@tangle-network/agent-eval 0.125.0 → 0.126.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (82) hide show
  1. package/CHANGELOG.md +62 -35
  2. package/README.md +270 -189
  3. package/dist/analyst/index.d.ts +15 -145
  4. package/dist/analyst/index.js +33 -47
  5. package/dist/analyst/index.js.map +1 -1
  6. package/dist/benchmarks/index.d.ts +45 -162
  7. package/dist/benchmarks/index.js +8 -9
  8. package/dist/campaign/index.d.ts +3674 -5393
  9. package/dist/campaign/index.js +21 -95
  10. package/dist/{chunk-R226UZOI.js → chunk-474LBSOX.js} +2 -2
  11. package/dist/{chunk-HM6V7F3M.js → chunk-FO7HEH76.js} +3 -3
  12. package/dist/chunk-IILEIWGW.js +635 -0
  13. package/dist/chunk-IILEIWGW.js.map +1 -0
  14. package/dist/{chunk-EQUK3RFS.js → chunk-J5SQWP6Y.js} +8 -5
  15. package/dist/chunk-J5SQWP6Y.js.map +1 -0
  16. package/dist/{chunk-W5B3ZGP3.js → chunk-KE2VWPZX.js} +8 -6
  17. package/dist/{chunk-W5B3ZGP3.js.map → chunk-KE2VWPZX.js.map} +1 -1
  18. package/dist/{chunk-DT7OXY3C.js → chunk-LUNF2SEL.js} +538 -851
  19. package/dist/chunk-LUNF2SEL.js.map +1 -0
  20. package/dist/chunk-NGUYT5CI.js +4637 -0
  21. package/dist/chunk-NGUYT5CI.js.map +1 -0
  22. package/dist/{chunk-QFQZ3U3X.js → chunk-OCFJACJU.js} +2 -2
  23. package/dist/{chunk-GID26AN4.js → chunk-P22LJ3Y2.js} +4 -6
  24. package/dist/{chunk-GID26AN4.js.map → chunk-P22LJ3Y2.js.map} +1 -1
  25. package/dist/{chunk-SJT4OBVL.js → chunk-SDPM6554.js} +3 -3
  26. package/dist/{chunk-D5JZ7UDZ.js → chunk-UCLVDLCH.js} +136 -50
  27. package/dist/chunk-UCLVDLCH.js.map +1 -0
  28. package/dist/chunk-VMUENW6F.js +7274 -0
  29. package/dist/chunk-VMUENW6F.js.map +1 -0
  30. package/dist/{chunk-JKDNAOF5.js → chunk-W4L6C2XT.js} +2 -2
  31. package/dist/chunk-WGXIEX7P.js +116 -0
  32. package/dist/chunk-WGXIEX7P.js.map +1 -0
  33. package/dist/{chunk-GRCDRKII.js → chunk-WS3NZZQQ.js} +58 -20
  34. package/dist/chunk-WS3NZZQQ.js.map +1 -0
  35. package/dist/cli.js +3 -3
  36. package/dist/contract/index.d.ts +3220 -3094
  37. package/dist/contract/index.js +173 -42
  38. package/dist/contract/index.js.map +1 -1
  39. package/dist/control.js +2 -3
  40. package/dist/fuzz.d.ts +14 -1
  41. package/dist/fuzz.js +1 -1
  42. package/dist/hosted/index.d.ts +8 -1
  43. package/dist/index.d.ts +71 -687
  44. package/dist/index.js +178 -497
  45. package/dist/index.js.map +1 -1
  46. package/dist/openapi.json +1 -1
  47. package/dist/rl.d.ts +5 -100
  48. package/dist/rl.js +4 -5
  49. package/dist/rl.js.map +1 -1
  50. package/dist/{run-campaign-I3JXKVAK.js → run-campaign-LVFKZCEU.js} +3 -3
  51. package/dist/traces.js +2 -3
  52. package/dist/wire/index.d.ts +14 -1
  53. package/dist/wire/index.js +3 -3
  54. package/docs/campaign-proposers.md +363 -168
  55. package/docs/design/loop-taxonomy.md +142 -190
  56. package/docs/design.md +1 -1
  57. package/docs/distributed-driver.md +8 -11
  58. package/docs/feature-guide.md +20 -19
  59. package/docs/knowledge-readiness.md +2 -5
  60. package/docs/multi-shot-optimization.md +35 -27
  61. package/docs/rollout.md +5 -5
  62. package/package.json +4 -4
  63. package/dist/chunk-A62YMFWA.js +0 -9269
  64. package/dist/chunk-A62YMFWA.js.map +0 -1
  65. package/dist/chunk-A6GT67HT.js +0 -550
  66. package/dist/chunk-A6GT67HT.js.map +0 -1
  67. package/dist/chunk-D5JZ7UDZ.js.map +0 -1
  68. package/dist/chunk-DT7OXY3C.js.map +0 -1
  69. package/dist/chunk-EQUK3RFS.js.map +0 -1
  70. package/dist/chunk-GC4ATIKK.js +0 -317
  71. package/dist/chunk-GC4ATIKK.js.map +0 -1
  72. package/dist/chunk-GRCDRKII.js.map +0 -1
  73. package/dist/chunk-LOW3U7JZ.js +0 -328
  74. package/dist/chunk-LOW3U7JZ.js.map +0 -1
  75. package/dist/chunk-PMITBABE.js +0 -3841
  76. package/dist/chunk-PMITBABE.js.map +0 -1
  77. /package/dist/{chunk-R226UZOI.js.map → chunk-474LBSOX.js.map} +0 -0
  78. /package/dist/{chunk-HM6V7F3M.js.map → chunk-FO7HEH76.js.map} +0 -0
  79. /package/dist/{chunk-QFQZ3U3X.js.map → chunk-OCFJACJU.js.map} +0 -0
  80. /package/dist/{chunk-SJT4OBVL.js.map → chunk-SDPM6554.js.map} +0 -0
  81. /package/dist/{chunk-JKDNAOF5.js.map → chunk-W4L6C2XT.js.map} +0 -0
  82. /package/dist/{run-campaign-I3JXKVAK.js.map → run-campaign-LVFKZCEU.js.map} +0 -0
@@ -190,44 +190,7 @@ interface RunScoreWeights {
190
190
  wallSeconds: number;
191
191
  }
192
192
 
193
- /**
194
- * Error taxonomy for `@tangle-network/agent-eval`.
195
- *
196
- * Every error this package throws as part of its *public contract* extends
197
- * `AgentEvalError`. Consumers can pattern-match by `instanceof <Subclass>` or
198
- * by the stable string `code` carried on the base class.
199
- *
200
- * The codes are stable across minor versions; new codes can be added, but
201
- * existing codes never change meaning. New subclasses are non-breaking.
202
- *
203
- * Internal invariant guards (`throw new Error('this should never happen')`)
204
- * remain plain `Error`s on purpose — they're programmer-mistake assertions,
205
- * not consumer-catchable contract failures.
206
- */
207
- type AgentEvalErrorCode = 'validation' | 'not_found' | 'config' | 'capture_integrity' | 'judge' | 'verification' | 'replay' | 'backend_integrity' | 'profile_matrix';
208
- /**
209
- * Base class for every contract error this package throws — carries the stable
210
- * string `code` taxonomy so consumers can `instanceof`-match or switch on `code`.
211
- */
212
- declare class AgentEvalError extends Error {
213
- /** Stable string code. Survives minification; safe to switch on. */
214
- readonly code: AgentEvalErrorCode;
215
- constructor(code: AgentEvalErrorCode, message: string, options?: {
216
- cause?: unknown;
217
- });
218
- }
219
- /** Caller passed invalid arguments (out of range, mutually-exclusive options, bad shape). */
220
- declare class ValidationError extends AgentEvalError {
221
- constructor(message: string, options?: {
222
- cause?: unknown;
223
- });
224
- }
225
-
226
193
  type AgentProfileCellSchemaVersion = 'agent-profile-cell/v1';
227
- type AgentProfileJsonObject = {
228
- [key: string]: AgentProfileJson;
229
- };
230
- type AgentProfileJson = string | number | boolean | null | AgentProfileJson[] | AgentProfileJsonObject;
231
194
  type AgentProfileDimensionValue = string | number | boolean | null;
232
195
  interface AgentProfileSource {
233
196
  /** Runtime/profile contract being fingerprinted, e.g. `agent-interface-profile`. */
@@ -831,29 +794,40 @@ interface CostReceipt extends CostCallBase, CostUsage {
831
794
  costUsd: number;
832
795
  costUnknown: boolean;
833
796
  usageUnknown?: boolean;
797
+ /** Rates used to estimate cost locally. Absent when cost is provider-reported or unknown. */
834
798
  pricing?: {
835
799
  inputUsdPerThousand: number;
800
+ cachedInputUsdPerThousand?: number;
801
+ cacheWriteUsdPerThousand?: number;
836
802
  outputUsdPerThousand: number;
837
803
  };
804
+ /** Cost reported by the provider, not a local token-price calculation. */
838
805
  actualCostUsd?: number;
839
806
  error?: string;
840
807
  }
841
808
  interface CostReceiptInput extends CostUsage {
842
809
  model: string;
810
+ /** Caller-supplied rates for a local estimate when the provider does not report billed cost. */
811
+ customTokenPricing?: CustomTokenPricing;
843
812
  actualCostUsd?: number;
844
813
  costUnknown?: boolean;
845
814
  usageUnknown?: boolean;
846
815
  }
847
816
  /** Per-million token rates for a model or endpoint not covered by package pricing. */
848
817
  interface CustomTokenPricing {
818
+ /** Non-cached input tokens. */
849
819
  inputUsdPerMillion: number;
820
+ /** Cache-read tokens. Falls back to the normal input rate when omitted. */
821
+ cachedInputUsdPerMillion?: number;
822
+ /** Cache-creation or cache-write tokens. Falls back to the normal input rate when omitted. */
823
+ cacheWriteUsdPerMillion?: number;
850
824
  outputUsdPerMillion: number;
851
825
  }
852
826
  type MaximumCharge = {
853
827
  externallyEnforcedMaximumUsd: number;
854
828
  } | ({
855
829
  customTokenPricing: CustomTokenPricing;
856
- } & Pick<CostUsage, 'inputTokens' | 'outputTokens'>) | ({
830
+ } & Pick<CostUsage, 'inputTokens' | 'outputTokens' | 'cachedTokens' | 'cacheWriteTokens'>) | ({
857
831
  model: string;
858
832
  } & CostUsage);
859
833
  interface RunPaidCallInput<T> {
@@ -921,6 +895,8 @@ interface CostLedgerFilter {
921
895
  interface CostLedgerWaitOptions {
922
896
  /** Maximum time to wait for active provider calls. Default 5 seconds. */
923
897
  timeoutMs?: number;
898
+ /** Wait only for calls matching this attribution filter. */
899
+ filter?: CostLedgerFilter;
924
900
  }
925
901
  /** Append-only storage. `append` must atomically reject stale revisions. */
926
902
  interface CostLedgerPersistence {
@@ -2926,112 +2902,6 @@ declare function coerceJson(text: string): unknown;
2926
2902
  */
2927
2903
  declare function coerceToFindingRows(raw: unknown): unknown[];
2928
2904
 
2929
- type PolicyEditSchemaVersion = 'policy-edit/v1';
2930
- declare const POLICY_EDIT_AXES: readonly ["carrier", "representation", "budget", "sampling", "output_contract", "tool_contract", "routing", "memory", "agent_profile", "deployment_target"];
2931
- type PolicyEditAxis = (typeof POLICY_EDIT_AXES)[number];
2932
- declare const POLICY_EDIT_TARGET_SURFACES: readonly ["prompt", "tool-contract", "runtime-config", "memory", "agent-profile", "code", "deployment"];
2933
- type PolicyEditTargetSurface = (typeof POLICY_EDIT_TARGET_SURFACES)[number];
2934
- type PolicyEditRisk = 'low' | 'medium' | 'high' | 'unknown';
2935
- type PolicyEditGainDirection = 'increase' | 'decrease';
2936
- type PolicyEditGainUnit = 'absolute' | 'relative' | 'percent' | 'score';
2937
- interface PolicyEditTarget {
2938
- surface: PolicyEditTargetSurface;
2939
- /** Stable path inside the target surface, for example `system-prompt:tools`
2940
- * or `budget.maxTurns`. */
2941
- path?: string;
2942
- /** Optional canonical deployment identity. Store the existing cell, not a
2943
- * local profile shape. */
2944
- agentProfileCell?: AgentProfileCell;
2945
- /** Human label when the path is not enough for a readable audit trail. */
2946
- label?: string;
2947
- }
2948
- type PolicyEditChange = {
2949
- kind: 'text';
2950
- mode: 'append' | 'prepend' | 'replace';
2951
- value: string;
2952
- /** Required when `mode === 'replace'`; exact match only. */
2953
- find?: string;
2954
- } | {
2955
- kind: 'json';
2956
- mode: 'set' | 'merge' | 'remove';
2957
- path: string;
2958
- value?: AgentProfileJson;
2959
- };
2960
- interface PolicyEditExpectedGain {
2961
- /** Metric this edit is expected to move, e.g. `holdout.composite`. */
2962
- metric: string;
2963
- direction: PolicyEditGainDirection;
2964
- /** Positive magnitude in the metric's native units. */
2965
- amount: number;
2966
- unit?: PolicyEditGainUnit;
2967
- rationale?: string;
2968
- }
2969
- interface PolicyEditSource {
2970
- findingIds: string[];
2971
- analystIds: string[];
2972
- evidenceRefs: EvidenceRef[];
2973
- /** Mirrors `AnalystFinding.derived_from_judge`; admission rejects it. */
2974
- derivedFromJudge?: boolean;
2975
- }
2976
- interface PolicyEdit {
2977
- schemaVersion: PolicyEditSchemaVersion;
2978
- editId: string;
2979
- axis: PolicyEditAxis;
2980
- target: PolicyEditTarget;
2981
- change: PolicyEditChange;
2982
- claim: string;
2983
- expectedGain: PolicyEditExpectedGain;
2984
- confidence: number;
2985
- risk: PolicyEditRisk;
2986
- source: PolicyEditSource;
2987
- rationale?: string;
2988
- validationPlan?: string;
2989
- metadata?: Record<string, unknown>;
2990
- }
2991
- declare const POLICY_EDIT_CANDIDATE_RECORD_SCHEMA: "tangle.policy-edit-candidate.v1";
2992
- /** JSON-safe attribution carried with a measured candidate and its scores. */
2993
- interface PolicyEditCandidateRecord {
2994
- schema: typeof POLICY_EDIT_CANDIDATE_RECORD_SCHEMA;
2995
- policyEdit: PolicyEdit;
2996
- }
2997
- type PolicyEditInit = Omit<PolicyEdit, 'schemaVersion' | 'editId'> & {
2998
- schemaVersion?: PolicyEditSchemaVersion;
2999
- editId?: string;
3000
- };
3001
- declare class PolicyEditValidationError extends ValidationError {
3002
- readonly path: string;
3003
- constructor(message: string, path?: string);
3004
- }
3005
- interface FindingToPolicyEditOptions {
3006
- expectedGain?: PolicyEditExpectedGain | ((finding: AnalystFinding) => PolicyEditExpectedGain | null | undefined);
3007
- risk?: PolicyEditRisk | ((finding: AnalystFinding) => PolicyEditRisk);
3008
- defaultAxis?: PolicyEditAxis;
3009
- defaultTargetSurface?: PolicyEditTargetSurface;
3010
- }
3011
- interface PolicyEditAdmissionOptions {
3012
- minScore?: number;
3013
- minExpectedGain?: number;
3014
- allowHighRisk?: boolean;
3015
- requireEvidence?: boolean;
3016
- }
3017
- interface PolicyEditAdmission {
3018
- edit: PolicyEdit;
3019
- decision: 'admit' | 'reject';
3020
- score: number;
3021
- reasons: string[];
3022
- }
3023
- declare function makePolicyEdit(init: PolicyEditInit): PolicyEdit;
3024
- declare function computePolicyEditId(edit: Omit<PolicyEdit, 'editId'> | PolicyEdit): string;
3025
- declare function validatePolicyEdit(input: unknown): PolicyEdit;
3026
- declare function makePolicyEditCandidateRecord(edit: PolicyEdit): PolicyEditCandidateRecord;
3027
- declare function validatePolicyEditCandidateRecord(input: unknown): PolicyEditCandidateRecord;
3028
- declare function isPolicyEdit(input: unknown): input is PolicyEdit;
3029
- declare function policyEditsFromFindings(findings: ReadonlyArray<AnalystFinding>, opts?: FindingToPolicyEditOptions): PolicyEdit[];
3030
- declare function policyEditFromFinding(finding: AnalystFinding, opts?: FindingToPolicyEditOptions): PolicyEdit | null;
3031
- declare function scorePolicyEditReadiness(edit: PolicyEdit, opts?: PolicyEditAdmissionOptions): number;
3032
- declare function admitPolicyEdit(edit: PolicyEdit, opts?: PolicyEditAdmissionOptions): PolicyEditAdmission;
3033
- declare function applyPolicyEditToSurface(surface: unknown, edit: PolicyEdit): unknown;
3034
-
3035
2905
  /** DESCRIPTIVE predicate: does the finding cite at least one observable
3036
2906
  * (span/event/artifact) evidence ref. Useful for ranking evidence quality or
3037
2907
  * rendering — it is NOT the steer gate. Evidence presence is the WRONG
@@ -3147,4 +3017,4 @@ type TraceToolGroupName =
3147
3017
  */
3148
3018
  declare function buildTraceToolsForGroup(group: TraceToolGroupName, store: TraceAnalysisStore): AxFunction[];
3149
3019
 
3150
- export { ANALYST_SEVERITIES, type Analyst, type AnalystContext, type AnalystCost, type AnalystFinding, type AnalystHooks, type AnalystInputKind, AnalystRegistry, type AnalystRegistryOptions, type AnalystRequirements, type AnalystRunEvent, type AnalystRunInputs, type AnalystRunResult, type AnalystRunSummary, type AnalystSeverity, type AnalystUsageReceipt, type BudgetPolicy, type CanonicalRawAnalystFinding, CanonicalRawAnalystFindingSchema, type ChatCallOpts, type ChatClient, type ChatRequest, type ChatResponse, type ChatTransport, type CliBridgeTransportOpts, type CreateAnalystAiConfig, type CreateChatClientOpts, type CreateTraceAnalystKindOpts, DEFAULT_TRACE_ANALYST_KINDS, type DefaultAnalystRegistryOptions, type DiffPolicy, type DirectProviderTransportOpts, type EvidenceRef, FAILURE_MODE_KIND_SPEC, FINDING_SUBJECT_GRAMMAR_PROMPT, FINDING_SUBJECT_KINDS, FINDING_SUBJECT_SYNTAX, type FindingSubject, type FindingSubjectKind, FindingSubjectStringSchema, type FindingToPolicyEditOptions, type FindingsDiff, FindingsStore, IMPROVEMENT_KIND_SPEC, type JudgeAdapterOpts, KIND_EXPECTED_SUBJECTS, KNOWLEDGE_GAP_KIND_SPEC, KNOWLEDGE_POISONING_KIND_SPEC, type MockTransportOpts, POLICY_EDIT_AXES, POLICY_EDIT_CANDIDATE_RECORD_SCHEMA, POLICY_EDIT_TARGET_SURFACES, type PersistedFinding, type PolicyEdit, type PolicyEditAdmission, type PolicyEditAdmissionOptions, type PolicyEditAxis, type PolicyEditCandidateRecord, type PolicyEditChange, type PolicyEditExpectedGain, type PolicyEditGainDirection, type PolicyEditGainUnit, type PolicyEditInit, type PolicyEditRisk, type PolicyEditSchemaVersion, type PolicyEditSource, type PolicyEditTarget, type PolicyEditTargetSurface, PolicyEditValidationError, RAW_FINDING_SCHEMA_PROMPT, type RawAnalystEvidence, RawAnalystEvidenceSchema, type RawAnalystFinding, RawAnalystFindingSchema, type RegistryRunOpts, type RouterTransportOpts, type RunCriticAdapterOpts, SKILL_USAGE_ANALYST, type SandboxSdkTransportOpts, type SemanticConceptJudgeAdapterOpts, SkillUsageAnalyst, type SkillUsageRecord, type SkillUsageReport, type SkillUsageScanConfig, type StructureFindingsOptions, type StructureFindingsResult, type TraceAnalystGolden, type TraceAnalystKindSpec, type TraceToolGroupName, type VerifierAdapterOpts, admitPolicyEdit, applyPolicyEditToSurface, assertNoJudgeVerdict, behavioralAnalyst, buildDefaultAnalystRegistry, buildSkillUsageReport, buildTraceToolsForGroup, coerceJson, coerceToFindingRows, computeFindingId, computePolicyEditId, createAnalystAi, createChatClient, createJudgeAdapter, createRunCriticAdapter, createSemanticConceptJudgeAdapter, createTraceAnalystKind, createVerifierAdapter, defaultIsMaterial, deriveEfficiencyFindings, diffFindings, emitSkillUsageFindings, evidenceRefsFromRawFinding, findingSubjectGrammarPromptFor, isJudgeVerdict, isPolicyEdit, isTraceObservable, liftSeverity, makeFinding, makePolicyEdit, makePolicyEditCandidateRecord, parseCanonicalRawFinding, parseFindingSubject, parseRawFinding, policyEditFromFinding, policyEditsFromFindings, renderFindingSubject, renderPriorFindings, renderUpstreamFindings, scorePolicyEditReadiness, stripCodeFences, structureFindings, validatePolicyEdit, validatePolicyEditCandidateRecord };
3020
+ export { ANALYST_SEVERITIES, type Analyst, type AnalystContext, type AnalystCost, type AnalystFinding, type AnalystHooks, type AnalystInputKind, AnalystRegistry, type AnalystRegistryOptions, type AnalystRequirements, type AnalystRunEvent, type AnalystRunInputs, type AnalystRunResult, type AnalystRunSummary, type AnalystSeverity, type AnalystUsageReceipt, type BudgetPolicy, type CanonicalRawAnalystFinding, CanonicalRawAnalystFindingSchema, type ChatCallOpts, type ChatClient, type ChatRequest, type ChatResponse, type ChatTransport, type CliBridgeTransportOpts, type CreateAnalystAiConfig, type CreateChatClientOpts, type CreateTraceAnalystKindOpts, DEFAULT_TRACE_ANALYST_KINDS, type DefaultAnalystRegistryOptions, type DiffPolicy, type DirectProviderTransportOpts, type EvidenceRef, FAILURE_MODE_KIND_SPEC, FINDING_SUBJECT_GRAMMAR_PROMPT, FINDING_SUBJECT_KINDS, FINDING_SUBJECT_SYNTAX, type FindingSubject, type FindingSubjectKind, FindingSubjectStringSchema, type FindingsDiff, FindingsStore, IMPROVEMENT_KIND_SPEC, type JudgeAdapterOpts, KIND_EXPECTED_SUBJECTS, KNOWLEDGE_GAP_KIND_SPEC, KNOWLEDGE_POISONING_KIND_SPEC, type MockTransportOpts, type PersistedFinding, RAW_FINDING_SCHEMA_PROMPT, type RawAnalystEvidence, RawAnalystEvidenceSchema, type RawAnalystFinding, RawAnalystFindingSchema, type RegistryRunOpts, type RouterTransportOpts, type RunCriticAdapterOpts, SKILL_USAGE_ANALYST, type SandboxSdkTransportOpts, type SemanticConceptJudgeAdapterOpts, SkillUsageAnalyst, type SkillUsageRecord, type SkillUsageReport, type SkillUsageScanConfig, type StructureFindingsOptions, type StructureFindingsResult, type TraceAnalystGolden, type TraceAnalystKindSpec, type TraceToolGroupName, type VerifierAdapterOpts, assertNoJudgeVerdict, behavioralAnalyst, buildDefaultAnalystRegistry, buildSkillUsageReport, buildTraceToolsForGroup, coerceJson, coerceToFindingRows, computeFindingId, createAnalystAi, createChatClient, createJudgeAdapter, createRunCriticAdapter, createSemanticConceptJudgeAdapter, createTraceAnalystKind, createVerifierAdapter, defaultIsMaterial, deriveEfficiencyFindings, diffFindings, emitSkillUsageFindings, evidenceRefsFromRawFinding, findingSubjectGrammarPromptFor, isJudgeVerdict, isTraceObservable, liftSeverity, makeFinding, parseCanonicalRawFinding, parseFindingSubject, parseRawFinding, renderFindingSubject, renderPriorFindings, renderUpstreamFindings, stripCodeFences, structureFindings };
@@ -9,14 +9,7 @@ import {
9
9
  diffFindings,
10
10
  emitSkillUsageFindings,
11
11
  runSemanticConceptJudge
12
- } from "../chunk-W5B3ZGP3.js";
13
- import {
14
- behavioralAnalyst,
15
- buildDefaultAnalystRegistry,
16
- createChatClient,
17
- deriveEfficiencyFindings
18
- } from "../chunk-A6GT67HT.js";
19
- import "../chunk-HHWE3POT.js";
12
+ } from "../chunk-KE2VWPZX.js";
20
13
  import {
21
14
  ANALYST_SEVERITIES,
22
15
  AnalystRegistry,
@@ -31,56 +24,42 @@ import {
31
24
  KIND_EXPECTED_SUBJECTS,
32
25
  KNOWLEDGE_GAP_KIND_SPEC,
33
26
  KNOWLEDGE_POISONING_KIND_SPEC,
34
- POLICY_EDIT_AXES,
35
- POLICY_EDIT_CANDIDATE_RECORD_SCHEMA,
36
- POLICY_EDIT_TARGET_SURFACES,
37
- PolicyEditValidationError,
38
27
  RAW_FINDING_SCHEMA_PROMPT,
39
28
  RawAnalystEvidenceSchema,
40
29
  RawAnalystFindingSchema,
41
- admitPolicyEdit,
42
- applyPolicyEditToSurface,
43
- assertNoJudgeVerdict,
30
+ behavioralAnalyst,
31
+ buildDefaultAnalystRegistry,
44
32
  buildTraceToolsForGroup,
45
33
  coerceJson,
46
34
  coerceToFindingRows,
47
35
  computeFindingId,
48
- computePolicyEditId,
49
36
  createAnalystAi,
37
+ createChatClient,
50
38
  createTraceAnalystKind,
39
+ deriveEfficiencyFindings,
51
40
  evidenceRefsFromRawFinding,
52
41
  findingSubjectGrammarPromptFor,
53
- isJudgeVerdict,
54
- isPolicyEdit,
55
- isTraceObservable,
56
42
  makeFinding,
57
- makePolicyEdit,
58
- makePolicyEditCandidateRecord,
59
43
  parseCanonicalRawFinding,
60
44
  parseFindingSubject,
61
45
  parseRawFinding,
62
- policyEditFromFinding,
63
- policyEditsFromFindings,
64
46
  renderFindingSubject,
65
47
  renderPriorFindings,
66
48
  renderUpstreamFindings,
67
- scorePolicyEditReadiness,
68
49
  settleUsageReceiptFromCostLedger,
69
50
  stripCodeFences,
70
51
  structureFindings,
71
- validatePolicyEdit,
72
- validatePolicyEditCandidateRecord,
73
52
  validateUsageSettlementTimeout
74
- } from "../chunk-DT7OXY3C.js";
75
- import "../chunk-EQUK3RFS.js";
53
+ } from "../chunk-LUNF2SEL.js";
54
+ import "../chunk-HHWE3POT.js";
55
+ import "../chunk-WGXIEX7P.js";
56
+ import "../chunk-J5SQWP6Y.js";
76
57
  import {
77
58
  CostLedger
78
- } from "../chunk-GRCDRKII.js";
59
+ } from "../chunk-WS3NZZQQ.js";
79
60
  import "../chunk-VI2UW6B6.js";
80
61
  import "../chunk-IR3KBHOY.js";
81
62
  import "../chunk-PC4UYEBM.js";
82
- import "../chunk-GC4ATIKK.js";
83
- import "../chunk-VSMTAMNK.js";
84
63
  import "../chunk-ONWEPEDO.js";
85
64
  import "../chunk-K4DBDHLK.js";
86
65
  import "../chunk-PZ5AY32C.js";
@@ -330,6 +309,28 @@ function createSemanticConceptJudgeAdapter(opts = {}) {
330
309
  }
331
310
  };
332
311
  }
312
+
313
+ // src/analyst/steer-firewall.ts
314
+ var OBSERVABLE_KINDS = /* @__PURE__ */ new Set([
315
+ "span",
316
+ "event",
317
+ "artifact"
318
+ ]);
319
+ function isTraceObservable(finding) {
320
+ return finding.evidence_refs.some((ref) => OBSERVABLE_KINDS.has(ref.kind));
321
+ }
322
+ function isJudgeVerdict(finding) {
323
+ return finding.derived_from_judge === true;
324
+ }
325
+ function assertNoJudgeVerdict(findings, context = "steer") {
326
+ const leaks = findings.filter(isJudgeVerdict);
327
+ if (leaks.length > 0) {
328
+ throw new Error(
329
+ `${context}: a judge verdict cannot be admitted as steering input \u2014 that is the held-out judge leaking into the loop. Offending judge-derived findings: [${leaks.map((f) => f.finding_id).join(", ")}]. Steering consumes observations of behavior, never acceptance verdicts.`
330
+ );
331
+ }
332
+ return findings;
333
+ }
333
334
  export {
334
335
  ANALYST_SEVERITIES,
335
336
  AnalystRegistry,
@@ -345,17 +346,11 @@ export {
345
346
  KIND_EXPECTED_SUBJECTS,
346
347
  KNOWLEDGE_GAP_KIND_SPEC,
347
348
  KNOWLEDGE_POISONING_KIND_SPEC,
348
- POLICY_EDIT_AXES,
349
- POLICY_EDIT_CANDIDATE_RECORD_SCHEMA,
350
- POLICY_EDIT_TARGET_SURFACES,
351
- PolicyEditValidationError,
352
349
  RAW_FINDING_SCHEMA_PROMPT,
353
350
  RawAnalystEvidenceSchema,
354
351
  RawAnalystFindingSchema,
355
352
  SKILL_USAGE_ANALYST,
356
353
  SkillUsageAnalyst,
357
- admitPolicyEdit,
358
- applyPolicyEditToSurface,
359
354
  assertNoJudgeVerdict,
360
355
  behavioralAnalyst,
361
356
  buildDefaultAnalystRegistry,
@@ -364,7 +359,6 @@ export {
364
359
  coerceJson,
365
360
  coerceToFindingRows,
366
361
  computeFindingId,
367
- computePolicyEditId,
368
362
  createAnalystAi,
369
363
  createChatClient,
370
364
  createJudgeAdapter,
@@ -379,24 +373,16 @@ export {
379
373
  evidenceRefsFromRawFinding,
380
374
  findingSubjectGrammarPromptFor,
381
375
  isJudgeVerdict,
382
- isPolicyEdit,
383
376
  isTraceObservable,
384
377
  liftSeverity,
385
378
  makeFinding,
386
- makePolicyEdit,
387
- makePolicyEditCandidateRecord,
388
379
  parseCanonicalRawFinding,
389
380
  parseFindingSubject,
390
381
  parseRawFinding,
391
- policyEditFromFinding,
392
- policyEditsFromFindings,
393
382
  renderFindingSubject,
394
383
  renderPriorFindings,
395
384
  renderUpstreamFindings,
396
- scorePolicyEditReadiness,
397
385
  stripCodeFences,
398
- structureFindings,
399
- validatePolicyEdit,
400
- validatePolicyEditCandidateRecord
386
+ structureFindings
401
387
  };
402
388
  //# sourceMappingURL=index.js.map
@@ -1 +1 @@
1
- {"version":3,"sources":["../../src/analyst/adapters.ts"],"sourcesContent":["/**\n * Adapter factories — lift each existing agent-eval primitive into the\n * Analyst contract without re-implementing it.\n *\n * Five primitives, five factories. Each one:\n * - Builds an Analyst with a stable id (caller chooses; defaults\n * given), a sensible default `inputKind`, a version derived from\n * the wrapped primitive's version + an adapter revision, and an\n * `analyze()` that calls the primitive and lifts its output to\n * AnalystFinding[] using `makeFinding()`.\n * - Maps severities: the existing `Severity` ('critical' | 'major' |\n * 'minor' | 'info') projects onto AnalystSeverity ('critical' |\n * 'high' | 'medium' | 'low' | 'info'); 'major' → 'high', 'minor' →\n * 'medium'. Domain analysts that want finer-grained mapping override.\n *\n * Adapters never own state. Calling the same factory twice with the\n * same primitive instance is safe.\n */\n\nimport { CostLedger } from '../cost-ledger'\nimport type {\n Finding as LayerFinding,\n Severity as LayerSeverity,\n MultiLayerVerifier,\n VerifyOptions,\n} from '../multi-layer-verifier'\nimport { RunCritic, type RunTrace } from '../run-critic'\nimport {\n runSemanticConceptJudge,\n SEMANTIC_CONCEPT_JUDGE_VERSION,\n type SemanticConceptJudgeInput,\n type SemanticConceptJudgeOptions,\n type SemanticConceptJudgeResult,\n} from '../semantic-concept-judge'\nimport type { JudgeFn, JudgeInput, JudgeScore, TCloud } from '../types'\nimport type { Analyst, AnalystFinding, AnalystSeverity } from './types'\nimport { makeFinding } from './types'\nimport { settleUsageReceiptFromCostLedger, validateUsageSettlementTimeout } from './usage-receipt'\n\nconst ADAPTER_REV = '1'\n\n// ── Severity bridges ───────────────────────────────────────────────\n\nexport function liftSeverity(s: LayerSeverity): AnalystSeverity {\n switch (s) {\n case 'critical':\n return 'critical'\n case 'major':\n return 'high'\n case 'minor':\n return 'medium'\n case 'info':\n return 'info'\n }\n}\n\n// ── 1. MultiLayerVerifier → Analyst ─────────────────────────────────\n\nexport interface VerifierAdapterOpts<Env> {\n id?: string\n area?: string\n verifier: MultiLayerVerifier<Env>\n /**\n * The verifier expects an `env` per run. Adapters take it from\n * `AnalystRunInputs.custom[<id>]` via the registry's 'custom' routing.\n */\n options?: Omit<VerifyOptions<Env>, 'env'>\n}\n\nexport function createVerifierAdapter<Env>(opts: VerifierAdapterOpts<Env>): Analyst<Env> {\n const id = opts.id ?? 'multi-layer-verifier'\n const area = opts.area ?? 'verification'\n return {\n id,\n description:\n \"Runs a MultiLayerVerifier and lifts each layer's findings into the analyst envelope.\",\n inputKind: 'custom',\n cost: { kind: 'deterministic' },\n version: `verifier-${ADAPTER_REV}`,\n async analyze(env, ctx) {\n const report = await opts.verifier.run({ env, ...opts.options })\n const out: AnalystFinding[] = []\n for (const layer of report.layers) {\n for (const finding of layer.findings) {\n out.push(liftLayerFinding(id, area, layer.layer, finding))\n }\n // Layer-level signal: a failed/error layer is itself a finding\n // even if it didn't emit per-finding rows.\n if (layer.status === 'fail' || layer.status === 'error' || layer.status === 'timeout') {\n out.push(\n makeFinding({\n analyst_id: id,\n area,\n subject: layer.layer,\n claim: `layer \"${layer.layer}\" ${layer.status}: ${layer.reason ?? 'no reason given'}`,\n severity:\n layer.status === 'error' ? 'high' : layer.status === 'timeout' ? 'medium' : 'high',\n confidence: 1,\n evidence_refs: [],\n metadata: {\n layer_status: layer.status,\n duration_ms: layer.durationMs,\n score: layer.score,\n diagnostics: layer.diagnostics,\n },\n }),\n )\n }\n }\n ctx.log?.('verifier complete', {\n layers: report.layers.length,\n blended: report.blendedScore,\n all_pass: report.allPass,\n })\n return out\n },\n }\n}\n\nfunction liftLayerFinding(\n analyst_id: string,\n area: string,\n layer: string,\n f: LayerFinding,\n): AnalystFinding {\n return makeFinding({\n analyst_id,\n area,\n subject: f.layer ?? layer,\n claim: f.message,\n severity: liftSeverity(f.severity),\n confidence: 0.85,\n evidence_refs: f.evidence\n ? [{ kind: 'artifact', uri: 'inline:evidence', excerpt: f.evidence }]\n : [],\n metadata: f.detail,\n })\n}\n\n// ── 2. RunCritic → Analyst ──────────────────────────────────────────\n\nexport interface RunCriticAdapterOpts {\n id?: string\n area?: string\n critic?: RunCritic\n /** Optional threshold below which a dimension is reported as a finding. Default 0.5. */\n threshold?: number\n}\n\nexport function createRunCriticAdapter(opts: RunCriticAdapterOpts = {}): Analyst<RunTrace> {\n const id = opts.id ?? 'run-critic'\n const area = opts.area ?? 'run-quality'\n const critic = opts.critic ?? new RunCritic()\n const threshold = opts.threshold ?? 0.5\n return {\n id,\n description:\n 'Scores a single run across success / grounding / drift / tool-quality and surfaces below-threshold dimensions.',\n inputKind: 'custom',\n cost: { kind: 'deterministic' },\n version: `run-critic-${ADAPTER_REV}`,\n async analyze(trace) {\n const score = critic.scoreTrace(trace)\n const out: AnalystFinding[] = []\n const dims: Array<[keyof typeof score, AnalystSeverity, string]> = [\n ['success', 'critical', 'run did not complete successfully'],\n ['goalProgress', 'high', 'goal progress is low'],\n ['repoGroundedness', 'high', 'output is poorly grounded in the repository'],\n ['toolUseQuality', 'medium', 'tool use quality is low'],\n ['patchQuality', 'medium', 'no real patch/edit evidence'],\n ['testReality', 'high', 'no real test/build evidence'],\n ['finalGate', 'critical', 'final gate is blocking'],\n ]\n for (const [dim, sev, msg] of dims) {\n const value = score[dim] as number\n if (typeof value === 'number' && value < threshold) {\n out.push(\n makeFinding({\n analyst_id: id,\n area,\n subject: dim,\n claim: msg,\n rationale: `${dim}=${value.toFixed(2)} below threshold ${threshold}`,\n severity: sev,\n confidence: 1,\n evidence_refs: [],\n metadata: { dimension: dim, value, threshold, run_id: trace.run.runId },\n }),\n )\n }\n }\n // Drift penalty is high → surface as a finding (inverse threshold).\n if (score.driftPenalty > 1 - threshold) {\n out.push(\n makeFinding({\n analyst_id: id,\n area,\n subject: 'drift',\n claim: 'agent output drifted from repository signal',\n rationale: `driftPenalty=${score.driftPenalty.toFixed(2)}`,\n severity: 'medium',\n confidence: 0.9,\n evidence_refs: [],\n metadata: { drift_penalty: score.driftPenalty, notes: score.notes },\n }),\n )\n }\n return out\n },\n }\n}\n\n// ── 3. JudgeFn → Analyst ────────────────────────────────────────────\n\nexport interface JudgeAdapterOpts {\n id?: string\n area?: string\n judge: JudgeFn\n /** TCloud handle the JudgeFn calls. */\n tcloud: TCloud\n /** Optional cost classification — most judges call an LLM. */\n cost?: Analyst['cost']\n /** Optional threshold below which a JudgeScore becomes a finding. Default 6 (on 0-10 scale). */\n threshold?: number\n}\n\nexport function createJudgeAdapter(opts: JudgeAdapterOpts): Analyst<JudgeInput> {\n const id = opts.id ?? 'judge'\n const area = opts.area ?? 'judge'\n const threshold = opts.threshold ?? 6\n return {\n id,\n description:\n 'Wraps an agent-eval JudgeFn into an analyst; below-threshold dimensions surface as findings.',\n inputKind: 'judge-input',\n cost: opts.cost ?? { kind: 'llm' },\n version: `judge-${ADAPTER_REV}`,\n async analyze(input) {\n const scores = await opts.judge(opts.tcloud, input)\n return scores\n .filter((s) => normalize10(s.score) < threshold)\n .map((s) => liftJudgeScore(id, area, s))\n },\n }\n}\n\nfunction normalize10(s: number): number {\n // JudgeScore convention is 0-10 but some judges emit 0-1. Coerce to 0-10.\n return s <= 1 ? s * 10 : s\n}\n\nfunction liftJudgeScore(analyst_id: string, area: string, s: JudgeScore): AnalystFinding {\n const score10 = normalize10(s.score)\n const severity: AnalystSeverity =\n score10 < 3 ? 'critical' : score10 < 5 ? 'high' : score10 < 7 ? 'medium' : 'low'\n return makeFinding({\n analyst_id,\n area,\n subject: s.dimension,\n claim: `${s.judgeName}/${s.dimension} scored ${score10.toFixed(1)}/10`,\n rationale: s.reasoning,\n severity,\n confidence: 0.8,\n evidence_refs: s.evidence\n ? [{ kind: 'artifact', uri: 'inline:evidence', excerpt: s.evidence }]\n : [],\n // Provenance: this finding IS a judge verdict (an acceptance score), not an\n // observation of behavior. The steer firewall (assertNoJudgeVerdict) rejects\n // it from steering — even when it cites an artifact above — because letting a\n // verdict steer the next attempt is the held-out judge leaking into the loop.\n derived_from_judge: true,\n metadata: { judge_name: s.judgeName, dimension: s.dimension, score_10: score10 },\n })\n}\n\n// ── 4. SemanticConceptJudge → Analyst ──────────────────────────────\n\nexport interface SemanticConceptJudgeAdapterOpts {\n id?: string\n area?: string\n /** Registry context owns cancellation and the per-analyst cost ledger. */\n options?: Omit<SemanticConceptJudgeOptions, 'costLedger' | 'signal'>\n /** Maximum post-cancellation wait for a provider receipt. Default 5 seconds. */\n settlementTimeoutMs?: number\n}\n\nexport function createSemanticConceptJudgeAdapter(\n opts: SemanticConceptJudgeAdapterOpts = {},\n): Analyst<SemanticConceptJudgeInput> {\n const id = opts.id ?? 'semantic-concept-judge'\n const area = opts.area ?? 'concept-coverage'\n const settlementTimeoutMs = validateUsageSettlementTimeout(opts.settlementTimeoutMs)\n return {\n id,\n description:\n 'Runs the semantic-concept judge and surfaces missing / weak concepts as findings.',\n inputKind: 'custom',\n cost: {\n kind: 'llm',\n models: opts.options?.model ? [opts.options.model] : undefined,\n settlement_timeout_ms: settlementTimeoutMs,\n },\n version: `${SEMANTIC_CONCEPT_JUDGE_VERSION}-adapter-${ADAPTER_REV}`,\n async analyze(input, ctx) {\n const costLedger = new CostLedger(ctx.budgetUsd)\n let result: SemanticConceptJudgeResult\n try {\n result = await runSemanticConceptJudge(input, {\n ...opts.options,\n costLedger,\n signal: ctx.signal,\n })\n } finally {\n const usage = await settleUsageReceiptFromCostLedger(costLedger, {\n channel: 'judge',\n timeoutMs: settlementTimeoutMs,\n })\n if (!usage.settled) {\n ctx.log?.('semantic-concept judge provider settlement timed out', {\n pending_calls: usage.pendingCalls,\n timeout_ms: settlementTimeoutMs,\n })\n }\n ctx.recordUsage?.(usage.receipt)\n }\n if (!result.available) {\n return [\n makeFinding({\n analyst_id: id,\n area,\n claim: 'semantic-concept judge unavailable',\n rationale: result.error,\n severity: 'info',\n confidence: 1,\n evidence_refs: [],\n metadata: { reason: result.error },\n }),\n ]\n }\n const out: AnalystFinding[] = []\n for (const f of result.findings) {\n // Only surface gaps: missing concepts or low scores. Concepts at\n // 7+/10 with present=true are not findings — they're successes.\n if (f.present && f.score >= 7) continue\n out.push(\n makeFinding({\n analyst_id: id,\n area,\n subject: f.concept,\n claim: f.present\n ? `concept \"${f.concept}\" is weak (${f.score}/10)`\n : `concept \"${f.concept}\" is missing`,\n rationale: f.evidence,\n severity: liftSeverity(f.severity),\n confidence: 0.85,\n evidence_refs: [{ kind: 'artifact', uri: 'inline:evidence', excerpt: f.evidence }],\n metadata: {\n concept: f.concept,\n present: f.present,\n score_10: f.score,\n },\n }),\n )\n }\n return out\n },\n }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuCA,IAAM,cAAc;AAIb,SAAS,aAAa,GAAmC;AAC9D,UAAQ,GAAG;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,EACX;AACF;AAeO,SAAS,sBAA2B,MAA8C;AACvF,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,gBAAgB;AAAA,IAC9B,SAAS,YAAY,WAAW;AAAA,IAChC,MAAM,QAAQ,KAAK,KAAK;AACtB,YAAM,SAAS,MAAM,KAAK,SAAS,IAAI,EAAE,KAAK,GAAG,KAAK,QAAQ,CAAC;AAC/D,YAAM,MAAwB,CAAC;AAC/B,iBAAW,SAAS,OAAO,QAAQ;AACjC,mBAAW,WAAW,MAAM,UAAU;AACpC,cAAI,KAAK,iBAAiB,IAAI,MAAM,MAAM,OAAO,OAAO,CAAC;AAAA,QAC3D;AAGA,YAAI,MAAM,WAAW,UAAU,MAAM,WAAW,WAAW,MAAM,WAAW,WAAW;AACrF,cAAI;AAAA,YACF,YAAY;AAAA,cACV,YAAY;AAAA,cACZ;AAAA,cACA,SAAS,MAAM;AAAA,cACf,OAAO,UAAU,MAAM,KAAK,KAAK,MAAM,MAAM,KAAK,MAAM,UAAU,iBAAiB;AAAA,cACnF,UACE,MAAM,WAAW,UAAU,SAAS,MAAM,WAAW,YAAY,WAAW;AAAA,cAC9E,YAAY;AAAA,cACZ,eAAe,CAAC;AAAA,cAChB,UAAU;AAAA,gBACR,cAAc,MAAM;AAAA,gBACpB,aAAa,MAAM;AAAA,gBACnB,OAAO,MAAM;AAAA,gBACb,aAAa,MAAM;AAAA,cACrB;AAAA,YACF,CAAC;AAAA,UACH;AAAA,QACF;AAAA,MACF;AACA,UAAI,MAAM,qBAAqB;AAAA,QAC7B,QAAQ,OAAO,OAAO;AAAA,QACtB,SAAS,OAAO;AAAA,QAChB,UAAU,OAAO;AAAA,MACnB,CAAC;AACD,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAEA,SAAS,iBACP,YACA,MACA,OACA,GACgB;AAChB,SAAO,YAAY;AAAA,IACjB;AAAA,IACA;AAAA,IACA,SAAS,EAAE,SAAS;AAAA,IACpB,OAAO,EAAE;AAAA,IACT,UAAU,aAAa,EAAE,QAAQ;AAAA,IACjC,YAAY;AAAA,IACZ,eAAe,EAAE,WACb,CAAC,EAAE,MAAM,YAAY,KAAK,mBAAmB,SAAS,EAAE,SAAS,CAAC,IAClE,CAAC;AAAA,IACL,UAAU,EAAE;AAAA,EACd,CAAC;AACH;AAYO,SAAS,uBAAuB,OAA6B,CAAC,GAAsB;AACzF,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,SAAS,KAAK,UAAU,IAAI,UAAU;AAC5C,QAAM,YAAY,KAAK,aAAa;AACpC,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,gBAAgB;AAAA,IAC9B,SAAS,cAAc,WAAW;AAAA,IAClC,MAAM,QAAQ,OAAO;AACnB,YAAM,QAAQ,OAAO,WAAW,KAAK;AACrC,YAAM,MAAwB,CAAC;AAC/B,YAAM,OAA6D;AAAA,QACjE,CAAC,WAAW,YAAY,mCAAmC;AAAA,QAC3D,CAAC,gBAAgB,QAAQ,sBAAsB;AAAA,QAC/C,CAAC,oBAAoB,QAAQ,6CAA6C;AAAA,QAC1E,CAAC,kBAAkB,UAAU,yBAAyB;AAAA,QACtD,CAAC,gBAAgB,UAAU,6BAA6B;AAAA,QACxD,CAAC,eAAe,QAAQ,6BAA6B;AAAA,QACrD,CAAC,aAAa,YAAY,wBAAwB;AAAA,MACpD;AACA,iBAAW,CAAC,KAAK,KAAK,GAAG,KAAK,MAAM;AAClC,cAAM,QAAQ,MAAM,GAAG;AACvB,YAAI,OAAO,UAAU,YAAY,QAAQ,WAAW;AAClD,cAAI;AAAA,YACF,YAAY;AAAA,cACV,YAAY;AAAA,cACZ;AAAA,cACA,SAAS;AAAA,cACT,OAAO;AAAA,cACP,WAAW,GAAG,GAAG,IAAI,MAAM,QAAQ,CAAC,CAAC,oBAAoB,SAAS;AAAA,cAClE,UAAU;AAAA,cACV,YAAY;AAAA,cACZ,eAAe,CAAC;AAAA,cAChB,UAAU,EAAE,WAAW,KAAK,OAAO,WAAW,QAAQ,MAAM,IAAI,MAAM;AAAA,YACxE,CAAC;AAAA,UACH;AAAA,QACF;AAAA,MACF;AAEA,UAAI,MAAM,eAAe,IAAI,WAAW;AACtC,YAAI;AAAA,UACF,YAAY;AAAA,YACV,YAAY;AAAA,YACZ;AAAA,YACA,SAAS;AAAA,YACT,OAAO;AAAA,YACP,WAAW,gBAAgB,MAAM,aAAa,QAAQ,CAAC,CAAC;AAAA,YACxD,UAAU;AAAA,YACV,YAAY;AAAA,YACZ,eAAe,CAAC;AAAA,YAChB,UAAU,EAAE,eAAe,MAAM,cAAc,OAAO,MAAM,MAAM;AAAA,UACpE,CAAC;AAAA,QACH;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAgBO,SAAS,mBAAmB,MAA6C;AAC9E,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,YAAY,KAAK,aAAa;AACpC,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,KAAK,QAAQ,EAAE,MAAM,MAAM;AAAA,IACjC,SAAS,SAAS,WAAW;AAAA,IAC7B,MAAM,QAAQ,OAAO;AACnB,YAAM,SAAS,MAAM,KAAK,MAAM,KAAK,QAAQ,KAAK;AAClD,aAAO,OACJ,OAAO,CAAC,MAAM,YAAY,EAAE,KAAK,IAAI,SAAS,EAC9C,IAAI,CAAC,MAAM,eAAe,IAAI,MAAM,CAAC,CAAC;AAAA,IAC3C;AAAA,EACF;AACF;AAEA,SAAS,YAAY,GAAmB;AAEtC,SAAO,KAAK,IAAI,IAAI,KAAK;AAC3B;AAEA,SAAS,eAAe,YAAoB,MAAc,GAA+B;AACvF,QAAM,UAAU,YAAY,EAAE,KAAK;AACnC,QAAM,WACJ,UAAU,IAAI,aAAa,UAAU,IAAI,SAAS,UAAU,IAAI,WAAW;AAC7E,SAAO,YAAY;AAAA,IACjB;AAAA,IACA;AAAA,IACA,SAAS,EAAE;AAAA,IACX,OAAO,GAAG,EAAE,SAAS,IAAI,EAAE,SAAS,WAAW,QAAQ,QAAQ,CAAC,CAAC;AAAA,IACjE,WAAW,EAAE;AAAA,IACb;AAAA,IACA,YAAY;AAAA,IACZ,eAAe,EAAE,WACb,CAAC,EAAE,MAAM,YAAY,KAAK,mBAAmB,SAAS,EAAE,SAAS,CAAC,IAClE,CAAC;AAAA;AAAA;AAAA;AAAA;AAAA,IAKL,oBAAoB;AAAA,IACpB,UAAU,EAAE,YAAY,EAAE,WAAW,WAAW,EAAE,WAAW,UAAU,QAAQ;AAAA,EACjF,CAAC;AACH;AAaO,SAAS,kCACd,OAAwC,CAAC,GACL;AACpC,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,sBAAsB,+BAA+B,KAAK,mBAAmB;AACnF,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM;AAAA,MACJ,MAAM;AAAA,MACN,QAAQ,KAAK,SAAS,QAAQ,CAAC,KAAK,QAAQ,KAAK,IAAI;AAAA,MACrD,uBAAuB;AAAA,IACzB;AAAA,IACA,SAAS,GAAG,8BAA8B,YAAY,WAAW;AAAA,IACjE,MAAM,QAAQ,OAAO,KAAK;AACxB,YAAM,aAAa,IAAI,WAAW,IAAI,SAAS;AAC/C,UAAI;AACJ,UAAI;AACF,iBAAS,MAAM,wBAAwB,OAAO;AAAA,UAC5C,GAAG,KAAK;AAAA,UACR;AAAA,UACA,QAAQ,IAAI;AAAA,QACd,CAAC;AAAA,MACH,UAAE;AACA,cAAM,QAAQ,MAAM,iCAAiC,YAAY;AAAA,UAC/D,SAAS;AAAA,UACT,WAAW;AAAA,QACb,CAAC;AACD,YAAI,CAAC,MAAM,SAAS;AAClB,cAAI,MAAM,wDAAwD;AAAA,YAChE,eAAe,MAAM;AAAA,YACrB,YAAY;AAAA,UACd,CAAC;AAAA,QACH;AACA,YAAI,cAAc,MAAM,OAAO;AAAA,MACjC;AACA,UAAI,CAAC,OAAO,WAAW;AACrB,eAAO;AAAA,UACL,YAAY;AAAA,YACV,YAAY;AAAA,YACZ;AAAA,YACA,OAAO;AAAA,YACP,WAAW,OAAO;AAAA,YAClB,UAAU;AAAA,YACV,YAAY;AAAA,YACZ,eAAe,CAAC;AAAA,YAChB,UAAU,EAAE,QAAQ,OAAO,MAAM;AAAA,UACnC,CAAC;AAAA,QACH;AAAA,MACF;AACA,YAAM,MAAwB,CAAC;AAC/B,iBAAW,KAAK,OAAO,UAAU;AAG/B,YAAI,EAAE,WAAW,EAAE,SAAS,EAAG;AAC/B,YAAI;AAAA,UACF,YAAY;AAAA,YACV,YAAY;AAAA,YACZ;AAAA,YACA,SAAS,EAAE;AAAA,YACX,OAAO,EAAE,UACL,YAAY,EAAE,OAAO,cAAc,EAAE,KAAK,SAC1C,YAAY,EAAE,OAAO;AAAA,YACzB,WAAW,EAAE;AAAA,YACb,UAAU,aAAa,EAAE,QAAQ;AAAA,YACjC,YAAY;AAAA,YACZ,eAAe,CAAC,EAAE,MAAM,YAAY,KAAK,mBAAmB,SAAS,EAAE,SAAS,CAAC;AAAA,YACjF,UAAU;AAAA,cACR,SAAS,EAAE;AAAA,cACX,SAAS,EAAE;AAAA,cACX,UAAU,EAAE;AAAA,YACd;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;","names":[]}
1
+ {"version":3,"sources":["../../src/analyst/adapters.ts","../../src/analyst/steer-firewall.ts"],"sourcesContent":["/**\n * Adapter factories — lift each existing agent-eval primitive into the\n * Analyst contract without re-implementing it.\n *\n * Five primitives, five factories. Each one:\n * - Builds an Analyst with a stable id (caller chooses; defaults\n * given), a sensible default `inputKind`, a version derived from\n * the wrapped primitive's version + an adapter revision, and an\n * `analyze()` that calls the primitive and lifts its output to\n * AnalystFinding[] using `makeFinding()`.\n * - Maps severities: the existing `Severity` ('critical' | 'major' |\n * 'minor' | 'info') projects onto AnalystSeverity ('critical' |\n * 'high' | 'medium' | 'low' | 'info'); 'major' → 'high', 'minor' →\n * 'medium'. Domain analysts that want finer-grained mapping override.\n *\n * Adapters never own state. Calling the same factory twice with the\n * same primitive instance is safe.\n */\n\nimport { CostLedger } from '../cost-ledger'\nimport type {\n Finding as LayerFinding,\n Severity as LayerSeverity,\n MultiLayerVerifier,\n VerifyOptions,\n} from '../multi-layer-verifier'\nimport { RunCritic, type RunTrace } from '../run-critic'\nimport {\n runSemanticConceptJudge,\n SEMANTIC_CONCEPT_JUDGE_VERSION,\n type SemanticConceptJudgeInput,\n type SemanticConceptJudgeOptions,\n type SemanticConceptJudgeResult,\n} from '../semantic-concept-judge'\nimport type { JudgeFn, JudgeInput, JudgeScore, TCloud } from '../types'\nimport type { Analyst, AnalystFinding, AnalystSeverity } from './types'\nimport { makeFinding } from './types'\nimport { settleUsageReceiptFromCostLedger, validateUsageSettlementTimeout } from './usage-receipt'\n\nconst ADAPTER_REV = '1'\n\n// ── Severity bridges ───────────────────────────────────────────────\n\nexport function liftSeverity(s: LayerSeverity): AnalystSeverity {\n switch (s) {\n case 'critical':\n return 'critical'\n case 'major':\n return 'high'\n case 'minor':\n return 'medium'\n case 'info':\n return 'info'\n }\n}\n\n// ── 1. MultiLayerVerifier → Analyst ─────────────────────────────────\n\nexport interface VerifierAdapterOpts<Env> {\n id?: string\n area?: string\n verifier: MultiLayerVerifier<Env>\n /**\n * The verifier expects an `env` per run. Adapters take it from\n * `AnalystRunInputs.custom[<id>]` via the registry's 'custom' routing.\n */\n options?: Omit<VerifyOptions<Env>, 'env'>\n}\n\nexport function createVerifierAdapter<Env>(opts: VerifierAdapterOpts<Env>): Analyst<Env> {\n const id = opts.id ?? 'multi-layer-verifier'\n const area = opts.area ?? 'verification'\n return {\n id,\n description:\n \"Runs a MultiLayerVerifier and lifts each layer's findings into the analyst envelope.\",\n inputKind: 'custom',\n cost: { kind: 'deterministic' },\n version: `verifier-${ADAPTER_REV}`,\n async analyze(env, ctx) {\n const report = await opts.verifier.run({ env, ...opts.options })\n const out: AnalystFinding[] = []\n for (const layer of report.layers) {\n for (const finding of layer.findings) {\n out.push(liftLayerFinding(id, area, layer.layer, finding))\n }\n // Layer-level signal: a failed/error layer is itself a finding\n // even if it didn't emit per-finding rows.\n if (layer.status === 'fail' || layer.status === 'error' || layer.status === 'timeout') {\n out.push(\n makeFinding({\n analyst_id: id,\n area,\n subject: layer.layer,\n claim: `layer \"${layer.layer}\" ${layer.status}: ${layer.reason ?? 'no reason given'}`,\n severity:\n layer.status === 'error' ? 'high' : layer.status === 'timeout' ? 'medium' : 'high',\n confidence: 1,\n evidence_refs: [],\n metadata: {\n layer_status: layer.status,\n duration_ms: layer.durationMs,\n score: layer.score,\n diagnostics: layer.diagnostics,\n },\n }),\n )\n }\n }\n ctx.log?.('verifier complete', {\n layers: report.layers.length,\n blended: report.blendedScore,\n all_pass: report.allPass,\n })\n return out\n },\n }\n}\n\nfunction liftLayerFinding(\n analyst_id: string,\n area: string,\n layer: string,\n f: LayerFinding,\n): AnalystFinding {\n return makeFinding({\n analyst_id,\n area,\n subject: f.layer ?? layer,\n claim: f.message,\n severity: liftSeverity(f.severity),\n confidence: 0.85,\n evidence_refs: f.evidence\n ? [{ kind: 'artifact', uri: 'inline:evidence', excerpt: f.evidence }]\n : [],\n metadata: f.detail,\n })\n}\n\n// ── 2. RunCritic → Analyst ──────────────────────────────────────────\n\nexport interface RunCriticAdapterOpts {\n id?: string\n area?: string\n critic?: RunCritic\n /** Optional threshold below which a dimension is reported as a finding. Default 0.5. */\n threshold?: number\n}\n\nexport function createRunCriticAdapter(opts: RunCriticAdapterOpts = {}): Analyst<RunTrace> {\n const id = opts.id ?? 'run-critic'\n const area = opts.area ?? 'run-quality'\n const critic = opts.critic ?? new RunCritic()\n const threshold = opts.threshold ?? 0.5\n return {\n id,\n description:\n 'Scores a single run across success / grounding / drift / tool-quality and surfaces below-threshold dimensions.',\n inputKind: 'custom',\n cost: { kind: 'deterministic' },\n version: `run-critic-${ADAPTER_REV}`,\n async analyze(trace) {\n const score = critic.scoreTrace(trace)\n const out: AnalystFinding[] = []\n const dims: Array<[keyof typeof score, AnalystSeverity, string]> = [\n ['success', 'critical', 'run did not complete successfully'],\n ['goalProgress', 'high', 'goal progress is low'],\n ['repoGroundedness', 'high', 'output is poorly grounded in the repository'],\n ['toolUseQuality', 'medium', 'tool use quality is low'],\n ['patchQuality', 'medium', 'no real patch/edit evidence'],\n ['testReality', 'high', 'no real test/build evidence'],\n ['finalGate', 'critical', 'final gate is blocking'],\n ]\n for (const [dim, sev, msg] of dims) {\n const value = score[dim] as number\n if (typeof value === 'number' && value < threshold) {\n out.push(\n makeFinding({\n analyst_id: id,\n area,\n subject: dim,\n claim: msg,\n rationale: `${dim}=${value.toFixed(2)} below threshold ${threshold}`,\n severity: sev,\n confidence: 1,\n evidence_refs: [],\n metadata: { dimension: dim, value, threshold, run_id: trace.run.runId },\n }),\n )\n }\n }\n // Drift penalty is high → surface as a finding (inverse threshold).\n if (score.driftPenalty > 1 - threshold) {\n out.push(\n makeFinding({\n analyst_id: id,\n area,\n subject: 'drift',\n claim: 'agent output drifted from repository signal',\n rationale: `driftPenalty=${score.driftPenalty.toFixed(2)}`,\n severity: 'medium',\n confidence: 0.9,\n evidence_refs: [],\n metadata: { drift_penalty: score.driftPenalty, notes: score.notes },\n }),\n )\n }\n return out\n },\n }\n}\n\n// ── 3. JudgeFn → Analyst ────────────────────────────────────────────\n\nexport interface JudgeAdapterOpts {\n id?: string\n area?: string\n judge: JudgeFn\n /** TCloud handle the JudgeFn calls. */\n tcloud: TCloud\n /** Optional cost classification — most judges call an LLM. */\n cost?: Analyst['cost']\n /** Optional threshold below which a JudgeScore becomes a finding. Default 6 (on 0-10 scale). */\n threshold?: number\n}\n\nexport function createJudgeAdapter(opts: JudgeAdapterOpts): Analyst<JudgeInput> {\n const id = opts.id ?? 'judge'\n const area = opts.area ?? 'judge'\n const threshold = opts.threshold ?? 6\n return {\n id,\n description:\n 'Wraps an agent-eval JudgeFn into an analyst; below-threshold dimensions surface as findings.',\n inputKind: 'judge-input',\n cost: opts.cost ?? { kind: 'llm' },\n version: `judge-${ADAPTER_REV}`,\n async analyze(input) {\n const scores = await opts.judge(opts.tcloud, input)\n return scores\n .filter((s) => normalize10(s.score) < threshold)\n .map((s) => liftJudgeScore(id, area, s))\n },\n }\n}\n\nfunction normalize10(s: number): number {\n // JudgeScore convention is 0-10 but some judges emit 0-1. Coerce to 0-10.\n return s <= 1 ? s * 10 : s\n}\n\nfunction liftJudgeScore(analyst_id: string, area: string, s: JudgeScore): AnalystFinding {\n const score10 = normalize10(s.score)\n const severity: AnalystSeverity =\n score10 < 3 ? 'critical' : score10 < 5 ? 'high' : score10 < 7 ? 'medium' : 'low'\n return makeFinding({\n analyst_id,\n area,\n subject: s.dimension,\n claim: `${s.judgeName}/${s.dimension} scored ${score10.toFixed(1)}/10`,\n rationale: s.reasoning,\n severity,\n confidence: 0.8,\n evidence_refs: s.evidence\n ? [{ kind: 'artifact', uri: 'inline:evidence', excerpt: s.evidence }]\n : [],\n // Provenance: this finding IS a judge verdict (an acceptance score), not an\n // observation of behavior. The steer firewall (assertNoJudgeVerdict) rejects\n // it from steering — even when it cites an artifact above — because letting a\n // verdict steer the next attempt is the held-out judge leaking into the loop.\n derived_from_judge: true,\n metadata: { judge_name: s.judgeName, dimension: s.dimension, score_10: score10 },\n })\n}\n\n// ── 4. SemanticConceptJudge → Analyst ──────────────────────────────\n\nexport interface SemanticConceptJudgeAdapterOpts {\n id?: string\n area?: string\n /** Registry context owns cancellation and the per-analyst cost ledger. */\n options?: Omit<SemanticConceptJudgeOptions, 'costLedger' | 'signal'>\n /** Maximum post-cancellation wait for a provider receipt. Default 5 seconds. */\n settlementTimeoutMs?: number\n}\n\nexport function createSemanticConceptJudgeAdapter(\n opts: SemanticConceptJudgeAdapterOpts = {},\n): Analyst<SemanticConceptJudgeInput> {\n const id = opts.id ?? 'semantic-concept-judge'\n const area = opts.area ?? 'concept-coverage'\n const settlementTimeoutMs = validateUsageSettlementTimeout(opts.settlementTimeoutMs)\n return {\n id,\n description:\n 'Runs the semantic-concept judge and surfaces missing / weak concepts as findings.',\n inputKind: 'custom',\n cost: {\n kind: 'llm',\n models: opts.options?.model ? [opts.options.model] : undefined,\n settlement_timeout_ms: settlementTimeoutMs,\n },\n version: `${SEMANTIC_CONCEPT_JUDGE_VERSION}-adapter-${ADAPTER_REV}`,\n async analyze(input, ctx) {\n const costLedger = new CostLedger(ctx.budgetUsd)\n let result: SemanticConceptJudgeResult\n try {\n result = await runSemanticConceptJudge(input, {\n ...opts.options,\n costLedger,\n signal: ctx.signal,\n })\n } finally {\n const usage = await settleUsageReceiptFromCostLedger(costLedger, {\n channel: 'judge',\n timeoutMs: settlementTimeoutMs,\n })\n if (!usage.settled) {\n ctx.log?.('semantic-concept judge provider settlement timed out', {\n pending_calls: usage.pendingCalls,\n timeout_ms: settlementTimeoutMs,\n })\n }\n ctx.recordUsage?.(usage.receipt)\n }\n if (!result.available) {\n return [\n makeFinding({\n analyst_id: id,\n area,\n claim: 'semantic-concept judge unavailable',\n rationale: result.error,\n severity: 'info',\n confidence: 1,\n evidence_refs: [],\n metadata: { reason: result.error },\n }),\n ]\n }\n const out: AnalystFinding[] = []\n for (const f of result.findings) {\n // Only surface gaps: missing concepts or low scores. Concepts at\n // 7+/10 with present=true are not findings — they're successes.\n if (f.present && f.score >= 7) continue\n out.push(\n makeFinding({\n analyst_id: id,\n area,\n subject: f.concept,\n claim: f.present\n ? `concept \"${f.concept}\" is weak (${f.score}/10)`\n : `concept \"${f.concept}\" is missing`,\n rationale: f.evidence,\n severity: liftSeverity(f.severity),\n confidence: 0.85,\n evidence_refs: [{ kind: 'artifact', uri: 'inline:evidence', excerpt: f.evidence }],\n metadata: {\n concept: f.concept,\n present: f.present,\n score_10: f.score,\n },\n }),\n )\n }\n return out\n },\n }\n}\n","// The realness-oracle firewall (docs/learning-flywheel.md, \"The steer is f(trace)\").\n//\n// A realness/authenticity signal has TWO legitimate roles that must stay\n// separated by a firewall:\n// (a) anchor judge J — write-only: scores the chosen output, gates promotion,\n// NEVER seen by the worker/optimizer mid-run (else the loop games it).\n// (b) steer f(trace) — an analyst observes the agent's OWN behavior in the\n// trace (\"imported a stub\", \"used a non-crypto PRNG where encryption was\n// required\") and steers the next attempt. Legitimate, because it is derived\n// from OBSERVABLE BEHAVIOR, not from J's held-out verdict.\n//\n// The correct discriminator is PROVENANCE, not evidence presence. A judge verdict\n// lifted into a finding (createJudgeAdapter → liftJudgeScore) is a verdict even\n// when it cites an artifact; an evidence-less trace-analyst bullet is an\n// observation even though it cites nothing. So the firewall keys on\n// `AnalystFinding.derived_from_judge` (set at the judge lift site), NOT on whether\n// evidence_refs is populated. The instant a verdict steers the next attempt it is\n// a back-channel for J and the loop Goodharts realness exactly as it would\n// Goodhart pass-rate.\n\nimport type { AnalystFinding, EvidenceRef } from './types'\n\n/** Evidence grounded in the agent's OWN execution: OTLP trace elements\n * (`span`/`event`) or the artifact it produced (`artifact`). */\nconst OBSERVABLE_KINDS: ReadonlySet<EvidenceRef['kind']> = new Set<EvidenceRef['kind']>([\n 'span',\n 'event',\n 'artifact',\n])\n\n/** DESCRIPTIVE predicate: does the finding cite at least one observable\n * (span/event/artifact) evidence ref. Useful for ranking evidence quality or\n * rendering — it is NOT the steer gate. Evidence presence is the WRONG\n * discriminator for steering: a legitimate trace-analyst observation may cite\n * nothing (it would be wrongly rejected), and a judge verdict may cite an\n * artifact (it would be wrongly admitted). Use `assertNoJudgeVerdict` to gate\n * steering; use this only where \"is this grounded in observable evidence\" is the\n * literal question. */\nexport function isTraceObservable(finding: AnalystFinding): boolean {\n return finding.evidence_refs.some((ref) => OBSERVABLE_KINDS.has(ref.kind))\n}\n\n/** True iff the finding is a JUDGE VERDICT (an acceptance score lifted into a\n * finding), identified by provenance set at the lift site — independent of\n * whatever evidence it cites. */\nexport function isJudgeVerdict(finding: AnalystFinding): boolean {\n return finding.derived_from_judge === true\n}\n\n/**\n * THE steer firewall. Fail-loud guard for any path that admits analyst findings\n * as STEERING input (the `f(trace)` role): rejects — naming the offenders — any\n * finding whose provenance is a judge verdict, rather than let `J` leak into the\n * loop. Returns the findings unchanged for chaining.\n *\n * Call this at the chokepoint where a detector that ALSO scores/gates has its\n * findings turned into a steer (the judge-and-steer dual-role case). It keys on\n * provenance, so it correctly admits evidence-less trace-analyst observations and\n * correctly rejects an artifact-citing judge verdict — the cases an evidence\n * check gets backwards.\n *\n * It is necessary, not sufficient: it stops PROVENANCE-tagged verdicts. A judge\n * whose output is laundered through a hand-built finding with no provenance flag\n * is out of its reach — provenance must be honestly set at every judge→finding\n * lift (today: createJudgeAdapter). That is why the integrity rule lives at the\n * lift site, and why ProposeContext.judgeScores?: never is the complementary\n * compile-time tripwire on the obvious direct channel.\n */\nexport function assertNoJudgeVerdict(\n findings: ReadonlyArray<AnalystFinding>,\n context = 'steer',\n): ReadonlyArray<AnalystFinding> {\n const leaks = findings.filter(isJudgeVerdict)\n if (leaks.length > 0) {\n throw new Error(\n `${context}: a judge verdict cannot be admitted as steering input — that is the ` +\n `held-out judge leaking into the loop. Offending judge-derived findings: [${leaks\n .map((f) => f.finding_id)\n .join(', ')}]. Steering consumes observations of behavior, never acceptance verdicts.`,\n )\n }\n return findings\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuCA,IAAM,cAAc;AAIb,SAAS,aAAa,GAAmC;AAC9D,UAAQ,GAAG;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,EACX;AACF;AAeO,SAAS,sBAA2B,MAA8C;AACvF,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,gBAAgB;AAAA,IAC9B,SAAS,YAAY,WAAW;AAAA,IAChC,MAAM,QAAQ,KAAK,KAAK;AACtB,YAAM,SAAS,MAAM,KAAK,SAAS,IAAI,EAAE,KAAK,GAAG,KAAK,QAAQ,CAAC;AAC/D,YAAM,MAAwB,CAAC;AAC/B,iBAAW,SAAS,OAAO,QAAQ;AACjC,mBAAW,WAAW,MAAM,UAAU;AACpC,cAAI,KAAK,iBAAiB,IAAI,MAAM,MAAM,OAAO,OAAO,CAAC;AAAA,QAC3D;AAGA,YAAI,MAAM,WAAW,UAAU,MAAM,WAAW,WAAW,MAAM,WAAW,WAAW;AACrF,cAAI;AAAA,YACF,YAAY;AAAA,cACV,YAAY;AAAA,cACZ;AAAA,cACA,SAAS,MAAM;AAAA,cACf,OAAO,UAAU,MAAM,KAAK,KAAK,MAAM,MAAM,KAAK,MAAM,UAAU,iBAAiB;AAAA,cACnF,UACE,MAAM,WAAW,UAAU,SAAS,MAAM,WAAW,YAAY,WAAW;AAAA,cAC9E,YAAY;AAAA,cACZ,eAAe,CAAC;AAAA,cAChB,UAAU;AAAA,gBACR,cAAc,MAAM;AAAA,gBACpB,aAAa,MAAM;AAAA,gBACnB,OAAO,MAAM;AAAA,gBACb,aAAa,MAAM;AAAA,cACrB;AAAA,YACF,CAAC;AAAA,UACH;AAAA,QACF;AAAA,MACF;AACA,UAAI,MAAM,qBAAqB;AAAA,QAC7B,QAAQ,OAAO,OAAO;AAAA,QACtB,SAAS,OAAO;AAAA,QAChB,UAAU,OAAO;AAAA,MACnB,CAAC;AACD,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAEA,SAAS,iBACP,YACA,MACA,OACA,GACgB;AAChB,SAAO,YAAY;AAAA,IACjB;AAAA,IACA;AAAA,IACA,SAAS,EAAE,SAAS;AAAA,IACpB,OAAO,EAAE;AAAA,IACT,UAAU,aAAa,EAAE,QAAQ;AAAA,IACjC,YAAY;AAAA,IACZ,eAAe,EAAE,WACb,CAAC,EAAE,MAAM,YAAY,KAAK,mBAAmB,SAAS,EAAE,SAAS,CAAC,IAClE,CAAC;AAAA,IACL,UAAU,EAAE;AAAA,EACd,CAAC;AACH;AAYO,SAAS,uBAAuB,OAA6B,CAAC,GAAsB;AACzF,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,SAAS,KAAK,UAAU,IAAI,UAAU;AAC5C,QAAM,YAAY,KAAK,aAAa;AACpC,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,gBAAgB;AAAA,IAC9B,SAAS,cAAc,WAAW;AAAA,IAClC,MAAM,QAAQ,OAAO;AACnB,YAAM,QAAQ,OAAO,WAAW,KAAK;AACrC,YAAM,MAAwB,CAAC;AAC/B,YAAM,OAA6D;AAAA,QACjE,CAAC,WAAW,YAAY,mCAAmC;AAAA,QAC3D,CAAC,gBAAgB,QAAQ,sBAAsB;AAAA,QAC/C,CAAC,oBAAoB,QAAQ,6CAA6C;AAAA,QAC1E,CAAC,kBAAkB,UAAU,yBAAyB;AAAA,QACtD,CAAC,gBAAgB,UAAU,6BAA6B;AAAA,QACxD,CAAC,eAAe,QAAQ,6BAA6B;AAAA,QACrD,CAAC,aAAa,YAAY,wBAAwB;AAAA,MACpD;AACA,iBAAW,CAAC,KAAK,KAAK,GAAG,KAAK,MAAM;AAClC,cAAM,QAAQ,MAAM,GAAG;AACvB,YAAI,OAAO,UAAU,YAAY,QAAQ,WAAW;AAClD,cAAI;AAAA,YACF,YAAY;AAAA,cACV,YAAY;AAAA,cACZ;AAAA,cACA,SAAS;AAAA,cACT,OAAO;AAAA,cACP,WAAW,GAAG,GAAG,IAAI,MAAM,QAAQ,CAAC,CAAC,oBAAoB,SAAS;AAAA,cAClE,UAAU;AAAA,cACV,YAAY;AAAA,cACZ,eAAe,CAAC;AAAA,cAChB,UAAU,EAAE,WAAW,KAAK,OAAO,WAAW,QAAQ,MAAM,IAAI,MAAM;AAAA,YACxE,CAAC;AAAA,UACH;AAAA,QACF;AAAA,MACF;AAEA,UAAI,MAAM,eAAe,IAAI,WAAW;AACtC,YAAI;AAAA,UACF,YAAY;AAAA,YACV,YAAY;AAAA,YACZ;AAAA,YACA,SAAS;AAAA,YACT,OAAO;AAAA,YACP,WAAW,gBAAgB,MAAM,aAAa,QAAQ,CAAC,CAAC;AAAA,YACxD,UAAU;AAAA,YACV,YAAY;AAAA,YACZ,eAAe,CAAC;AAAA,YAChB,UAAU,EAAE,eAAe,MAAM,cAAc,OAAO,MAAM,MAAM;AAAA,UACpE,CAAC;AAAA,QACH;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAgBO,SAAS,mBAAmB,MAA6C;AAC9E,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,YAAY,KAAK,aAAa;AACpC,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,KAAK,QAAQ,EAAE,MAAM,MAAM;AAAA,IACjC,SAAS,SAAS,WAAW;AAAA,IAC7B,MAAM,QAAQ,OAAO;AACnB,YAAM,SAAS,MAAM,KAAK,MAAM,KAAK,QAAQ,KAAK;AAClD,aAAO,OACJ,OAAO,CAAC,MAAM,YAAY,EAAE,KAAK,IAAI,SAAS,EAC9C,IAAI,CAAC,MAAM,eAAe,IAAI,MAAM,CAAC,CAAC;AAAA,IAC3C;AAAA,EACF;AACF;AAEA,SAAS,YAAY,GAAmB;AAEtC,SAAO,KAAK,IAAI,IAAI,KAAK;AAC3B;AAEA,SAAS,eAAe,YAAoB,MAAc,GAA+B;AACvF,QAAM,UAAU,YAAY,EAAE,KAAK;AACnC,QAAM,WACJ,UAAU,IAAI,aAAa,UAAU,IAAI,SAAS,UAAU,IAAI,WAAW;AAC7E,SAAO,YAAY;AAAA,IACjB;AAAA,IACA;AAAA,IACA,SAAS,EAAE;AAAA,IACX,OAAO,GAAG,EAAE,SAAS,IAAI,EAAE,SAAS,WAAW,QAAQ,QAAQ,CAAC,CAAC;AAAA,IACjE,WAAW,EAAE;AAAA,IACb;AAAA,IACA,YAAY;AAAA,IACZ,eAAe,EAAE,WACb,CAAC,EAAE,MAAM,YAAY,KAAK,mBAAmB,SAAS,EAAE,SAAS,CAAC,IAClE,CAAC;AAAA;AAAA;AAAA;AAAA;AAAA,IAKL,oBAAoB;AAAA,IACpB,UAAU,EAAE,YAAY,EAAE,WAAW,WAAW,EAAE,WAAW,UAAU,QAAQ;AAAA,EACjF,CAAC;AACH;AAaO,SAAS,kCACd,OAAwC,CAAC,GACL;AACpC,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,sBAAsB,+BAA+B,KAAK,mBAAmB;AACnF,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM;AAAA,MACJ,MAAM;AAAA,MACN,QAAQ,KAAK,SAAS,QAAQ,CAAC,KAAK,QAAQ,KAAK,IAAI;AAAA,MACrD,uBAAuB;AAAA,IACzB;AAAA,IACA,SAAS,GAAG,8BAA8B,YAAY,WAAW;AAAA,IACjE,MAAM,QAAQ,OAAO,KAAK;AACxB,YAAM,aAAa,IAAI,WAAW,IAAI,SAAS;AAC/C,UAAI;AACJ,UAAI;AACF,iBAAS,MAAM,wBAAwB,OAAO;AAAA,UAC5C,GAAG,KAAK;AAAA,UACR;AAAA,UACA,QAAQ,IAAI;AAAA,QACd,CAAC;AAAA,MACH,UAAE;AACA,cAAM,QAAQ,MAAM,iCAAiC,YAAY;AAAA,UAC/D,SAAS;AAAA,UACT,WAAW;AAAA,QACb,CAAC;AACD,YAAI,CAAC,MAAM,SAAS;AAClB,cAAI,MAAM,wDAAwD;AAAA,YAChE,eAAe,MAAM;AAAA,YACrB,YAAY;AAAA,UACd,CAAC;AAAA,QACH;AACA,YAAI,cAAc,MAAM,OAAO;AAAA,MACjC;AACA,UAAI,CAAC,OAAO,WAAW;AACrB,eAAO;AAAA,UACL,YAAY;AAAA,YACV,YAAY;AAAA,YACZ;AAAA,YACA,OAAO;AAAA,YACP,WAAW,OAAO;AAAA,YAClB,UAAU;AAAA,YACV,YAAY;AAAA,YACZ,eAAe,CAAC;AAAA,YAChB,UAAU,EAAE,QAAQ,OAAO,MAAM;AAAA,UACnC,CAAC;AAAA,QACH;AAAA,MACF;AACA,YAAM,MAAwB,CAAC;AAC/B,iBAAW,KAAK,OAAO,UAAU;AAG/B,YAAI,EAAE,WAAW,EAAE,SAAS,EAAG;AAC/B,YAAI;AAAA,UACF,YAAY;AAAA,YACV,YAAY;AAAA,YACZ;AAAA,YACA,SAAS,EAAE;AAAA,YACX,OAAO,EAAE,UACL,YAAY,EAAE,OAAO,cAAc,EAAE,KAAK,SAC1C,YAAY,EAAE,OAAO;AAAA,YACzB,WAAW,EAAE;AAAA,YACb,UAAU,aAAa,EAAE,QAAQ;AAAA,YACjC,YAAY;AAAA,YACZ,eAAe,CAAC,EAAE,MAAM,YAAY,KAAK,mBAAmB,SAAS,EAAE,SAAS,CAAC;AAAA,YACjF,UAAU;AAAA,cACR,SAAS,EAAE;AAAA,cACX,SAAS,EAAE;AAAA,cACX,UAAU,EAAE;AAAA,YACd;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;;;ACvVA,IAAM,mBAAqD,oBAAI,IAAyB;AAAA,EACtF;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAUM,SAAS,kBAAkB,SAAkC;AAClE,SAAO,QAAQ,cAAc,KAAK,CAAC,QAAQ,iBAAiB,IAAI,IAAI,IAAI,CAAC;AAC3E;AAKO,SAAS,eAAe,SAAkC;AAC/D,SAAO,QAAQ,uBAAuB;AACxC;AAqBO,SAAS,qBACd,UACA,UAAU,SACqB;AAC/B,QAAM,QAAQ,SAAS,OAAO,cAAc;AAC5C,MAAI,MAAM,SAAS,GAAG;AACpB,UAAM,IAAI;AAAA,MACR,GAAG,OAAO,sJACoE,MACzE,IAAI,CAAC,MAAM,EAAE,UAAU,EACvB,KAAK,IAAI,CAAC;AAAA,IACjB;AAAA,EACF;AACA,SAAO;AACT;","names":[]}