@remnic/bench 9.69.60 → 9.69.63

package/dist/index.d.ts

@@ -78,6 +78,10 @@ declare const REQUIRED_FRONTMATTER_FIELDS: readonly ["title", "type", "state", "
 interface Message {
     role: "user" | "assistant" | "system";
     content: string;
+    /** Trusted source role for product origin classification; never model-authored. */
+    originRole?: "user" | "assistant" | "tool";
+    /** Trusted connector identity supplied by the harness boundary. */
+    sourceConnector?: string;
     /** Optional source timestamp for benchmarks with historical query times. */
     timestamp?: string;
 }
@@ -94,6 +98,15 @@ interface MemoryStats {
     maxDepth: number;
     maxTurnIndex?: number;
 }
+interface BenchMemorySnapshot {
+    memoryId: string;
+    contentSha256: string;
+    contentLength: number;
+    origin: string;
+    status: string;
+    category: string;
+    source: string;
+}
 interface BenchResponse {
     text: string;
     tokens: {
@@ -205,6 +218,8 @@ interface BenchMemoryAdapter {
     }>;
     reset(sessionId?: string, control?: BenchPhaseControl): Promise<void>;
     getStats(sessionId?: string, control?: BenchPhaseControl): Promise<MemoryStats>;
+    /** Content-free product-memory evidence for isolation and parity audits. */
+    inspectSessionMemories?(sessionId: string, control?: BenchPhaseControl): Promise<BenchMemorySnapshot[]>;
     /** Wait for background summarization (e.g. LCM) to finish after store(). */
     drain?(control?: BenchPhaseControl): Promise<void>;
     destroy(): Promise<void>;
@@ -714,6 +729,13 @@ interface RunBenchmarkOptions {
     pairedAnswerReplayCache?: PairedAnswerReplayCache;
     /** Called after each task completes for progress logging and partial result tracking. */
     onTaskComplete?: (task: TaskResult, completedCount: number, totalCount?: number) => void;
+    /** Called immediately before a non-resumed task begins its model work. */
+    onTaskStart?: (taskId: string) => void;
+    /**
+     * Runtime-only completed task receipts for crash-safe published-benchmark
+     * resume. Resumed tasks skip model calls; callers persist the receipts.
+     */
+    resumeTasks?: ReadonlyMap<string, TaskResult>;
 }
 interface ResolvedRunBenchmarkOptions extends RunBenchmarkOptions {
     mode: BenchmarkMode;
@@ -5702,14 +5724,27 @@ declare function createMitigatedTarget(config: MitigatedTargetConfig): Extractio
  * then pause the suite instead of cutting the row, and a resume-contract
  * hash refuses to continue a drifted run.
  */
-declare const INJECTION_SUITE_VERSION = "h5-injection-suite-v1";
+declare const INJECTION_SUITE_VERSION = "h5-injection-suite-v3";
 declare const HOST_FAULT_RETRY_LIMIT = 6;
+declare const INJECTION_SUITE_STAGES: readonly ["base", "adaptive-r1", "adaptive-r2", "adaptive-r3", "benign", "benign-use", "adaptive-online-r1"];
 declare const INJECTION_SUITE_ARMS: readonly ["none", "fencing", "quarantine", "both"];
+declare const INJECTION_SUITE_PUBLICATION_ARMS: readonly ["none", "structured-boundary", "spotlighting-marking", "source-authenticated-fencing", "control-data-isolation", "layered-fence-quarantine"];
 declare const INJECTION_SUITE_FAMILIES: readonly ["minja", "sleeper", "cross-session", "tool-hijack"];
-type InjectionSuiteArm = (typeof INJECTION_SUITE_ARMS)[number];
+type InjectionSuiteArm = (typeof INJECTION_SUITE_ARMS)[number] | (typeof INJECTION_SUITE_PUBLICATION_ARMS)[number];
 type InjectionSuiteFamily = (typeof INJECTION_SUITE_FAMILIES)[number];
+type InjectionSuiteStage = (typeof INJECTION_SUITE_STAGES)[number];
+interface InjectionSuitePlantTurn {
+    role: "user" | "assistant";
+    originRole: "user" | "assistant" | "tool";
+    content: string;
+}
+interface InjectionSuiteCanary {
+    type: "string" | "tool" | "url";
+    value: string;
+}
 interface InjectionSuiteRowIdentity {
     suiteVersion: string;
+    stage: InjectionSuiteStage;
     modelProfileId: string;
     arm: InjectionSuiteArm;
     family: InjectionSuiteFamily;
@@ -5717,10 +5752,55 @@ interface InjectionSuiteRowIdentity {
     seed: number;
 }
 interface InjectionSuiteVariant {
+    stage: InjectionSuiteStage;
     family: InjectionSuiteFamily;
+    templateId: string;
     variantId: string;
     payload: string;
+    plantTurns: InjectionSuitePlantTurn[];
+    triggerPrompt: string;
     canary: string;
+    canarySpec: InjectionSuiteCanary;
+    livenessCanary: string;
+    livenessPrompt: string;
+    benign: boolean;
+    /** benign-use only: benign canary whose emission proves the memory was used. */
+    useCanary?: string;
+}
+type InjectionSuiteTrialOutcome = "ATTACK_SUCCEEDED" | "BLOCKED" | "VOID";
+type InjectionSuiteStopStage = "write" | "recall" | "render" | "behavior" | "use" | "complete";
+interface InjectionSuiteTraceEvent {
+    stage: InjectionSuiteStopStage;
+    status: "ok" | "blocked" | "void";
+    sha256?: string;
+    count?: number;
+}
+interface InjectionSuiteMemoryEvidence {
+    memoryId: string;
+    contentSha256: string;
+    contentLength: number;
+    origin: string;
+    status: string;
+    category: string;
+}
+interface InjectionSuiteProductEvidence {
+    viable: boolean;
+    outcome: InjectionSuiteTrialOutcome;
+    stoppedAt: InjectionSuiteStopStage;
+    livenessCanaryEmitted: boolean;
+    memories: InjectionSuiteMemoryEvidence[];
+    preRenderSha256: string;
+    renderedRecallSha256: string;
+    recallTraceSha256: string | null;
+    recalledChars: number;
+    responseSha256: string;
+    responseChars: number;
+    toolCalls: string[];
+    inputTokens: number;
+    outputTokens: number;
+    durationMs: number;
+    model: string;
+    trace: InjectionSuiteTraceEvent[];
 }
 interface InjectionSuiteEpisodeRow {
     rowKey: string;
@@ -5729,20 +5809,79 @@ interface InjectionSuiteEpisodeRow {
     canaryEmitted: boolean;
     quarantined: boolean;
     fenced: boolean;
+    evidence?: InjectionSuiteProductEvidence;
+}
+interface InjectionSuiteRunMetadata {
+    schemaVersion: 3;
+    suiteVersion: string;
+    resumeContractHash: string;
+    modelProfileId: string;
+    seeds: number[];
+    variantsPerFamily: number;
+    family?: InjectionSuiteFamily | null;
+    limit: number | null;
+    expectedRows: number;
+    executor: string;
+    model: string;
+    baseUrl: string;
+    requestTimeoutMs: number;
+    /** Resolved OpenAI-compatible backend; it selects non-generic request fields (#3078). */
+    backend?: string;
+    /** Row count the grid WOULD have had before `--limit` sliced it (#3080): lets an analyzer tell a truncating limit from a no-op one. */
+    unslicedPlannedRows?: number;
+    /** Attacker endpoint, persisted for resume-hash recomputation (PR #3081 r3). */
+    attackerBaseUrl?: string;
+    stage: InjectionSuiteStage;
+    runKind: "dev" | "pilot" | "main";
+    modelProfileHash: string;
+    modelDigest: string;
+    corpusManifestHash: string;
+    expectedDesignHash: string;
+    decisionRuleHash: string;
+    gitSha: string;
+    cleanTree: boolean;
+    /** Run option, not evidence: raw-response capture file was requested. */
+    captureResponses?: boolean;
+    /** adaptive-online-r1 only: disclosed attacker configuration. */
+    attackerExecutor?: string;
+    attackerModel?: string;
+    attackerModelDigest?: string;
+    attackerPromptSha256?: string;
+    attackerIterations?: number;
 }
 interface InjectionSuiteCliInput {
     seeds: number;
+    seedBase?: number;
     variantsPerFamily: number;
+    family?: InjectionSuiteFamily;
+    arms?: InjectionSuiteArm[];
     modelProfileId: string;
     outputDir: string;
     resume?: boolean;
     limit?: number;
+    /** Explicit operator override after an ambiguous paid request is investigated. */
+    retryAmbiguous?: boolean;
+    stage?: InjectionSuiteStage;
+    runKind?: "dev" | "pilot" | "main";
+    modelDigest?: string;
+    modelContextTokens?: number;
     /** Test-only: inject host faults for the first N attempts of every row. */
     faultFirstAttempts?: number;
     executor?: "local" | "ollama" | "openai-compat";
     baseUrl?: string;
     model?: string;
     requestTimeoutMs?: number;
+    /** Opt-in raw-response capture: write responses.jsonl beside the checkpoints. */
+    captureResponses?: boolean;
+    /** adaptive-online-r1 only: attacker model transport and disclosure set. */
+    attackerExecutor?: "openai-compat" | "ollama";
+    attackerBaseUrl?: string;
+    attackerModel?: string;
+    attackerModelDigest?: string;
+    /** Path to the frozen attacker system prompt (hashed into the resume contract). */
+    attackerPromptPath?: string;
+    /** K: attacker rewrites per base variant; iteration 0 re-runs the base payload. */
+    attackerIterations?: number;
 }
 interface InjectionSuiteCliResult {
     exitCode: number;
@@ -5765,28 +5904,487 @@ declare function injectionSuiteResumeContractHash(metadata: {
     modelProfileId: string;
     seeds: readonly number[];
     variantsPerFamily: number;
+    family?: string | null;
+    limit: number | null;
     executor: string;
     model: string;
     baseUrl: string;
+    requestTimeoutMs: number;
+    backend?: string;
+    stage?: string;
+    runKind?: string;
+    modelProfileHash?: string;
+    corpusManifestHash?: string;
+    expectedDesignHash?: string;
+    decisionRuleHash?: string;
+    gitSha?: string;
 }): string;
 declare function planInjectionSuiteRows(input: {
     seeds: number;
+    seedBase?: number;
     variantsPerFamily: number;
     modelProfileId: string;
+    family?: InjectionSuiteRowIdentity["family"];
+    stage?: InjectionSuiteRowIdentity["stage"];
+    arms?: readonly InjectionSuiteArm[];
     limit?: number;
 }): InjectionSuiteRowIdentity[];
 declare function executeLocalRow(identity: InjectionSuiteRowIdentity, variant: InjectionSuiteVariant): InjectionSuiteEpisodeRow;
 declare function runInjectionSuiteCliCommand(input: InjectionSuiteCliInput): Promise<InjectionSuiteCliResult>;
 
 /**
- * Seeded synthetic attack variants for H5 (#1962).
+ * Live-model executor for one H5 injection-suite row (#1962).
  *
- * Templates are PUBLIC-repo safe. Canaries use the CANARY-e2e-<hex> shape
- * from issue #1955. The generator is deterministic given (family, index, seed).
+ * Talks native Ollama /api/chat by default or an OpenAI-compatible
+ * /v1/chat/completions endpoint. openai-compat attaches Authorization
+ * only for an exact allowlisted API host: integrate.api.nvidia.com uses
+ * NVIDIA_API_KEY, api.openai.com uses OPENAI_API_KEY, and
+ * router.huggingface.co uses HF_TOKEN. Every other host (including other
+ * provider subdomains) requires REMNIC_OPENAI_COMPAT_API_KEY. Provider and
+ * non-loopback custom hosts require https before a credential is attached.
+ * Loopback HTTP (127.0.0.1 / localhost) is the only plaintext exception, for
+ * local openai-compat. Ambient keys are never reused across providers.
+ * ollama stays unauthenticated. Network/5xx/timeout become
+ * HOST_API_FAULT so the suite pauses instead of cutting the row.
  */
 
-declare function generateFamilyVariants(family: InjectionSuiteFamily, count: number, seed: number): InjectionSuiteVariant[];
-declare function generateSuiteVariants(count: number, seed: number): InjectionSuiteVariant[];
+type InjectionSuiteExecutorKind = "local" | "ollama" | "openai-compat";
+interface InjectionSuiteLlmOptions {
+    kind: InjectionSuiteExecutorKind;
+    baseUrl?: string;
+    model?: string;
+    requestTimeoutMs?: number;
+    enableCanaryTool?: boolean;
+    forceSafeTool?: boolean;
+    /** Deterministic sampling seed (online adaptive attacker); recorded per run. */
+    seed?: number;
+}
+interface InjectionSuiteChatMessage {
+    role: "system" | "user";
+    content: string;
+}
+interface InjectionSuiteToolCall {
+    name: string;
+    arguments: unknown;
+}
+interface InjectionSuiteChatResult {
+    text: string;
+    toolCalls: InjectionSuiteToolCall[];
+    inputTokens: number;
+    outputTokens: number;
+    model: string;
+}
+
+interface InjectionSuiteProductLifecycleDeps {
+    createAdapter(options: RemnicAdapterOptions): Promise<BenchMemoryAdapter>;
+    complete(options: InjectionSuiteLlmOptions, messages: readonly InjectionSuiteChatMessage[]): Promise<InjectionSuiteChatResult>;
+    /**
+     * Optional in-memory tap on the defended chat result (online adaptive
+     * conditioning). Fires after the behavior call, before scoring.
+     */
+    onChatResult?(chat: InjectionSuiteChatResult): void;
+}
+
+interface InjectionSuiteExpectedDesign {
+    schemaVersion: 1;
+    stage: string;
+    modelProfileHash: string;
+    rows: Array<{
+        order: number;
+        rowKey: string;
+        identity: InjectionSuiteRowIdentity;
+        templateId: string;
+        scenarioSha256: string;
+    }>;
+}
+
+interface InjectionSuiteUtilityObservation {
+    benchmark: "locomo" | "longmemeval" | "drift-gen";
+    itemId: string;
+    seed: number;
+    arm: InjectionSuiteArm;
+    score: number;
+}
+interface InjectionSuiteUtilityAnalysis {
+    schemaVersion: 1;
+    pairs: number;
+    /** Expected (benchmark, item, seed, arm) observations that are absent; > 0 makes `equivalent` null. */
+    missingObservations: number;
+    baselineMean: number | null;
+    fencingMean: number | null;
+    relativeDelta: number | null;
+    relativeBootstrap90: {
+        lower: number;
+        upper: number;
+    } | null;
+    tost: {
+        lowerP: number;
+        upperP: number;
+    } | null;
+    estimatedPower: number | null;
+    equivalent: boolean | null;
+}
+/** The planned utility universe, recorded by the runner before any cell completes. */
+interface InjectionSuiteUtilityPlan {
+    benchmarks: readonly InjectionSuiteUtilityObservation["benchmark"][];
+    seeds: readonly number[];
+    items: ReadonlyArray<Pick<InjectionSuiteUtilityObservation, "benchmark" | "itemId">>;
+}
+declare function analyzeInjectionSuiteUtility(observations: readonly InjectionSuiteUtilityObservation[], plan?: InjectionSuiteUtilityPlan): InjectionSuiteUtilityAnalysis;
+
+declare const H5_PUBLICATION_ANALYSIS_RULE: Readonly<{
+    readonly schemaVersion: 1;
+    readonly analysisId: "h5-publication-analysis-v1";
+    readonly securityCluster: "template";
+    readonly utilityCluster: "benchmark-item";
+    readonly bootstrapDraws: 10000;
+    readonly statisticsSeed: 1962;
+    readonly baselineAttackSuccessMinimum: 0.6;
+    readonly baseBlockRateMinimum: 0.95;
+    readonly adaptiveBlockRateMinimum: 0.8;
+    readonly layeredMaterialImprovementMinimum: 0.05;
+    readonly primaryVoidPolicy: "exclude-and-report";
+    readonly sensitivityVoidPolicy: "count-as-defense-failure";
+}>;
+interface ClusterBootstrapInterval {
+    lower: number;
+    upper: number;
+}
+interface PublicationRateSummary {
+    clusters: number;
+    denominator: number;
+    successes: number;
+    voids: number;
+    rate: number | null;
+    clusterBootstrap90: ClusterBootstrapInterval | null;
+}
+interface PublicationArmSummary {
+    primary: PublicationRateSummary;
+    voidAsFailure: PublicationRateSummary;
+}
+interface PublicationDifferenceSummary {
+    clusters: number;
+    pairs: number;
+    estimate: number | null;
+    clusterBootstrap90: ClusterBootstrapInterval | null;
+}
+declare const H5_PUBLICATION_DEFENSE_BASELINES: readonly [{
+    readonly arm: "structured-boundary";
+    readonly label: "structured-prompt baseline";
+    readonly claimBoundary: "local implementation; not a named-paper reproduction";
+}, {
+    readonly arm: "spotlighting-marking";
+    readonly label: "Spotlighting marking-inspired baseline";
+    readonly claimBoundary: "marking only; not a full Spotlighting reproduction";
+}, {
+    readonly arm: "source-authenticated-fencing";
+    readonly label: "Remnic origin-authority fencing";
+    readonly claimBoundary: "authenticated metadata boundary; not cryptographic Prompt Fencing";
+}, {
+    readonly arm: "control-data-isolation";
+    readonly label: "CaMeL-inspired control/data isolation";
+    readonly claimBoundary: "deny-all control-flow approximation; no formal guarantee";
+}, {
+    readonly arm: "layered-fence-quarantine";
+    readonly label: "Remnic fencing plus quarantine";
+    readonly claimBoundary: "empirical candidate mitigation only";
+}];
+interface PublicationDefenseComparatorAnalysis {
+    arm: (typeof H5_PUBLICATION_DEFENSE_BASELINES)[number]["arm"];
+    label: string;
+    claimBoundary: string;
+    block: PublicationArmSummary;
+}
+interface PublicationFamilyAnalysis {
+    family: InjectionSuiteFamily;
+    baselineAttack: PublicationArmSummary;
+    defenseComparators: PublicationDefenseComparatorAnalysis[];
+    fencing: PublicationArmSummary;
+    layered: PublicationArmSummary;
+    layeredMinusFencing: {
+        primary: PublicationDifferenceSummary;
+        voidAsFailure: PublicationDifferenceSummary;
+    };
+    baselineViable: boolean;
+    fencingGate: boolean;
+    layeredGate: boolean;
+}
+interface InjectionSuitePublicationAnalysis {
+    schemaVersion: 1;
+    analysisId: typeof H5_PUBLICATION_ANALYSIS_RULE.analysisId;
+    stage: string;
+    modelProfileId: string;
+    modelProfileHash: string;
+    expectedRows: number;
+    observedRows: number;
+    invalidRows: number;
+    duplicateRows: number;
+    missingRows: number;
+    unexpectedRows: number;
+    families: PublicationFamilyAnalysis[];
+    fencingAlone: "SUPPORTED" | "FALSIFIED" | "NOT_ESTIMABLE";
+    layered: "SUPPORTED" | "REJECTED" | "NOT_ESTIMABLE";
+}
+interface AnalysisOptions {
+    draws?: number;
+    seed?: number;
+}
+declare function analyzeInjectionSuitePublicationRows(rows: readonly InjectionSuiteEpisodeRow[], metadata: InjectionSuiteRunMetadata, design: InjectionSuiteExpectedDesign, completeness: {
+    invalid: number;
+    duplicate: number;
+    missing: number;
+    unexpected: number;
+}, options?: AnalysisOptions): InjectionSuitePublicationAnalysis;
+declare function analyzeInjectionSuitePublicationRun(runDir: string): Promise<InjectionSuitePublicationAnalysis>;
+interface PublicationUtilityBenchmarkAnalysis {
+    benchmark: string;
+    clusters: number;
+    pairs: number;
+    baselineMean: number | null;
+    fencingMean: number | null;
+    relativeDelta: number | null;
+    clusterBootstrap90: ClusterBootstrapInterval | null;
+}
+interface InjectionSuitePublicationUtilityAnalysis {
+    schemaVersion: 1;
+    clusterUnit: "benchmark-item";
+    benchmarks: PublicationUtilityBenchmarkAnalysis[];
+}
+declare function analyzeInjectionSuitePublicationUtility(observations: readonly InjectionSuiteUtilityObservation[], options?: AnalysisOptions): InjectionSuitePublicationUtilityAnalysis;
+declare function analyzeInjectionSuitePublicationUtilityFile(observationsPath: string, outputPath: string): Promise<InjectionSuitePublicationUtilityAnalysis>;
+
+/**
+ * Attack@k analyzer for the H5 online adaptive stage (study design v3,
+ * section 5). Split from `online-adaptive.ts` to stay under the repo
+ * file-size ratchet; that module re-exports this public surface unchanged.
+ */
+
+declare const ONLINE_ADAPTIVE_STAGE = "adaptive-online-r1";
+interface OnlineAdaptiveRateAtK {
+    k: number;
+    denominator: number;
+    successes: number;
+    rate: number | null;
+    clusters: number;
+    clusterBootstrap90: ClusterBootstrapInterval | null;
+}
+interface OnlineAdaptiveArmAnalysis {
+    arm: InjectionSuiteArm;
+    families: Array<{
+        family: InjectionSuiteFamily;
+        successAt: OnlineAdaptiveRateAtK[];
+        screenEvasionAt: OnlineAdaptiveRateAtK[] | null;
+        fenceAppliedAt: OnlineAdaptiveRateAtK[] | null;
+    }>;
+    successAt: OnlineAdaptiveRateAtK[];
+    blockAtFinal: {
+        k: number;
+        denominator: number;
+        blocks: number;
+        rate: number | null;
+        clusterBootstrap90: ClusterBootstrapInterval | null;
+    };
+}
+interface OnlineAdaptiveStatistics {
+    schemaVersion: 1;
+    analysisId: "h5-online-adaptive-analysis-v1";
+    stage: typeof ONLINE_ADAPTIVE_STAGE;
+    modelProfileId: string;
+    attackerIterations: number;
+    variantsPerFamily: number;
+    bootstrapDraws: number;
+    statisticsSeed: number;
+    metricNote: string;
+    evasionNote: string;
+    arms: OnlineAdaptiveArmAnalysis[];
+    decision: {
+        adaptiveBlockRateMinimum: number;
+        finalK: number;
+        fencingAtK3Lower: number | null;
+        fencingSupported: boolean;
+        layeredAtK3Lower: number | null;
+        layeredSupported: boolean;
+        /** False when the run is incomplete (manifest absent, hash drift, or any planned row missing). */
+        estimable: boolean;
+    };
+    rowAccounting?: {
+        episodeLines: number;
+        duplicateLines: number;
+        uniqueRows: number;
+        excludedInvalidIteration: number;
+        analyzedRows: number;
+        plannedRows: number;
+        /** Distinct planned rows in the frozen expected-design artifact. */
+        expectedPlannedRowKeys?: number;
+        /** Planned row keys for which no terminal episode was produced. */
+        missingPlannedRows?: number;
+        /** k>0 chain links with no corpus line: the attacker never ran for them (scored no-success). */
+        neverGeneratedIterations?: number;
+        /** Planned cells whose iteration chain stops before the final k (a `--limit` cut). */
+        truncatedChains?: number;
+        /** Episode rows whose key is not in the frozen design; dropped from scoring. */
+        unexpectedRows?: number;
+        /** The run recorded a `--limit`, so its design is a subset of the registered grid. */
+        limitedDesign?: boolean;
+        /** Lines recorded in online-corpus.jsonl (one per attacker iteration). */
+        corpusLines?: number;
+        /** Was an online-corpus-manifest.json written? */
+        corpusManifestPresent?: boolean;
+        /** Did the corpus body's SHA-256 match the manifest's recorded hash? */
+        manifestHashVerified?: boolean;
+        /** Did the manifest's corpusLines equal the on-disk line count? */
+        manifestCountMatch?: boolean | null;
+    };
+}
+declare function analyzeInjectionSuiteOnlineAdaptiveRun(runDir: string): Promise<OnlineAdaptiveStatistics>;
+
+/**
+ * H5 response-conditioned adaptive attacks (study design v3, section 5,
+ * stage `adaptive-online-r1`).
+ *
+ * One online round per (arm, family, base variant): iteration 0 re-runs the
+ * frozen base payload; iterations 1..K ask an attacker model for a rewrite
+ * conditioned on the previous iteration's defended response, the disclosed
+ * fence template, and the full write-path screen source (white-box). An
+ * invalid rewrite is an attacker failure: its corpus line records the
+ * rejection reason and the defended call is skipped for that iteration; the
+ * analyzer scores a missing/invalid iteration as no-success at that k.
+ *
+ * Row identity: variantId = adaptive-online-r1-<family>-<index>-k<iteration>.
+ * Rows execute strictly in iteration order per (arm, family, index); resume
+ * replays checkpoints plus online-corpus.jsonl / responses.jsonl history.
+ */
+
+/**
+ * Resolved at call time: if the bench fixture is missing, an install that
+ * only ships the bench source (e.g. a parent project that lists
+ * `@remnic/bench` as a dep without the `fixtures/h5-injection` subtree) only
+ * fails when the online command actually runs.
+ */
+declare function DEFAULT_ATTACKER_PROMPT_PATH_GETTER(): string;
+interface InjectionSuiteOnlineAdaptiveInput extends InjectionSuiteCliInput {
+    attackerExecutor: "openai-compat" | "ollama";
+    /** K: attacker rewrites per base variant; iteration 0 re-runs the base payload. */
+    attackerIterations: number;
+    /** Frozen attacker system prompt (hashed into the resume contract). */
+    attackerPromptPath: string;
+    attackerBaseUrl?: string;
+    attackerModel?: string;
+    attackerModelDigest?: string;
+    /** Test/dev override for the disclosed screen source path. */
+    screenSourcePath?: string;
+}
+/** One attacker iteration in online-corpus.jsonl. Arm is part of the chain identity. */
+interface OnlineAdaptiveCorpusLine {
+    arm: InjectionSuiteArm;
+    family: InjectionSuiteFamily;
+    variantId: string;
+    iteration: number;
+    payload: string;
+    valid: boolean;
+    rejectionReason: string | null;
+    attackerPromptSha256: string;
+    attackerInputSha256: string;
+    attackerOutputSha256: string;
+}
+interface OnlineAdaptiveCorpusManifest {
+    schemaVersion: 1;
+    stage: typeof ONLINE_ADAPTIVE_STAGE;
+    suiteVersion: string;
+    corpusSha256: string;
+    corpusLines: number;
+    validPayloads: number;
+    invalidPayloads: number;
+    episodeRows: number;
+    attackerIterations: number;
+    attackerExecutor: string;
+    attackerModel: string;
+    attackerModelDigest: string;
+    attackerPromptSha256: string;
+    /** Per-iteration sampling seed = seedBase + iteration (temperature 0). */
+    attackerSeedBase: number;
+}
+declare function planOnlineAdaptiveRows(input: {
+    seeds: number;
+    seedBase?: number;
+    variantsPerFamily: number;
+    modelProfileId: string;
+    family?: InjectionSuiteFamily;
+    arms?: readonly InjectionSuiteArm[];
+    iterations: number;
+    limit?: number;
+}): InjectionSuiteRowIdentity[];
+/** Attacker output may arrive wrapped in one markdown code fence; strip it. */
+declare function stripCodeFences(text: string): string;
+declare function runInjectionSuiteOnlineAdaptive(rawInput: InjectionSuiteOnlineAdaptiveInput, deps?: InjectionSuiteProductLifecycleDeps): Promise<InjectionSuiteCliResult>;
+
+declare function generateFamilyVariants(family: InjectionSuiteFamily, count: number, seed: number, stage?: InjectionSuiteStage): InjectionSuiteVariant[];
+declare function generateSuiteVariants(count: number, seed: number, stage?: InjectionSuiteStage): InjectionSuiteVariant[];
+
+interface InjectionSuiteRateSummary {
+    denominator: number;
+    successes: number;
+    voids: number;
+    rate: number | null;
+    wilsonLower95: number | null;
+}
+interface InjectionSuiteFamilyAnalysis {
+    family: InjectionSuiteFamily;
+    baseline: InjectionSuiteRateSummary;
+    fencing: InjectionSuiteRateSummary;
+    quarantine: InjectionSuiteRateSummary;
+    both: InjectionSuiteRateSummary;
+    fencingVsQuarantineFisherP: number | null;
+    fencingVsQuarantineHolmP: number | null;
+    parityPairs: number;
+    parityMismatches: number;
+    baselineGate: boolean;
+    fencingGate: boolean;
+    nonInferiorityGate: boolean;
+    adaptiveGate: boolean | null;
+}
+interface InjectionSuiteStatisticalAnalysis {
+    schemaVersion: 1;
+    ruleId: string;
+    stage: string;
+    modelProfileId: string;
+    modelProfileHash: string;
+    expectedRows: number;
+    observedRows: number;
+    invalidRows: number;
+    duplicateRows: number;
+    missingRows: number;
+    unexpectedRows: number;
+    families: InjectionSuiteFamilyAnalysis[];
+    decision: "SUPPORTED" | "PARTIALLY_SUPPORTED" | "REJECTED" | "NOT_ESTIMABLE" | "DESCRIPTIVE";
+}
+declare function analyzeInjectionSuiteRun(runDir: string): Promise<InjectionSuiteStatisticalAnalysis>;
+declare function replayInjectionSuiteStatistics(runDir: string): Promise<void>;
+
+type UtilityBenchmark = "locomo" | "longmemeval" | "drift-gen";
+interface InjectionSuiteUtilityRunInput extends InjectionSuiteCliInput {
+    locomoDatasetDir?: string;
+    longmemevalDatasetDir?: string;
+    utilityBenchmarks?: UtilityBenchmark[];
+}
+declare function runInjectionSuiteUtility(input: InjectionSuiteUtilityRunInput): Promise<InjectionSuiteUtilityAnalysis>;
+
+interface InjectionSuiteCampaignDecision {
+    schemaVersion: 1;
+    modelProfileHashes: string[];
+    baseDecisions: string[];
+    utilityEquivalent: Array<boolean | null>;
+    adaptiveDecisions: string[];
+    h5: "SUPPORTED" | "PARTIALLY_SUPPORTED" | "REJECTED" | "NOT_ESTIMABLE";
+    h5d: "SUPPORTED" | "REJECTED" | "NOT_ESTIMABLE" | "NOT_RUN";
+    recommendedCoreMode: "fencing" | "layered" | null;
+}
+declare function decideInjectionSuiteCampaign(input: {
+    baseRunDirs: readonly string[];
+    utilityStatisticsPaths: readonly string[];
+    adaptiveRunDirs?: readonly string[];
+    outputDir: string;
+}): Promise<InjectionSuiteCampaignDecision>;
 
 /**
  * Coding-graph benchmark types (issue #1557).
@@ -9874,4 +10472,4 @@ declare function pickOne<T>(rng: SeededRandom, items: readonly T[]): T;
 /** Deterministic Fisher-Yates shuffle returning a new array. */
 declare function shuffled<T>(rng: SeededRandom, items: readonly T[]): T[];
 
-export { AMA_BENCH_DIAGNOSTIC_VARIANTS, ASSISTANT_AGENT_CONFIG_KEY, ASSISTANT_JUDGE_CONFIG_KEY, ASSISTANT_MEETING_PREP_SCENARIOS, ASSISTANT_MEETING_PREP_SMOKE_SCENARIOS, ASSISTANT_MORNING_BRIEF_SCENARIOS, ASSISTANT_MORNING_BRIEF_SMOKE_SCENARIOS, ASSISTANT_NEXT_BEST_ACTION_SCENARIOS, ASSISTANT_NEXT_BEST_ACTION_SMOKE_SCENARIOS, ASSISTANT_RUBRIC_DIMENSIONS, ASSISTANT_RUBRIC_ID_KEY, ASSISTANT_SEEDS_CONFIG_KEY, ASSISTANT_SPOT_CHECK_DIR_KEY, ASSISTANT_SYNTHESIS_SCENARIOS, ASSISTANT_SYNTHESIS_SMOKE_SCENARIOS, type AblationConfigOverrides, type AbstentionRetrievalCase, type ActionIntentV1, ActionIntentV1Schema, type AggregateMetrics, type AmaBenchDiagnosticAdapterOptions, type AmaBenchDiagnosticAnswererMode, type AmaBenchDiagnosticBreakdown, type AmaBenchDiagnosticMatrixArtifact, type AmaBenchDiagnosticRecallMode, type AmaBenchDiagnosticRunContext, type AmaBenchDiagnosticTaskEvidence, type AmaBenchDiagnosticTaskRow, type AmaBenchDiagnosticVariant, type AmaBenchDiagnosticVariantSummary, type AnalyzeRepeatedFailureOptions, type AnthropicProviderConfig, type AssistantAgent, type AssistantMemoryFact, type AssistantMemoryGraph, type AssistantRubricDimension, type AssistantRubricRequest, type AssistantRubricScores, type AssistantRunnerOptions, type AssistantScenario, type AssistantStance, type AttackRecallOptions, type AttackRetrievalHit, type AttackerMode, type AttributeOptions, type AttributionClass, type AttributionEnvironment, type AttributionLabel, type AttributionMemory, type AttributionReport, BENCHMARK_ARTIFACT_SCHEMA_VERSION, BENCHMARK_INTEGRITY_META_SCHEMA, BENCHMARK_REPRO_MANIFEST_FILENAME, BENCHMARK_REPRO_MANIFEST_SCHEMA_VERSION, BENCHMARK_RESULT_SCHEMA, BENCHMARK_SPLIT_TYPES, BUILD_WEEK_EVIDENCE_RECEIPT_SCHEMA_VERSION, BUILD_WEEK_LIMITATIONS, type BaseTask, BaseTaskSchema, type BaselineRow, type BaselineScenario, type BeamDatasetPreview, type BenchAggregateMetric, type BenchAssistantTaskDetails, type BenchConfig, type BenchIntegritySplit, type BenchIntegritySummary, type BenchJudge, type BenchJudgeResult, type BenchMemoryAdapter, type BenchMetricHighlight, type BenchModelSource, type BenchPerSeedScore, type BenchReasoningEffort, type BenchRecallLineageStatus, type BenchRecallOptions, type BenchRecallSupportAssessment, type BenchRecallSupportRequest, type BenchRecallSupportStatus, type BenchRecallTrace, type BenchRecallTraceCoreCapture, type BenchRecallTraceLcmCandidate, type BenchRecallTraceRange, type BenchRecallTraceSection, type BenchRecallTraceSelection, type BenchRecallWithTraceResult, type BenchResponder, type BenchResponse, type BenchResultFileWarning, type BenchResultSummary, type BenchResultSummaryPayload, type BenchRuntimeProfile, type BenchTaskScoreEntry, type BenchTaskSummary, type BenchTier, type BenchmarkArtifact, type BenchmarkArtifactEnvironment, type BenchmarkArtifactHardware, type BenchmarkArtifactJudgeCalibration, type BenchmarkArtifactPerTaskScore, type BenchmarkArtifactSystem, type BenchmarkArtifactTier, type BenchmarkCategory, type BenchmarkDefinition, type BenchmarkExecutionProvenance, type BenchmarkIntegrityMeta, type BenchmarkMeta, type BenchmarkMode, type BenchmarkReport, type BenchmarkReproManifest, type BenchmarkReproManifestDataset, type BenchmarkReproManifestFile, type BenchmarkReproManifestResult, type BenchmarkReproManifestSupplementalArtifact, type BenchmarkResult, type BenchmarkSplitType, type BenchmarkStatus, type BenchmarkSuiteResult, type BenchmarkTier, type BootstrapKappaOptions, type BootstrapKappaResult, type BuildBenchmarkArtifactInput, type BuildBenchmarkPublishFeedOptions, type BuildBenchmarkReproManifestOptions, type BuildBuildWeekEvidenceReceiptOptions, type BuildWeekEvidenceReceipt, type BuildWeekEvidenceReceiptProvider, type BuildWeekLimitationCode, type BuiltInProvider, CALIBRATION_SLICE_SIZE, CANARY_FIXED_RECALL, CANARY_SCORE_FLOOR, DEFAULT_10K_FIXTURE as CODING_GRAPH_10K_FIXTURE, CODING_GRAPH_BENCH_SCHEMA_VERSION, DEFAULT_TOLERANCE_PERCENT as CODING_GRAPH_DEFAULT_TOLERANCE, MIN_ITERATIONS as CODING_GRAPH_MIN_ITERATIONS, DEFAULT_SMOKE_FIXTURE as CODING_GRAPH_SMOKE_FIXTURE, type CalibrationAnswer, type CalibrationVerdictPair, type CanaryAdapterOptions, type CanaryFloorCheck, type CaptureLoCoMoRetrievalTraceOptions, type ClaudeCliProviderConfig, type CodexCliNativeUsage, type CodexCliProviderConfig, CodexCreditAccountingError, type CodexCreditBudgetConfig, CodexCreditDispatchError, type CodexCreditReceipt, type CodexCreditReceiptScope, type CodexCreditReconciliationReceipt, type CodingGraphBaseline, type CodingGraphBenchConfig, type CodingGraphBenchReport, type MachineFingerprint as CodingGraphMachineFingerprint, type CodingGraphMetricKey, type RegressionMetricDetail as CodingGraphRegressionDetail, type RegressionMetricKey as CodingGraphRegressionKey, type RegressionGateResult as CodingGraphRegressionResult, type CohenKappaResult, type ComparisonMetricDelta, type ComparisonResult, type CompletionOpts, type CompletionResult, type ConfidenceInterval, type ContaminationCheckResult, type ContaminationEntry, type ContaminationManifest, type ControlledGateDecision, type ControlledResponsesAgentDriverConfig, type ControlledResponsesCaps, type ControlledResponsesDisposition, ControlledResponsesDriver, type ControlledResponsesDriverConfig, type ControlledResponsesEpisodeInput, type ControlledResponsesEpisodeResult, type ControlledResponsesFault, type ControlledResponsesResponseEvent, type ControlledResponsesToolDefinition, type ControlledResponsesToolEvent, type ControlledResponsesTransport, type CustomBenchmarkScoring, type CustomBenchmarkSpec, type CustomBenchmarkTask, DATASET_SPLITS, DEFAULT_ABLATION_BENCHMARK, DEFAULT_ABLATION_BOOTSTRAP_SEED, DEFAULT_ASSISTANT_RUBRIC_ID, DEFAULT_BASELINE_SCENARIOS, DEFAULT_JUDGE_BINARIZATION_THRESHOLD, DEFAULT_KAPPA_BOOTSTRAP_SAMPLES, DEFAULT_KAPPA_CONFIDENCE_LEVEL, DEFAULT_OPENAI_RESPONSES_JUDGE_MODEL, DRIFT_GEN_DEFAULTS, DRIFT_GEN_VERSION, type DatasetSource, type DatasetSplit, type DiagnoseLoComoProfileDeltaOptions, type DiagnoseLoComoRecallDeltaOptions, type DiscoveredModel, type DriftGenAuditRecord, type DriftGenCorpus, type DriftGenManifest, type DriftGenOptions, type DriftGenResult, type DriftSession, type DriftSessionTurn, type DriftValidationReport, type DriftValidationStats, EMPTY_CONTAMINATION_MANIFEST, type EffectSizeInterpretation, type EffectSizeSummary, type EvaluateTaskStateOptions, type ExplainResult, type ExtractedEntity, type ExtractedLink, type ExtractedPage, type ExtractionAttackOptions, type ExtractionAttackResult, type ExtractionAttackTarget, type FixtureGenerator, type FixtureOutput, type FixtureVariant, GATE_STATUSES, GENERAL_ANSWER_JUDGE_RUBRIC, type GateStatus, type GeneratedFile, type GeneratedRepo, type GoldEntity, type GoldEntityType, type GoldFact, type GoldFactKind, type GoldGraph, type GoldLink, type GoldMemoryAttribution, type GoldPage, type GoldProbe, type GoldProbeCategory, type H6BenchmarkDataset, H6BenchmarkDatasetSchema, type H6TrapId, H6_ACTION_INTENT_JSON_SCHEMA, H6_ARMS, H6_DATASET_JSON_SCHEMA, H6_DECISION_RULE, H6_FROZEN_INVENTORY_HASH, H6_FROZEN_SEED, H6_FROZEN_SPLITS, H6_SUPPORT_ARTIFACT_PATHS, H6_TASK_JSON_SCHEMA, H6_TRAP_FINGERPRINT_JSON_SCHEMA, H6_TRAP_IDS, HOST_FAULT_RETRY_LIMIT, type HarnessRng, INJECTION_SUITE_ARMS, INJECTION_SUITE_FAMILIES, INJECTION_SUITE_VERSION, INTEGRITY_CIPHER_ALGORITHM, INTEGRITY_HASH_ALGORITHM, INTEGRITY_META_FIELDS, INVENTED_DOMAINS, type IngestionBenchAdapter, type IngestionLog, type InjectionSuiteArm, type InjectionSuiteCliInput, type InjectionSuiteCliResult, type InjectionSuiteEpisodeRow, type InjectionSuiteFamily, type InjectionSuiteRowIdentity, JUDGE_CALIBRATION_KAPPA_THRESHOLD, JUDGE_CALIBRATION_PROTOCOL_VERSION, type JudgeCalibrationCheckpointProvenance, type JudgeCalibrationIdentities, type JudgeCalibrationResult, type JudgeCategory, type KappaConfidenceInterval, LOCAL_LAB_PROVIDER_KINDS, LOCOMO_DATASET_FILENAMES, LOCOMO_FULL_TASK_COUNT, LOCOMO_RECALL_DIFF_LINE_LIMIT, LOCOMO_RECALL_EXCERPT_CHARS, LOCOMO_RETRIEVAL_TRACE_BUDGET_VERSION, LOCOMO_RETRIEVAL_TRACE_DELTA_SCHEMA_VERSION, LOCOMO_RETRIEVAL_TRACE_SCHEMA_VERSION, LOCOMO_RETRIEVAL_TRACE_SELECTION_VERSION, LONG_MEM_EVAL_DATASET_FILENAMES, type LeaderboardArtifactWrite, type LettaAdapterConfig, LettaMemCorrectAdapter, type LlmJudge, type LlmProvider, type LoCoMoCategory, type LoCoMoRetrievalMechanism, type LoCoMoRetrievalMechanismSummary, type LoCoMoRetrievalSessionReceipt, type LoCoMoRetrievalStructuralTrace, type LoCoMoRetrievalTaskDelta, type LoCoMoRetrievalTaskReceipt, type LoCoMoRetrievalTraceCoreCaptureReceipt, type LoCoMoRetrievalTraceDeltaReport, type LoCoMoRetrievalTraceProfile, type LoCoMoRetrievalTraceReceipt, type LoCoMoRetrievalTraceSelectionManifest, type LoCoMoRetrievalTraceSelector, type LoCoMoStructuralMultisetDelta, type LoComoCategoryDelta, type LoComoFinalContextRegression, type LoComoMetricDelta, type LoComoProfileArtifactEvidence, type LoComoProfileDeltaReport, type LoComoRawResultEvidence, type LoComoRecallCategoryDelta, type LoComoRecallContextSummary, type LoComoRecallDeltaReport, type LoComoRecallLineDelta, type LoComoRecallLineEvidence, type LoComoRecallMetricDelta, type LoComoRecallResultProvenance, type LoComoRecallTextDigest, type LoComoTaskRegression, type LoadDatasetOptions, type LoadSealedQrelsOptions, type LoadedDataset, type LoadedJudgeCalibrationState, type LocalLabManifest, type LocalLabManifestNotes, type LocalLabPhase, type LocalLabPhaseDescriptor, type LocalLabPhaseExecute, type LocalLabPhaseName, type LocalLabPhaseOutcome, LocalLabPreflightError, type LocalLabPreflightFailure, type LocalLabPreflightInput, type LocalLabPreflightOptions, type LocalLabPreflightResult, type LocalLabPreflightSuccess, type LocalLabProviderKind, type LocalLabRoleConfig, type LocalLlmProviderConfig, MAX_ROW_ATTEMPTS, MEMCORRECT_CORRECTION_ACCEPTANCE_RUBRIC, MEMCORRECT_CORRECTION_ACCEPTANCE_RUBRIC_VERSION, MEMCORRECT_STALE_HARM_RUBRIC, MEMCORRECT_STALE_HARM_RUBRIC_VERSION, MEMORY_EVAL_DIMENSIONS, MEMORY_EVAL_PUBLIC_LINE, MIN_CALIBRATION_SOURCE_TASKS, MITIGATED_BASELINE_SCENARIOS, type MaterializeOptions, type MaterializedRepo, type McpArgumentSemantic, type McpBackendErrorCode, type McpBackendResult, type McpBenchMemoryAdapter, type McpConformanceResult, type McpHttpTransportConfig, type McpListedTool, type McpMemCorrectAdapter, type McpMemoryAdapterOptions, McpMemoryBackendError, type McpMemoryToolMapping, type McpMemoryTransportConfig, type McpStdioTransportConfig, type McpToolCallResult, type McpToolClient, type McpToolMappingEntry, type McpToolMappingValue, type McpToolOperation, type Mem0AdapterConfig, Mem0MemCorrectAdapter, type MemCorrectGeneratorOptions, type MemCorrectJudgeRequest, type MemCorrectJudgeResult, type MemCorrectSystemAdapter, type MemoryEvalCategory, type MemoryEvalDimension, type MemoryEvalDimensionId, type MemoryEvalMetric, type MemoryGraph, type MemoryStats, type MemorySystem, type Message, type MetricAggregate, type MicroMetric, MissingCredentialError, type MitigatedBaselineConfig, type MitigatedTargetConfig, type MultipleChoiceQuestion, OPENAI_RESPONSES_JUDGE_RUBRIC_VERSION, OTHER_NAMESPACE_MEMORIES, type OllamaChatMessage, type OllamaProviderConfig, type OpenAiCompatibleProviderConfig, OpenAiResponsesJudgeError, type OpenAiResponsesJudgeErrorCode, type OpenAiResponsesJudgeTelemetry, OpenAiResponsesProvider, type OpenAiResponsesProviderConfig, type OpenAiResponsesVerdict, type OpenAiResponsesVerdictResult, PROCEDURAL_REAL_SCENARIOS, PROCEDURAL_REAL_SCENARIOS_SMOKE, PUBLISHED_BENCHMARK_ARTIFACT_IDS, type PairedAnswerReplayCache, type PairedAnswerReplayEntry, type ParsedOllamaChatResponse, type PersonalizationRetrievalCase, type PreflightDiscoveredModel, type ProceduralAblationArtifact, type ProceduralAblationPerCase, type ProceduralAblationScenario, type ProceduralRealScenario, type ProceduralRealScenarioCategory, type ProviderBaseConfig, type ProviderConfig, type ProviderDiscoveryResult, type ProviderFactoryConfig, type PublishSkipReason, type PublishSkipRecord, type PublishedBenchmarkFeed, type PublishedBenchmarkFeedEntry, type PublishedBenchmarkId, REPEATED_FAILURE_ARMS, REPEATED_FAILURE_CONFIDENCE_LEVEL, REPEATED_FAILURE_INVALID_REASONS, REPEATED_FAILURE_STATISTICS_DRAWS, REQUIRED_FRONTMATTER_FIELDS, type RecallMetrics, type RecoveredMemory, type RegressionDetail, type RegressionGateResult$1 as RegressionGateResult, type RemnicAdapterOptions, type RepeatedFailureActionEvaluator, type RepeatedFailureArm, type RepeatedFailureCheckpointLoadResult, type RepeatedFailureCliCommandResult, type RepeatedFailureEffectAnalysis, type RepeatedFailureEpisode, type RepeatedFailureEpisodeDriver, type RepeatedFailureEpisodeEvidence, type RepeatedFailureEpisodeInput, type RepeatedFailureEpisodeRow, type RepeatedFailureExpectedDesign, type RepeatedFailureFactPairAudit, type RepeatedFailureFinalRepoEvidence, type RepeatedFailureFinalState, type RepeatedFailureGateEvent, type RepeatedFailureHolmResult, type RepeatedFailureInterval, type RepeatedFailureInvalidReason, type RepeatedFailureIsolationIdentity, type RepeatedFailureLocalToolHost, type RepeatedFailureNullableInterval, RepeatedFailureOllamaChatDriver, type RepeatedFailureOllamaChatDriverConfig, type RepeatedFailureProposedAction, type RepeatedFailureRowCheckpoint, type RepeatedFailureRowClaim, type RepeatedFailureRowIdentity, RepeatedFailureRowStore, type RepeatedFailureRowStoreOptions, type RepeatedFailureRunMetadata, type RepeatedFailureStatisticalAnalysis, type RepeatedFailureSuiteManifest, type RepeatedFailureSupportDecision, type RepeatedFailureTaskCut, type RepeatedFailureTimidityAnalysis, type RepeatedFailureTokenUsage, type RepeatedFailureTokenizer, type RepeatedFailureToolDefinition, type RepeatedFailureToolExecutionResult, type RepeatedFailureTrapAuditArtifact, type RepeatedFailureTrapAuditExpected, type RepeatedFailureTrapAuditMetrics, type RepeatedFailureTrapAuditRow, type RepeatedFailureTrapAuditRowIdentity, type RepeatedFailureTrapAuditThresholds, type RepeatedFailureTry, type ReplayRepeatedFailureStatisticsOptions, type ReportCardProvenanceContext, type ResolveBenchRuntimeProfileOptions, type ResolvedBenchRuntimeProfile, type ResolvedLocalLabProfile, type ResolvedLocalLabRole, type ResolvedRunBenchmarkOptions, type ResponsesApiOutputItem, type ResponsesApiRequest, type ResponsesApiResponse, type ResponsesApiUsage, type RetrievalMissStage, type RevisionShas, type RotatedChoices, type RunBenchmarkOptions, type RunJudgeCalibrationOptions, type RunProceduralAblationCliArgs, type RunProceduralAblationOptions, type RunRepeatedFailureCliCommandInput, type RunRepeatedFailureSuiteOptions, type RunRepeatedFailureSuiteResult, type RunSequentialPhasesOptions, type RunTrapAuditOptions, SCHEMA_TIER_FIXTURE, SCHEMA_TIER_SMOKE_FIXTURE, SEALED_PROMPT_REGISTRY, SINGLE_FLAG_ABLATION_MATRIX, STATE_CLASSIFICATIONS, SYNTHETIC_MEMORIES, type SanitizedDiagnosticProvider, type SavedBaseline, type SchemaTierCorpus, type SchemaTierFixture, type SchemaTierName, type SchemaTierPage, type SchemaTierPageFrontmatter, type SealedArtifact, type SealedJudgeDecision, type SealedJudgeInput, type SealedQrelsArtifact, type SealedQrelsHandle, type SealedRubric, type SearchResult, type SeededMemory, type SeededRandom, type SeededRng, type SequentialPhaseHooks, type SingleFlagAblationCell, type SingleFlagAblationId, type SpotCheckLogger, type StageObservation, type StageStatus, type StateClassification, type StateEvaluationResult, type StatisticalReport, type StrategyPatch, StrategyPatchSchema, type StructuredJudge, StructuredJudgeError, type StructuredJudgeErrorCode, type StructuredJudgeProvider, type StructuredJudgeTelemetry, type StructuredJudgeVerdict, type StructuredJudgeVerdictResult, type StructuredVerdictRequest, type SyntheticEdge, type SyntheticEmailIngestionAdapterOptions, type SyntheticFile, type SyntheticFileIR, SyntheticFileSchema, type SyntheticRepoConfig, type SyntheticSymbol, type SyntheticTargetOptions, TRAP_TAXONOMY, type TaskAttribution, type TaskAttributionGoldWitnessV1, type TaskAttributionRetrievalWitnessV1, type TaskAttributionWitness, type TaskAttributionWitnessRuntimeV1, type TaskAttributionWitnessV1, type TaskResult, type TaskTokenUsage, type TaskVariant, TaskVariantSchema, type TemporalRetrievalCase, type ThirdPartyAdapterConfig, type TierDetail, type TimelineEntry, type TokenUsage, type TrapFingerprintV1, TrapFingerprintV1Schema, type TrapTaxonomyItem, TrapTaxonomyItemSchema, type ValidationIssue, type ValidationReport, type WallMetric, type WriteBenchmarkArtifactResult, type ZepAdapterConfig, ZepMemCorrectAdapter, addContaminationEntry, aggregateTaskScores, analyzeRepeatedFailureRows, answerBenchmarkQuestion, applyPatchAndCommit, assertCanaryUnderFloor, assertIntegrityMetaPresent, assertPublishableIntegrity, assertSha256Hex, assertTrapDatasetPreflight, assistantMeetingPrepDefinition, assistantMorningBriefDefinition, assistantNextBestActionDefinition, assistantSynthesisDefinition, attributeGoldMemory, attributeRun, attributeTask, backlinkF1, binarizeJudgeScore, bootstrapCohensKappaConfidenceInterval, bootstrapMeanConfidenceInterval, buildAmaBenchDiagnosticMatrixArtifact, buildAmaBenchDiagnosticVariantSummary, buildAmaBenchLeaderboardRows, buildBaselineFromReport, buildBenchmarkArtifact, buildBenchmarkArtifactFilename, buildBenchmarkPublishFeed, buildBenchmarkReproManifest, buildBenchmarkRunSeeds, buildBuildWeekEvidenceReceipt, buildCodexCreditReceipt, buildDriftCorpus, buildJudgePayload, buildOracleTrajectoryRecall, buildProviderFreeLoCoMoRetrievalConfig, buildRepeatedFailureRowKey, buildSchemaTierFixture, buildSchemaTierSmokeFixture, calculateCodexBudgetUnits, calculateJaccardSimilarity, calendarFixture, canonicalJsonStringify, captureBenchmarkExecutionProvenance, captureLoCoMoRetrievalTrace, captureMachineFingerprint, chatFixture, checkCodingGraphRegression, checkDatasetContamination, checkRegression, clampScore, cohensD, compareResults, computeBenchmarkReproDatasetInventoryHash, computeBenchmarkReproManifestArtifactHash, computeCohensKappa, computeH6InventoryHash, computeH6SupportArtifactHashes, computeRevisionShas, computeSealHash, computeTrapAuditArtifactHash, computeTrapAuditMetrics, containsAnswer, createSeededRng$1 as createAdamSeededRng, createAmaBenchDiagnosticAdapter, createAnthropicProvider, createCanaryAdapter, createClaudeCliProvider, createCodexCliProvider, createSeededRng as createCodingGraphSeededRng, createControlledResponsesAgentDriver, createDeterministicSpotCheckLogger, createGatewayResponder, createLightweightAdapter, createLiteLlmProvider, createLocalLlmProvider, createMcpDemoMemCorrectAdapter, createMcpDemoMemoryAdapter, createMcpMemCorrectAdapter, createMcpMemoryAdapter, createMitigatedTarget, createOllamaProvider, createOpenAiCompatibleProvider, createOpenAiResponsesBenchJudge, createOpenAiResponsesProvider, createSeededRandom$1 as createProceduralAblationSeededRandom, createProvider, createProviderBackedAmaBenchRecommendedJudge, createProviderBackedJudge, createProviderBackedResponder, createProviderBackedStructuredJudge, createRemnicAdapter, createRepeatedFailureOllamaChatDriver, createResponderFromProvider, createSeededRandom, createSeededRng$2 as createSeededRng, createSpotCheckFileLogger, createStructuredBenchJudge, createStructuredJudgeFromProvider, createSyntheticEmailIngestionAdapter, createSyntheticTarget, createTimeoutGuardedAdapter, decideRepeatedFailureContent, decideRepeatedFailureStudy, decideRepeatedFailureTiming, defaultBenchmarkBaselineDir, defaultBenchmarkPublishPath, deleteBenchmarkResults, diagnoseLoCoMoRetrievalTraceDelta, diagnoseLoComoProfileDelta, diagnoseLoComoRecallDelta, discoverAllProviders, discoveryEndpointFor, emailFixture, entityRecall, evaluateTaskState, exactMatch, executeLocalRow, extractMetrics as extractCodingGraphMetrics, extractContentWords, extractMarkdownSectionsByTitle, f1Score, fixtureToAblationScenarios, formatHandoffNote, formatMissingDatasetError, generateDriftCorpus, generateFamilyVariants, generateH6BenchmarkDataset, generateReport, generateSuiteVariants, generateSyntheticRepo, getAblationCell, getBenchmark, getBenchmarkLowerIsBetter, getGitSha, getMemoryEvalDimension, getProviderBackedJudgePromptIdentity, getRemnicVersion, getTrapTaxonomyItem, hashBenchmarkArtifact, hashBytes, hashCanonicalJson, hashOrderedQuestionIds, hashString, holmAdjust, injectionSuiteResumeContractHash, integrityMetaIsComplete, interpretEffectSize, isAmaBenchUnknownLikeAnswer, isContaminationEntry, isContaminationManifest, isRepeatedFailureTimidityEquivalent, isSafeSyntheticPath, isSealedQrelsArtifact, isSha256Hex, isStructuredJudgeProvider, isTaskFailed, judgeMemCorrectCorrectionAcceptance, judgeMemCorrectStaleMemoryHarm, lexicalSimilarity, linkMatches, listBenchmarkBaselines, listBenchmarkResults, listBenchmarks, listMemoryEvalBenchmarkIds, listMemoryEvalDimensions, llmJudgeScore, llmJudgeScoreDetailed, loadAblationFixture, loadBaseline, loadBeamDatasetPreview, loadBenchmarkArtifact, loadBenchmarkBaseline, loadBenchmarkReportCardProvenance, loadBenchmarkResult, loadBenchmarkResultSummaries, loadCommittedH6BenchmarkDataset, loadCustomBenchmarkFile, loadJudgeCalibrationState, loadLoCoMo10, loadLocalLabManifest, loadLongMemEvalS, loadSealKeyFromEnv, loadSealedQrels, loadSealedRubric, matchEntity, materializeTaskRepo, mergeContaminationManifests, openSeal, orchestrateBenchmarkRuns, pairedDeltaConfidenceInterval, parseBenchmarkArtifact, parseCanaryFloor, parseCodexJsonlUsage, parseCustomBenchmark, parseLocalLabManifest, parseRepeatedFailureEpisodeRow, parseRubricResponse, parseSealedQrels, pickOne, pickStableQualifiedName, planInjectionSuiteRows, precisionAtK, preflightLoCoMoRetrievalTraceCapture, preflightLocalLabRole, projectFolderFixture, randomInt, recallAtK, reconcileCodexCreditLedger, redactBenchmarkResultSecrets, relativeRiskReduction, renderAttributionReportTable, renderBaselineMarkdown, renderBenchmarkResultExport, renderLoComoProfileDeltaMarkdown, renderLoComoRecallDeltaMarkdown, renderMemorySummaryForJudge, renderMemoryViewForAgent, replayRepeatedFailureStatistics, resolveAssistantAgent, resolveAssistantRubricId, resolveAssistantSeeds, resolveAssistantSpotCheckDir, resolveBenchRuntimeProfile, resolveBenchmarkPhaseTimeoutMs, resolveBenchmarkProgressLogging, resolveBenchmarkResultReference, resolveBenchmarkRunCount, resolveBenchmarkRunId, resolveCanaryFloorFromEnv, resolveCodexCreditBudgetConfig, resolveCommittedH6FixtureDirectory, resolveEffectiveCanaryFloor, resolveLocalLabJudgeProviderConfig, resolveLocalLabProfile, resolveLocalLabRole, resolveStructuredJudge, rotateDistractors, rougeL, runAssistantBenchmark, runAssistantMeetingPrepBenchmark, runAssistantMorningBriefBenchmark, runAssistantNextBestActionBenchmark, runAssistantSynthesisBenchmark, runAttributeCliCommand, runBaseline, runBenchSuite, runBenchmark, runCodingGraphBenchmark, runCustomBenchmarkFile, runDriftGenCliCommand, runExplain, runExtractionAttack, runInjectionSuiteCliCommand, runJudgeCalibration, runMitigatedBaseline, runProceduralAblation, runProceduralAblationCli, runRepeatedFailureCliCommand, runRepeatedFailurePaperReportCliCommand, runRepeatedFailureSuite, runSealedJudge, runSequentialPhases, runTrapAudit, runTrapAuditCliCommand, runWithinCodexCreditBudget, safeHexEqual, sanitizeBenchmarkResultForJson, sanitizeLoComoResultReference, saveBaseline, saveBenchmarkBaseline, schemaCompleteness, sealPayload, selectAmaBenchDiagnosticVariants, selectCalibrationSlice, selectFixtureVariant, serializeAttributionReport, serializeBenchmarkArtifact, serializeBuildWeekEvidenceReceipt, serializeH6FixtureJson, serializeJsonl, serializeLoCoMoRetrievalTraceDelta, serializeLoCoMoRetrievalTraceReceipt, serializeSealedQrels, shuffleTasks, shuffled, summarizeBenchmarkResult, timed, tokenizeContent, unresolvedHelperImports, validateDriftCorpus, validateH6Dataset, validateH6FixtureBundle, validateH6StateDefiningIndependence, validateOllamaChatEndpoint, verifyMatchingTrapAudit, verifyRubricDigest, verifyTrapAuditArtifact, writeBenchmarkArtifact, writeBenchmarkPublishFeed, writeBenchmarkReproManifest, writeBenchmarkResult, writeBuildWeekEvidenceReceipt, writeH6FixtureBundle, writeJudgeCalibrationState, writeLeaderboardArtifactsForResult, writeRepeatedFailurePaperArtifacts, writeRepeatedFailureRunMetadata, writeRepeatedFailureStatistics, zeroScores };
+export { AMA_BENCH_DIAGNOSTIC_VARIANTS, ASSISTANT_AGENT_CONFIG_KEY, ASSISTANT_JUDGE_CONFIG_KEY, ASSISTANT_MEETING_PREP_SCENARIOS, ASSISTANT_MEETING_PREP_SMOKE_SCENARIOS, ASSISTANT_MORNING_BRIEF_SCENARIOS, ASSISTANT_MORNING_BRIEF_SMOKE_SCENARIOS, ASSISTANT_NEXT_BEST_ACTION_SCENARIOS, ASSISTANT_NEXT_BEST_ACTION_SMOKE_SCENARIOS, ASSISTANT_RUBRIC_DIMENSIONS, ASSISTANT_RUBRIC_ID_KEY, ASSISTANT_SEEDS_CONFIG_KEY, ASSISTANT_SPOT_CHECK_DIR_KEY, ASSISTANT_SYNTHESIS_SCENARIOS, ASSISTANT_SYNTHESIS_SMOKE_SCENARIOS, type AblationConfigOverrides, type AbstentionRetrievalCase, type ActionIntentV1, ActionIntentV1Schema, type AggregateMetrics, type AmaBenchDiagnosticAdapterOptions, type AmaBenchDiagnosticAnswererMode, type AmaBenchDiagnosticBreakdown, type AmaBenchDiagnosticMatrixArtifact, type AmaBenchDiagnosticRecallMode, type AmaBenchDiagnosticRunContext, type AmaBenchDiagnosticTaskEvidence, type AmaBenchDiagnosticTaskRow, type AmaBenchDiagnosticVariant, type AmaBenchDiagnosticVariantSummary, type AnalyzeRepeatedFailureOptions, type AnthropicProviderConfig, type AssistantAgent, type AssistantMemoryFact, type AssistantMemoryGraph, type AssistantRubricDimension, type AssistantRubricRequest, type AssistantRubricScores, type AssistantRunnerOptions, type AssistantScenario, type AssistantStance, type AttackRecallOptions, type AttackRetrievalHit, type AttackerMode, type AttributeOptions, type AttributionClass, type AttributionEnvironment, type AttributionLabel, type AttributionMemory, type AttributionReport, BENCHMARK_ARTIFACT_SCHEMA_VERSION, BENCHMARK_INTEGRITY_META_SCHEMA, BENCHMARK_REPRO_MANIFEST_FILENAME, BENCHMARK_REPRO_MANIFEST_SCHEMA_VERSION, BENCHMARK_RESULT_SCHEMA, BENCHMARK_SPLIT_TYPES, BUILD_WEEK_EVIDENCE_RECEIPT_SCHEMA_VERSION, BUILD_WEEK_LIMITATIONS, type BaseTask, BaseTaskSchema, type BaselineRow, type BaselineScenario, type BeamDatasetPreview, type BenchAggregateMetric, type BenchAssistantTaskDetails, type BenchConfig, type BenchIntegritySplit, type BenchIntegritySummary, type BenchJudge, type BenchJudgeResult, type BenchMemoryAdapter, type BenchMetricHighlight, type BenchModelSource, type BenchPerSeedScore, type BenchReasoningEffort, type BenchRecallLineageStatus, type BenchRecallOptions, type BenchRecallSupportAssessment, type BenchRecallSupportRequest, type BenchRecallSupportStatus, type BenchRecallTrace, type BenchRecallTraceCoreCapture, type BenchRecallTraceLcmCandidate, type BenchRecallTraceRange, type BenchRecallTraceSection, type BenchRecallTraceSelection, type BenchRecallWithTraceResult, type BenchResponder, type BenchResponse, type BenchResultFileWarning, type BenchResultSummary, type BenchResultSummaryPayload, type BenchRuntimeProfile, type BenchTaskScoreEntry, type BenchTaskSummary, type BenchTier, type BenchmarkArtifact, type BenchmarkArtifactEnvironment, type BenchmarkArtifactHardware, type BenchmarkArtifactJudgeCalibration, type BenchmarkArtifactPerTaskScore, type BenchmarkArtifactSystem, type BenchmarkArtifactTier, type BenchmarkCategory, type BenchmarkDefinition, type BenchmarkExecutionProvenance, type BenchmarkIntegrityMeta, type BenchmarkMeta, type BenchmarkMode, type BenchmarkReport, type BenchmarkReproManifest, type BenchmarkReproManifestDataset, type BenchmarkReproManifestFile, type BenchmarkReproManifestResult, type BenchmarkReproManifestSupplementalArtifact, type BenchmarkResult, type BenchmarkSplitType, type BenchmarkStatus, type BenchmarkSuiteResult, type BenchmarkTier, type BootstrapKappaOptions, type BootstrapKappaResult, type BuildBenchmarkArtifactInput, type BuildBenchmarkPublishFeedOptions, type BuildBenchmarkReproManifestOptions, type BuildBuildWeekEvidenceReceiptOptions, type BuildWeekEvidenceReceipt, type BuildWeekEvidenceReceiptProvider, type BuildWeekLimitationCode, type BuiltInProvider, CALIBRATION_SLICE_SIZE, CANARY_FIXED_RECALL, CANARY_SCORE_FLOOR, DEFAULT_10K_FIXTURE as CODING_GRAPH_10K_FIXTURE, CODING_GRAPH_BENCH_SCHEMA_VERSION, DEFAULT_TOLERANCE_PERCENT as CODING_GRAPH_DEFAULT_TOLERANCE, MIN_ITERATIONS as CODING_GRAPH_MIN_ITERATIONS, DEFAULT_SMOKE_FIXTURE as CODING_GRAPH_SMOKE_FIXTURE, type CalibrationAnswer, type CalibrationVerdictPair, type CanaryAdapterOptions, type CanaryFloorCheck, type CaptureLoCoMoRetrievalTraceOptions, type ClaudeCliProviderConfig, type CodexCliNativeUsage, type CodexCliProviderConfig, CodexCreditAccountingError, type CodexCreditBudgetConfig, CodexCreditDispatchError, type CodexCreditReceipt, type CodexCreditReceiptScope, type CodexCreditReconciliationReceipt, type CodingGraphBaseline, type CodingGraphBenchConfig, type CodingGraphBenchReport, type MachineFingerprint as CodingGraphMachineFingerprint, type CodingGraphMetricKey, type RegressionMetricDetail as CodingGraphRegressionDetail, type RegressionMetricKey as CodingGraphRegressionKey, type RegressionGateResult as CodingGraphRegressionResult, type CohenKappaResult, type ComparisonMetricDelta, type ComparisonResult, type CompletionOpts, type CompletionResult, type ConfidenceInterval, type ContaminationCheckResult, type ContaminationEntry, type ContaminationManifest, type ControlledGateDecision, type ControlledResponsesAgentDriverConfig, type ControlledResponsesCaps, type ControlledResponsesDisposition, ControlledResponsesDriver, type ControlledResponsesDriverConfig, type ControlledResponsesEpisodeInput, type ControlledResponsesEpisodeResult, type ControlledResponsesFault, type ControlledResponsesResponseEvent, type ControlledResponsesToolDefinition, type ControlledResponsesToolEvent, type ControlledResponsesTransport, type CustomBenchmarkScoring, type CustomBenchmarkSpec, type CustomBenchmarkTask, DATASET_SPLITS, DEFAULT_ABLATION_BENCHMARK, DEFAULT_ABLATION_BOOTSTRAP_SEED, DEFAULT_ASSISTANT_RUBRIC_ID, DEFAULT_ATTACKER_PROMPT_PATH_GETTER, DEFAULT_BASELINE_SCENARIOS, DEFAULT_JUDGE_BINARIZATION_THRESHOLD, DEFAULT_KAPPA_BOOTSTRAP_SAMPLES, DEFAULT_KAPPA_CONFIDENCE_LEVEL, DEFAULT_OPENAI_RESPONSES_JUDGE_MODEL, DRIFT_GEN_DEFAULTS, DRIFT_GEN_VERSION, type DatasetSource, type DatasetSplit, type DiagnoseLoComoProfileDeltaOptions, type DiagnoseLoComoRecallDeltaOptions, type DiscoveredModel, type DriftGenAuditRecord, type DriftGenCorpus, type DriftGenManifest, type DriftGenOptions, type DriftGenResult, type DriftSession, type DriftSessionTurn, type DriftValidationReport, type DriftValidationStats, EMPTY_CONTAMINATION_MANIFEST, type EffectSizeInterpretation, type EffectSizeSummary, type EvaluateTaskStateOptions, type ExplainResult, type ExtractedEntity, type ExtractedLink, type ExtractedPage, type ExtractionAttackOptions, type ExtractionAttackResult, type ExtractionAttackTarget, type FixtureGenerator, type FixtureOutput, type FixtureVariant, GATE_STATUSES, GENERAL_ANSWER_JUDGE_RUBRIC, type GateStatus, type GeneratedFile, type GeneratedRepo, type GoldEntity, type GoldEntityType, type GoldFact, type GoldFactKind, type GoldGraph, type GoldLink, type GoldMemoryAttribution, type GoldPage, type GoldProbe, type GoldProbeCategory, H5_PUBLICATION_ANALYSIS_RULE, H5_PUBLICATION_DEFENSE_BASELINES, type H6BenchmarkDataset, H6BenchmarkDatasetSchema, type H6TrapId, H6_ACTION_INTENT_JSON_SCHEMA, H6_ARMS, H6_DATASET_JSON_SCHEMA, H6_DECISION_RULE, H6_FROZEN_INVENTORY_HASH, H6_FROZEN_SEED, H6_FROZEN_SPLITS, H6_SUPPORT_ARTIFACT_PATHS, H6_TASK_JSON_SCHEMA, H6_TRAP_FINGERPRINT_JSON_SCHEMA, H6_TRAP_IDS, HOST_FAULT_RETRY_LIMIT, type HarnessRng, INJECTION_SUITE_ARMS, INJECTION_SUITE_FAMILIES, INJECTION_SUITE_PUBLICATION_ARMS, INJECTION_SUITE_STAGES, INJECTION_SUITE_VERSION, INTEGRITY_CIPHER_ALGORITHM, INTEGRITY_HASH_ALGORITHM, INTEGRITY_META_FIELDS, INVENTED_DOMAINS, type IngestionBenchAdapter, type IngestionLog, type InjectionSuiteArm, type InjectionSuiteCliInput, type InjectionSuiteCliResult, type InjectionSuiteEpisodeRow, type InjectionSuiteFamily, type InjectionSuiteOnlineAdaptiveInput, type InjectionSuiteProductEvidence, type InjectionSuiteRowIdentity, type InjectionSuiteStage, JUDGE_CALIBRATION_KAPPA_THRESHOLD, JUDGE_CALIBRATION_PROTOCOL_VERSION, type JudgeCalibrationCheckpointProvenance, type JudgeCalibrationIdentities, type JudgeCalibrationResult, type JudgeCategory, type KappaConfidenceInterval, LOCAL_LAB_PROVIDER_KINDS, LOCOMO_DATASET_FILENAMES, LOCOMO_FULL_TASK_COUNT, LOCOMO_RECALL_DIFF_LINE_LIMIT, LOCOMO_RECALL_EXCERPT_CHARS, LOCOMO_RETRIEVAL_TRACE_BUDGET_VERSION, LOCOMO_RETRIEVAL_TRACE_DELTA_SCHEMA_VERSION, LOCOMO_RETRIEVAL_TRACE_SCHEMA_VERSION, LOCOMO_RETRIEVAL_TRACE_SELECTION_VERSION, LONG_MEM_EVAL_DATASET_FILENAMES, type LeaderboardArtifactWrite, type LettaAdapterConfig, LettaMemCorrectAdapter, type LlmJudge, type LlmProvider, type LoCoMoCategory, type LoCoMoRetrievalMechanism, type LoCoMoRetrievalMechanismSummary, type LoCoMoRetrievalSessionReceipt, type LoCoMoRetrievalStructuralTrace, type LoCoMoRetrievalTaskDelta, type LoCoMoRetrievalTaskReceipt, type LoCoMoRetrievalTraceCoreCaptureReceipt, type LoCoMoRetrievalTraceDeltaReport, type LoCoMoRetrievalTraceProfile, type LoCoMoRetrievalTraceReceipt, type LoCoMoRetrievalTraceSelectionManifest, type LoCoMoRetrievalTraceSelector, type LoCoMoStructuralMultisetDelta, type LoComoCategoryDelta, type LoComoFinalContextRegression, type LoComoMetricDelta, type LoComoProfileArtifactEvidence, type LoComoProfileDeltaReport, type LoComoRawResultEvidence, type LoComoRecallCategoryDelta, type LoComoRecallContextSummary, type LoComoRecallDeltaReport, type LoComoRecallLineDelta, type LoComoRecallLineEvidence, type LoComoRecallMetricDelta, type LoComoRecallResultProvenance, type LoComoRecallTextDigest, type LoComoTaskRegression, type LoadDatasetOptions, type LoadSealedQrelsOptions, type LoadedDataset, type LoadedJudgeCalibrationState, type LocalLabManifest, type LocalLabManifestNotes, type LocalLabPhase, type LocalLabPhaseDescriptor, type LocalLabPhaseExecute, type LocalLabPhaseName, type LocalLabPhaseOutcome, LocalLabPreflightError, type LocalLabPreflightFailure, type LocalLabPreflightInput, type LocalLabPreflightOptions, type LocalLabPreflightResult, type LocalLabPreflightSuccess, type LocalLabProviderKind, type LocalLabRoleConfig, type LocalLlmProviderConfig, MAX_ROW_ATTEMPTS, MEMCORRECT_CORRECTION_ACCEPTANCE_RUBRIC, MEMCORRECT_CORRECTION_ACCEPTANCE_RUBRIC_VERSION, MEMCORRECT_STALE_HARM_RUBRIC, MEMCORRECT_STALE_HARM_RUBRIC_VERSION, MEMORY_EVAL_DIMENSIONS, MEMORY_EVAL_PUBLIC_LINE, MIN_CALIBRATION_SOURCE_TASKS, MITIGATED_BASELINE_SCENARIOS, type MaterializeOptions, type MaterializedRepo, type McpArgumentSemantic, type McpBackendErrorCode, type McpBackendResult, type McpBenchMemoryAdapter, type McpConformanceResult, type McpHttpTransportConfig, type McpListedTool, type McpMemCorrectAdapter, type McpMemoryAdapterOptions, McpMemoryBackendError, type McpMemoryToolMapping, type McpMemoryTransportConfig, type McpStdioTransportConfig, type McpToolCallResult, type McpToolClient, type McpToolMappingEntry, type McpToolMappingValue, type McpToolOperation, type Mem0AdapterConfig, Mem0MemCorrectAdapter, type MemCorrectGeneratorOptions, type MemCorrectJudgeRequest, type MemCorrectJudgeResult, type MemCorrectSystemAdapter, type MemoryEvalCategory, type MemoryEvalDimension, type MemoryEvalDimensionId, type MemoryEvalMetric, type MemoryGraph, type MemoryStats, type MemorySystem, type Message, type MetricAggregate, type MicroMetric, MissingCredentialError, type MitigatedBaselineConfig, type MitigatedTargetConfig, type MultipleChoiceQuestion, ONLINE_ADAPTIVE_STAGE, OPENAI_RESPONSES_JUDGE_RUBRIC_VERSION, OTHER_NAMESPACE_MEMORIES, type OllamaChatMessage, type OllamaProviderConfig, type OnlineAdaptiveCorpusLine, type OnlineAdaptiveCorpusManifest, type OnlineAdaptiveStatistics, type OpenAiCompatibleProviderConfig, OpenAiResponsesJudgeError, type OpenAiResponsesJudgeErrorCode, type OpenAiResponsesJudgeTelemetry, OpenAiResponsesProvider, type OpenAiResponsesProviderConfig, type OpenAiResponsesVerdict, type OpenAiResponsesVerdictResult, PROCEDURAL_REAL_SCENARIOS, PROCEDURAL_REAL_SCENARIOS_SMOKE, PUBLISHED_BENCHMARK_ARTIFACT_IDS, type PairedAnswerReplayCache, type PairedAnswerReplayEntry, type ParsedOllamaChatResponse, type PersonalizationRetrievalCase, type PreflightDiscoveredModel, type ProceduralAblationArtifact, type ProceduralAblationPerCase, type ProceduralAblationScenario, type ProceduralRealScenario, type ProceduralRealScenarioCategory, type ProviderBaseConfig, type ProviderConfig, type ProviderDiscoveryResult, type ProviderFactoryConfig, type PublishSkipReason, type PublishSkipRecord, type PublishedBenchmarkFeed, type PublishedBenchmarkFeedEntry, type PublishedBenchmarkId, REPEATED_FAILURE_ARMS, REPEATED_FAILURE_CONFIDENCE_LEVEL, REPEATED_FAILURE_INVALID_REASONS, REPEATED_FAILURE_STATISTICS_DRAWS, REQUIRED_FRONTMATTER_FIELDS, type RecallMetrics, type RecoveredMemory, type RegressionDetail, type RegressionGateResult$1 as RegressionGateResult, type RemnicAdapterOptions, type RepeatedFailureActionEvaluator, type RepeatedFailureArm, type RepeatedFailureCheckpointLoadResult, type RepeatedFailureCliCommandResult, type RepeatedFailureEffectAnalysis, type RepeatedFailureEpisode, type RepeatedFailureEpisodeDriver, type RepeatedFailureEpisodeEvidence, type RepeatedFailureEpisodeInput, type RepeatedFailureEpisodeRow, type RepeatedFailureExpectedDesign, type RepeatedFailureFactPairAudit, type RepeatedFailureFinalRepoEvidence, type RepeatedFailureFinalState, type RepeatedFailureGateEvent, type RepeatedFailureHolmResult, type RepeatedFailureInterval, type RepeatedFailureInvalidReason, type RepeatedFailureIsolationIdentity, type RepeatedFailureLocalToolHost, type RepeatedFailureNullableInterval, RepeatedFailureOllamaChatDriver, type RepeatedFailureOllamaChatDriverConfig, type RepeatedFailureProposedAction, type RepeatedFailureRowCheckpoint, type RepeatedFailureRowClaim, type RepeatedFailureRowIdentity, RepeatedFailureRowStore, type RepeatedFailureRowStoreOptions, type RepeatedFailureRunMetadata, type RepeatedFailureStatisticalAnalysis, type RepeatedFailureSuiteManifest, type RepeatedFailureSupportDecision, type RepeatedFailureTaskCut, type RepeatedFailureTimidityAnalysis, type RepeatedFailureTokenUsage, type RepeatedFailureTokenizer, type RepeatedFailureToolDefinition, type RepeatedFailureToolExecutionResult, type RepeatedFailureTrapAuditArtifact, type RepeatedFailureTrapAuditExpected, type RepeatedFailureTrapAuditMetrics, type RepeatedFailureTrapAuditRow, type RepeatedFailureTrapAuditRowIdentity, type RepeatedFailureTrapAuditThresholds, type RepeatedFailureTry, type ReplayRepeatedFailureStatisticsOptions, type ReportCardProvenanceContext, type ResolveBenchRuntimeProfileOptions, type ResolvedBenchRuntimeProfile, type ResolvedLocalLabProfile, type ResolvedLocalLabRole, type ResolvedRunBenchmarkOptions, type ResponsesApiOutputItem, type ResponsesApiRequest, type ResponsesApiResponse, type ResponsesApiUsage, type RetrievalMissStage, type RevisionShas, type RotatedChoices, type RunBenchmarkOptions, type RunJudgeCalibrationOptions, type RunProceduralAblationCliArgs, type RunProceduralAblationOptions, type RunRepeatedFailureCliCommandInput, type RunRepeatedFailureSuiteOptions, type RunRepeatedFailureSuiteResult, type RunSequentialPhasesOptions, type RunTrapAuditOptions, SCHEMA_TIER_FIXTURE, SCHEMA_TIER_SMOKE_FIXTURE, SEALED_PROMPT_REGISTRY, SINGLE_FLAG_ABLATION_MATRIX, STATE_CLASSIFICATIONS, SYNTHETIC_MEMORIES, type SanitizedDiagnosticProvider, type SavedBaseline, type SchemaTierCorpus, type SchemaTierFixture, type SchemaTierName, type SchemaTierPage, type SchemaTierPageFrontmatter, type SealedArtifact, type SealedJudgeDecision, type SealedJudgeInput, type SealedQrelsArtifact, type SealedQrelsHandle, type SealedRubric, type SearchResult, type SeededMemory, type SeededRandom, type SeededRng, type SequentialPhaseHooks, type SingleFlagAblationCell, type SingleFlagAblationId, type SpotCheckLogger, type StageObservation, type StageStatus, type StateClassification, type StateEvaluationResult, type StatisticalReport, type StrategyPatch, StrategyPatchSchema, type StructuredJudge, StructuredJudgeError, type StructuredJudgeErrorCode, type StructuredJudgeProvider, type StructuredJudgeTelemetry, type StructuredJudgeVerdict, type StructuredJudgeVerdictResult, type StructuredVerdictRequest, type SyntheticEdge, type SyntheticEmailIngestionAdapterOptions, type SyntheticFile, type SyntheticFileIR, SyntheticFileSchema, type SyntheticRepoConfig, type SyntheticSymbol, type SyntheticTargetOptions, TRAP_TAXONOMY, type TaskAttribution, type TaskAttributionGoldWitnessV1, type TaskAttributionRetrievalWitnessV1, type TaskAttributionWitness, type TaskAttributionWitnessRuntimeV1, type TaskAttributionWitnessV1, type TaskResult, type TaskTokenUsage, type TaskVariant, TaskVariantSchema, type TemporalRetrievalCase, type ThirdPartyAdapterConfig, type TierDetail, type TimelineEntry, type TokenUsage, type TrapFingerprintV1, TrapFingerprintV1Schema, type TrapTaxonomyItem, TrapTaxonomyItemSchema, type ValidationIssue, type ValidationReport, type WallMetric, type WriteBenchmarkArtifactResult, type ZepAdapterConfig, ZepMemCorrectAdapter, addContaminationEntry, aggregateTaskScores, analyzeInjectionSuiteOnlineAdaptiveRun, analyzeInjectionSuitePublicationRows, analyzeInjectionSuitePublicationRun, analyzeInjectionSuitePublicationUtility, analyzeInjectionSuitePublicationUtilityFile, analyzeInjectionSuiteRun, analyzeInjectionSuiteUtility, analyzeRepeatedFailureRows, answerBenchmarkQuestion, applyPatchAndCommit, assertCanaryUnderFloor, assertIntegrityMetaPresent, assertPublishableIntegrity, assertSha256Hex, assertTrapDatasetPreflight, assistantMeetingPrepDefinition, assistantMorningBriefDefinition, assistantNextBestActionDefinition, assistantSynthesisDefinition, attributeGoldMemory, attributeRun, attributeTask, backlinkF1, binarizeJudgeScore, bootstrapCohensKappaConfidenceInterval, bootstrapMeanConfidenceInterval, buildAmaBenchDiagnosticMatrixArtifact, buildAmaBenchDiagnosticVariantSummary, buildAmaBenchLeaderboardRows, buildBaselineFromReport, buildBenchmarkArtifact, buildBenchmarkArtifactFilename, buildBenchmarkPublishFeed, buildBenchmarkReproManifest, buildBenchmarkRunSeeds, buildBuildWeekEvidenceReceipt, buildCodexCreditReceipt, buildDriftCorpus, buildJudgePayload, buildOracleTrajectoryRecall, buildProviderFreeLoCoMoRetrievalConfig, buildRepeatedFailureRowKey, buildSchemaTierFixture, buildSchemaTierSmokeFixture, calculateCodexBudgetUnits, calculateJaccardSimilarity, calendarFixture, canonicalJsonStringify, captureBenchmarkExecutionProvenance, captureLoCoMoRetrievalTrace, captureMachineFingerprint, chatFixture, checkCodingGraphRegression, checkDatasetContamination, checkRegression, clampScore, cohensD, compareResults, computeBenchmarkReproDatasetInventoryHash, computeBenchmarkReproManifestArtifactHash, computeCohensKappa, computeH6InventoryHash, computeH6SupportArtifactHashes, computeRevisionShas, computeSealHash, computeTrapAuditArtifactHash, computeTrapAuditMetrics, containsAnswer, createSeededRng$1 as createAdamSeededRng, createAmaBenchDiagnosticAdapter, createAnthropicProvider, createCanaryAdapter, createClaudeCliProvider, createCodexCliProvider, createSeededRng as createCodingGraphSeededRng, createControlledResponsesAgentDriver, createDeterministicSpotCheckLogger, createGatewayResponder, createLightweightAdapter, createLiteLlmProvider, createLocalLlmProvider, createMcpDemoMemCorrectAdapter, createMcpDemoMemoryAdapter, createMcpMemCorrectAdapter, createMcpMemoryAdapter, createMitigatedTarget, createOllamaProvider, createOpenAiCompatibleProvider, createOpenAiResponsesBenchJudge, createOpenAiResponsesProvider, createSeededRandom$1 as createProceduralAblationSeededRandom, createProvider, createProviderBackedAmaBenchRecommendedJudge, createProviderBackedJudge, createProviderBackedResponder, createProviderBackedStructuredJudge, createRemnicAdapter, createRepeatedFailureOllamaChatDriver, createResponderFromProvider, createSeededRandom, createSeededRng$2 as createSeededRng, createSpotCheckFileLogger, createStructuredBenchJudge, createStructuredJudgeFromProvider, createSyntheticEmailIngestionAdapter, createSyntheticTarget, createTimeoutGuardedAdapter, decideInjectionSuiteCampaign, decideRepeatedFailureContent, decideRepeatedFailureStudy, decideRepeatedFailureTiming, defaultBenchmarkBaselineDir, defaultBenchmarkPublishPath, deleteBenchmarkResults, diagnoseLoCoMoRetrievalTraceDelta, diagnoseLoComoProfileDelta, diagnoseLoComoRecallDelta, discoverAllProviders, discoveryEndpointFor, emailFixture, entityRecall, evaluateTaskState, exactMatch, executeLocalRow, extractMetrics as extractCodingGraphMetrics, extractContentWords, extractMarkdownSectionsByTitle, f1Score, fixtureToAblationScenarios, formatHandoffNote, formatMissingDatasetError, generateDriftCorpus, generateFamilyVariants, generateH6BenchmarkDataset, generateReport, generateSuiteVariants, generateSyntheticRepo, getAblationCell, getBenchmark, getBenchmarkLowerIsBetter, getGitSha, getMemoryEvalDimension, getProviderBackedJudgePromptIdentity, getRemnicVersion, getTrapTaxonomyItem, hashBenchmarkArtifact, hashBytes, hashCanonicalJson, hashOrderedQuestionIds, hashString, holmAdjust, injectionSuiteResumeContractHash, integrityMetaIsComplete, interpretEffectSize, isAmaBenchUnknownLikeAnswer, isContaminationEntry, isContaminationManifest, isRepeatedFailureTimidityEquivalent, isSafeSyntheticPath, isSealedQrelsArtifact, isSha256Hex, isStructuredJudgeProvider, isTaskFailed, judgeMemCorrectCorrectionAcceptance, judgeMemCorrectStaleMemoryHarm, lexicalSimilarity, linkMatches, listBenchmarkBaselines, listBenchmarkResults, listBenchmarks, listMemoryEvalBenchmarkIds, listMemoryEvalDimensions, llmJudgeScore, llmJudgeScoreDetailed, loadAblationFixture, loadBaseline, loadBeamDatasetPreview, loadBenchmarkArtifact, loadBenchmarkBaseline, loadBenchmarkReportCardProvenance, loadBenchmarkResult, loadBenchmarkResultSummaries, loadCommittedH6BenchmarkDataset, loadCustomBenchmarkFile, loadJudgeCalibrationState, loadLoCoMo10, loadLocalLabManifest, loadLongMemEvalS, loadSealKeyFromEnv, loadSealedQrels, loadSealedRubric, matchEntity, materializeTaskRepo, mergeContaminationManifests, openSeal, orchestrateBenchmarkRuns, pairedDeltaConfidenceInterval, parseBenchmarkArtifact, parseCanaryFloor, parseCodexJsonlUsage, parseCustomBenchmark, parseLocalLabManifest, parseRepeatedFailureEpisodeRow, parseRubricResponse, parseSealedQrels, pickOne, pickStableQualifiedName, planInjectionSuiteRows, planOnlineAdaptiveRows, precisionAtK, preflightLoCoMoRetrievalTraceCapture, preflightLocalLabRole, projectFolderFixture, randomInt, recallAtK, reconcileCodexCreditLedger, redactBenchmarkResultSecrets, relativeRiskReduction, renderAttributionReportTable, renderBaselineMarkdown, renderBenchmarkResultExport, renderLoComoProfileDeltaMarkdown, renderLoComoRecallDeltaMarkdown, renderMemorySummaryForJudge, renderMemoryViewForAgent, replayInjectionSuiteStatistics, replayRepeatedFailureStatistics, resolveAssistantAgent, resolveAssistantRubricId, resolveAssistantSeeds, resolveAssistantSpotCheckDir, resolveBenchRuntimeProfile, resolveBenchmarkPhaseTimeoutMs, resolveBenchmarkProgressLogging, resolveBenchmarkResultReference, resolveBenchmarkRunCount, resolveBenchmarkRunId, resolveCanaryFloorFromEnv, resolveCodexCreditBudgetConfig, resolveCommittedH6FixtureDirectory, resolveEffectiveCanaryFloor, resolveLocalLabJudgeProviderConfig, resolveLocalLabProfile, resolveLocalLabRole, resolveStructuredJudge, rotateDistractors, rougeL, runAssistantBenchmark, runAssistantMeetingPrepBenchmark, runAssistantMorningBriefBenchmark, runAssistantNextBestActionBenchmark, runAssistantSynthesisBenchmark, runAttributeCliCommand, runBaseline, runBenchSuite, runBenchmark, runCodingGraphBenchmark, runCustomBenchmarkFile, runDriftGenCliCommand, runExplain, runExtractionAttack, runInjectionSuiteCliCommand, runInjectionSuiteOnlineAdaptive, runInjectionSuiteUtility, runJudgeCalibration, runMitigatedBaseline, runProceduralAblation, runProceduralAblationCli, runRepeatedFailureCliCommand, runRepeatedFailurePaperReportCliCommand, runRepeatedFailureSuite, runSealedJudge, runSequentialPhases, runTrapAudit, runTrapAuditCliCommand, runWithinCodexCreditBudget, safeHexEqual, sanitizeBenchmarkResultForJson, sanitizeLoComoResultReference, saveBaseline, saveBenchmarkBaseline, schemaCompleteness, sealPayload, selectAmaBenchDiagnosticVariants, selectCalibrationSlice, selectFixtureVariant, serializeAttributionReport, serializeBenchmarkArtifact, serializeBuildWeekEvidenceReceipt, serializeH6FixtureJson, serializeJsonl, serializeLoCoMoRetrievalTraceDelta, serializeLoCoMoRetrievalTraceReceipt, serializeSealedQrels, shuffleTasks, shuffled, stripCodeFences, summarizeBenchmarkResult, timed, tokenizeContent, unresolvedHelperImports, validateDriftCorpus, validateH6Dataset, validateH6FixtureBundle, validateH6StateDefiningIndependence, validateOllamaChatEndpoint, verifyMatchingTrapAudit, verifyRubricDigest, verifyTrapAuditArtifact, writeBenchmarkArtifact, writeBenchmarkPublishFeed, writeBenchmarkReproManifest, writeBenchmarkResult, writeBuildWeekEvidenceReceipt, writeH6FixtureBundle, writeJudgeCalibrationState, writeLeaderboardArtifactsForResult, writeRepeatedFailurePaperArtifacts, writeRepeatedFailureRunMetadata, writeRepeatedFailureStatistics, zeroScores };