@tangle-network/agent-eval 0.93.0 → 0.95.0

package/dist/analyst/findings-store.d.ts

@@ -0,0 +1,75 @@
+/**
+ * FindingsStore — durable persistence for AnalystFinding rows + a diff
+ * helper so we can answer "what changed since the last run?" without
+ * recomputing analysts.
+ *
+ * On-disk shape is JSONL: one finding per line, append-only, locked via
+ * LockedJsonlAppender. Operators get crash-safety (no partial JSON),
+ * cheap reads (sequential parse), and trivial backup (rsync the file).
+ *
+ * Reads are non-locking: a reader sees a consistent snapshot of all
+ * fully-written lines and skips an incomplete trailing line if the
+ * writer is mid-append. Cross-process locking is intentionally out of
+ * scope (see locked-jsonl-appender.ts).
+ *
+ * The store is run-scoped: callers pass `runId` on append and on load,
+ * which keeps multi-run files cleanly partitioned. The `diffFindings`
+ * helper compares two run-id sets using stable `finding_id` semantics —
+ * the diff is the cross-run signal the regression dashboard renders.
+ */
+import type { AnalystFinding } from './types';
+/**
+ * One persisted row. We attach `run_id` on disk so a single file can
+ * hold multiple runs and the diff helper can query without re-walking
+ * separate files.
+ */
+export interface PersistedFinding extends AnalystFinding {
+    run_id: string;
+}
+export declare class FindingsStore {
+    readonly path: string;
+    private readonly appender;
+    constructor(path: string);
+    append(runId: string, findings: AnalystFinding[]): Promise<void>;
+    /** Load every persisted finding. Discards malformed trailing lines silently. */
+    loadAll(): PersistedFinding[];
+    /** Filter to a single run. */
+    loadRun(runId: string): PersistedFinding[];
+}
+export interface FindingsDiff {
+    /** New finding ids in `current` that weren't in `previous`. */
+    appeared: PersistedFinding[];
+    /** Finding ids in `previous` that aren't in `current`. */
+    disappeared: PersistedFinding[];
+    /** Same finding id present in both runs and unchanged per the materiality test. */
+    persisted: PersistedFinding[];
+    /**
+     * Same finding id in both runs but at least one non-identity field
+     * shifted per `DiffPolicy.isMaterial`. Reported as [previous, current].
+     */
+    changed: Array<{
+        previous: PersistedFinding;
+        current: PersistedFinding;
+    }>;
+}
+export interface DiffPolicy {
+    /**
+     * Predicate that decides whether two findings (same finding_id) count
+     * as a material change. Defaults to {@link defaultIsMaterial}: severity
+     * shift, confidence Δ > 0.05, or evidence count change. Compliance /
+     * perf consumers MAY supply a stricter predicate (e.g. rationale text
+     * diff, metric Δ thresholds).
+     */
+    isMaterial?: (previous: AnalystFinding, current: AnalystFinding) => boolean;
+}
+/**
+ * Default materiality test. Deliberately narrow so LLM-reword churn
+ * doesn't flood the diff. Stricter tests are opt-in via DiffPolicy.
+ */
+export declare function defaultIsMaterial(a: AnalystFinding, b: AnalystFinding): boolean;
+/**
+ * Diff two findings sets by stable finding_id. Callers typically load
+ * the two run-id slices from the same store and pass them in.
+ */
+export declare function diffFindings(previous: PersistedFinding[], current: PersistedFinding[], policy?: DiffPolicy): FindingsDiff;
+//# sourceMappingURL=findings-store.d.ts.map

package/dist/analyst/findings-store.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"findings-store.d.ts","sourceRoot":"","sources":["../../src/analyst/findings-store.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAKH,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,SAAS,CAAA;AAE7C;;;;GAIG;AACH,MAAM,WAAW,gBAAiB,SAAQ,cAAc;IACtD,MAAM,EAAE,MAAM,CAAA;CACf;AAED,qBAAa,aAAa;aAGI,IAAI,EAAE,MAAM;IAFxC,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAqB;gBAElB,IAAI,EAAE,MAAM;IAIlC,MAAM,CAAC,KAAK,EAAE,MAAM,EAAE,QAAQ,EAAE,cAAc,EAAE,GAAG,OAAO,CAAC,IAAI,CAAC;IAOtE,gFAAgF;IAChF,OAAO,IAAI,gBAAgB,EAAE;IAiB7B,8BAA8B;IAC9B,OAAO,CAAC,KAAK,EAAE,MAAM,GAAG,gBAAgB,EAAE;CAG3C;AAID,MAAM,WAAW,YAAY;IAC3B,+DAA+D;IAC/D,QAAQ,EAAE,gBAAgB,EAAE,CAAA;IAC5B,0DAA0D;IAC1D,WAAW,EAAE,gBAAgB,EAAE,CAAA;IAC/B,mFAAmF;IACnF,SAAS,EAAE,gBAAgB,EAAE,CAAA;IAC7B;;;OAGG;IACH,OAAO,EAAE,KAAK,CAAC;QAAE,QAAQ,EAAE,gBAAgB,CAAC;QAAC,OAAO,EAAE,gBAAgB,CAAA;KAAE,CAAC,CAAA;CAC1E;AAED,MAAM,WAAW,UAAU;IACzB;;;;;;OAMG;IACH,UAAU,CAAC,EAAE,CAAC,QAAQ,EAAE,cAAc,EAAE,OAAO,EAAE,cAAc,KAAK,OAAO,CAAA;CAC5E;AAED;;;GAGG;AACH,wBAAgB,iBAAiB,CAAC,CAAC,EAAE,cAAc,EAAE,CAAC,EAAE,cAAc,GAAG,OAAO,CAK/E;AAED;;;GAGG;AACH,wBAAgB,YAAY,CAC1B,QAAQ,EAAE,gBAAgB,EAAE,EAC5B,OAAO,EAAE,gBAAgB,EAAE,EAC3B,MAAM,GAAE,UAAe,GACtB,YAAY,CA0Bd"}

package/dist/analyst/index.d.ts

@@ -1,256 +1,28 @@
-import { AxAIService, AxFunction } from '@ax-llm/ax';
-import { M as MultiLayerVerifier, V as VerifyOptions, S as Severity } from '../multi-layer-verifier-DUZXrPDA.js';
-import { c as RunCritic, a as RunTrace } from '../run-critic-BAIjX99r.js';
-import { S as SemanticConceptJudgeOptions, a as SemanticConceptJudgeInput, B as BehavioralMetrics } from '../semantic-concept-judge-Dn8Z6KEG.js';
-export { C as CreateAnalystAiConfig, D as DEFAULT_TRACE_ANALYST_KINDS, b as DiffPolicy, F as FAILURE_MODE_KIND_SPEC, c as FINDING_SUBJECT_GRAMMAR_PROMPT, d as FINDING_SUBJECT_KINDS, e as FindingSubject, f as FindingSubjectKind, g as FindingSubjectStringSchema, h as FindingsDiff, i as FindingsStore, I as IMPROVEMENT_KIND_SPEC, K as KIND_EXPECTED_SUBJECTS, j as KNOWLEDGE_GAP_KIND_SPEC, k as KNOWLEDGE_POISONING_KIND_SPEC, P as PersistedFinding, l as SKILL_USAGE_ANALYST, m as SkillUsageAnalyst, n as SkillUsageRecord, o as SkillUsageReport, p as SkillUsageScanConfig, q as buildSkillUsageReport, r as createAnalystAi, s as defaultIsMaterial, t as diffFindings, u as emitSkillUsageFindings, v as parseFindingSubject, w as renderFindingSubject } from '../semantic-concept-judge-Dn8Z6KEG.js';
-import { A as AnalyzeTracesOptions } from '../analyst-C8HHvfJp.js';
-import { T as TraceAnalysisStore } from '../store-C1YxJDEK.js';
-import { b as JudgeFn, a as JudgeInput } from '../types-Croy5h7V.js';
-import { A as Analyst, h as AnalystSeverity, c as AnalystFinding } from '../types-2VVIL04s.js';
-export { a as AnalystContext, g as AnalystCost, i as AnalystInputKind, j as AnalystRequirements, f as AnalystRunEvent, e as AnalystRunInputs, d as AnalystRunResult, b as AnalystRunSummary, k as ChatCallOpts, C as ChatClient, l as ChatRequest, m as ChatResponse, n as ChatTransport, o as CliBridgeTransportOpts, p as CreateChatClientOpts, D as DirectProviderTransportOpts, E as EvidenceRef, M as MockTransportOpts, R as RouterTransportOpts, S as SandboxSdkTransportOpts, q as computeFindingId, r as createChatClient, s as makeFinding } from '../types-2VVIL04s.js';
-import { TCloud } from '@tangle-network/tcloud';
-export { a as AnalystHooks, A as AnalystRegistry, c as AnalystRegistryOptions, B as BudgetPolicy, D as DefaultAnalystRegistryOptions, R as RegistryRunOpts, b as buildDefaultAnalystRegistry } from '../default-registry-zoGHUQEH.js';
-export { A as ANALYST_SEVERITIES, C as CreateTraceAnalystKindOpts, R as RAW_FINDING_SCHEMA_PROMPT, a as RawAnalystFinding, b as RawAnalystFindingSchema, c as TraceAnalystGolden, T as TraceAnalystKindSpec, d as createTraceAnalystKind, p as parseRawFinding, r as renderPriorFindings } from '../kind-factory-5b7xXXOr.js';
-import { L as LlmClientOptions } from '../llm-client-BeEcAokY.js';
-import '../verdict-C9MlYujm.js';
-import '../schema-m0gsnbt3.js';
-import '../store-CKUAgsJz.js';
-import 'zod';
-import '../run-record-e7vj1uZQ.js';
-import '../errors-CzMUYo7b.js';
-import '../raw-provider-sink-C46HDghv.js';
-
-/**
- * Adapter factories — lift each existing agent-eval primitive into the
- * Analyst contract without re-implementing it.
- *
- * Five primitives, five factories. Each one:
- *   - Builds an Analyst with a stable id (caller chooses; defaults
- *     given), a sensible default `inputKind`, a version derived from
- *     the wrapped primitive's version + an adapter revision, and an
- *     `analyze()` that calls the primitive and lifts its output to
- *     AnalystFinding[] using `makeFinding()`.
- *   - Maps severities: the existing `Severity` ('critical' | 'major' |
- *     'minor' | 'info') projects onto AnalystSeverity ('critical' |
- *     'high' | 'medium' | 'low' | 'info'); 'major' → 'high', 'minor' →
- *     'medium'. Domain analysts that want finer-grained mapping override.
- *
- * Adapters never own state. Calling the same factory twice with the
- * same primitive instance is safe.
- */
-
-declare function liftSeverity(s: Severity): AnalystSeverity;
-interface TraceAnalystAdapterOpts {
-    id?: string;
-    area?: string;
-    /** The natural-language question(s) put to the analyst. One finding per question. */
-    questions: string[];
-    /** Caller-provided AxAI service — same one trace-analyst.ts expects. */
-    ai: AxAIService;
-    model?: string;
-    /** Forwarded to analyzeTraces. */
-    extra?: Omit<AnalyzeTracesOptions, 'source' | 'ai' | 'model'>;
-}
-/**
- * @deprecated Prefer `createTraceAnalystKind` + one of the failure /
- * improvement kinds from `./kinds`. This adapter wraps the legacy
- * `analyzeTraces` flow whose output is `findings:string[]` — every
- * bullet gets flat-defaulted severity `medium` / confidence `0.6`,
- * which loses the per-finding grading kinds provide via Ax structured
- * output + Zod validation. Kept for one minor while consumers migrate.
- */
-declare function createTraceAnalystAdapter(opts: TraceAnalystAdapterOpts): Analyst<TraceAnalysisStore>;
-interface VerifierAdapterOpts<Env> {
-    id?: string;
-    area?: string;
-    verifier: MultiLayerVerifier<Env>;
-    /**
-     * The verifier expects an `env` per run. Adapters take it from
-     * `AnalystRunInputs.custom[<id>]` via the registry's 'custom' routing.
-     */
-    options?: Omit<VerifyOptions<Env>, 'env'>;
-}
-declare function createVerifierAdapter<Env>(opts: VerifierAdapterOpts<Env>): Analyst<Env>;
-interface RunCriticAdapterOpts {
-    id?: string;
-    area?: string;
-    critic?: RunCritic;
-    /** Optional threshold below which a dimension is reported as a finding. Default 0.5. */
-    threshold?: number;
-}
-declare function createRunCriticAdapter(opts?: RunCriticAdapterOpts): Analyst<RunTrace>;
-interface JudgeAdapterOpts {
-    id?: string;
-    area?: string;
-    judge: JudgeFn;
-    /** TCloud handle the JudgeFn calls. */
-    tcloud: TCloud;
-    /** Optional cost classification — most judges call an LLM. */
-    cost?: Analyst['cost'];
-    /** Optional threshold below which a JudgeScore becomes a finding. Default 6 (on 0-10 scale). */
-    threshold?: number;
-}
-declare function createJudgeAdapter(opts: JudgeAdapterOpts): Analyst<JudgeInput>;
-interface SemanticConceptJudgeAdapterOpts {
-    id?: string;
-    area?: string;
-    options?: SemanticConceptJudgeOptions;
-}
-declare function createSemanticConceptJudgeAdapter(opts?: SemanticConceptJudgeAdapterOpts): Analyst<SemanticConceptJudgeInput>;
-
-/**
- * `behavioralAnalyst` — a DETERMINISTIC analyst (cost.kind = 'deterministic',
- * never calls the LLM). It produces the efficiency/behavioral findings a
- * tolerant agentic analyzer (HALO) re-derives per run inside the model —
- * context bloat, output decay, tool monoculture, missing self-verification —
- * directly from arithmetic over spans (`computeTraceMetrics`).
- *
- * Why it matters: these findings are model-agnostic BY CONSTRUCTION (no model
- * in the loop), so they cannot return 0 on a weak model the way the Ax-RLM
- * does — and they are strictly more reliable than HALO, which spends tokens
- * re-deriving the same numbers and can hallucinate the trend. The agentic
- * RLM kinds remain for SEMANTIC findings that genuinely need a model; this
- * analyst owns the behavioral class.
- */
-
-/**
- * Map computed signals → structured AnalystFindings. Pure: no LLM, no clock
- * dependence beyond `produced_at` (overridable for deterministic tests).
- */
-declare function deriveEfficiencyFindings(metrics: BehavioralMetrics, opts?: {
-    analystId?: string;
-    producedAt?: string;
-}): AnalystFinding[];
-/** The deterministic behavioral/efficiency analyst (no LLM, any-model). */
-declare function behavioralAnalyst(): Analyst<TraceAnalysisStore>;
-
-/**
- * Forgiving pre-parse for analyst findings. Weak models routinely emit
- * schema-correct content in an unusable wrapper — fenced ```json blocks, a
- * single object where an array is expected, trailing commas. Measured: GPT-4o
- * drops to 0% usable output purely from markdown-fence wrapping
- * (arXiv:2605.02363). A five-line de-fence recovers most of it. This module is
- * the de-fence/coerce step that runs BEFORE Zod, so a recoverable finding is
- * repaired, not dropped.
- *
- * Pure + deterministic. No model, no network.
- */
-/** Strip a ```lang ... ``` (or bare ``` ... ```) code fence, if the string is one. */
-declare function stripCodeFences(text: string): string;
-/**
- * Best-effort parse of a string into JSON. De-fences, drops trailing commas,
- * then `JSON.parse`. Returns `undefined` (never throws) when unrecoverable.
- */
-declare function coerceJson(text: string): unknown;
-/**
- * Coerce arbitrary actor/structurer output into an array of candidate finding
- * rows: a JSON string → parse; a single object → 1-element array; an array →
- * as-is; anything else → []. Callers still run each row through Zod
- * (`parseRawFinding`) — this only fixes the SHAPE, never invents fields.
- */
-declare function coerceToFindingRows(raw: unknown): unknown[];
-
-/** DESCRIPTIVE predicate: does the finding cite at least one observable
- *  (span/event/artifact) evidence ref. Useful for ranking evidence quality or
- *  rendering — it is NOT the steer gate. Evidence presence is the WRONG
- *  discriminator for steering: a legitimate trace-analyst observation may cite
- *  nothing (it would be wrongly rejected), and a judge verdict may cite an
- *  artifact (it would be wrongly admitted). Use `assertNoJudgeVerdict` to gate
- *  steering; use this only where "is this grounded in observable evidence" is the
- *  literal question. */
-declare function isTraceObservable(finding: AnalystFinding): boolean;
-/** True iff the finding is a JUDGE VERDICT (an acceptance score lifted into a
- *  finding), identified by provenance set at the lift site — independent of
- *  whatever evidence it cites. */
-declare function isJudgeVerdict(finding: AnalystFinding): boolean;
-/**
- * THE steer firewall. Fail-loud guard for any path that admits analyst findings
- * as STEERING input (the `f(trace)` role): rejects — naming the offenders — any
- * finding whose provenance is a judge verdict, rather than let `J` leak into the
- * loop. Returns the findings unchanged for chaining.
- *
- * Call this at the chokepoint where a detector that ALSO scores/gates has its
- * findings turned into a steer (the judge-and-steer dual-role case). It keys on
- * provenance, so it correctly admits evidence-less trace-analyst observations and
- * correctly rejects an artifact-citing judge verdict — the cases an evidence
- * check gets backwards.
- *
- * It is necessary, not sufficient: it stops PROVENANCE-tagged verdicts. A judge
- * whose output is laundered through a hand-built finding with no provenance flag
- * is out of its reach — provenance must be honestly set at every judge→finding
- * lift (today: createJudgeAdapter). That is why the integrity rule lives at the
- * lift site, and why ProposeContext.judgeScores?: never is the complementary
- * compile-time tripwire on the obvious direct channel.
- */
-declare function assertNoJudgeVerdict(findings: ReadonlyArray<AnalystFinding>, context?: string): ReadonlyArray<AnalystFinding>;
-
-/**
- * `structureFindings` — the deferred structuring pass (DSPy TwoStepAdapter /
- * HALO `synthesize_traces` analog). The agentic actor reasons FREE-FORM and
- * emits a prose `report` (which any model does reliably); this separate, cheap
- * call's ONLY job is to turn that report into `AnalystFinding[]`. Decoupling
- * reasoning from structuring is what makes the SEMANTIC findings model-agnostic
- * — the reasoning model never has to satisfy a strict typed-array contract
- * while it diagnoses.
- *
- * Forgiving: the response runs through `coerceToFindingRows` (de-fence, lift
- * single→array) before Zod, and on a zero-finding extraction from a substantive
- * report it reasks ONCE with the schema restated. Returns a typed outcome so a
- * legitimate "nothing to report" is distinguishable from a failed extraction
- * (no silent empty).
- */
-
-interface StructureFindingsOptions {
-    /** The actor's free-form diagnosis prose. */
-    report: string;
-    analystId: string;
-    /** Coarse classification stamped on every extracted finding. */
-    area: string;
-    model: string;
-    baseUrl: string;
-    apiKey?: string;
-    /** Max reask attempts after a zero/invalid extraction. Default 1. */
-    maxReasks?: number;
-    /** Test seam: inject a fetch (no network in unit tests). */
-    fetchImpl?: LlmClientOptions['fetch'];
-}
-interface StructureFindingsResult {
-    findings: AnalystFinding[];
-    outcome: 'ok' | 'extraction_failed';
-}
-declare function structureFindings(opts: StructureFindingsOptions): Promise<StructureFindingsResult>;
-
-/**
- * Pre-curated tool subsets for analyst kinds.
- *
- * The full trace-analyst tool set is seven functions. Most kinds only
- * need three or four. Picking from named groups instead of importing
- * the whole bundle keeps every kind's actor-context budget tight and
- * makes "what can this analyst see?" obvious at registration time.
- *
- * Each function in the group keeps its full `name`/`description` from
- * `buildTraceAnalystTools` — we filter, we don't re-implement.
- */
-
-/** Named tool sets. Kinds pass `tools: TRACE_TOOL_GROUPS.failureForensics` etc. */
-type TraceToolGroupName = 
-/** All seven tools. Use for open-ended discovery kinds. */
-'all'
-/** Overview + paginated query + count. No deep reads. Cheap. */
- | 'discovery'
-/** Discovery + viewTrace + viewSpans. Deep-read but no regex search. */
- | 'discoveryAndRead'
-/** Discovery + search tools. For pattern-matching across many traces. */
- | 'discoveryAndSearch'
-/** Discovery + viewSpans + searchSpan. Targeted-span work after another kind narrows down. */
- | 'targeted';
-/**
- * Build the tool set for a named group bound to a specific trace store.
- *
- * `all` returns every tool. Other groups filter `buildTraceAnalystTools`
- * by name to the documented subset. An unrecognised group name throws —
- * silently returning all tools would defeat the cost-control point.
- */
-declare function buildTraceToolsForGroup(group: TraceToolGroupName, store: TraceAnalysisStore): AxFunction[];
-
-export { Analyst, AnalystFinding, AnalystSeverity, type JudgeAdapterOpts, type RunCriticAdapterOpts, type SemanticConceptJudgeAdapterOpts, type StructureFindingsOptions, type StructureFindingsResult, type TraceAnalystAdapterOpts, type TraceToolGroupName, type VerifierAdapterOpts, assertNoJudgeVerdict, behavioralAnalyst, buildTraceToolsForGroup, coerceJson, coerceToFindingRows, createJudgeAdapter, createRunCriticAdapter, createSemanticConceptJudgeAdapter, createTraceAnalystAdapter, createVerifierAdapter, deriveEfficiencyFindings, isJudgeVerdict, isTraceObservable, liftSeverity, stripCodeFences, structureFindings };
+export type { JudgeAdapterOpts, RunCriticAdapterOpts, SemanticConceptJudgeAdapterOpts, VerifierAdapterOpts, } from './adapters';
+export { createJudgeAdapter, createRunCriticAdapter, createSemanticConceptJudgeAdapter, createVerifierAdapter, liftSeverity, } from './adapters';
+export { type CreateAnalystAiConfig, createAnalystAi } from './ax-service';
+export { behavioralAnalyst, deriveEfficiencyFindings } from './behavioral-analyst';
+export type { ChatCallOpts, ChatClient, ChatRequest, ChatResponse, ChatTransport, CliBridgeTransportOpts, CreateChatClientOpts, DirectProviderTransportOpts, MockTransportOpts, RouterTransportOpts, SandboxSdkTransportOpts, } from './chat-client';
+export { createChatClient } from './chat-client';
+export { buildDefaultAnalystRegistry, type DefaultAnalystRegistryOptions, } from './default-registry';
+export type { RawAnalystFinding } from './finding-signature';
+export { ANALYST_SEVERITIES, parseRawFinding, RAW_FINDING_SCHEMA_PROMPT, RawAnalystFindingSchema, } from './finding-signature';
+export type { FindingSubject, FindingSubjectKind } from './finding-subject';
+export { FINDING_SUBJECT_GRAMMAR_PROMPT, FINDING_SUBJECT_KINDS, FindingSubjectStringSchema, KIND_EXPECTED_SUBJECTS, parseFindingSubject, renderFindingSubject, } from './finding-subject';
+export type { DiffPolicy, FindingsDiff, PersistedFinding } from './findings-store';
+export { defaultIsMaterial, diffFindings, FindingsStore } from './findings-store';
+export type { CreateTraceAnalystKindOpts, TraceAnalystGolden, TraceAnalystKindSpec, } from './kind-factory';
+export { createTraceAnalystKind, renderPriorFindings } from './kind-factory';
+export { DEFAULT_TRACE_ANALYST_KINDS, FAILURE_MODE_KIND_SPEC, IMPROVEMENT_KIND_SPEC, KNOWLEDGE_GAP_KIND_SPEC, KNOWLEDGE_POISONING_KIND_SPEC, } from './kinds';
+export type { SkillUsageRecord, SkillUsageReport, SkillUsageScanConfig, } from './kinds/skill-usage';
+export { buildSkillUsageReport, emitSkillUsageFindings, SKILL_USAGE_ANALYST, SkillUsageAnalyst, } from './kinds/skill-usage';
+export { coerceJson, coerceToFindingRows, stripCodeFences } from './parse-tolerant';
+export type { AnalystHooks, AnalystRegistryOptions, BudgetPolicy, RegistryRunOpts, } from './registry';
+export { AnalystRegistry } from './registry';
+export { assertNoJudgeVerdict, isJudgeVerdict, isTraceObservable } from './steer-firewall';
+export { type StructureFindingsOptions, type StructureFindingsResult, structureFindings, } from './structure-findings';
+export type { TraceToolGroupName } from './tool-groups';
+export { buildTraceToolsForGroup } from './tool-groups';
+export type { Analyst, AnalystContext, AnalystCost, AnalystFinding, AnalystInputKind, AnalystRequirements, AnalystRunEvent, AnalystRunInputs, AnalystRunResult, AnalystRunSummary, AnalystSeverity, EvidenceRef, } from './types';
+export { computeFindingId, makeFinding } from './types';
+//# sourceMappingURL=index.d.ts.map

package/dist/analyst/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/analyst/index.ts"],"names":[],"mappings":"AAQA,YAAY,EACV,gBAAgB,EAChB,oBAAoB,EACpB,+BAA+B,EAC/B,mBAAmB,GACpB,MAAM,YAAY,CAAA;AACnB,OAAO,EACL,kBAAkB,EAClB,sBAAsB,EACtB,iCAAiC,EACjC,qBAAqB,EACrB,YAAY,GACb,MAAM,YAAY,CAAA;AACnB,OAAO,EAAE,KAAK,qBAAqB,EAAE,eAAe,EAAE,MAAM,cAAc,CAAA;AAC1E,OAAO,EAAE,iBAAiB,EAAE,wBAAwB,EAAE,MAAM,sBAAsB,CAAA;AAClF,YAAY,EACV,YAAY,EACZ,UAAU,EACV,WAAW,EACX,YAAY,EACZ,aAAa,EACb,sBAAsB,EACtB,oBAAoB,EACpB,2BAA2B,EAC3B,iBAAiB,EACjB,mBAAmB,EACnB,uBAAuB,GACxB,MAAM,eAAe,CAAA;AACtB,OAAO,EAAE,gBAAgB,EAAE,MAAM,eAAe,CAAA;AAChD,OAAO,EACL,2BAA2B,EAC3B,KAAK,6BAA6B,GACnC,MAAM,oBAAoB,CAAA;AAC3B,YAAY,EAAE,iBAAiB,EAAE,MAAM,qBAAqB,CAAA;AAC5D,OAAO,EACL,kBAAkB,EAClB,eAAe,EACf,yBAAyB,EACzB,uBAAuB,GACxB,MAAM,qBAAqB,CAAA;AAC5B,YAAY,EAAE,cAAc,EAAE,kBAAkB,EAAE,MAAM,mBAAmB,CAAA;AAC3E,OAAO,EACL,8BAA8B,EAC9B,qBAAqB,EACrB,0BAA0B,EAC1B,sBAAsB,EACtB,mBAAmB,EACnB,oBAAoB,GACrB,MAAM,mBAAmB,CAAA;AAC1B,YAAY,EAAE,UAAU,EAAE,YAAY,EAAE,gBAAgB,EAAE,MAAM,kBAAkB,CAAA;AAClF,OAAO,EAAE,iBAAiB,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,kBAAkB,CAAA;AACjF,YAAY,EACV,0BAA0B,EAC1B,kBAAkB,EAClB,oBAAoB,GACrB,MAAM,gBAAgB,CAAA;AACvB,OAAO,EAAE,sBAAsB,EAAE,mBAAmB,EAAE,MAAM,gBAAgB,CAAA;AAC5E,OAAO,EACL,2BAA2B,EAC3B,sBAAsB,EACtB,qBAAqB,EACrB,uBAAuB,EACvB,6BAA6B,GAC9B,MAAM,SAAS,CAAA;AAChB,YAAY,EACV,gBAAgB,EAChB,gBAAgB,EAChB,oBAAoB,GACrB,MAAM,qBAAqB,CAAA;AAC5B,OAAO,EACL,qBAAqB,EACrB,sBAAsB,EACtB,mBAAmB,EACnB,iBAAiB,GAClB,MAAM,qBAAqB,CAAA;AAC5B,OAAO,EAAE,UAAU,EAAE,mBAAmB,EAAE,eAAe,EAAE,MAAM,kBAAkB,CAAA;AACnF,YAAY,EACV,YAAY,EACZ,sBAAsB,EACtB,YAAY,EACZ,eAAe,GAChB,MAAM,YAAY,CAAA;AACnB,OAAO,EAAE,eAAe,EAAE,MAAM,YAAY,CAAA;AAC5C,OAAO,EAAE,oBAAoB,EAAE,cAAc,EAAE,iBAAiB,EAAE,MAAM,kBAAkB,CAAA;AAC1F,OAAO,EACL,KAAK,wBAAwB,EAC7B,KAAK,uBAAuB,EAC5B,iBAAiB,GAClB,MAAM,sBAAsB,CAAA;AAC7B,YAAY,EAAE,kBAAkB,EAAE,MAAM,eAAe,CAAA;AACvD,OAAO,EAAE,uBAAuB,EAAE,MAAM,eAAe,CAAA;AACvD,YAAY,EACV,OAAO,EACP,cAAc,EACd,WAAW,EACX,cAAc,EACd,gBAAgB,EAChB,mBAAmB,EACnB,eAAe,EACf,gBAAgB,EAChB,gBAAgB,EAChB,iBAAiB,EACjB,eAAe,EACf,WAAW,GACZ,MAAM,SAAS,CAAA;AAChB,OAAO,EAAE,gBAAgB,EAAE,WAAW,EAAE,MAAM,SAAS,CAAA"}

package/dist/analyst/index.js

@@ -11,12 +11,13 @@ import {
   diffFindings,
   emitSkillUsageFindings,
   runSemanticConceptJudge
-} from "../chunk-XY4DDNEG.js";
+} from "../chunk-PSWWQXHF.js";
 import {
   behavioralAnalyst,
   buildDefaultAnalystRegistry,
   deriveEfficiencyFindings
-} from "../chunk-UMMZHCPB.js";
+} from "../chunk-AQ5WQAIV.js";
+import "../chunk-DJWX3GVS.js";
 import {
   ANALYST_SEVERITIES,
   AnalystRegistry,
@@ -41,16 +42,13 @@ import {
   renderPriorFindings,
   stripCodeFences,
   structureFindings
-} from "../chunk-D3V5B42D.js";
-import "../chunk-CVVHBFGN.js";
+} from "../chunk-S4SYLDFX.js";
+import "../chunk-CWNP4DV4.js";
 import {
   computeFindingId,
   makeFinding
 } from "../chunk-45EEMHTC.js";
-import {
-  analyzeTraces
-} from "../chunk-UHMJT4T7.js";
-import "../chunk-QAY5UIJO.js";
+import "../chunk-YBIGNSCZ.js";
 import "../chunk-PC4UYEBM.js";
 import "../chunk-3BFEG2F6.js";
 import "../chunk-PZ5AY32C.js";
@@ -69,63 +67,6 @@ function liftSeverity(s) {
       return "info";
   }
 }
-function createTraceAnalystAdapter(opts) {
-  const id = opts.id ?? "trace-analyst";
-  const area = opts.area ?? "agent-reasoning";
-  return {
-    id,
-    description: "Runs the agent-eval trace analyst over an OTLP trace store and lifts its bulleted findings.",
-    inputKind: "trace-store",
-    cost: { kind: "llm", models: opts.model ? [opts.model] : void 0 },
-    version: `trace-analyst-${ADAPTER_REV}`,
-    async analyze(store, ctx) {
-      const out = [];
-      for (const question of opts.questions) {
-        if (ctx.signal?.aborted) break;
-        const result = await analyzeTraces(
-          { question },
-          { source: store, ai: opts.ai, model: opts.model, ...opts.extra }
-        );
-        const subject = ctx.tags?.subject ?? question.slice(0, 60);
-        if (result.findings.length === 0) {
-          out.push(
-            makeFinding({
-              analyst_id: id,
-              area,
-              subject,
-              claim: result.answer.slice(0, 200),
-              rationale: result.answer,
-              severity: "info",
-              confidence: 0.5,
-              evidence_refs: [],
-              metadata: {
-                actor_prompt_version: result.actorPromptVersion,
-                turns: result.turnCount
-              }
-            })
-          );
-          continue;
-        }
-        result.findings.forEach((claim, i) => {
-          out.push(
-            makeFinding({
-              analyst_id: id,
-              area,
-              subject,
-              claim,
-              rationale: i === 0 ? result.answer : void 0,
-              severity: "medium",
-              confidence: 0.6,
-              evidence_refs: [],
-              metadata: { question, turns: result.turnCount, finding_index: i }
-            })
-          );
-        });
-      }
-      return out;
-    }
-  };
-}
 function createVerifierAdapter(opts) {
   const id = opts.id ?? "multi-layer-verifier";
   const area = opts.area ?? "verification";
@@ -385,7 +326,6 @@ export {
   createJudgeAdapter,
   createRunCriticAdapter,
   createSemanticConceptJudgeAdapter,
-  createTraceAnalystAdapter,
   createTraceAnalystKind,
   createVerifierAdapter,
   defaultIsMaterial,

package/dist/analyst/index.js.map

@@ -1 +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 type { AxAIService } from '@ax-llm/ax'\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} from '../semantic-concept-judge'\nimport { type AnalyzeTracesOptions, analyzeTraces } from '../trace-analyst/analyst'\nimport type { TraceAnalysisStore } from '../trace-analyst/store'\nimport type { JudgeFn, JudgeInput, JudgeScore, TCloud } from '../types'\nimport type { Analyst, AnalystFinding, AnalystSeverity } from './types'\nimport { makeFinding } from './types'\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. analyzeTraces → Analyst ─────────────────────────────────────\n\nexport interface TraceAnalystAdapterOpts {\n  id?: string\n  area?: string\n  /** The natural-language question(s) put to the analyst. One finding per question. */\n  questions: string[]\n  /** Caller-provided AxAI service — same one trace-analyst.ts expects. */\n  ai: AxAIService\n  model?: string\n  /** Forwarded to analyzeTraces. */\n  extra?: Omit<AnalyzeTracesOptions, 'source' | 'ai' | 'model'>\n}\n\n/**\n * @deprecated Prefer `createTraceAnalystKind` + one of the failure /\n * improvement kinds from `./kinds`. This adapter wraps the legacy\n * `analyzeTraces` flow whose output is `findings:string[]` — every\n * bullet gets flat-defaulted severity `medium` / confidence `0.6`,\n * which loses the per-finding grading kinds provide via Ax structured\n * output + Zod validation. Kept for one minor while consumers migrate.\n */\nexport function createTraceAnalystAdapter(\n  opts: TraceAnalystAdapterOpts,\n): Analyst<TraceAnalysisStore> {\n  const id = opts.id ?? 'trace-analyst'\n  const area = opts.area ?? 'agent-reasoning'\n  return {\n    id,\n    description:\n      'Runs the agent-eval trace analyst over an OTLP trace store and lifts its bulleted findings.',\n    inputKind: 'trace-store',\n    cost: { kind: 'llm', models: opts.model ? [opts.model] : undefined },\n    version: `trace-analyst-${ADAPTER_REV}`,\n    async analyze(store, ctx) {\n      const out: AnalystFinding[] = []\n      for (const question of opts.questions) {\n        if (ctx.signal?.aborted) break\n        const result = await analyzeTraces(\n          { question },\n          { source: store, ai: opts.ai, model: opts.model, ...opts.extra },\n        )\n        const subject = ctx.tags?.subject ?? question.slice(0, 60)\n        // The responder produces a list of bullet strings. Each becomes\n        // one finding; the prose answer is attached as rationale on the\n        // first (so renderers that show only top-N still get context).\n        if (result.findings.length === 0) {\n          out.push(\n            makeFinding({\n              analyst_id: id,\n              area,\n              subject,\n              claim: result.answer.slice(0, 200),\n              rationale: result.answer,\n              severity: 'info',\n              confidence: 0.5,\n              evidence_refs: [],\n              metadata: {\n                actor_prompt_version: result.actorPromptVersion,\n                turns: result.turnCount,\n              },\n            }),\n          )\n          continue\n        }\n        result.findings.forEach((claim, i) => {\n          out.push(\n            makeFinding({\n              analyst_id: id,\n              area,\n              subject,\n              claim,\n              rationale: i === 0 ? result.answer : undefined,\n              severity: 'medium',\n              confidence: 0.6,\n              evidence_refs: [],\n              metadata: { question, turns: result.turnCount, finding_index: i },\n            }),\n          )\n        })\n      }\n      return out\n    },\n  }\n}\n\n// ── 2. 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}: ${layer.reason ?? 'no reason given'}`,\n              severity:\n                layer.status === 'error' ? 'high' : layer.status === 'timeout' ? 'medium' : 'high',\n              confidence: 1,\n              evidence_refs: [],\n              metadata: {\n                layer_status: layer.status,\n                duration_ms: layer.durationMs,\n                score: layer.score,\n                diagnostics: layer.diagnostics,\n              },\n            }),\n          )\n        }\n      }\n      ctx.log?.('verifier complete', {\n        layers: report.layers.length,\n        blended: report.blendedScore,\n        all_pass: report.allPass,\n      })\n      return out\n    },\n  }\n}\n\nfunction liftLayerFinding(\n  analyst_id: string,\n  area: string,\n  layer: string,\n  f: LayerFinding,\n): AnalystFinding {\n  return makeFinding({\n    analyst_id,\n    area,\n    subject: f.layer ?? layer,\n    claim: f.message,\n    severity: liftSeverity(f.severity),\n    confidence: 0.85,\n    evidence_refs: f.evidence\n      ? [{ kind: 'artifact', uri: 'inline:evidence', excerpt: f.evidence }]\n      : [],\n    metadata: f.detail,\n  })\n}\n\n// ── 3. RunCritic → Analyst ──────────────────────────────────────────\n\nexport interface RunCriticAdapterOpts {\n  id?: string\n  area?: string\n  critic?: RunCritic\n  /** Optional threshold below which a dimension is reported as a finding. Default 0.5. */\n  threshold?: number\n}\n\nexport function createRunCriticAdapter(opts: RunCriticAdapterOpts = {}): Analyst<RunTrace> {\n  const id = opts.id ?? 'run-critic'\n  const area = opts.area ?? 'run-quality'\n  const critic = opts.critic ?? new RunCritic()\n  const threshold = opts.threshold ?? 0.5\n  return {\n    id,\n    description:\n      'Scores a single run across success / grounding / drift / tool-quality and surfaces below-threshold dimensions.',\n    inputKind: 'custom',\n    cost: { kind: 'deterministic' },\n    version: `run-critic-${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// ── 4. 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 ${score10.toFixed(1)}/10`,\n    rationale: s.reasoning,\n    severity,\n    confidence: 0.8,\n    evidence_refs: s.evidence\n      ? [{ kind: 'artifact', uri: 'inline:evidence', excerpt: s.evidence }]\n      : [],\n    // Provenance: this finding IS a judge verdict (an acceptance score), not an\n    // observation of behavior. The steer firewall (assertNoJudgeVerdict) rejects\n    // it from steering — even when it cites an artifact above — because letting a\n    // verdict steer the next attempt is the held-out judge leaking into the loop.\n    derived_from_judge: true,\n    metadata: { judge_name: s.judgeName, dimension: s.dimension, score_10: score10 },\n  })\n}\n\n// ── 5. SemanticConceptJudge → Analyst ──────────────────────────────\n\nexport interface SemanticConceptJudgeAdapterOpts {\n  id?: string\n  area?: string\n  options?: SemanticConceptJudgeOptions\n}\n\nexport function createSemanticConceptJudgeAdapter(\n  opts: SemanticConceptJudgeAdapterOpts = {},\n): Analyst<SemanticConceptJudgeInput> {\n  const id = opts.id ?? 'semantic-concept-judge'\n  const area = opts.area ?? 'concept-coverage'\n  return {\n    id,\n    description:\n      'Runs the semantic-concept judge and surfaces missing / weak concepts as findings.',\n    inputKind: 'custom',\n    cost: { kind: 'llm', models: opts.options?.model ? [opts.options.model] : undefined },\n    version: `${SEMANTIC_CONCEPT_JUDGE_VERSION}-adapter-${ADAPTER_REV}`,\n    async analyze(input) {\n      const result = await runSemanticConceptJudge(input, opts.options)\n      if (!result.available) {\n        return [\n          makeFinding({\n            analyst_id: id,\n            area,\n            claim: 'semantic-concept judge unavailable',\n            rationale: result.error,\n            severity: 'info',\n            confidence: 1,\n            evidence_refs: [],\n            metadata: { reason: result.error },\n          }),\n        ]\n      }\n      const out: AnalystFinding[] = []\n      for (const f of result.findings) {\n        // Only surface gaps: missing concepts or low scores. Concepts at\n        // 7+/10 with present=true are not findings — they're successes.\n        if (f.present && f.score >= 7) continue\n        out.push(\n          makeFinding({\n            analyst_id: id,\n            area,\n            subject: f.concept,\n            claim: f.present\n              ? `concept \"${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              cost_usd: result.costUsd ?? undefined,\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;AAwBO,SAAS,0BACd,MAC6B;AAC7B,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,OAAO,QAAQ,KAAK,QAAQ,CAAC,KAAK,KAAK,IAAI,OAAU;AAAA,IACnE,SAAS,iBAAiB,WAAW;AAAA,IACrC,MAAM,QAAQ,OAAO,KAAK;AACxB,YAAM,MAAwB,CAAC;AAC/B,iBAAW,YAAY,KAAK,WAAW;AACrC,YAAI,IAAI,QAAQ,QAAS;AACzB,cAAM,SAAS,MAAM;AAAA,UACnB,EAAE,SAAS;AAAA,UACX,EAAE,QAAQ,OAAO,IAAI,KAAK,IAAI,OAAO,KAAK,OAAO,GAAG,KAAK,MAAM;AAAA,QACjE;AACA,cAAM,UAAU,IAAI,MAAM,WAAW,SAAS,MAAM,GAAG,EAAE;AAIzD,YAAI,OAAO,SAAS,WAAW,GAAG;AAChC,cAAI;AAAA,YACF,YAAY;AAAA,cACV,YAAY;AAAA,cACZ;AAAA,cACA;AAAA,cACA,OAAO,OAAO,OAAO,MAAM,GAAG,GAAG;AAAA,cACjC,WAAW,OAAO;AAAA,cAClB,UAAU;AAAA,cACV,YAAY;AAAA,cACZ,eAAe,CAAC;AAAA,cAChB,UAAU;AAAA,gBACR,sBAAsB,OAAO;AAAA,gBAC7B,OAAO,OAAO;AAAA,cAChB;AAAA,YACF,CAAC;AAAA,UACH;AACA;AAAA,QACF;AACA,eAAO,SAAS,QAAQ,CAAC,OAAO,MAAM;AACpC,cAAI;AAAA,YACF,YAAY;AAAA,cACV,YAAY;AAAA,cACZ;AAAA,cACA;AAAA,cACA;AAAA,cACA,WAAW,MAAM,IAAI,OAAO,SAAS;AAAA,cACrC,UAAU;AAAA,cACV,YAAY;AAAA,cACZ,eAAe,CAAC;AAAA,cAChB,UAAU,EAAE,UAAU,OAAO,OAAO,WAAW,eAAe,EAAE;AAAA,YAClE,CAAC;AAAA,UACH;AAAA,QACF,CAAC;AAAA,MACH;AACA,aAAO;AAAA,IACT;AAAA,EACF;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;AAUO,SAAS,kCACd,OAAwC,CAAC,GACL;AACpC,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,OAAO,QAAQ,KAAK,SAAS,QAAQ,CAAC,KAAK,QAAQ,KAAK,IAAI,OAAU;AAAA,IACpF,SAAS,GAAG,8BAA8B,YAAY,WAAW;AAAA,IACjE,MAAM,QAAQ,OAAO;AACnB,YAAM,SAAS,MAAM,wBAAwB,OAAO,KAAK,OAAO;AAChE,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,cACZ,UAAU,OAAO,WAAW;AAAA,YAC9B;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;;;AClZA,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":[]}
+{"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 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} from '../semantic-concept-judge'\nimport type { JudgeFn, JudgeInput, JudgeScore, TCloud } from '../types'\nimport type { Analyst, AnalystFinding, AnalystSeverity } from './types'\nimport { makeFinding } from './types'\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 ${score10.toFixed(1)}/10`,\n    rationale: s.reasoning,\n    severity,\n    confidence: 0.8,\n    evidence_refs: s.evidence\n      ? [{ kind: 'artifact', uri: 'inline:evidence', excerpt: s.evidence }]\n      : [],\n    // Provenance: this finding IS a judge verdict (an acceptance score), not an\n    // observation of behavior. The steer firewall (assertNoJudgeVerdict) rejects\n    // it from steering — even when it cites an artifact above — because letting a\n    // verdict steer the next attempt is the held-out judge leaking into the loop.\n    derived_from_judge: true,\n    metadata: { judge_name: s.judgeName, dimension: s.dimension, score_10: score10 },\n  })\n}\n\n// ── 4. SemanticConceptJudge → Analyst ──────────────────────────────\n\nexport interface SemanticConceptJudgeAdapterOpts {\n  id?: string\n  area?: string\n  options?: SemanticConceptJudgeOptions\n}\n\nexport function createSemanticConceptJudgeAdapter(\n  opts: SemanticConceptJudgeAdapterOpts = {},\n): Analyst<SemanticConceptJudgeInput> {\n  const id = opts.id ?? 'semantic-concept-judge'\n  const area = opts.area ?? 'concept-coverage'\n  return {\n    id,\n    description:\n      'Runs the semantic-concept judge and surfaces missing / weak concepts as findings.',\n    inputKind: 'custom',\n    cost: { kind: 'llm', models: opts.options?.model ? [opts.options.model] : undefined },\n    version: `${SEMANTIC_CONCEPT_JUDGE_VERSION}-adapter-${ADAPTER_REV}`,\n    async analyze(input) {\n      const result = await runSemanticConceptJudge(input, opts.options)\n      if (!result.available) {\n        return [\n          makeFinding({\n            analyst_id: id,\n            area,\n            claim: 'semantic-concept judge unavailable',\n            rationale: result.error,\n            severity: 'info',\n            confidence: 1,\n            evidence_refs: [],\n            metadata: { reason: result.error },\n          }),\n        ]\n      }\n      const out: AnalystFinding[] = []\n      for (const f of result.findings) {\n        // Only surface gaps: missing concepts or low scores. Concepts at\n        // 7+/10 with present=true are not findings — they're successes.\n        if (f.present && f.score >= 7) continue\n        out.push(\n          makeFinding({\n            analyst_id: id,\n            area,\n            subject: f.concept,\n            claim: f.present\n              ? `concept \"${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              cost_usd: result.costUsd ?? undefined,\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":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAoCA,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;AAUO,SAAS,kCACd,OAAwC,CAAC,GACL;AACpC,QAAM,KAAK,KAAK,MAAM;AACtB,QAAM,OAAO,KAAK,QAAQ;AAC1B,SAAO;AAAA,IACL;AAAA,IACA,aACE;AAAA,IACF,WAAW;AAAA,IACX,MAAM,EAAE,MAAM,OAAO,QAAQ,KAAK,SAAS,QAAQ,CAAC,KAAK,QAAQ,KAAK,IAAI,OAAU;AAAA,IACpF,SAAS,GAAG,8BAA8B,YAAY,WAAW;AAAA,IACjE,MAAM,QAAQ,OAAO;AACnB,YAAM,SAAS,MAAM,wBAAwB,OAAO,KAAK,OAAO;AAChE,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,cACZ,UAAU,OAAO,WAAW;AAAA,YAC9B;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;;;ACzTA,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":[]}