@tangle-network/agent-eval 0.200.0 → 0.201.0

package/dist/trace-repair/index.d.ts

@@ -7,10 +7,10 @@ import { t as PrimeBridgeTransport } from "../prime-bridge-transport-6feEglLf.js
 import { a as RecordedTrajectoryStep, h as isRecordedTimeout } from "../steps-Bg5-KVx1.js";
 //#region src/trace-repair/mini-swe-scaffold.d.ts
 /** Whole-line marker that ends a run. The first output line must equal it and the command must exit 0. */
-declare const SUBMIT_SENTINEL = "COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT";
+export declare const SUBMIT_SENTINEL = "COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT";
 /** Outputs at or above this length are elided head+tail instead of shown whole. */
-declare const OUTPUT_ELISION_THRESHOLD = 10000;
-declare const MINI_SWE_SYSTEM_MESSAGE = "You are a helpful assistant that can interact with a computer.\n\nYour response must contain exactly ONE bash code block with ONE command (or commands connected with && or ||).\nInclude a THOUGHT section before your command where you explain your reasoning process.\nFormat your response as shown in <format_example>.\n\n<format_example>\nYour reasoning and analysis here. Explain why you want to perform the action.\n\n```bash\nyour_command_here\n```\n</format_example>\n\nFailure to follow these rules will cause your response to be rejected.\n";
+export declare const OUTPUT_ELISION_THRESHOLD = 10000;
+export declare const MINI_SWE_SYSTEM_MESSAGE = "You are a helpful assistant that can interact with a computer.\n\nYour response must contain exactly ONE bash code block with ONE command (or commands connected with && or ||).\nInclude a THOUGHT section before your command where you explain your reasoning process.\nFormat your response as shown in <format_example>.\n\n<format_example>\nYour reasoning and analysis here. Explain why you want to perform the action.\n\n```bash\nyour_command_here\n```\n</format_example>\n\nFailure to follow these rules will cause your response to be rejected.\n";
 interface InstanceMessageInput {
   /** The task statement, substituted verbatim. */
   task: string;
@@ -21,7 +21,7 @@ interface InstanceMessageInput {
   systemInformation: string;
 }
 /** The second message of every recorded run: task plus workflow rules. */
-declare function renderInstanceMessage(input: InstanceMessageInput): string;
+export declare function renderInstanceMessage(input: InstanceMessageInput): string;
 type ParsedAction = {
   kind: 'action';
   command: string;
@@ -34,7 +34,7 @@ type ParsedAction = {
  * The scaffold trims the command, so a block padded with blank lines executes
  * the same command as an unpadded one.
  */
-declare function parseAction(assistantMessage: string): ParsedAction;
+export declare function parseAction(assistantMessage: string): ParsedAction;
 interface CommandOutput {
   returncode: number;
   output: string;
@@ -46,17 +46,17 @@ interface CommandOutput {
  * whole; long ones keep the first and last `OUTPUT_ELISION_WINDOW` characters
  * with the dropped count between them.
  */
-declare function renderObservation(output: CommandOutput): string;
+export declare function renderObservation(output: CommandOutput): string;
 /** The observation after the environment killed a command for exceeding its timeout. */
-declare function renderTimeoutObservation(command: string, partialOutput: string): string;
+export declare function renderTimeoutObservation(command: string, partialOutput: string): string;
 /** The observation after a turn that did not contain exactly one bash block. */
-declare function renderFormatErrorObservation(actionCount: number): string;
+export declare function renderFormatErrorObservation(actionCount: number): string;
 /**
  * The submission text when this output ends the run, `null` otherwise.
  * A non-zero exit does not submit even when the sentinel is echoed, so an
  * agent cannot end the run through a command that failed.
  */
-declare function submissionOf(output: CommandOutput): string | null;
+export declare function submissionOf(output: CommandOutput): string | null;
 //#endregion
 //#region src/trace-repair/action-budget.d.ts
 /**
@@ -92,14 +92,14 @@ interface InterventionBudget {
   readonly maxHeredocs: number;
 }
 /** The scaffold's own per-action budget, pre-registered for the campaign. */
-declare const SCAFFOLD_INTERVENTION_BUDGET: InterventionBudget;
+export declare const SCAFFOLD_INTERVENTION_BUDGET: InterventionBudget;
 /**
  * Actions whose only effect is to consume a turn. They are rejected before a
  * container opens: an intervention that changes nothing is measurably
  * identical to the no-op control, and paying rollouts to rediscover that
  * wastes the corpus.
  */
-declare const NO_OP_ACTIONS: readonly string[];
+export declare const NO_OP_ACTIONS: readonly string[];
 /**
  * Split a shell script into top-level statements and count its heredocs.
  *
@@ -109,13 +109,13 @@ declare const NO_OP_ACTIONS: readonly string[];
  * action reads as one oversized statement and is rejected on bytes rather
  * than silently accepted as one clean action.
  */
-declare function scanShellAction(script: string): {
+export declare function scanShellAction(script: string): {
   statements: string[];
   heredocs: number;
 };
 /** Payload shape of an action: `edit` when it authors file content through a
  *  heredoc, `shell` otherwise. */
-declare function classifyActionPayload(action: string): ActionPayloadKind;
+export declare function classifyActionPayload(action: string): ActionPayloadKind;
 /** Why an action fails the budget. Every value is a rejection the report
  *  counts by name; none of them is a scoring judgement. */
 type BudgetViolation = 'empty' | 'over-byte-cap' | 'multiple-statements' | 'multiple-heredocs' | 'no-op-action' | 'submit-instead-of-repair';
@@ -150,13 +150,13 @@ type BudgetCheck = {
  * how it described a repair rather than on the repair. A reader who wants the
  * mismatch counts `declared` against `payload`.
  */
-declare function checkInterventionBudget(action: string, declaredKind: ActionPayloadKind, budget?: InterventionBudget): BudgetCheck;
+export declare function checkInterventionBudget(action: string, declaredKind: ActionPayloadKind, budget?: InterventionBudget): BudgetCheck;
 /**
  * Whitespace-insensitive comparison used to reject an intervention that is
  * the recorded action again. Trailing whitespace and blank lines are the only
  * differences a re-proposal can carry without changing what runs.
  */
-declare function normalizeActionForComparison(action: string): string;
+export declare function normalizeActionForComparison(action: string): string;
 //#endregion
 //#region src/trace-repair/continuation-records.d.ts
 /** Which arm produced a rollout. The policy is identical across all three. */
@@ -272,9 +272,9 @@ interface RecordedStep {
  * with no answer keeps `obs: null`, which is how the corpus records a run that
  * ended on its last command.
  */
-declare function toRecordedSteps(messages: readonly ContinuationMessage[]): RecordedStep[];
+export declare function toRecordedSteps(messages: readonly ContinuationMessage[]): RecordedStep[];
 /** Corpus steps for the continuation alone, excluding the prefix it inherited. */
-declare function rolloutRecordedSteps(rollout: ContinuationRollout): RecordedStep[];
+export declare function rolloutRecordedSteps(rollout: ContinuationRollout): RecordedStep[];
 /**
  * Hash over everything the policy determines: the actions taken, the
  * observations they produced, the seeds, the exit, and the usage.
@@ -283,11 +283,11 @@ declare function rolloutRecordedSteps(rollout: ContinuationRollout): RecordedSte
  * two rollouts with the same digest did the same work, whatever they cost in
  * seconds. Use it to assert determinism, never to assert equal latency.
  */
-declare function rolloutDigest(rollout: ContinuationRollout): string;
+export declare function rolloutDigest(rollout: ContinuationRollout): string;
 //#endregion
 //#region src/trace-repair/control-policy.d.ts
 /** The declared control cannot produce the outcome the criteria screen for. */
-declare class UncalibratedControlError extends CaptureIntegrityError {}
+export declare class UncalibratedControlError extends CaptureIntegrityError {}
 /**
  * What the criteria expect of the control.
  *
@@ -299,7 +299,7 @@ declare class UncalibratedControlError extends CaptureIntegrityError {}
  *   disagreed with itself about identical state, and the row leaves as that.
  */
 type ControlScreening = 'enforced' | 'declared-inert';
-declare const CONTROL_SCREENING_MODES: readonly ControlScreening[];
+export declare const CONTROL_SCREENING_MODES: readonly ControlScreening[];
 interface ControlPolicyInput {
   /** Stable name for the frozen configuration. */
   readonly id: string;
@@ -342,8 +342,8 @@ interface ControlPolicy extends ControlPolicyInput, ControlCapability {
  * it executes only what a model call asks for. At a zero budget it grades the
  * bytes it was handed.
  */
-declare function controlCanRescue(stepBudget: number): boolean;
-declare function defineControlPolicy(input: ControlPolicyInput): ControlPolicy;
+export declare function controlCanRescue(stepBudget: number): boolean;
+export declare function defineControlPolicy(input: ControlPolicyInput): ControlPolicy;
 /**
  * Refuse a configuration whose control and screening mode contradict.
  *
@@ -352,13 +352,13 @@ declare function defineControlPolicy(input: ControlPolicyInput): ControlPolicy;
  * fire. A control declared inert that can in fact act hides a real screen
  * behind a label that says nothing was screened.
  */
-declare function assertControlCalibrated(policy: ControlCapability, screening: ControlScreening): void;
+export declare function assertControlCalibrated(policy: ControlCapability, screening: ControlScreening): void;
 //#endregion
 //#region src/trace-repair/admission-records.d.ts
 /** A campaign scored rows the pre-pass did not admit, or dropped rows it did. */
-declare class AdmissionDenominatorError extends CaptureIntegrityError {}
+export declare class AdmissionDenominatorError extends CaptureIntegrityError {}
 /** A row handed to the pre-pass carried a field only an analyst could produce. */
-declare class AdmissionIndependenceError extends CaptureIntegrityError {}
+export declare class AdmissionIndependenceError extends CaptureIntegrityError {}
 /**
  * One corpus row, described by what the recording holds and nothing else.
  *
@@ -392,7 +392,7 @@ interface AdmissionRow {
  * names, because the failure to catch is "some new analyst output leaked into
  * the gate", and only a closed shape catches the fields nobody thought of.
  */
-declare function assertAnalystIndependent(rows: readonly AdmissionRow[]): void;
+export declare function assertAnalystIndependent(rows: readonly AdmissionRow[]): void;
 /**
  * Failure population a row belongs to.
  *
@@ -405,9 +405,9 @@ declare function assertAnalystIndependent(rows: readonly AdmissionRow[]): void;
  *   so this population stays separate from the other two.
  */
 type AdmissionStratum = 'clean-exit' | 'command-error' | 'signal-kill';
-declare const ADMISSION_STRATA: readonly AdmissionStratum[];
+export declare const ADMISSION_STRATA: readonly AdmissionStratum[];
 /** `null` when the recording holds no parseable final return code. */
-declare function stratumOf(finalReturncode: number | null): AdmissionStratum | null;
+export declare function stratumOf(finalReturncode: number | null): AdmissionStratum | null;
 /**
  * Why a row left the funnel.
  *
@@ -416,9 +416,9 @@ declare function stratumOf(finalReturncode: number | null): AdmissionStratum | n
  * first container, and the control rollouts run last.
  */
 type AdmissionExclusionReason = 'no-recorded-commands' | 'unparseable-final-returncode' | 'stratum-not-admitted' | 'task-oracle-uncertified' | 'task-oracle-nondeterministic' | 'prefix-replay-error' | 'prefix-replay-empty' | 'prefix-replay-truncated' | 'prefix-divergence-above-threshold' | 'end-state-oracle-error' | 'end-state-tests-pass' | 'no-fix-control-error' | 'no-fix-control-rescued' | 'no-op-control-error' | 'no-op-control-rescued';
-declare const ADMISSION_EXCLUSION_ORDER: readonly AdmissionExclusionReason[];
+export declare const ADMISSION_EXCLUSION_ORDER: readonly AdmissionExclusionReason[];
 /** One line of prose per reason, for the rendered chain. */
-declare const ADMISSION_EXCLUSION_MEANING: Readonly<Record<AdmissionExclusionReason, string>>;
+export declare const ADMISSION_EXCLUSION_MEANING: Readonly<Record<AdmissionExclusionReason, string>>;
 type AdmissionControlArm = 'no-fix-control' | 'no-op-control';
 /** The inert action the no-op control substitutes, and where it goes. */
 interface AdmissionNoOpInjection {
@@ -517,15 +517,15 @@ interface DenominatorChainArtifact {
  * and the rows that survived, so `input = admitted + sum(excluded)` can be read
  * off the table instead of trusted.
  */
-declare function buildDenominatorChain(verdicts: readonly AdmissionRowVerdict[], admitStrata: readonly AdmissionStratum[]): DenominatorChainArtifact;
+export declare function buildDenominatorChain(verdicts: readonly AdmissionRowVerdict[], admitStrata: readonly AdmissionStratum[]): DenominatorChainArtifact;
 /** A chain that does not add up is a broken denominator, so this throws. */
-declare function assertChainReconciles(artifact: DenominatorChainArtifact): void;
+export declare function assertChainReconciles(artifact: DenominatorChainArtifact): void;
 //#endregion
 //#region src/trace-repair/continuation-policy.d.ts
 /** A rollout ran outside the pinned policy, so its evidence cannot be used. */
-declare class ContinuationPolicyViolationError extends CaptureIntegrityError {}
+export declare class ContinuationPolicyViolationError extends CaptureIntegrityError {}
 /** Two arms did not run the same policy, so their difference is not the intervention. */
-declare class ContinuationSymmetryError extends CaptureIntegrityError {}
+export declare class ContinuationSymmetryError extends CaptureIntegrityError {}
 interface PinnedContinuationPolicy {
   /** Stable name for the frozen configuration, recorded on every rollout. */
   readonly id: string;
@@ -551,18 +551,18 @@ interface DefineContinuationPolicyInput extends Partial<Omit<PinnedContinuationP
   /** Required: a default seed would make two campaigns silently share a draw. */
   seed: number;
 }
-declare function definePinnedContinuationPolicy(input: DefineContinuationPolicyInput): PinnedContinuationPolicy;
+export declare function definePinnedContinuationPolicy(input: DefineContinuationPolicyInput): PinnedContinuationPolicy;
 /**
  * Hash over the policy and the scaffold text it renders. A changed template
  * changes the digest, so rollouts recorded before and after an edit cannot be
  * pooled by accident.
  */
-declare function continuationPolicyDigest(policy: PinnedContinuationPolicy): string;
+export declare function continuationPolicyDigest(policy: PinnedContinuationPolicy): string;
 /**
  * Per-rollout seed. It reads the policy seed, the row, and the rollout index —
  * deliberately not the arm, so paired rollouts across arms draw identically.
  */
-declare function continuationSeed(policySeed: number, rowId: string, rolloutIndex: number): number;
+export declare function continuationSeed(policySeed: number, rowId: string, rolloutIndex: number): number;
 interface ContinuationModelRequest {
   model: string;
   messages: ContinuationMessage[];
@@ -644,35 +644,35 @@ interface RunContinuationOptions {
  * mutates the container it runs in and the next one must start from the same
  * state, not from the previous rollout's leftovers.
  */
-declare function runContinuation(options: RunContinuationOptions): Promise<ContinuationRollout[]>;
+export declare function runContinuation(options: RunContinuationOptions): Promise<ContinuationRollout[]>;
 /**
  * Sum only what the provider reported. A call with no usage raises
  * `callsWithUsage` short of `calls` and clears `captured`, so a partially
  * reported rollout can never read as a fully measured one.
  */
-declare function totalUsage(steps: readonly ContinuationStepRecord[]): ContinuationUsageTotals;
+export declare function totalUsage(steps: readonly ContinuationStepRecord[]): ContinuationUsageTotals;
 /**
  * One unpriced call makes the rollout's cost unknown. Summing the rest would
  * report a number smaller than what was spent.
  */
-declare function totalCost(steps: readonly ContinuationStepRecord[]): CostProvenance;
+export declare function totalCost(steps: readonly ContinuationStepRecord[]): CostProvenance;
 /**
  * Prove the arms ran the same policy. Rollouts paired by row and index must
  * carry the same policy digest and the same seed; anything else means the
  * measured difference includes a policy change, not only the intervention.
  */
-declare function assertArmSymmetry(rollouts: readonly ContinuationRollout[]): void;
+export declare function assertArmSymmetry(rollouts: readonly ContinuationRollout[]): void;
 //#endregion
 //#region src/trace-repair/oracle-determinism.d.ts
 /** A task graded a repair by something other than the state under test. */
-declare class NondeterministicOracleError extends CaptureIntegrityError {}
+export declare class NondeterministicOracleError extends CaptureIntegrityError {}
 /** The two states certification produces: the published image, and that image
  *  after the task's own reference solution ran. */
 type OracleStateLabel = 'unsolved' | 'solved';
 /** Machine load the replicate ran under. */
 type OracleLoad = 'idle' | 'contended';
 /** The unit a whole-suite verdict is counted under when nothing finer exists. */
-declare const SUITE_REWARD_UNIT = "suite-reward";
+export declare const SUITE_REWARD_UNIT = "suite-reward";
 interface OracleAssertionResult {
   /** Test identifier the suite reported, verbatim. */
   readonly id: string;
@@ -766,7 +766,7 @@ interface OracleDeterminismVerdict {
  * counts here, so the rule that decides stability is one rule and a reviewer
  * can re-derive any verdict from the recorded replicates.
  */
-declare function oracleDeterminism(evidence: OracleDeterminismEvidence): OracleDeterminismVerdict;
+export declare function oracleDeterminism(evidence: OracleDeterminismEvidence): OracleDeterminismVerdict;
 /**
  * Certified verdicts by task name.
  *
@@ -787,10 +787,10 @@ interface TaskOracleRegistryDocument {
   readonly version: 1;
   readonly measurements: readonly OracleDeterminismEvidence[];
 }
-declare function parseTaskOracleRegistry(document: unknown): TaskOracleRegistry;
-declare function taskOracleRegistry(verdicts: readonly OracleDeterminismVerdict[]): TaskOracleRegistry;
+export declare function parseTaskOracleRegistry(document: unknown): TaskOracleRegistry;
+export declare function taskOracleRegistry(verdicts: readonly OracleDeterminismVerdict[]): TaskOracleRegistry;
 /** Throws unless the task graded the same bytes the same way every replicate. */
-declare function assertDeterministicOracle(verdict: OracleDeterminismVerdict): void;
+export declare function assertDeterministicOracle(verdict: OracleDeterminismVerdict): void;
 //#endregion
 //#region src/trace-repair/admission.d.ts
 /**
@@ -883,15 +883,15 @@ interface AdmissionConfig {
   readonly controlScreening: ControlScreening;
   readonly concurrency: number;
 }
-declare const ADMISSION_CONFIG_DEFAULTS: AdmissionConfig;
-declare function resolveAdmissionConfig(input?: AdmissionConfigInput): AdmissionConfig;
+export declare const ADMISSION_CONFIG_DEFAULTS: AdmissionConfig;
+export declare function resolveAdmissionConfig(input?: AdmissionConfigInput): AdmissionConfig;
 /**
  * The recorded command the no-op control replaces, drawn per rollout.
  *
  * The draw reads the policy seed, the row, and the rollout index, so it is
  * reproducible from the artifact and cannot depend on when the pre-pass ran.
  */
-declare function noOpInjectionStep(policySeed: number, rowId: string, rolloutIndex: number, recordedCommands: number): number;
+export declare function noOpInjectionStep(policySeed: number, rowId: string, rolloutIndex: number, recordedCommands: number): number;
 interface AdmissionProvenance {
   replayerId: string;
   oracleId: string;
@@ -947,10 +947,10 @@ interface RunAdmissionOptions {
  * proves it measured this denominator by passing the report back to
  * `assertDenominatorIntact`.
  */
-declare function runAdmission(options: RunAdmissionOptions): Promise<AdmissionReport>;
+export declare function runAdmission(options: RunAdmissionOptions): Promise<AdmissionReport>;
 /** Admitted row ids in one stratum. No call returns them pooled. */
-declare function admittedRowIds(report: AdmissionReport, stratum: AdmissionStratum): readonly string[];
-declare function admittedCount(report: AdmissionReport): number;
+export declare function admittedRowIds(report: AdmissionReport, stratum: AdmissionStratum): readonly string[];
+export declare function admittedCount(report: AdmissionReport): number;
 interface DenominatorIntactInput {
   report: AdmissionReport;
   /** Strata the campaign drew from. Rows outside them are not expected. */
@@ -969,7 +969,7 @@ interface DenominatorIntactInput {
  * can cause on its own — declining a row it cannot solve — so a sampled row
  * with no outcome is an error, not a smaller `n`.
  */
-declare function assertDenominatorIntact(input: DenominatorIntactInput): void;
+export declare function assertDenominatorIntact(input: DenominatorIntactInput): void;
 //#endregion
 //#region src/trace-repair/admission-contract.d.ts
 /** Pre-registered admission thresholds. */
@@ -983,7 +983,7 @@ interface AdmissionCriteria {
   /** How a control pass is read. `enforced` requires a control that can act. */
   readonly controlScreening: ControlScreening;
 }
-declare const TB_REPAIR_ADMISSION_CRITERIA: AdmissionCriteria;
+export declare const TB_REPAIR_ADMISSION_CRITERIA: AdmissionCriteria;
 interface PrefixFidelityEvidence {
   /** Recorded steps that were replayed. */
   readonly stepsReplayed: number;
@@ -1094,7 +1094,7 @@ type AdmissionDecision = {
  * of the row, so it must stop the run instead of producing one verdict per row
  * that reads as if a check had been applied.
  */
-declare function admitRow(evidence: AdmissionEvidence, criteria?: AdmissionCriteria): AdmissionDecision;
+export declare function admitRow(evidence: AdmissionEvidence, criteria?: AdmissionCriteria): AdmissionDecision;
 //#endregion
 //#region src/trace-repair/admission-report.d.ts
 interface AdmissionArtifact {
@@ -1116,17 +1116,17 @@ interface AdmissionArtifact {
   rows: readonly AdmissionRowVerdict[];
 }
 /** Plain JSON. `JSON.stringify` of this object is the machine-readable artifact. */
-declare function admissionArtifact(report: AdmissionReport): AdmissionArtifact;
+export declare function admissionArtifact(report: AdmissionReport): AdmissionArtifact;
 interface RenderAdmissionOptions {
   /** Rows listed in the per-row appendix. Default 0, which omits the appendix. */
   rowLimit?: number;
 }
 /** Markdown for the campaign report. Every number here is also in the artifact. */
-declare function renderAdmissionReport(artifact: AdmissionArtifact, options?: RenderAdmissionOptions): string;
+export declare function renderAdmissionReport(artifact: AdmissionArtifact, options?: RenderAdmissionOptions): string;
 //#endregion
 //#region src/trace-repair/analyst-response.d.ts
 /** The literal an analyst returns when no single step carries the failure. */
-declare const NO_DECISIVE_FAILURE = "no-decisive-failure";
+export declare const NO_DECISIVE_FAILURE = "no-decisive-failure";
 interface RepairIntervention {
   /** What the action is: a shell command, or an edit that authors file content. */
   readonly kind: ActionPayloadKind;
@@ -1164,9 +1164,9 @@ type ParseAnalystResponseOutcome = {
  * is untrusted text, so every field is checked and nothing is defaulted: a
  * missing `k` is a parse failure, never step 1.
  */
-declare function parseAnalystResponse(reply: string): ParseAnalystResponseOutcome;
+export declare function parseAnalystResponse(reply: string): ParseAnalystResponseOutcome;
 /** Build a finding directly, for callers that already hold typed fields. */
-declare function repairFinding(input: {
+export declare function repairFinding(input: {
   k: number;
   failureClaim: string;
   intervention: RepairIntervention;
@@ -1194,12 +1194,12 @@ interface BlindTrajectoryOptions {
    *  recording, which is what the recorded agent itself saw before it stopped. */
   readonly throughStep?: number;
 }
-declare function blindTrajectory(row: AdmittedRow, options?: BlindTrajectoryOptions): BlindedTrajectoryPrefix;
+export declare function blindTrajectory(row: AdmittedRow, options?: BlindTrajectoryOptions): BlindedTrajectoryPrefix;
 /**
  * Fields an analyst prompt must never carry. Exported so a consumer that
  * builds its own prompt can assert against the same list this module honours.
  */
-declare const BLINDED_FIELDS: readonly string[];
+export declare const BLINDED_FIELDS: readonly string[];
 //#endregion
 //#region src/trace-repair/analyst-arm.d.ts
 /**
@@ -1361,7 +1361,7 @@ interface AskRepairArmOptions {
  * Blinding, prompt identity and budget measurement happen here, once, for
  * every arm. An arm that wants a different question does not get one.
  */
-declare function askRepairArm(options: AskRepairArmOptions): Promise<RepairArmAnswer>;
+export declare function askRepairArm(options: AskRepairArmOptions): Promise<RepairArmAnswer>;
 /**
  * The answer in the grammar the grader consumes.
  *
@@ -1369,7 +1369,7 @@ declare function askRepairArm(options: AskRepairArmOptions): Promise<RepairArmAn
  * turning it into a decline would credit an arm with an honest null it never
  * produced.
  */
-declare function repairArmResponse(answer: RepairArmAnswer): AnalystResponse | null;
+export declare function repairArmResponse(answer: RepairArmAnswer): AnalystResponse | null;
 /** One arm's difference from the set, as a report renders it. */
 interface RepairArmAsymmetry {
   readonly armId: string;
@@ -1411,7 +1411,7 @@ interface RepairArmAsymmetryReport {
  * to none — a mixed set measures the optimisation, then reports it as the
  * harness.
  */
-declare function repairArmAsymmetries(arms: readonly RepairArm[], options?: {
+export declare function repairArmAsymmetries(arms: readonly RepairArm[], options?: {
   readonly budget?: InterventionBudget;
 }): RepairArmAsymmetryReport;
 //#endregion
@@ -1434,7 +1434,7 @@ interface CompletionRepairArmOptions {
   readonly affordances: readonly RepairArmAffordance[];
   readonly certification: RepairArmCertification;
 }
-declare function createCompletionRepairArm(options: CompletionRepairArmOptions): RepairArm;
+export declare function createCompletionRepairArm(options: CompletionRepairArmOptions): RepairArm;
 //#endregion
 //#region src/trace-repair/arm-dspy.d.ts
 /**
@@ -1444,9 +1444,9 @@ declare function createCompletionRepairArm(options: CompletionRepairArmOptions):
  * inside them rather than by adding a field every unrelated caller would have
  * to set. The same mechanism selects the CodeTraceBench typed signature.
  */
-declare const DSPY_REPAIR_TASK_TOKEN = "tb-repair-typed-";
+export declare const DSPY_REPAIR_TASK_TOKEN = "tb-repair-typed-";
 /** Signature the bridge reports for a repair analysis. */
-declare const DSPY_REPAIR_SIGNATURE = "tb-repair-typed-v1";
+export declare const DSPY_REPAIR_SIGNATURE = "tb-repair-typed-v1";
 /**
  * The repair contract restated for the typed signature.
  *
@@ -1455,7 +1455,7 @@ declare const DSPY_REPAIR_SIGNATURE = "tb-repair-typed-v1";
  * `repairTaskPolicy`, so the execution rules an arm is held to cannot drift
  * between arms.
  */
-declare function dspyRepairInstructions(budget?: InterventionBudget): string;
+export declare function dspyRepairInstructions(budget?: InterventionBudget): string;
 interface DspyRepairArmOptions {
   /** The DSPy engine. Injected, so a test drives the arm with no subprocess. */
   readonly engine: TraceAnalysisEngine;
@@ -1468,7 +1468,7 @@ interface DspyRepairArmOptions {
   readonly id?: string;
   readonly log?: (message: string, fields?: Record<string, unknown>) => void;
 }
-declare function createDspyRepairArm(options: DspyRepairArmOptions): RepairArm;
+export declare function createDspyRepairArm(options: DspyRepairArmOptions): RepairArm;
 interface ReadRepairPayloadOptions {
   /** The step ids the blinded prefix holds. A row naming any other k is the
    *  model's mistake, not the bridge's, so it lands in `dropped` — the same
@@ -1477,7 +1477,7 @@ interface ReadRepairPayloadOptions {
   readonly validSteps: readonly number[];
 }
 /** Thrown by callers that build a repair arm with a non-DSPy engine. */
-declare function assertDspyRepairEngine(engine: TraceAnalysisEngine): void;
+export declare function assertDspyRepairEngine(engine: TraceAnalysisEngine): void;
 //#endregion
 //#region src/trace-repair/degenerate-strategies.d.ts
 /**
@@ -1505,7 +1505,7 @@ interface DegenerateStrategy {
   /** Module that carries the mechanism. */
   readonly enforcedIn: string;
 }
-declare const DEGENERATE_STRATEGIES: readonly [{
+export declare const DEGENERATE_STRATEGIES: readonly [{
   readonly id: 'point-at-any-nonzero-exit-step';
   readonly strategy: 'Name the first step with a nonzero returncode, claim it failed, and let the reproduction gate confirm it.';
   readonly defeat: 'Reproduction is a gate, not a tier that pays. The credit vector has no term for it, so a reproduced step with no working intervention scores exactly what an unreproduced one scores: nothing.';
@@ -1788,9 +1788,9 @@ interface RepairCredit {
   /** Share of repair rollouts whose held-out suite passed. */
   readonly repairRate: number;
 }
-declare const CREDIT_TERMS: readonly ["executes", "localFlip", "repairRate"];
+export declare const CREDIT_TERMS: readonly ["executes", "localFlip", "repairRate"];
 /** What an answer earned. Every outcome but `measured` earns nothing. */
-declare function repairCredit(grade: RepairGrade): RepairCredit;
+export declare function repairCredit(grade: RepairGrade): RepairCredit;
 interface RepairFunnelCounts {
   readonly rows: number;
   readonly rejected: number;
@@ -1804,7 +1804,7 @@ interface RepairFunnelCounts {
   /** Rows where every repair rollout passed. */
   readonly t4RepairFlipAll: number;
 }
-declare function countFunnel(grades: readonly RepairGrade[]): RepairFunnelCounts;
+export declare function countFunnel(grades: readonly RepairGrade[]): RepairFunnelCounts;
 //#endregion
 //#region src/trace-repair/grade.d.ts
 interface GradeRepairOptions {
@@ -1855,7 +1855,7 @@ interface RepairRowResult extends DefaultVerdict {
   readonly delta: number;
   readonly wallMs: number;
 }
-declare function gradeRepairRow(options: GradeRepairOptions): Promise<RepairRowResult>;
+export declare function gradeRepairRow(options: GradeRepairOptions): Promise<RepairRowResult>;
 //#endregion
 //#region src/trace-repair/delta-repair.d.ts
 interface DeltaRepairOptions {
@@ -1898,10 +1898,10 @@ interface DeltaRepairReport {
   readonly rowResults: readonly RepairRowResult[];
   readonly threats: readonly RepairThreat[];
 }
-declare function deltaRepair(rowResults: readonly RepairRowResult[], options?: DeltaRepairOptions): DeltaRepairReport;
+export declare function deltaRepair(rowResults: readonly RepairRowResult[], options?: DeltaRepairOptions): DeltaRepairReport;
 /** Markdown report: provenance, the funnel, every per-row column, the
  *  distribution, and the threats. */
-declare function renderDeltaRepairReport(report: DeltaRepairReport): string;
+export declare function renderDeltaRepairReport(report: DeltaRepairReport): string;
 //#endregion
 //#region src/trace-repair/docker-environment.d.ts
 interface ProcessRequest {
@@ -1923,7 +1923,7 @@ type ProcessRunner = (request: ProcessRequest) => Promise<ProcessResult>;
  * and kills the whole process group on timeout so a killed command leaves no
  * children running in the container's namespace.
  */
-declare const nodeProcessRunner: ProcessRunner;
+export declare const nodeProcessRunner: ProcessRunner;
 interface DockerContinuationEnvironmentOptions {
   /** Container id or name holding the post-step-k state. */
   containerRef: string;
@@ -1945,33 +1945,33 @@ interface DockerContinuationEnvironmentOptions {
  * recorded run could not, and the arms would no longer differ only by the
  * intervention.
  */
-declare function dockerRunArgs(input: {
+export declare function dockerRunArgs(input: {
   image: string;
   name: string;
   cwd: string;
   containerLifetime?: string;
   executable?: string;
 }): string[];
-declare function createDockerContinuationEnvironment(options: DockerContinuationEnvironmentOptions): ContinuationEnvironment;
+export declare function createDockerContinuationEnvironment(options: DockerContinuationEnvironmentOptions): ContinuationEnvironment;
 //#endregion
 //#region src/trace-repair/repair-prompt.d.ts
-declare const REPAIR_QUESTION = "This coding agent ran and did not finish the task. Name the ONE recorded step whose action you would replace, and give the single shell action to run instead of it.";
+export declare const REPAIR_QUESTION = "This coding agent ran and did not finish the task. Name the ONE recorded step whose action you would replace, and give the single shell action to run instead of it.";
 /**
  * The task definition. Parameterised by the budget so the stated caps are the
  * enforced caps.
  */
-declare function repairTaskPolicy(budget?: InterventionBudget): readonly string[];
+export declare function repairTaskPolicy(budget?: InterventionBudget): readonly string[];
 /** The reply grammar, stated to an arm that answers in one JSON object. */
-declare const REPAIR_CONTRACT_LINES: readonly string[];
+export declare const REPAIR_CONTRACT_LINES: readonly string[];
 /** The same grammar restated for the bounded repair turn, which never carries
  *  the trajectory again. */
-declare const REPAIR_REPAIR_CONTRACT_LINES: readonly string[];
+export declare const REPAIR_REPAIR_CONTRACT_LINES: readonly string[];
 /** The trajectory as an arm sees it: actions, observations, nothing about grading. */
-declare function renderRepairTrajectory(prefix: BlindedTrajectoryPrefix): string;
-declare function repairTrajectoryHeader(prefix: BlindedTrajectoryPrefix): string;
+export declare function renderRepairTrajectory(prefix: BlindedTrajectoryPrefix): string;
+export declare function repairTrajectoryHeader(prefix: BlindedTrajectoryPrefix): string;
 /** The task definition an arm receives: the policy plus the statement the
  *  recorded agent was given. */
-declare function repairTaskDefinition(prefix: BlindedTrajectoryPrefix, budget?: InterventionBudget): string;
+export declare function repairTaskDefinition(prefix: BlindedTrajectoryPrefix, budget?: InterventionBudget): string;
 /**
  * Digest of the question and task policy every arm shares.
  *
@@ -1980,7 +1980,7 @@ declare function repairTaskDefinition(prefix: BlindedTrajectoryPrefix, budget?:
  * An arm's own contract text — its output grammar, its typed-signature
  * instructions — is deliberately outside it, because arms differ there.
  */
-declare function repairQuestionSha256(budget?: InterventionBudget): string;
+export declare function repairQuestionSha256(budget?: InterventionBudget): string;
 /**
  * Digest of the composed question one arm answered.
  *
@@ -1989,13 +1989,13 @@ declare function repairQuestionSha256(budget?: InterventionBudget): string;
  * versus a typed SUBMIT signature — stamp different digests, and two arms
  * that ask the identical composed question share one.
  */
-declare function repairArmPromptSha256(budget: InterventionBudget, contract: readonly string[]): string;
+export declare function repairArmPromptSha256(budget: InterventionBudget, contract: readonly string[]): string;
 //#endregion
 //#region src/trace-repair/test-oracle.d.ts
 /** The suite the oracle read back is not the suite it uploaded. */
-declare class TestSuiteTamperedError extends CaptureIntegrityError {}
+export declare class TestSuiteTamperedError extends CaptureIntegrityError {}
 /** The oracle could not place or run the suite, so it graded nothing. */
-declare class TestOracleError extends CaptureIntegrityError {}
+export declare class TestOracleError extends CaptureIntegrityError {}
 interface TestSuiteFile {
   /** Absolute path inside the container. */
   readonly path: string;
@@ -2018,8 +2018,8 @@ interface InjectedTestOracleOptions {
  * Content digest of the suite: sha256 over each path and its bytes, in path
  * order. Two suites with the same digest are the same suite.
  */
-declare function testSuiteDigest(files: readonly TestSuiteFile[]): string;
-declare function injectedTestOracle(options: InjectedTestOracleOptions): TestOracle;
+export declare function testSuiteDigest(files: readonly TestSuiteFile[]): string;
+export declare function injectedTestOracle(options: InjectedTestOracleOptions): TestOracle;
 //#endregion
-export { ADMISSION_CONFIG_DEFAULTS, ADMISSION_EXCLUSION_MEANING, ADMISSION_EXCLUSION_ORDER, ADMISSION_STRATA, type ActionPayloadKind, type AdmissionArtifact, type AdmissionCheckRecord, type AdmissionConfig, type AdmissionConfigInput, type AdmissionControlArm, type AdmissionControlObservation, type AdmissionControlRequest, type AdmissionControlRunner, type AdmissionCriteria, type AdmissionDecision, AdmissionDenominatorError, type AdmissionDivergence, type AdmissionEndStateOracle, type AdmissionEvidence, type AdmissionExclusionReason, AdmissionIndependenceError, type AdmissionNoOpInjection, type AdmissionOutcome, type AdmissionPrefixReplay, type AdmissionPrefixReplayer, type AdmissionProvenance, type AdmissionRejection, type AdmissionReport, type AdmissionRolloutSummary, type AdmissionRow, type AdmissionRowVerdict, type AdmissionScreeningRecord, type AdmissionStratum, type AdmissionTestVerdict, type AdmittedRow, type AnalystResponse, type AskRepairArmOptions, BLINDED_FIELDS, type BlindTrajectoryOptions, type BlindedStep, type BlindedTrajectoryPrefix, type BudgetCheck, type BudgetMeasurement, type BudgetViolation, CONTROL_SCREENING_MODES, CREDIT_TERMS, type CommandOutput, type CompletionRepairArmOptions, type ContinuationArm, type ContinuationEnvironment, type ContinuationEnvironmentDescription, type ContinuationEnvironmentFactory, type ContinuationEnvironmentRequest, type ContinuationExecRecord, type ContinuationExecResult, type ContinuationExitStatus, type ContinuationMessage, type ContinuationModel, type ContinuationModelCall, type ContinuationModelRequest, type ContinuationModelResponse, ContinuationPolicyViolationError, type ContinuationRollout, type ContinuationStepRecord, ContinuationSymmetryError, type ContinuationUsageTotals, type ControlArmEvidence, type ControlCapability, type ControlPolicy, type ControlPolicyInput, type ControlScreening, DEGENERATE_STRATEGIES, DSPY_REPAIR_SIGNATURE, DSPY_REPAIR_TASK_TOKEN, type DefineContinuationPolicyInput, type DegenerateDefeatKind, type DegenerateStrategy, type DegenerateStrategyId, type DeltaRepairInterval, type DeltaRepairOptions, type DeltaRepairReport, type DenominatorChain, type DenominatorChainArtifact, type DenominatorIntactInput, type DenominatorStage, type DockerContinuationEnvironmentOptions, type DspyRepairArmOptions, type GradeRepairOptions, type InjectedAction, type InjectedTestOracleOptions, type InterventionBudget, type InterventionExecution, MINI_SWE_SYSTEM_MESSAGE, NO_DECISIVE_FAILURE, NO_OP_ACTIONS, NondeterministicOracleError, OUTPUT_ELISION_THRESHOLD, type OracleAssertionResult, type OracleDeterminismEvidence, type OracleDeterminismVerdict, type OracleFlippedUnit, type OracleLoad, type OracleReplicate, type OracleReplicateGroup, type OracleStateLabel, type OracleStateVerdict, type ParseAnalystResponseOutcome, type ParsedAction, type PinnedContinuationPolicy, type PrefixFidelityEvidence, type PrefixReplayEvidence, type ProcessRequest, type ProcessResult, type ProcessRunner, REPAIR_CONTRACT_LINES, REPAIR_QUESTION, REPAIR_REPAIR_CONTRACT_LINES, type ReadRepairPayloadOptions, type RecordedStep, type RecordedToolCall, type RenderAdmissionOptions, type RepairArm as RepairAnalystArm, type RepairArm$1 as RepairArm, type RepairArmAffordance, type RepairArmAnswer, type RepairArmAsymmetry, type RepairArmAsymmetryReport, type RepairArmBudgetRecord, type RepairArmCertification, type RepairArmDeclaration, type RepairArmEvidence, type RepairArmRejectedRow, type RepairArmRepairTurn, type RepairArmReply, type RepairArmRequest, type RepairArmUsage, type RepairContinuationOutcome, type RepairContinuationRequest, type RepairContinuationRunner, type RepairCredit, type RepairDeclined, type RepairExecResult, type RepairFinding, type RepairFunnelCounts, type RepairGrade, type RepairIntervention, type RepairRejection, type RepairRolloutEvidence, type RepairRowResult, type RepairSession, type RepairSessionFactory, type RepairSessionRequest, type RepairThreat, type RepairThreatId, type ReproductionEvidence, type ResponseParseFailure, type RunAdmissionOptions, type RunContinuationOptions, SCAFFOLD_INTERVENTION_BUDGET, SUBMIT_SENTINEL, SUITE_REWARD_UNIT, TB_REPAIR_ADMISSION_CRITERIA, type TaskOracleRegistry, type TaskOracleRegistryDocument, type TestOracle, type TestOracleContext, TestOracleError, type TestOracleOutcome, type TestRunEvidence, type TestSuiteFile, TestSuiteTamperedError, UncalibratedControlError, admissionArtifact, admitRow, admittedCount, admittedRowIds, askRepairArm, assertAnalystIndependent, assertArmSymmetry, assertChainReconciles, assertControlCalibrated, assertDenominatorIntact, assertDeterministicOracle, assertDspyRepairEngine, blindTrajectory, buildDenominatorChain, checkInterventionBudget, classifyActionPayload, continuationPolicyDigest, continuationSeed, controlCanRescue, countFunnel, createCompletionRepairArm, createDockerContinuationEnvironment, createDspyRepairArm, defineControlPolicy, definePinnedContinuationPolicy, deltaRepair, dockerRunArgs, dspyRepairInstructions, gradeRepairRow, injectedTestOracle, isRecordedTimeout, noOpInjectionStep, nodeProcessRunner, normalizeActionForComparison, oracleDeterminism, parseAction, parseAnalystResponse, parseTaskOracleRegistry, renderAdmissionReport, renderDeltaRepairReport, renderFormatErrorObservation, renderInstanceMessage, renderObservation, renderRepairTrajectory, renderTimeoutObservation, repairArmAsymmetries, repairArmPromptSha256, repairArmResponse, repairCredit, repairFinding, repairQuestionSha256, repairTaskDefinition, repairTaskPolicy, repairTrajectoryHeader, resolveAdmissionConfig, rolloutDigest, rolloutRecordedSteps, runAdmission, runContinuation, scanShellAction, stratumOf, submissionOf, taskOracleRegistry, testSuiteDigest, toRecordedSteps, totalCost, totalUsage };
+export { type ActionPayloadKind, type AdmissionArtifact, type AdmissionCheckRecord, type AdmissionConfig, type AdmissionConfigInput, type AdmissionControlArm, type AdmissionControlObservation, type AdmissionControlRequest, type AdmissionControlRunner, type AdmissionCriteria, type AdmissionDecision, type AdmissionDivergence, type AdmissionEndStateOracle, type AdmissionEvidence, type AdmissionExclusionReason, type AdmissionNoOpInjection, type AdmissionOutcome, type AdmissionPrefixReplay, type AdmissionPrefixReplayer, type AdmissionProvenance, type AdmissionRejection, type AdmissionReport, type AdmissionRolloutSummary, type AdmissionRow, type AdmissionRowVerdict, type AdmissionScreeningRecord, type AdmissionStratum, type AdmissionTestVerdict, type AdmittedRow, type AnalystResponse, type AskRepairArmOptions, type BlindTrajectoryOptions, type BlindedStep, type BlindedTrajectoryPrefix, type BudgetCheck, type BudgetMeasurement, type BudgetViolation, type CommandOutput, type CompletionRepairArmOptions, type ContinuationArm, type ContinuationEnvironment, type ContinuationEnvironmentDescription, type ContinuationEnvironmentFactory, type ContinuationEnvironmentRequest, type ContinuationExecRecord, type ContinuationExecResult, type ContinuationExitStatus, type ContinuationMessage, type ContinuationModel, type ContinuationModelCall, type ContinuationModelRequest, type ContinuationModelResponse, type ContinuationRollout, type ContinuationStepRecord, type ContinuationUsageTotals, type ControlArmEvidence, type ControlCapability, type ControlPolicy, type ControlPolicyInput, type ControlScreening, type DefineContinuationPolicyInput, type DegenerateDefeatKind, type DegenerateStrategy, type DegenerateStrategyId, type DeltaRepairInterval, type DeltaRepairOptions, type DeltaRepairReport, type DenominatorChain, type DenominatorChainArtifact, type DenominatorIntactInput, type DenominatorStage, type DockerContinuationEnvironmentOptions, type DspyRepairArmOptions, type GradeRepairOptions, type InjectedAction, type InjectedTestOracleOptions, type InterventionBudget, type InterventionExecution, type OracleAssertionResult, type OracleDeterminismEvidence, type OracleDeterminismVerdict, type OracleFlippedUnit, type OracleLoad, type OracleReplicate, type OracleReplicateGroup, type OracleStateLabel, type OracleStateVerdict, type ParseAnalystResponseOutcome, type ParsedAction, type PinnedContinuationPolicy, type PrefixFidelityEvidence, type PrefixReplayEvidence, type ProcessRequest, type ProcessResult, type ProcessRunner, type ReadRepairPayloadOptions, type RecordedStep, type RecordedToolCall, type RenderAdmissionOptions, type RepairArm as RepairAnalystArm, type RepairArm$1 as RepairArm, type RepairArmAffordance, type RepairArmAnswer, type RepairArmAsymmetry, type RepairArmAsymmetryReport, type RepairArmBudgetRecord, type RepairArmCertification, type RepairArmDeclaration, type RepairArmEvidence, type RepairArmRejectedRow, type RepairArmRepairTurn, type RepairArmReply, type RepairArmRequest, type RepairArmUsage, type RepairContinuationOutcome, type RepairContinuationRequest, type RepairContinuationRunner, type RepairCredit, type RepairDeclined, type RepairExecResult, type RepairFinding, type RepairFunnelCounts, type RepairGrade, type RepairIntervention, type RepairRejection, type RepairRolloutEvidence, type RepairRowResult, type RepairSession, type RepairSessionFactory, type RepairSessionRequest, type RepairThreat, type RepairThreatId, type ReproductionEvidence, type ResponseParseFailure, type RunAdmissionOptions, type RunContinuationOptions, type TaskOracleRegistry, type TaskOracleRegistryDocument, type TestOracle, type TestOracleContext, type TestOracleOutcome, type TestRunEvidence, type TestSuiteFile, isRecordedTimeout };
 //# sourceMappingURL=index.d.ts.map