchmarkDatasetItem, type BenchmarkEvaluation, type BenchmarkFamily, type BenchmarkReport$1 as BenchmarkReport, type BenchmarkResponder, BenchmarkRunner, type BenchmarkRunnerConfig, type BenchmarkScenario, type BenchmarkSource, type BenchmarkTaskKind, type BisectOptions, type BisectResult, type BisectStep, type BlendWeights, type BootstrapOptions, type BootstrapResult, BudgetBreachError, BudgetGuard, type BudgetLedgerEntry, type BudgetPolicy, type BudgetSpec, CODING_HARNESSES, type CachedJudge, type CachedJudgeOptions, type CalibrationResult, CallExpectation, CallbackResearcher, type CallbackResearcherOptions, type CampaignFactoryParams, type CampaignIntegrityPolicy, type CampaignRunContext, type CampaignRunOutcome, type CampaignRunner, type CampaignScenario, type CampaignVariant, type CanaryAlert, type CanaryKind, type CanaryLeak, type CanaryOptions, type CanaryReport, type CanarySeverity, type CandidateComparison, type CandidateScenario, type CandidateScore, type CanonicalRawAnalystFinding, type CapabilityHeadroomOptions, type CapabilityHeadroomResult, type CaptureFetchContext, type CaptureFetchOptions, CaptureIntegrityError, type CausalAttributionReport, type CellVerdict, type ChannelRollup, type ChatCallOpts, type ChatClient, type ChatMessage, type ChatRequest, type ChatResponse, type ChatToolCall, type ChatTransport, type CheckResult, type CliBridgeTransportOpts, type CliffsMagnitude, type ClusterBootstrapInterval, type ClusterSignFlipAlternative, type ClusterSignFlipResult, type ClusteredBinaryCluster, type ClusteredMatchedPair, type ClusteredPairedBinaryOptions, type ClusteredPairedBinaryResult, type ClusteredPairedBinaryStatistics, type CollectedArtifacts, type CommandRunner, type ComparePairedArmsOptions, type CompletionCriterion, type CompletionRequirement, type CompletionVerdict, type ConceptComplexity, type ConceptFinding, type ConceptSpec, type ConceptWeightStrategy, ConfigError, type ContinuityCheck, type ContinuityCheckResult, type ContinuityReport, type ContinuitySnapshotPair, type ContinuousAgreement, type ContinuousAgreementOptions, type ContinuousCalibrationResult, type ContractCheckResult, type ContractJudgeOptions, type ContractMetric, type ContractReport, type ContractRule, type ContractRuleKind, type ContractSpan, type ContractVerdict, type ContractViolation, type ControlActionFailureMode, type ControlActionOutcome, type ControlBudget, type ControlContext, type ControlDecision, type ControlEvalResult, type ControlRunResult, type ControlRunToRunRecordOptions, type ControlRuntimeConfig, type ControlRuntimeError, type ControlSeverity, type ControlStep, type ControlStopPolicies, ConvergenceTracker, type CorpusAgreementOptions, type CorpusAgreementPerDimension, type CorpusAgreementReport, type CorpusScoreRecord, type CorrectnessChecker, CostAccountingIncompleteError, CostCallConflictError, CostCeilingReachedError, type CostChannel, type CostEntry, CostLedger, type CostLedgerEntry, type CostLedgerFilter, type CostLedgerHandle, type CostLedgerOptions, type CostLedgerPersistence, CostLedgerPersistenceError, type CostLedgerSummary, type CostReceipt, CostReceiptCaptureError, type CostReceiptInput, type CostReport, CostReservationExceededError, type CostResult, type CostSummary, CostTracker, type CostUsage, type CounterfactualContext, type CounterfactualMutation, type CounterfactualResult, type CounterfactualRunner, type CreateAnalystAiConfig, type CreateChatClientOpts, type CreateDefaultReviewerOptions, type CreateExperimentInput, type CreateSandboxPoolOpts, type CreateTraceAnalystKindOpts, CrossFamilyError, type CrossTraceDiff, type CrossTraceDiffOptions, type CustomTokenPricing, DEFAULT_AGENT_SLOS, DEFAULT_COMPLEXITY_WEIGHTS, DEFAULT_RULES as DEFAULT_FAILURE_RULES, DEFAULT_FINDERS, DEFAULT_HARNESS_OBJECTIVES, DEFAULT_MUTATION_PRIMITIVES, DEFAULT_MUTATORS, DEFAULT_PR_REVIEW_SCORE_WEIGHTS, DEFAULT_REDACTION_RULES, DEFAULT_RED_TEAM_CORPUS, DEFAULT_RUN_SCORE_WEIGHTS, DEFAULT_SEVERITY_WEIGHTS, DEFAULT_TRACE_ANALYST_BUDGETS, DEFAULT_TRACE_ANALYST_KINDS, type DataAcquisitionPlan, Dataset, type DatasetDifficulty, type DatasetManifest, type DatasetOverview, type DatasetProvenance, type DatasetScenario, type DatasetSplit, type DecideNextUserTurnOpts, type DefaultAnalystRegistryOptions, type DefaultVerdict, type DeployFamily, type DeployGateLayerInput, type DeployRunResult, type DeployRunner, type DescriptionLengthCandidate, type DescriptionLengthConfig, type DescriptionLengthDecision, type DescriptionLengthEvidence, DescriptionLengthGate, type DescriptionLengthRejectionCode, type DetectorEvent, type DetectorSeverity, type DetectorSignal, type DiffPolicy, type DiffScorecardOptions, type DirEntry, type DirectProviderTransportOpts, type Direction, type DiscoverPersonasOptions, type DiscoveredPersona, DockerSandboxDriver, type DriverResult, type DriverState, DualAgentBench, type DualAgentBenchConfig, type DualAgentReport, type DualAgentRound, type DualAgentScenario, type DualAgentScenarioResult, type EProcess, type EProcessOptions, type EProcessState, type EProcessStep, ERROR_COUNT_PATTERNS, type EnsembleAggregate, type EnsembleJudgeOptions, type ErrorCluster, type ErrorCountPattern, type ErrorStreakOptions, type EvalCampaignOptions, type EvalCampaignResult, type EvalResult, type EvalToolDef, EvalTraceStore, type EventFilter, type EventKind, type EvidenceRef, type EvolutionRound, type ExecutorConfig, type Expectation, type Experiment, type ExperimentPlan, type ExperimentProvenance, type ExperimentRep, type ExperimentResult, type ExperimentStats, type ExperimentStore, ExperimentTracker, type ExperimentTrackerOptions, type ExperimentVerdict, type ExportableSpan, type ExportedRewardModel, type ExtractOptions, type ExtractResult, type ExtractUsageFromSseOptions, type ExtractedUsage, FAILURE_CLASSES, FAILURE_MODE_KIND_SPEC, type FactorContribution, type FactorialCell, type FailedRun, type FailureClass, type FailureClassification, type FailureContext, type FailureMode, type FailureRule, type FeedbackArtifactType, type FeedbackAttempt, type FeedbackLabel, type FeedbackLabelKind, type FeedbackLabelSource, type FeedbackOptimizerRow, type FeedbackOutcome, type FeedbackPattern, type FeedbackReplayAdapter, type FeedbackReplayResult, type FeedbackSeverity, type FeedbackSplitPolicy, type FeedbackTask, type FeedbackTrajectory, type FeedbackTrajectoryFilter, type FeedbackTrajectoryStore, type FieldDestination, type FileChange, FileSystemFeedbackTrajectoryStore, FileSystemRawProviderSink, type FileSystemRawProviderSinkOptions, FileSystemTraceStore, type FileSystemTraceStoreOptions, type Finding, type FindingSubject, type FindingSubjectKind, type FindingsDiff, FindingsStore, type FlattenOtlpOptions, type FlowAction, type FlowLayerEnv, type FlowLayerFactoryInput, type FlowRunner, type FlowRunnerStepResult, type FlowSpec, type FlowStep, type GainDistributionBin, type GainDistributionFigureSpec, type GainDistributionOptions, type GateDecision$1 as GateDecision, type GateEvidence, type GenericSpan, type GhCliClientOptions, type GoldenItem, type GoldenSeverity, type GoldenSpec, HARNESS_NATIVE_MODEL, type HarnessAdapter, type HarnessConfig, type HarnessExperimentConfig, type HarnessExperimentResult, type HarnessIntervention, type HarnessRunRequest, type HarnessRunResult, type HarnessScenario, type HarnessSelection, type HarnessVariant, type HarnessVariantReport, type HeadroomClass, type HeadroomInput, HeldOutGate, type HeldOutGateConfig, type HeldOutGateRejectionCode, type HeldOutPartition, type HiddenCriteriaGrader, type HiddenGradeResult, type HiddenLeak, HoldoutAuditor, HoldoutLockedError, type HttpGithubClientOptions, type HypothesisManifest, type HypothesisResult, IMPROVEMENT_KIND_SPEC, INPUT_VALUE, INTENT_MATCH_JUDGE_VERSION, type ImageData, type ImprovementThresholds, type ImprovementVerdictResult, InMemoryFeedbackTrajectoryStore, InMemoryRawProviderSink, type InMemoryRawProviderSinkOptions, InMemoryTraceStore, InMemoryWorkspaceInspector, type InferenceScorer, type InspectorContext, type IntentMatchInput, type IntentMatchOptions, type IntentMatchResult, type InteractionContribution, type InterimReleaseConfidence, type InterimReleaseConfidenceInput, type JudgeConfig$1 as JudgeConfig, JudgeError, type JudgeFamily, type JudgeFleetOptions, type JudgeFn, type JudgeInput, JudgeParseError, type JudgeReplayGateArgs, type JudgeReplayResult, type JudgeRetryOutcome, type JudgeRetryPolicy, type JudgeRubric, JudgeRunner, type JudgeScore$1 as JudgeScore, type JudgeScoreInput, type JudgeScoresRecord, type JudgeSpan, type JudgeVerdict, KNOWLEDGE_GAP_KIND_SPEC, KNOWLEDGE_POISONING_KIND_SPEC, type KeywordConceptSpec, type KeywordCoverageFinding, type KeywordCoverageOptions, type KeywordCoverageResult, type KnowledgeAcquisitionMode, type KnowledgeBundle, type KnowledgeFallbackPolicy, type KnowledgeFreshness, type KnowledgeImportance, type KnowledgeReadinessReport, type KnowledgeRecommendedAction, type KnowledgeRequirement, type KnowledgeRequirementCategory, type KnowledgeResponsibleSurface, type KnowledgeSensitivity, LLM_CACHED_TOKENS, LLM_CACHED_TOKEN_ATTR_KEYS, LLM_CACHE_WRITE_TOKENS, LLM_CACHE_WRITE_TOKEN_ATTR_KEYS, LLM_CONTEXT_TOKENS, LLM_COST_ATTR_KEYS, LLM_COST_USD, LLM_INPUT_TOKENS, LLM_INPUT_TOKEN_ATTR_KEYS, LLM_MODEL_ATTR_KEYS, LLM_MODEL_NAME, LLM_OUTPUT_TOKENS, LLM_OUTPUT_TOKEN_ATTR_KEYS, LLM_REASONING_TOKENS, LLM_REASONING_TOKEN_ATTR_KEYS, type LangfuseEnvelope, type LangfuseGeneration, type LangfuseScore, type Layer, type LayerResult, type LayerStatus, type LeaderboardOptions, type LeaderboardRow, type LiveProofArtifact, type LiveProofConfig, type LiveProofContext, type LiveProofResult, LlmCallError, type LlmCallMetadata, type LlmCallRequest, type LlmCallResult, LlmClient, type LlmClientOptions, type LlmCorrectnessCheckerOpts, type LlmJsonCall, type LlmJudgeDimension, type LlmJudgeOptions, type LlmMessage, LlmResponseError, type LlmReviewerConfig, LlmRouteAssertionError, type LlmRouteRequirements, type LlmSpan, type LlmSpanOtlpInput, type LlmUsage, LockedJsonlAppender, MODEL_PRICING, type MakeEvalToolsConfig, type MatchResult, type MatchedPair, type MatcherResult, type MaximumCharge, type McNemarResult, type Measured, type MeasurementPolicy, type MergeOptions, type Message, type MetricSamples, type MetricVerdict, MetricsCollector, type MintRolloutOptions, type MintRolloutResult, type MockTransportOpts, type ModelCostRollup, type ModelPreflight, type ModelSeats, ModelsUnreachableError, type MuffledFinder, type MuffledFinding, MultiLayerVerifier, type MultiToolchainLayerConfig, type Mutator, Mutex, type NoLeakOptions, type NoProgressOptions, NoopRawProviderSink, NoopResearcher, NotFoundError, OPENINFERENCE_SPAN_KIND, OTEL_AGENT_EVAL_SCOPE, OUTPUT_VALUE, type Objective, type Oracle, type OracleObservation, type OracleReport, type OracleResult, type OrthogonalityInput, type OrthogonalityResult, type OtelExportConfig, type OtelExporter, type OtelPipelineHandle, type OtelPipelineOptions, type OtlpExport, OtlpFileTraceStore, type OtlpFileTraceStoreOptions, type OtlpFlatLine, type OtlpResourceSpans, type OtlpSpan, type OtlpToRunRecordsOptions, type OtlpTraceRunRecord, type PaidCallResult, type PairArmsOptions, type PairArmsResult, type PairedArmRow, type PairedArmsComparison, type PairedBootstrapOptions, type PairedBootstrapResult, type PairedCorrectness, type PairedEvalueOptions, type PairedEvalueSequence, type PairedEvalueStep, type PairedMetricDelta, type PairedSignTestResult, PairwiseSteeringOptimizer, type ParaphraseRobustnessScenarioInput, type ParaphraseRobustnessScenarioResult, type ParetoFigureSpec, type ParetoPoint, type ParetoResult, type PartitionHeldOutOptions, type PendingCostCall, type PendingCostCallView, type PersistedFinding, type PersonaConfig, type PersonaRigor, type Playbook, type PlaybookEntry, type PoolSlot, type PositionalBiasResult, type PrReviewAuditCase, type PrReviewBenchmarkSummary, type PrReviewComment, type PrReviewMatchedFinding, type PrReviewOutcome, type PrReviewReferenceFinding, type PrReviewScore, type PrReviewScoreWeights, type PrReviewSeverity, type PrReviewSource, type PreferenceMemoryEntry, type PreflightModelsOptions, type PreflightOutcome, type ProducedProposal, type ProducedState, type ProductBenchmarkArm, type ProductBenchmarkArtifactPaths, type ProductBenchmarkBudgets, type ProductBenchmarkExportOptions, type ProductBenchmarkExportResult, type ProductBenchmarkManifest, type ProductBenchmarkProfileRef, type ProductBenchmarkRecord, type ProductBenchmarkRepoRef, type ProductBenchmarkRunInput, type ProductBenchmarkScenario, type ProductBenchmarkSingleRunExportOptions, type ProductBenchmarkSplit, type ProductBenchmarkSubstrateVersions, type ProductBenchmarkValidationReport, ProductClient, type ProductClientConfig, type ProfileAxisSpec, type ProjectRuntimeTrajectoryEvidenceOptions, type ProjectedOtlpSpan, type PromptHandle, PromptRegistry, type ProportionInterval, type ProposalEventLike, type ProposeAutomatedPullRequestInput, type ProposeAutomatedPullRequestResult, type ProposeFn, type ProposeInput, type ProposeOutput, type ProposeReviewConfig, type ProposeReviewControlAction, type ProposeReviewControlConfig, type ProposeReviewControlResult, type ProposeReviewControlState, type ProposeReviewReport, type ProposeReviewShot, type ProposedSideEffect, type ProvenanceReader, type ProviderRedactor, type QueryTracesPage, REDACTION_VERSION, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, RESEARCH_REPORT_HARD_PAIR_FLOOR, ROLLOUT_FORMAT, ROLLOUT_SCHEMA, RUN_COST_ATTR_KEYS, type RawAnalystEvidence, type RawAnalystFinding, type RawProviderDirection, type RawProviderEvent, type RawProviderSink, type RawProviderSinkFilter, type RecordRunsOptions, type RedTeamCase, type RedTeamCategory, type RedTeamFinding, type RedTeamPayload, type RedTeamReport, type RedactionReport, type RedactionRule, type ReferenceEquivalenceJudgeInput, type ReferenceEquivalenceJudgeOptions, type ReferenceEquivalenceJudgeResult, type ReferenceEquivalenceScenario, type ReferenceMatchResult, type ReferenceReplayAdapter, type ReferenceReplayAdapterFn, type ReferenceReplayAdapterLike, type ReferenceReplayAggregate, type ReferenceReplayCandidate, type ReferenceReplayCase, type ReferenceReplayCaseRun, type ReferenceReplayExecutionScenario, type ReferenceReplayItem, type ReferenceReplayMatch, type ReferenceReplayMatchStrategy, type ReferenceReplayMatcher, type ReferenceReplayPromotionDecision, type ReferenceReplayPromotionPolicy, type ReferenceReplayRun, type ReferenceReplayRunContext, type ReferenceReplayRunOptions, type ReferenceReplayRunStore, type ReferenceReplayScenario, type ReferenceReplayScenarioScore, type ReferenceReplayScore, type ReferenceReplayScoreOptions, type ReferenceReplaySplit, type ReferenceReplaySplitComparison, type ReferenceReplaySteeringRowsOptions, type ReflectionContext, type ReflectionProposal, type RegistryRunOpts, type ReleaseConfidenceAxis, type ReleaseConfidenceAxisName, type ReleaseConfidenceInput, type ReleaseConfidenceIssue, type ReleaseConfidenceMetrics, type ReleaseConfidenceScorecard, type ReleaseConfidenceStatus, type ReleaseConfidenceThresholds, type ReleaseTraceEvidence, type RenderReleaseReportOptions, type RepeatedActionOptions, ReplayCache, type ReplayCacheEntry, ReplayCacheMissError, type ReplayCacheStats, ReplayError, type ReplayFetchOptions, type RepoRef, type RequirementCheck, type ResearchReport, type ResearchReportCandidate, type ResearchReportDecision, type ResearchReportMethodology, type ResearchReportOptions, type ResearchReportRecommendation, type Researcher, type RetrievalSpan, type Review, type ReviewFn, type ReviewInput, type ReviewMemoryEntry, type ReviewMemoryStore, type ReviewerMemoryEntry, type ReviewerOutput, type ReviewerPromptInput, type ReviewerSoftFailDefaults, type ReviewerVerificationSummary, type RewardRow, type RiskDifferenceResult, type RobustnessResult, type RolloutCapture, type RolloutLine, type RolloutRole, type RolloutScrubber, type RolloutSplit, type RolloutStep, type RouteMap, type RoutedField, type RouterTransportOpts, type RubricDimension, type Run, type RunCommandInput, type RunCommandResult, type RunCompleteHook, type RunCompleteHookContext, type RunCostProvenance, RunCritic, type RunCriticOptions, type RunEvidenceMetadata, type RunFilter, RunIntegrityError, type RunIntegrityExpectations, type RunIntegrityIssue, type RunIntegrityIssueCode, type RunIntegrityReport, type RunJudgeMetadata, type RunLayer, type RunOutcome, type RunPaidCallInput, type RunRecord, type RunRecordBackend, type RunRecordFilter, RunRecordValidationError, type RunScore, type RunScoreWeights, type RunSplitTag, type RunStatus, type RunTokenUsage, type RunTrace, type RuntimeEventLike, type RuntimeResolution, type RuntimeTrajectoryEvidenceProjection, type RuntimeTrajectoryEvidenceSummary, type RuntimeTrajectoryHookEvent, type RuntimeTrajectoryRecord, type RuntimeTrajectoryRunRecord, SEMANTIC_CONCEPT_JUDGE_VERSION, SKILL_USAGE_ANALYST, SPAN_KIND_ATTR_KEYS, SUPERVISOR_RUN_SCHEMA, type SandboxDriver, SandboxHarness, type SandboxHarnessResult, type SandboxJudgeKind, type SandboxJudgeResult, type SandboxJudgeSpec, type SandboxPool, type SandboxResult, type SandboxSdkTransportOpts, type SandboxSpan, type SatisfiedBy, type ScanOptions, type Scenario$1 as Scenario, type ScenarioCost, type ScenarioFile, ScenarioRegistry, type ScenarioResult, type ScoreKnowledgeReadinessOptions, type Scorecard, type ScorecardCell, type ScorecardCellDiff, type ScorecardDiff, type ScorecardEntry, type ScorecardLogLine, type ScoredTarget, type SearchSpanResult, type SearchTraceResult, type SeatName, type SeatPresetName, SeatUnsetError, type SelfPlayOptions, type SelfPlayProposer, type SelfPlayScorer, type SelfPreferenceResult, type SemanticConceptJudgeInput, type SemanticConceptJudgeOptions, type SemanticConceptJudgeResult, type SequentialDecision, type SerializedRegex, type SeriesConvergenceOptions, type SeriesConvergenceResult, type Severity, type SftExportOptions, type SftRow, type SignTestAlternative, type SignedManifest, type SignedManifestAlgo, type SingleBackendDivergence, SingleBackendError, type SingleBackendField, type SingleBackendReport, SkillUsageAnalyst, type SliceOptions, type Slo, type SloCheckResult, type SloComparator, type SloReport, type SloSeverity, type SlopCategory, type SlotFactory, type SourceLimits, type Span, type SpanBase, type SpanFilter, type SpanHandle, type SpanKind, type SpanMatchRecord, SpanNotFoundError, type SpanPredicate, type SpanStatus, type SseUsageMode, type SteeringBundle, type SteeringChange, type SteeringDelta, type SteeringOptimizationResult, type SteeringOptimizationRow, type SteeringOptimizationSelector, type SteeringOptimizerBackend, type SteeringOptimizerConfig, type SteeringRolePrompt, type StepAttribution, type StopDecision, type StreamingDetector, type SuboptimalCode, type SuboptimalSignal, SubprocessSandboxDriver, type SubprocessSandboxDriverOptions, type SummaryTable, type SummaryTableOptions, type SummaryTableRow, type SupervisorRunReader, type SupervisorRunReport, type SupervisorRunRollup, type SupervisorRunSources, type SupervisorRunTree, type SynthesisReason, type SynthesisTarget, TOOL_ARGS_CAPTURED, TOOL_LATENCY_MS, TOOL_NAME, TOOL_NAME_ATTR_KEYS, TRACE_ANALYST_ACTOR_DESCRIPTION, TRACE_ANALYST_ACTOR_DESCRIPTION_VERSION, TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, TRACE_SCHEMA_VERSION, type TaskGold, type TaskHeadroom, type TestGradedRunOptions, type TestGradedRunResult, type TestGradedScenario, type TestOutputParser, type TestResult, type TextMatcher, type ThresholdContract, TokenCounter, type TokenSpec, type ToolCallEventLike, type ToolDef, type ToolMatcher, type ToolSpan, type ToolSpanOtlpInput, type ToolStats, type ToolUseMetrics, type ToolUseOptions, type TraceAggregate, type TraceAnalysisStore, type TraceAnalystByteBudgets, type TraceAnalystFilters, type TraceAnalystGolden, type TraceAnalystHookOptions, type TraceAnalystKindSpec, type TraceAnalystSpan, type TraceAnalystSpanKind, type TraceAnalystSpanStatus, type TraceAnalystTraceSummary, type TraceContract, TraceContractBuilder, TraceEmitter, type TraceEmitterOptions, type TraceEvent, TraceFileMissingError, type TraceInsightContext, type TraceInsightFinding, type TraceInsightPanelRole, type TraceInsightPromptInput, type TraceInsightQualityGate, type TraceInsightQuestion, type TraceInsightReadiness, type TraceInsightSuite, type TraceInsightTask, TraceNotFoundError, type TraceStore, type TraceStoreSource, type TraceStoreToOtlpOptions, type TracedAnalystOptions, type TracedJudgeOptions, type TracesToOtlpResult, type Trajectory, type TrajectoryStep, type TreatmentClass, type TreatmentGate, type TreatmentGateInput, type TreatmentGateOptions, type TrialTrace, type Turn, type TurnMetrics, type TurnResult, UI_FINDING_SEVERITIES, UI_LENSES, UNIVERSAL_FINDERS, type UiFinding, type UiFindingScreenshot, type UiFindingSeverity, type UiLens, type Unavailable, type UserQuestion, type ValidationContext, ValidationError, type ValidationIssue, type ValidationResult, type VerbosityBiasResult, type Verdict, type VerdictCacheStats, type VerdictCacheStore, type Verification, VerificationError, type VerificationReport, type VerifyContext, type VerifyFn, type VerifyOptions, type ViewSpansResult, type ViewTraceOversized, type ViewTraceResult, type VisualDiffOptions, type VisualDiffResult, type ViteDeployRunnerInput, type WeightedCompositeInput, type WeightedCompositeResult, type WorkerDriverContext, type WorkflowTopology, type WorkspaceAssertion, type WorkspaceAssertionResult, type WorkspaceInspector, type WorkspaceSnapshot, type WranglerDeployRunnerInput, acquisitionPlansForKnowledgeGaps, adversarialJudge, agentProfileCellHashMaterial, agentProfileCellKey, agentProfileHash, agentProfileId, agentProfileModelId, agentVisibleFields, aggregateJudgeVerdicts, aggregateLlm, aggregatePrReviewScore, aggregateRunScore, allCriticalPassed, analyzeAntiSlop, analyzeSeries, analyzeSupervisorRun, analyzeSupervisorRunSources, analyzeTraces, appendScorecard, applyLlmSpanOtlpAttributes, applyToolSpanOtlpAttributes, argHash, asNumber, asString, assertCapabilityHeadroom, assertCrossFamily, assertLlmRoute, assertModelsServed, assertNoHiddenLeak, assertProductBenchmarkRun, assertRealAgentReceipts, assertRealBackend, assertReleaseConfidence, assertRolloutLine, assertRunAgentProfileCell, assertRunCaptured, assertSingleBackend, assignFeedbackSplit, assignHeldOutTag, attachCostToReport, attest, attributeCounterfactuals, backoffMs, deterministicSplit as benchmarkDeterministicSplit, index$1 as benchmarks, benjaminiHochberg, bisect, blendHeldout, blockingKnowledgeEval, bonferroni, bootstrapCi, buildAgentInterfaceProfileCell, buildAgentProfileCell, buildDefaultAnalystRegistry, buildDriverSystemPrompt, buildProductBenchmarkManifest, buildReflectionPrompt, buildReviewerPrompt, buildTraceAnalystTools, buildTraceInsightContext, buildTraceInsightPrompt, buildTrajectory, buildWorkerDriverSystemPrompt, byteLengthRange, cachedJudge, calibrateJudge, calibrateJudgeContinuous, callLlm, callLlmJson, canaryLeakView, canonicalJson, canonicalize, capabilityHeadroom, captureFetchToRawSink, causalAttribution, checkBehavioralCanary, checkCanaries, checkSlos, checkTraceContracts, clamp01, classifyFailure, classifyTreatment, claudeCodeSupervisorRunReader, cliffsDelta, clusteredPairedBinary, codeExecutionJudge, cohensD, coherenceJudge, collectionPreserved, commentsForSource, commitBisect, comparePairedArms, compareReferenceReplay, compareToBaseline, compilerJudge, completionVerdict, composeParsers, composeValidators, computeExperimentStats, computeFindingId, computeToolUseMetrics, computeTraceMetrics, confidenceInterval, containsAll, contentHash, contextInputTokens, continuousAgreement, contractJudge, controlFailureClassFromVerification, controlRunToFeedbackTrajectory, controlRunToRunRecord, convertTraceStoresToOtlp, corpusInterRaterAgreement, corpusInterRaterAgreementFromJudgeScores, costForTokenPricing, costForUsage, costReceiptFromLlm, costReceiptFromLlmError, costReport, createAnalystAi, createAntiSlopJudge, createChatClient, createCustomJudge, createDefaultReviewer, createDomainExpertJudge, createFeedbackTrajectory, createIntentMatchJudge, createLlmCorrectnessChecker, createLlmReviewer, createOtelExporter, createOtelTracingStore, createReferenceEquivalenceJudge, createReplayFetch, createSandboxPool, createSemanticConceptJudge, createTokenRecallChecker, createTraceAnalystKind, crossTraceDiff, crowdingDistance, dataDescriptionBits, decideNextUserTurn, decideReferenceReplayPromotion, decideReferenceReplayRunPromotion, defaultBlendWeights, defaultIsMaterial, defaultJudges, defaultProviderRedactor, defaultReferenceReplayMatcher, defaultTraceInsightPanel, deployGateLayer, describeTraceInsightScope, diffFindings, diffScorecard, discoverPersonas, distillPlaybook, domainEvidencePattern, dominates, eProcess, ensembleJudge, errorStreakDetector, estimateCost, estimateTokens, evaluateActionPolicy, evaluateContract, evaluateHypothesis, evaluateInterimReleaseConfidence, evaluateOracles, evaluateReleaseConfidence, evaluateTraceContract, executeScenario, expandProfileAxes, expectAgent, exportProductBenchmark, exportProductBenchmarkRuns, exportRewardModel, exportRunAsOtlp, extractAssetUrls, extractErrorCount, extractOtlpAttributes, extractProducedState, extractUsage, extractUsageFromResponse, extractUsageFromSse, feedbackTrajectoriesToDatasetScenarios, feedbackTrajectoriesToOptimizerRows, feedbackTrajectoryToDatasetScenario, feedbackTrajectoryToOptimizerRow, fileContains, fileExists, fileExperimentStore, fileVerdictCache, findAutoMatchNoExpectation, findConstructorCwdDropped, findFallbackToPass, findLiteralTruePass, findProductBenchmarkArtifacts, findSkipCountsAsPass, firstNumberAttr, firstStringAttr, flattenOtlpExportToNdjson, flowLayer, fnv1a32, formatBenchmarkReport, formatDriverReport, formatFindings, formatScorecardDiff, gainHistogram, gateTreatmentApplied, gateTreatmentFromMetrics, gateTreatmentFromSpans, gateTreatmentFromToolSpans, ghCliClient, gitProvenanceReader, precision as goldenPrecision, gradeOnHidden, gradeSemanticStatus, groupBy, groupRunsByAgentProfileCell, harnessAxisOf, hasCapturedToolArgs, hashContent, hashJson, hashScenarios, hashToUnit, hiddenGrade, holm, htmlContainsElement, httpGithubClient, improvementVerdict, inMemoryExperimentStore, inMemoryReferenceReplayStore, inMemoryReviewStore, inMemoryRunRecordBackend, inMemoryVerdictCache, inferDomainKeywords, inferOtlpKind, interRaterReliability, interpretCliffs, iqr, isHiddenDestination, isJudgeSpan, isLlmSpan, isModelPriced, isOtelConfigured, isRetrievalSpan, isRolloutLine, isRunRecord, isSandboxSpan, isToolSpan, isTrainableSplit, isTransientLlmError, isUnavailable, iterateRawCalls, jestTestParser, jsonHasKeys, jsonShape, jsonlReferenceReplayStore, jsonlReviewStore, jsonlRunRecordBackend, judgeFamily, judgeReplayGate, judgeSpans, keyPreserved, knowledgeReadinessTracePayload, leaderboard, linterJudge, llmJudge, llmSpanFromProvider, llmSpans, loadScorecard, loadScorerFromGrader, localCommandRunner, lowercaseMutator, makeEvalTools, makeFinding, mannWhitneyU, mapConcurrent, matchGoldens, matchSpan, maximumChargeForLlmRequest, mcnemar, mcnemarPower, mcnemarRequiredN, mergeLayerResults, mergeSteeringBundle, mintRolloutRows, modelDescriptionBits, modelHasSnapshot, modelPriceKey, mulberry32, multiToolchainLayer, noProgressDetector, normalizeScores, notBlocked, objectiveEval, observeAll, otelRunCompleteHook, otlpRowsToRunRecords, otlpRowsToTraceRunRecords, otlpToRunRecords, otlpToTraceRunRecords, pairArms, pairedBootstrap, pairedEvalueSequence, pairedMde, pairedRiskDifference, pairedSignTest, pairedTTest, paraphraseRobustness, paraphraseRobustnessScenarios, paretoChart, paretoFrontier, paretoFrontierWithCrowding, parseCorrectnessResponse, parseFeedbackTrajectoriesJsonl, parseReflectionResponse, parseRunRecordSafe, parseRuntimeTrajectoryHookEvent, partialCredit, partitionHeldOut, passAtK, passOrthogonality, pearsonR, pixelDeltaRatio, planTraceInsightQuestions, politenessPrefixMutator, positionalBias, preflightModels, printDriverSummary, probeLlm, productBenchmarkIntegrityFailures, productBenchmarkMutableSurfaces, productBenchmarkRepoIdentity, productBenchmarkSplits, index as profile, projectOtlpFlatLine, projectRuntimeTrajectoryEvidence, promptBisect, proposeSynthesisTargets, providerFromBaseUrl, pytestTestParser, ranks, readClaudeCodeSupervisorRun, readOtlpStatus, readProductBenchmarkManifest, readProductBenchmarkRecords, recordRuns, recordRunsToScorecard, redTeamDataset, redTeamReport, redactString, redactValue, referenceReplayRunsToSteeringRows, referenceReplayScenarioToRunScore, regexMatch, regexMatches, renderMarkdownReport, renderPlaybookMarkdown, renderPreferenceMemoryMarkdown, renderPriorFindings, renderReleaseReport, renderSteeringText, renderSupervisorRunHeadline, renderSupervisorRunMarkdown, renderUpstreamFindings, repeatedActionDetector, replayFeedbackTrajectories, replayFeedbackTrajectory, replayScorerOverCorpus, replayTraceThroughJudge, requireAgentProfileCell, requiredSampleSize, researchReport, resolveModelPricing, resolveRunCostProvenance, resolveSeat, rolloutReward, rollupSupervisorRuns, roundTripRunRecord, routeFields, rowCount, rowWhere, runAgentControlLoop, runAssertions, runBehavioralCanaries, runCanaries, runCounterfactual, runE2EWorkflow, runEvalCampaign, runExpectations, runFailureClass, runHarnessExperiment, runIntentMatchJudge, runJudgeFleet, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runLiveProof, runProposeReview, runProposeReviewAsControlLoop, runRecordToProductBenchmarkRecord, runReferenceEquivalenceJudge, runReferenceReplay, runScore, runSelfPlay, runSemanticConceptJudge, runTestGradedScenario, runsForScenario, scalarScore, scanForMuffledGates, scoreContinuity, scoreFromEvals, scoreKnowledgeReadiness, scorePrReviewComments, scorePrReviewSource, scoreRedTeamOutput, scoreReferenceReplay, scoreTraceInsightReadiness, seatPresets, securityJudge, selectHarnessVariant, selfPreference, sentenceReorderMutator, serializeFeedbackTrajectoriesJsonl, showMeasured, signManifest, spearmanR, statusAdvanced, stopOnNoProgress, stopOnRepeatedAction, stringField, stripFencedJson, subjectiveEval, summarizeAgentReceiptIntegrity, summarizeBackendIntegrity, summarizeHarnessResults, summarizePrReviewBenchmark, summarizePreferenceMemory, summaryTable, supervisorRunRolloutLines, testJudge, textInSnapshot, throwIfRunIncomplete, toAgentProfileJson, toJsonl, toLangfuseEnvelope, toOpenAiTool, toPrometheusText, toRewardRows, toSftRows, tokenizeDomainWords, toolNamesForRun, toolSpans, traceAnalystFunctionGroup, traceAnalystOnRunComplete, traceContract, traceJudge, traceJudgeEnsemble, traceSpanKindToOpenInferenceKind, tracedAnalyzeTraces, typoMutator, urlContains, userQuestionsForKnowledgeGaps, validateAgentProfileCell, validateProductBenchmarkManifest, validateProductBenchmarkRecord, validateProductBenchmarkRun, validateRolloutLine, validateRunRecord, verbosityBias, verifyAgentProfileCell, verifyAttestation, verifyCompletion, verifyManifest, visualDiff, viteDeployRunner, vitestTestParser, weightedComposite, weightedMean, weightedRecall, welchsTTest, whitespaceCollapseMutator, wilcoxonSignedRank, wilson, withAssignedFeedbackSplit, withHeldoutBlend, withJudgeRetry, withOtelPipeline, wranglerDeployRunner, writeSupervisorRunReport };