@tangle-network/agent-eval 0.123.8 → 0.124.0

package/dist/index.d.ts

@@ -1512,6 +1512,7 @@ interface LlmMessage {
         };
     }>;
 }
+type LlmThinkingMode = 'enabled' | 'disabled';
 interface LlmCallRequest {
     model: string;
     messages: LlmMessage[];
@@ -1524,6 +1525,8 @@ interface LlmCallRequest {
     };
     temperature?: number;
     maxTokens?: number;
+    /** OpenAI-compatible reasoning mode. Omitted when the provider default should apply. */
+    thinking?: LlmThinkingMode;
     /** Per-call timeout, default 300s. */
     timeoutMs?: number;
 }
@@ -1531,7 +1534,7 @@ interface LlmCallRequest {
  * Returns undefined when output or multimodal input is not bounded, causing a
  * capped CostLedger to reject the call before execution. Pass
  * `customTokenPricing` when package pricing does not cover the model or endpoint. */
-declare function maximumChargeForLlmRequest(request: Pick<LlmCallRequest, 'model' | 'messages' | 'jsonSchema' | 'maxTokens'>, options?: LlmClientOptions): MaximumCharge | undefined;
+declare function maximumChargeForLlmRequest(request: Pick<LlmCallRequest, 'model' | 'messages' | 'jsonSchema' | 'maxTokens' | 'thinking'>, options?: LlmClientOptions): MaximumCharge | undefined;
 interface LlmUsage {
     promptTokens: number;
     completionTokens: number;
@@ -1642,6 +1645,8 @@ interface LlmClientOptions {
      * Default: `extract`.
      */
     jsonPayloadMode?: 'extract' | 'exact';
+    /** Default provider reasoning mode. A per-call request value takes precedence. */
+    thinking?: LlmThinkingMode;
     /** Fetch implementation — defaults to global `fetch`. Override for custom transport (e.g. tests). */
     fetch?: typeof fetch;
     /**
@@ -6147,30 +6152,82 @@ declare function formatDriverReport(results: DriverResult[]): string;
 declare function printDriverSummary(results: DriverResult[]): void;
 
 /**
- * Rollout export — `tangle.rollout.v1` training rows minted from the
- * records the substrate ALREADY keeps. There is no separate rollout
- * store: a rollout is the JOIN of a RunRecord (identity, provenance,
- * cost, outcome) with its trace (spans share `runId`), optionally
- * annotated with feedback labels. This module performs that join and
- * serializes it to fine-tune-ready formats.
- *
- * Composition, not duplication:
- *   - identity/provenance     → `RunRecord` (candidateId, splitTag, agentProfile, hashes)
- *   - step structure          → `buildTrajectory` over the shared TraceStore
- *   - preference-pair export  → `feedbackTrajectoryToOptimizerRow` (feedback-trajectory.ts)
- *   - PRM / reward-model      → `reward-model-export.ts`
- *   This module adds only the joined row + SFT/reward serialization,
- *   which none of the above produce.
- *
- * Anti-Goodhart invariant: a run whose `outcome.realness.gated` is true
- * is never exported with a positive reward — the gate travels into the
- * training data, so a fine-tune cannot learn from gamed successes.
+ * `tangle.rollout.v1` — THE canonical rollout serialization, owned by
+ * agent-eval. One JSONL line per agent invocation (a solo eval run, a
+ * supervisor episode, a worker session, a proposer shot, a judge call, an
+ * analyst pass), labeled with its task/split coordinates and a single
+ * scalar reward, carrying the FULL message transcript inline.
+ *
+ * This schema is the reconciliation of two prior producers:
+ *   - agent-eval's RunRecord-joined rollout rows (PR #410): identity,
+ *     provenance hashes, the realness gate travelling into the reward,
+ *     trace-derived steps.
+ *   - the bench rollout-ledger (agent-runtime PR #591): the wire shape —
+ *     role, task.split/rep, parent_rollout_id, policy provenance, capture
+ *     provenance, inline canonical chat-with-tools messages.
+ * Where the two conflicted, RunRecord-derived semantics won; the wire
+ * field names follow the ledger (snake_case). See `docs/rollout.md` for
+ * the field-by-field decision table.
+ *
+ * Messages are inlined — never referenced — because every harness store a
+ * rollout can be recovered from is mutable or garbage-collected. A line
+ * must stay a complete training/eval example on its own.
+ *
+ * `outcome.reward` is THE single scalar (null = no verdict exists — a
+ * labeled gap, never 0). `outcome.realness_gated` is the anti-Goodhart
+ * flag: a gated line must never export as a positive training example.
+ */
+declare const ROLLOUT_SCHEMA = "tangle.rollout.v1";
+/** @deprecated alias kept for consumers of the pre-unification constant name. */
+declare const ROLLOUT_FORMAT = "tangle.rollout.v1";
+/** `agent` = a solo evaluation run (no multi-agent topology). */
+type RolloutRole = 'agent' | 'supervisor' | 'worker' | 'proposer' | 'judge' | 'analyst';
+/**
+ * Split vocabulary follows `RunRecord.splitTag` ('search' is the pool the
+ * optimizer may read — the trainable split), extended with the ledger's
+ * 'canary'. 'train' is a legacy alias for 'search' emitted by
+ * pre-unification ledgers; it validates and counts as trainable, but new
+ * producers must emit 'search'.
+ */
+type RolloutSplit = 'search' | 'dev' | 'holdout' | 'canary' | 'train';
+declare function isTrainableSplit(split: RolloutSplit): boolean;
+/** 'mint' = joined live from RunRecord + trace by `mintRolloutRows`. */
+type RolloutCapture = 'mint' | 'settle-time' | 'backfill';
+type ChatRole = 'system' | 'user' | 'assistant' | 'tool';
+interface ChatToolCall {
+    id: string;
+    type: 'function';
+    function: {
+        name: string;
+        /** JSON-encoded argument object, exactly as the model emitted it. */
+        arguments: string;
+    };
+}
+interface ChatMessage {
+    role: ChatRole;
+    content: string | null;
+    /** Reasoning/thinking channel where the harness captured it (full fidelity). */
+    reasoning_content?: string;
+    tool_calls?: ChatToolCall[];
+    /** Required on role:"tool" — the ChatToolCall this result answers. */
+    tool_call_id?: string;
+    name?: string;
+}
+interface ToolDef {
+    type: 'function';
+    function: {
+        name: string;
+        description?: string;
+        parameters?: Record<string, unknown>;
+    };
+}
+/**
+ * Compact trace-span projection (llm/tool step) carried alongside the
+ * conversation when the line was minted from a trace. Optional: lines
+ * recovered from harness stores have no span structure.
  */
-
-declare const ROLLOUT_FORMAT: "tangle.rollout.v1";
-/** Compact, serialization-safe projection of one span for training rows. */
 interface RolloutStep {
-    kind: Span['kind'];
+    kind: string;
     name: string;
     /** llm: last-message summary · tool: stringified args. Scrubbed. */
     input?: string;
@@ -6179,87 +6236,217 @@ interface RolloutStep {
     status?: 'ok' | 'error';
     durationMs?: number;
 }
-interface RolloutRow {
-    format: typeof ROLLOUT_FORMAT;
-    runId: string;
-    experimentId: string;
-    candidateId: string;
-    scenarioId?: string;
-    splitTag: RunRecord['splitTag'];
-    model: string;
-    /** Canonical profile-cell identity when the record carries one. */
-    agentProfileCellId?: string;
-    /** holdoutScore ?? searchScore, forced to 0 when realness-gated. */
-    reward: number;
-    /** True when `outcome.realness.gated` — excluded from SFT positives. */
-    realnessGated: boolean;
-    costUsd: number;
-    totalTokens: number;
-    steps: RolloutStep[];
-    /** Full message history of the final llm span — the SFT conversation. */
-    conversation: Message[];
-}
-/** Redactor applied to every exported string (secrets, PII). Identity by default. */
-type RolloutScrubber = (text: string) => string;
-interface MintRolloutOptions {
-    scrub?: RolloutScrubber;
-    /** Cap steps per row (longest runs first drop middle steps). Default: no cap. */
-    maxSteps?: number;
-}
-interface MintRolloutResult {
-    rows: RolloutRow[];
-    /** runIds that had a RunRecord but no spans — surfaced, never silently dropped. */
-    missingTraces: string[];
+interface RolloutTask {
+    /** Benchmark/suite id (e.g. "swe-bench-verified") or the experiment id. */
+    suite: string;
+    instance_id: string;
+    split: RolloutSplit;
+    /** Sampling seed the campaign pinned; null = not recorded. */
+    seed: number | null;
+    /** Replicate index (0-based). */
+    rep: number;
 }
-declare function rolloutReward(record: RunRecord): {
-    reward: number;
-    gated: boolean;
-};
-/**
- * Join RunRecords with their traces into rollout rows. Records without
- * spans are reported in `missingTraces` — a capture gap is a finding,
- * not a silent omission.
+interface RolloutPolicy {
+    /** Harness that drove the invocation (e.g. "opencode", "claude", "pi-loops"). */
+    harness: string | null;
+    harness_version: string | null;
+    model: string | null;
+    provider: string | null;
+    /** Commit of the agent profile / candidate under evaluation. */
+    profile_commit: string | null;
+    /** sha256 of the effective prompt (post-steering), when recorded. */
+    prompt_hash?: string | null;
+    /** sha256 of the effective run config, when recorded. */
+    config_hash?: string | null;
+    /** Canonical agent-profile cell identity, when the run carries one. */
+    agent_profile_cell_id?: string | null;
+    /** Sampling params (temperature, top_p, max_tokens…); null = not recorded. */
+    sampling: Record<string, unknown> | null;
+}
+interface RolloutOutcome {
+    /**
+     * THE single scalar training signal — the official verdict.
+     * null = no verdict exists for this invocation (a labeled gap, never 0).
+     */
+    reward: number | null;
+    /** Where the reward came from (judge id; "/inherited" = parent episode's). */
+    reward_source: string | null;
+    /** Raw judge verdict record, verbatim. */
+    verdict: unknown;
+    /** Everything that is NOT the scalar reward. */
+    metrics: Record<string, unknown>;
+    is_completed: boolean;
+    is_truncated: boolean;
+    error: string | null;
+    /**
+     * Anti-Goodhart flag from `RunRecord.outcome.realness.gated`: the run
+     * faked its success signal. Reward is forced to 0 at mint time and the
+     * line never qualifies for SFT. Optional on the wire (absent = false)
+     * so pre-unification ledgers stay readable.
+     */
+    realness_gated?: boolean;
+}
+interface RolloutCostBlock {
+    usd: number | null;
+    tokens_in: number | null;
+    tokens_out: number | null;
+    tokens_reasoning: number | null;
+    cache_read: number | null;
+    cache_write: number | null;
+    wall_s: number | null;
+}
+interface RolloutArtifacts {
+    patch_path: string | null;
+    run_dir: string | null;
+    /** Source-of-truth transcript pointer (session id / jsonl path) for audit. */
+    transcript_ref: string | null;
+}
+interface RolloutProvenance {
+    captured_at: string;
+    capture: RolloutCapture;
+    /** Present on gap lines: why `messages` could not be recovered. */
+    gap?: string;
+}
+interface RolloutLine {
+    schema: typeof ROLLOUT_SCHEMA;
+    rollout_id: string;
+    /** Spawning invocation within the same episode (worker → supervisor). */
+    parent_rollout_id: string | null;
+    run_id: string;
+    /** Logical experiment grouping from `RunRecord.experimentId`. Optional on
+     *  the wire (pre-unification ledgers lack it); null = not recorded. */
+    experiment_id?: string | null;
+    /** Stable candidate identity from `RunRecord.candidateId`; null = not recorded. */
+    candidate_id?: string | null;
+    /** Improvement-loop generation (-1 = baseline); null = not an improvement loop. */
+    generation: number | null;
+    /** Improvement-loop candidate index (-1 = baseline); null = not an improvement loop. */
+    candidate_index: number | null;
+    role: RolloutRole;
+    task: RolloutTask;
+    policy: RolloutPolicy;
+    /** Full transcript, inline. [] = gap line (see provenance.gap). */
+    messages: ChatMessage[];
+    tool_defs: ToolDef[];
+    /** Trace-span projections, when minted from a trace. */
+    steps?: RolloutStep[];
+    outcome: RolloutOutcome;
+    cost: RolloutCostBlock;
+    artifacts: RolloutArtifacts;
+    provenance: RolloutProvenance;
+}
+declare function validateRolloutLine(value: unknown): string[];
+declare function assertRolloutLine(value: unknown, context?: string): asserts value is RolloutLine;
+declare function isRolloutLine(value: unknown): value is RolloutLine;
+
+/**
+ * Pure exporters over `tangle.rollout.v1` lines → the training-data shapes
+ * the improvement loops feed:
+ *   - SFT chat JSONL           (clean trainable successes, {messages, metadata})
+ *   - reward rows              (every scored line, success or failure, with steps)
+ *   - Prime Intellect verifiers RolloutOutput (prompt/completion split + reward)
+ *   - OpenAI RFT items         (prompt turns + verdict reference fields)
+ *
+ * All exporters are pure functions of the lines — filtering (never train on
+ * holdout, reward thresholds, the realness gate) happens HERE, on inline
+ * labels, no joins.
  */
-declare function mintRolloutRows(records: RunRecord[], store: TraceStore, options?: MintRolloutOptions): Promise<MintRolloutResult>;
+
 interface SftExportOptions {
-    /** Export only rows with reward ≥ this (default 1 = clean successes only). */
+    /** Export only lines with reward ≥ this (default 1 = clean successes only). */
     minReward?: number;
 }
 interface SftRow {
-    messages: Message[];
+    messages: ChatMessage[];
     metadata: {
-        runId: string;
-        candidateId: string;
-        scenarioId?: string;
+        rollout_id: string;
+        run_id: string;
+        candidate_id: string | null;
+        instance_id: string;
         reward: number;
     };
 }
 /**
- * Supervised fine-tune rows: the completed conversation of each
- * qualifying rollout. Realness-gated rows never qualify (reward forced
- * to 0 at mint time).
+ * Supervised fine-tune rows: the completed conversation of each qualifying
+ * line. Fail-closed filters: trainable split only (never holdout/canary),
+ * reward ≥ minReward, realness-gated lines never qualify, gap lines carry
+ * no trainable content.
  */
-declare function toSftRows(rows: RolloutRow[], options?: SftExportOptions): SftRow[];
+declare function toSftRows(lines: RolloutLine[], options?: SftExportOptions): SftRow[];
 interface RewardRow {
+    /** First user turn — the task prompt. */
     prompt: string;
     steps: RolloutStep[];
     reward: number;
     metadata: {
-        runId: string;
-        candidateId: string;
-        scenarioId?: string;
-        splitTag: RunRecord['splitTag'];
+        rollout_id: string;
+        run_id: string;
+        candidate_id: string | null;
+        instance_id: string;
+        split: RolloutSplit;
     };
 }
 /**
- * Reward-labeled rows (RFT / verifiers-style): every rollout, success or
- * failure, with its scalar reward. Failures are signal here — only the
- * realness-gate zeroing is applied, never filtering.
+ * Reward-labeled rows: every line with a scalar reward, success or
+ * failure. Failures are signal here — only the realness-gate zeroing
+ * (applied at mint time) touches the reward, never filtering. Lines with
+ * no verdict (reward null) are excluded: an unlabeled example is a gap,
+ * not a zero.
  */
-declare function toRewardRows(rows: RolloutRow[]): RewardRow[];
-/** One JSON object per line — the interchange format for every export. */
+declare function toRewardRows(lines: RolloutLine[]): RewardRow[];
 declare function toJsonl(rows: ReadonlyArray<unknown>): string;
 
+/**
+ * Rollout minting — `tangle.rollout.v1` lines joined from the records the
+ * substrate ALREADY keeps. There is no separate rollout store: a rollout
+ * is the JOIN of a RunRecord (identity, provenance, cost, outcome) with
+ * its trace (spans share `runId`), projected into the canonical line.
+ *
+ * Composition, not duplication:
+ *   - identity/provenance     → `RunRecord` (candidateId, splitTag, agentProfile, hashes)
+ *   - step structure          → `buildTrajectory` over the shared TraceStore
+ *   - preference-pair export  → `feedbackTrajectoryToOptimizerRow` (feedback-trajectory.ts)
+ *   - PRM / reward-model      → `reward-model-export.ts`
+ *
+ * Anti-Goodhart invariant: a run whose `outcome.realness.gated` is true
+ * is never exported with a positive reward — the gate travels into the
+ * training data (`reward` forced to 0, `realness_gated: true`), so a
+ * fine-tune cannot learn from gamed successes.
+ *
+ * Records without spans become labeled GAP LINES (messages: [],
+ * provenance.gap) — present in the output AND surfaced in
+ * `missingTraces`; a capture gap is a finding, never a silent omission.
+ */
+
+/** Redactor applied to every exported string (secrets, PII). Identity by default. */
+type RolloutScrubber = (text: string) => string;
+interface MintRolloutOptions {
+    scrub?: RolloutScrubber;
+    /** Cap steps per line (longest runs first drop middle steps). Default: no cap. */
+    maxSteps?: number;
+    /** Role recorded on every minted line. Default 'agent' (a solo eval run). */
+    role?: RolloutRole;
+    /** Task suite label. Default: the record's `experimentId`. */
+    suite?: string;
+    /** Injected clock for deterministic output. */
+    now?: () => Date;
+}
+interface MintRolloutResult {
+    rows: RolloutLine[];
+    /** runIds that had a RunRecord but no spans — emitted as gap lines AND listed here. */
+    missingTraces: string[];
+}
+declare function rolloutReward(record: RunRecord): {
+    reward: number;
+    gated: boolean;
+};
+/**
+ * Join RunRecords with their traces into canonical rollout lines. Records
+ * without spans are emitted as labeled gap lines and reported in
+ * `missingTraces` — a capture gap is a finding, not a silent omission.
+ */
+declare function mintRolloutRows(records: RunRecord[], store: TraceStore, options?: MintRolloutOptions): Promise<MintRolloutResult>;
+
 interface RunEvidenceMetadata {
     experimentId: string;
     candidateId: string;
@@ -6290,6 +6477,436 @@ interface ControlRunToRunRecordOptions extends RunEvidenceMetadata {
 declare function controlRunToRunRecord<TState, TAction, TActionResult, TEval extends ControlEvalResult = ControlEvalResult>(run: ControlRunResult<TState, TAction, TActionResult, TEval>, options: ControlRunToRunRecordOptions): RunRecord;
 declare function scoreFromEvals(evals: readonly ControlEvalResult[]): number | undefined;
 
+/**
+ * Supervisor-run analysis — the multi-agent analogue of single-rollout trace
+ * analysis. A solo rollout is one invocation with a transcript; a supervisor
+ * run is a TREE of invocations (a brain that spawns, steers, and settles
+ * workers) plus the event timeline that connects them. `src/trace-analyst`
+ * answers "what happened inside one session"; this module answers "what did
+ * the tree do" — did the brain steer anyone mid-task, how many spawn waves,
+ * how concurrent, how idle, what did each role cost, what came back.
+ *
+ * The nodes of that tree are NOT a new shape: they are `tangle.rollout.v1`
+ * rows (`src/rollout`), keyed by `parent_rollout_id`, with `role` already
+ * spanning `supervisor` / `worker`. `supervisorRunRolloutLines` mints them.
+ * What rollout rows deliberately do NOT carry is the inter-invocation event
+ * timeline (spawn/settle/steer instants), which is what every structural
+ * metric here is computed from — so the reader consumes the journal event
+ * stream and emits rollout rows, rather than maintaining a parallel node type.
+ *
+ * ## UNAVAILABLE ≠ ZERO
+ *
+ * Every metric whose backing artifact can be missing is typed
+ * `Measured<T> = T | { unavailable: reason }`. A supervisor that steered
+ * nobody reports `steers: 0`; a supervisor whose worker logs were never
+ * written reports `steers: unavailable — <reason>`. The two have driven
+ * opposite conclusions about the same architecture, so they never collapse.
+ */
+
+/** A metric that could not be computed, with the reason its artifact was missing. */
+interface Unavailable {
+    readonly unavailable: string;
+}
+/** A metric value, or the reason it is unknown. NEVER collapse `unavailable` to 0. */
+type Measured<T> = T | Unavailable;
+declare function isUnavailable(v: unknown): v is Unavailable;
+/** Render a measured scalar for the markdown/headline: `0` and `unavailable` stay distinct. */
+declare function showMeasured(v: Measured<number | string | boolean | null>): string;
+/** One worker's logs, as read. `null` = the artifact did not exist. */
+interface WorkerLogSource {
+    readonly label: string;
+    /** Worker event stream — started / progress / finished / message events (JSONL). */
+    readonly events: string | null;
+    /** The durable steer queue — one line per steer request (JSONL). */
+    readonly inbox: string | null;
+    /** Worker patch byte length, or null when absent. */
+    readonly patchBytes: number | null;
+}
+/**
+ * Everything the pure analyzer reads — already-read bytes, never paths. Each
+ * field is `null` when its artifact was absent, which is what turns the
+ * dependent metrics into `unavailable` rather than 0.
+ *
+ * This is the whole input contract. Any store that can produce these strings
+ * (an on-disk loops run, an object-store archive, a database, a test fixture)
+ * is a valid source; `loopsSupervisorRunReader` is ONE implementation.
+ */
+interface SupervisorRunSources {
+    /** Stable identity of the run being analyzed (a directory, a run id, a URL). */
+    readonly runRef: string;
+    readonly instanceId: string | null;
+    /** Which arm/variant of a comparison this run is, when the run belongs to one. */
+    readonly arm: string | null;
+    /** Identity of the supervision-tree store this was read from; null = none found. */
+    readonly supRunDir: string | null;
+    /** Supervision journal — spawned / settled / cancelled / metered events (JSONL). */
+    readonly journal: string | null;
+    /** Per-brain-call tap (JSONL): finish_reason, completion tokens, requested max tokens. */
+    readonly brainLog: string | null;
+    /** Supervisor state document (JSON). */
+    readonly state: string | null;
+    /** Supervisor progress stream (JSONL). */
+    readonly progress: string | null;
+    /** Per-worker logs; `null` = the worker log store itself was missing. */
+    readonly workers: readonly WorkerLogSource[] | null;
+    /** Why `workers` is null (only set when it is). */
+    readonly workersMissingReason: string | null;
+    /** Run result document (JSON). */
+    readonly result: string | null;
+    /**
+     * Judge verdict document (JSON), or the matching ledger row re-encoded as
+     * one. Runners that write the verdict straight to a ledger leave no judge
+     * document, so the ledger row is the same fact from the same run — not a
+     * substitute measurement.
+     */
+    readonly judge: string | null;
+    /** Where `judge` came from, for the report's provenance line. */
+    readonly judgeSource: string | null;
+    /** Delivered unified-diff patch text. */
+    readonly patch: string | null;
+    /** Outer-driver log (used for the driver's steer verbs + deadline evidence). */
+    readonly driverLog: string | null;
+    /**
+     * Worker tokens recovered from a harness session store; null = store unavailable.
+     * `store` names the store in the report's provenance line (e.g. `opencode`).
+     */
+    readonly harnessWorkerTokens: {
+        store: string;
+        sessions: number;
+        input: number;
+        output: number;
+    } | null;
+    readonly harnessMissingReason: string | null;
+}
+/**
+ * A source of supervisor-run bytes. Implementations own their storage layout;
+ * the analyzer only ever sees `SupervisorRunSources`.
+ */
+interface SupervisorRunReader {
+    /** Stable identity of what this reader points at (for logs and report labels). */
+    readonly runRef: string;
+    read(): Promise<SupervisorRunSources>;
+}
+declare const SUPERVISOR_RUN_SCHEMA = "tangle.supervisor-run@1";
+declare const SUPERVISOR_RUN_ROLLUP_SCHEMA = "tangle.supervisor-run-rollup@1";
+interface SteerBreakdown {
+    readonly worker: string;
+    /** Steer requests durably queued to this worker's inbox. */
+    readonly queued: number;
+    /** Steers the worker's executor actually accepted (control event `delivered:true`). */
+    readonly delivered: number;
+}
+interface OrchestrationMetrics {
+    readonly workersSpawned: Measured<number>;
+    readonly workersSettled: Measured<number>;
+    readonly workersCancelled: Measured<number>;
+    /** THE HEADLINE: mid-task steers the brain sent to live workers. 0 ≠ unavailable. */
+    readonly steers: Measured<number>;
+    readonly steersDelivered: Measured<number>;
+    readonly steersByWorker: Measured<readonly SteerBreakdown[]>;
+    /** Outer-driver `supervisor_steer` tool calls seen in the driver log (a second steer path). */
+    readonly driverSteerCalls: Measured<number>;
+    /**
+     * Spawn waves. A wave is a maximal run of worker spawns with no settle/cancel between
+     * them: wave N+1 begins at the first spawn issued after at least one worker from an
+     * earlier wave has settled. Structural, not a time threshold — no tunable constant.
+     */
+    readonly waves: Measured<number>;
+    readonly waveSizes: Measured<readonly number[]>;
+    readonly maxConcurrency: Measured<number>;
+    /** Worker spawns issued after the first settlement — the retry/respawn tail. */
+    readonly respawns: Measured<number>;
+    /** Labels spawned more than once (a literal retry of the same subtask). */
+    readonly repeatedLabels: Measured<readonly string[]>;
+    /** Longest parent chain below the root, in worker hops. */
+    readonly delegationDepth: Measured<number>;
+    readonly timeToFirstSpawnMs: Measured<number>;
+    readonly supervisorWallMs: Measured<number>;
+    /** Wall time inside the supervisor run with ZERO live workers. */
+    readonly idleMs: Measured<number>;
+    readonly idlePct: Measured<number>;
+    /** sum(worker wall) / supervisor wall. >1 means real parallelism. */
+    readonly workerUtilization: Measured<number>;
+}
+interface DecisionMetrics {
+    readonly settledByStatus: Measured<Record<string, number>>;
+    readonly settledVerdicts: Measured<Record<string, number>>;
+    /** Worker verified its own work green AND produced a patch. */
+    readonly accepted: Measured<number>;
+    /** Worker settled with a failing verify. */
+    readonly rejected: Measured<number>;
+    /** Worker verified green but delivered no patch bytes — output with nothing to accept. */
+    readonly emptyPass: Measured<number>;
+    /** Settlements the brain observed before issuing its next spawn (evidence→respawn). */
+    readonly observeThenRespawn: Measured<number>;
+    /** Respawns with no settled evidence in front of them. */
+    readonly respawnWithoutEvidence: Measured<number>;
+    /** Steer + question traffic on the live down/up legs — the only "review while running" signal. */
+    readonly reviewActions: Measured<number>;
+    readonly workerEvidenceBytes: Measured<number>;
+}
+interface RoleSpend {
+    readonly tokensIn: Measured<number>;
+    readonly tokensOut: 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;
+    readonly patchBytes: number | null;
+    readonly passed: boolean | null;
+}
+interface WallDistribution {
+    readonly n: number;
+    readonly min: number;
+    readonly p50: number;
+    readonly p90: number;
+    readonly max: number;
+    readonly sum: number;
+}
+interface EconomicsMetrics {
+    /** Driver/brain inference — journal `metered` events. */
+    readonly brain: RoleSpend;
+    /**
+     * Brain completions that came back `finish_reason: "length"` — output TRUNCATED. Any value
+     * above 0 means the supervisor planned into a wall and then acted on the half-written plan,
+     * which is a defect and not a cost figure. The journal's `metered` rows carry token counts
+     * but no finish reason, so this reads the per-call brain tap; a run whose supervisor
+     * predates that tap reports `unavailable`, never 0.
+     */
+    readonly brainTruncations: Measured<number>;
+    /** Worker inference — journal `settled` spend plus the harness session join. */
+    readonly workers: RoleSpend;
+    readonly totalUsd: Measured<number>;
+    /**
+     * Where `totalUsd` came from. CLI-backend workers never price their own inference into
+     * the journal, so on those arms the total is BRAIN-ONLY and the worker row's token
+     * counts (recovered from the harness store) are the honest worker-side figure.
+     */
+    readonly totalUsdSource: string;
+    readonly costPerAcceptedPatchUsd: Measured<number>;
+    readonly workerWallMsDistribution: Measured<WallDistribution>;
+    readonly perWorker: Measured<readonly PerWorkerRow[]>;
+}
+interface PatchStats {
+    readonly files: number;
+    readonly linesAdded: number;
+    readonly linesRemoved: number;
+    readonly testFilesTouched: readonly string[];
+}
+interface OutcomeMetrics {
+    readonly supStatus: Measured<string>;
+    readonly supVerdict: Measured<string>;
+    readonly delivered: Measured<boolean>;
+    readonly judgeResolved: Measured<boolean | null>;
+    readonly judgeScore: Measured<number | null>;
+    readonly judgePassed: Measured<number | null>;
+    readonly judgeTotal: Measured<number | null>;
+    readonly verifyPass: Measured<boolean>;
+    readonly verifyRc: Measured<number>;
+    readonly patch: Measured<PatchStats>;
+    /** Which document the judge fields came from (a judge file, a ledger row, or nothing). */
+    readonly judgeSource: string | null;
+}
+interface SupervisorRunReport {
+    readonly schema: typeof SUPERVISOR_RUN_SCHEMA;
+    /** The `runRef` of the sources this report was computed from. */
+    readonly runRef: string;
+    readonly instanceId: string | null;
+    readonly arm: string | null;
+    readonly supervisorId: Measured<string>;
+    readonly generatedAt: string;
+    readonly orchestration: OrchestrationMetrics;
+    readonly decision: DecisionMetrics;
+    readonly economics: EconomicsMetrics;
+    readonly outcome: OutcomeMetrics;
+    /** Artifacts that were missing, in read order — the provenance of every `unavailable`. */
+    readonly gaps: readonly string[];
+    /** The `traces` CLI command that covers the harness-session layer for this run. */
+    readonly traceCommand: string;
+}
+interface RollupCellRow {
+    readonly instanceId: string | null;
+    readonly arm: string | null;
+    readonly steers: Measured<number>;
+    readonly waves: Measured<number>;
+    readonly utilization: Measured<number>;
+    readonly idlePct: Measured<number>;
+    readonly resolved: Measured<boolean | null>;
+    readonly usd: Measured<number>;
+}
+interface SupervisorRunRollup {
+    readonly schema: typeof SUPERVISOR_RUN_ROLLUP_SCHEMA;
+    readonly cells: number;
+    readonly steersTotal: Measured<number>;
+    readonly cellsWithSteers: Measured<number>;
+    readonly cellsWithUnavailableSteers: number;
+    readonly wavesMean: Measured<number>;
+    readonly maxConcurrencyMax: Measured<number>;
+    readonly utilizationMean: Measured<number>;
+    readonly idlePctMean: Measured<number>;
+    readonly workersSpawnedTotal: Measured<number>;
+    readonly acceptedTotal: Measured<number>;
+    readonly usdTotal: Measured<number>;
+    readonly resolvedCount: Measured<number>;
+    readonly perCell: readonly RollupCellRow[];
+}
+/**
+ * A supervision tree expressed in the canonical rollout row type: one
+ * `RolloutLine` per invocation, joined by `parent_rollout_id`. The root row
+ * carries `role: 'supervisor'`; every spawned worker carries `role: 'worker'`
+ * with the root as its parent.
+ */
+interface SupervisorRunTree {
+    readonly rootId: string | null;
+    readonly nodes: readonly RolloutLine[];
+    /** Why a node could not be recovered, in read order. */
+    readonly gaps: readonly string[];
+}
+
+/**
+ * The pure analyzer. Takes already-read bytes (`SupervisorRunSources`) and
+ * returns the report — every metric derivable from a synthetic journal string
+ * with no filesystem, no process, and no network. All I/O lives in a reader
+ * (`loops-reader.ts` is one).
+ */
+
+/**
+ * Analyze already-read supervisor-run bytes. Pure and synchronous: same bytes
+ * in, same report out (modulo `generatedAt`, which `now` pins in tests).
+ */
+declare function analyzeSupervisorRunSources(src: SupervisorRunSources, now?: () => number): SupervisorRunReport;
+/**
+ * Aggregate many supervisor-run reports. A metric no run could measure stays
+ * `unavailable` rather than becoming a 0-valued mean, and cells whose steer
+ * count was unavailable are counted separately from cells that measured zero.
+ */
+declare function rollupSupervisorRuns(reports: readonly SupervisorRunReport[]): SupervisorRunRollup;
+
+/**
+ * ONE implementation of `SupervisorRunReader`: the on-disk layout the loops
+ * supervisor writes — `<runDir>/ws/.loops/supervisor/<id>/{journal.jsonl,
+ * state.json, progress.ndjson, workers/*.ndjson}` alongside the run's
+ * `result.json` / `judge.json` / `driver.log` / delivered patch.
+ *
+ * Nothing in `analyze.ts` knows this layout exists. A different store (an
+ * archive, an object bucket, a database) implements the same interface and
+ * gets the same report.
+ *
+ * Worker token recovery reuses the rollout module's opencode reader rather
+ * than opening a second sqlite path — one store client, one corruption policy.
+ */
+
+interface LoopsReaderOptions {
+    /** Override the workspace dir (default `<runDir>/ws`). */
+    readonly ws?: string;
+    /** Delivered patch path (default: `patchPath` from result.json). */
+    readonly patchPath?: string;
+    /** opencode sqlite store; set to `null` to skip the worker-token join entirely. */
+    readonly opencodeDb?: string | null;
+    /** Ledger to fall back to when the run has no `judge.json` (matched on iid + arm + runDir). */
+    readonly ledgerPath?: string;
+}
+/**
+ * Analyze a supervisor run. Accepts a run directory (read through the loops
+ * reader), any `SupervisorRunReader`, or already-read source bytes — so a
+ * caller with its own store never has to touch the filesystem layout.
+ */
+declare function analyzeSupervisorRun(input: string | SupervisorRunReader | SupervisorRunSources, opts?: LoopsReaderOptions): Promise<SupervisorRunReport>;
+interface WriteSupervisorRunOptions extends LoopsReaderOptions {
+    /** Append the headline block here (the experiment's run log). */
+    readonly appendHeadlineTo?: string;
+    /** Also console.log the headline (default true). */
+    readonly echo?: boolean;
+    /**
+     * Write `run-report.{json,md}` here instead of into the run dir. Set when
+     * reporting over a run directory that must stay READ-ONLY (a live run, an
+     * archived generation).
+     */
+    readonly reportDir?: string;
+}
+/**
+ * Read a completed run, write `run-report.json` + `run-report.md` beside its
+ * artifacts, and append the headline block to the run log. Never throws on a
+ * missing artifact — a run that produced nothing still yields a report whose
+ * every metric says why.
+ */
+declare function writeSupervisorRunReport(runDir: string, opts?: WriteSupervisorRunOptions): Promise<SupervisorRunReport>;
+
+/**
+ * Human-readable renderings of a supervisor-run report. Zero and unavailable
+ * render differently on purpose (`0` vs `unavailable — <reason>`), because the
+ * two have driven opposite conclusions about the same architecture.
+ */
+
+/**
+ * The block appended to a run log after every run — the answers an operator asks
+ * for, in the log tail, with no extra command.
+ */
+declare function renderSupervisorRunHeadline(r: SupervisorRunReport): string;
+declare function renderSupervisorRunMarkdown(r: SupervisorRunReport): string;
+
+/**
+ * The supervision tree as `tangle.rollout.v1` rows.
+ *
+ * A supervisor run IS a tree of rollouts, so its nodes are not a new shape:
+ * the root becomes one `RolloutLine` with `role: 'supervisor'`, every spawned
+ * worker becomes a `RolloutLine` with `role: 'worker'` and
+ * `parent_rollout_id` pointing at its spawner. The rows append to the same
+ * ledger as solo-agent rollouts and join to them with the same keys.
+ *
+ * What the journal CANNOT supply is the transcript: a worker's messages live
+ * in its harness store (opencode sqlite, Claude Code jsonl), which the
+ * `src/rollout/readers/*` intake readers own. Rows minted here are therefore
+ * GAP lines (`messages: []`, `provenance.gap` set) carrying identity,
+ * structure, outcome and cost; hydrating them with messages is the readers'
+ * job, keyed on `artifacts.transcript_ref`.
+ *
+ * Timing lives in `outcome.metrics` (`spawned_at` / `settled_at` / `wall_ms`)
+ * rather than a schema field: `tangle.rollout.v1` describes ONE invocation,
+ * and the inter-invocation event timeline — which is what waves, concurrency,
+ * idle and utilization are computed from — is a property of the journal, not
+ * of any single row. The analyzer reads that timeline; these rows carry the
+ * per-node facts.
+ */
+
+interface SupervisorRolloutOptions {
+    /** Benchmark/suite id for `task.suite`. Defaults to `'supervisor-run'`. */
+    readonly suite?: string;
+    /** `task.split`. Defaults to `'search'` (the trainable pool). */
+    readonly split?: RolloutSplit;
+    /** Replicate index. Defaults to 0. */
+    readonly rep?: number;
+    /** Sampling seed the campaign pinned. Defaults to null (not recorded). */
+    readonly seed?: number | null;
+    /** `run_id` for every node. Defaults to the supervisor root id, else `runRef`. */
+    readonly runId?: string;
+    /** Harness that drove the supervisor. */
+    readonly supervisorHarness?: string | null;
+    /** Harness that drove the workers. */
+    readonly workerHarness?: string | null;
+    /** Model the supervisor ran on. */
+    readonly supervisorModel?: string | null;
+    /** Model the workers ran on. */
+    readonly workerModel?: string | null;
+    readonly experimentId?: string | null;
+    readonly candidateId?: string | null;
+    readonly generation?: number | null;
+    readonly candidateIndex?: number | null;
+    /** Pins `provenance.captured_at`; defaults to now. */
+    readonly capturedAt?: string;
+}
+/**
+ * Mint the supervision tree as rollout rows. Returns the rows plus the gaps
+ * that made any of them incomplete — same unavailable-vs-zero discipline as
+ * the report: a row with no transcript says WHY, it never pretends to be empty.
+ */
+declare function supervisorRunRolloutLines(src: SupervisorRunSources, opts?: SupervisorRolloutOptions): SupervisorRunTree;
+
 /** Canonical OpenInference-over-OTLP attribute names used at trace boundaries. */
 declare const OPENINFERENCE_SPAN_KIND = "openinference.span.kind";
 declare const LLM_MODEL_NAME = "llm.model_name";
@@ -16930,4 +17547,4 @@ 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 ChatRequest, type ChatResponse, 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 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, 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 RolloutRow, type RolloutScrubber, 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, 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 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 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 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, analyzeTraces, appendScorecard, applyLlmSpanOtlpAttributes, applyPolicyEditToSurface, applyToolSpanOtlpAttributes, argHash, asNumber, asString, assertCapabilityHeadroom, assertCrossFamily, assertLlmRoute, assertModelsServed, assertNoHiddenLeak, assertProductBenchmarkRun, assertRealAgentReceipts, assertRealBackend, assertReleaseConfidence, 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, isRunRecord, isSandboxSpan, isToolSpan, isTransientLlmError, 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, renderUpstreamFindings, repeatedActionDetector, replayFeedbackTrajectories, replayFeedbackTrajectory, replayScorerOverCorpus, replayTraceThroughJudge, requireAgentProfileCell, requiredSampleSize, researchReport, resolveModelPricing, resolveRunCostProvenance, resolveSeat, rolloutReward, 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, signManifest, spearmanR, splitGold, statusAdvanced, stopOnNoProgress, stopOnRepeatedAction, stringField, stripFencedJson, subjectiveEval, summarizeAgentReceiptIntegrity, summarizeBackendIntegrity, summarizeHarnessResults, summarizePrReviewBenchmark, summarizePreferenceMemory, summaryTable, 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, validateRunRecord, verbosityBias, verifyAgentProfileCell, verifyAttestation, verifyCompletion, verifyManifest, visualDiff, viteDeployRunner, vitestTestParser, weightedComposite, weightedMean, weightedRecall, welchsTTest, whitespaceCollapseMutator, wilcoxonSignedRank, wilson, withAssignedFeedbackSplit, withHeldoutBlend, withJudgeRetry, withOtelPipeline, wranglerDeployRunner };
+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 };