@tangle-network/agent-eval 0.215.1 → 0.216.1

package/dist/{types-CGEmVYOk.d.ts → types-Cq8mSbq8.d.ts}

@@ -1,18 +1,20 @@
-import { S as PaidCallResult, T as RunPaidCallInput, a as CostChannel, c as CostLedgerHandle, f as CostLedgerSummary, p as CostProvenance } from "./cost-ledger-CO7UyEcQ.js";
-import { d as RunTokenUsage, s as RunSearchCoordinates } from "./run-record-D1Hw-Jml.js";
-import { _ as ProposalFinding } from "./types-DHb5TI36.js";
-import { M as LlmCallMetadata } from "./types-DBZ7keha.js";
+import { a as CostLedgerHandle, c as CostLedgerSummary, h as PaidCallResult, l as CostProvenance, r as CostChannel, v as RunPaidCallInput } from "./cost-ledger-Jk62OBfF.js";
+import { d as RunTokenUsage, s as RunSearchCoordinates } from "./run-record-C1E24G2A.js";
+import { b as ProposalFinding } from "./types-C1tnkEgM.js";
+import { j as LlmCallMetadata } from "./types-C02qtnG8.js";
 import { t as SeriesDistribution } from "./descriptive-B2iPaT9J.js";
 import { A as SearchLedgerHash, G as SearchSplit, I as SearchNodeStatus, L as SearchOpenedEvent, N as SearchNodeDecision, O as SearchLedgerEntry, P as SearchNodeRef, Q as SearchTaskOutcome, R as SearchOperationKind, U as SearchReservation, Z as SearchTask, b as SearchClosedEvent, d as SearchCellCancelledEvent, i as SearchArtifactRef, l as SearchCandidateSurface, o as SearchAudit, p as SearchCellStage, t as NodeEstimate, w as SearchEdgeRecordedEvent, z as SearchOperationRecordedEvent } from "./search-ledger-types-p2xKfiDJ.js";
 import { w as LedgerProjector } from "./index-DHKEAv3s.js";
-import { i as RolloutSearchLineage } from "./mint-DCGoBYhB.js";
-//#region src/campaign/search-state.d.ts
+import { i as RolloutSearchLineage } from "./mint-DNU8QPG4.js";
+//#region src/campaign/search-state-identity.d.ts
 /** Deterministic node id: identical content in one search is one node. */
 declare function searchNodeId(searchId: string, artifactDigest: string): string;
 /** Deterministic cell id: one node on one task in one split at one repeat. */
 declare function searchCellId(searchId: string, nodeId: string, taskId: string, split: SearchSplit, rep: number): string;
 /** Digest a split's task list in taskId order. */
 declare function searchTaskSetDigest(tasks: readonly SearchTask[]): SearchLedgerHash;
+//#endregion
+//#region src/campaign/search-state-types.d.ts
 interface SearchDecisionRecord {
   sequence: number;
   decision: SearchNodeDecision;
@@ -42,13 +44,6 @@ interface SearchScoredCell {
   attempt: number;
   score: number;
 }
-/**
- * Per-unit means of scored cells, in unitId order. Each unit sums its cells in
- * cellId order, so the result depends on the set of cells alone, never on the
- * order they settled in: a verifier that reads the same cells from a database
- * gets the same bits.
- */
-declare function searchUnitScores(cells: readonly SearchScoredCell[]): SearchUnitScore[];
 interface SearchNode {
   nodeId: string;
   /** Registration order; the root is 0. */
@@ -125,6 +120,74 @@ interface SearchCompletion {
   complete: boolean;
   reasons: string[];
 }
+//#endregion
+//#region src/campaign/search-state-view.d.ts
+/**
+ * Per-unit means of scored cells, in unitId order. Each unit sums its cells in
+ * cellId order, so the result depends on the set of cells alone, never on the
+ * order they settled in: a verifier that reads the same cells from a database
+ * gets the same bits.
+ */
+declare function searchUnitScores(cells: readonly SearchScoredCell[]): SearchUnitScore[];
+/**
+ * A read of `SearchState` at one ledger position. Header, head, audit and
+ * claim are values. Node, edge, cell and score reads go to the live indexes and
+ * throw once the state has applied a later entry.
+ */
+declare class SearchStateView {
+  readonly searchId: string;
+  readonly header: SearchOpenedEvent | null;
+  readonly closed: SearchClosedEvent | null;
+  readonly head: {
+    sequence: number;
+    entryHash: SearchLedgerHash;
+  } | null;
+  /** `occurredAt` of the head entry; a writer stamps later events at or after it. */
+  readonly lastOccurredAt: string | null;
+  readonly audit: SearchAudit;
+  readonly completion: SearchCompletion;
+  /**
+   * The admission rule's room: `headroomUsd` is the largest reservation a new
+   * non-claim cell or operation may hold (committed spend, open holds and the
+   * unspent claim reserve all count against the cap), or null when the search
+   * declares no cap. `claimReserveUsd` is what claim cells may still draw.
+   */
+  readonly budget: {
+    headroomUsd: number | null;
+    claimReserveUsd: number;
+  };
+  private readonly state;
+  private readonly version;
+  constructor(state: SearchState, version: number);
+  /** The node the search started from: the first registered. */
+  get rootNodeId(): string | null;
+  node(nodeId: string): SearchNode | undefined;
+  hasNode(nodeId: string): boolean;
+  /** Node ids in registration order. */
+  nodeIds(): string[];
+  /** Nodes in registration order. */
+  nodes(): SearchNode[];
+  nodeIdForDigest(artifactDigest: string): string | undefined;
+  edge(edgeId: string): SearchEdgeRecordedEvent | undefined;
+  /** Edges in ledger order. */
+  edges(): SearchEdgeRecordedEvent[];
+  cell(cellId: string): SearchCell | undefined;
+  /** Cells in allocation order, optionally of one node. */
+  cells(filter?: {
+    nodeId?: string;
+  }): SearchCell[];
+  operation(operationId: string): SearchOperation | undefined;
+  /** A node's scored cells on one split, in allocation order. Unscored cells
+   * (errored, cancelled or in flight) are absent, never zero. */
+  scoredCells(nodeId: string, split: SearchSplit): SearchScoredCell[];
+  /** Per-unit score means of a node on one split, in unitId order. */
+  unitScores(nodeId: string, split: SearchSplit): SearchUnitScore[];
+  /** The lineage a RunRecord mint needs for one cell attempt. */
+  lineage(coordinates: RunSearchCoordinates): RolloutSearchLineage;
+  private live;
+}
+//#endregion
+//#region src/campaign/search-state-projection.d.ts
 /**
  * The pure, incremental projection of one search ledger. `apply` refuses any
  * entry that breaks an invariant; the journal then discards this instance and
@@ -232,63 +295,6 @@ declare class SearchState implements LedgerProjector<SearchLedgerEntry, SearchSt
   /** @internal */
   readScoredCells(nodeId: string, split: SearchSplit): SearchScoredCell[];
 }
-/**
- * A read of `SearchState` at one ledger position. Header, head, audit and
- * claim are values. Node, edge, cell and score reads go to the live indexes and
- * throw once the state has applied a later entry.
- */
-declare class SearchStateView {
-  readonly searchId: string;
-  readonly header: SearchOpenedEvent | null;
-  readonly closed: SearchClosedEvent | null;
-  readonly head: {
-    sequence: number;
-    entryHash: SearchLedgerHash;
-  } | null;
-  /** `occurredAt` of the head entry; a writer stamps later events at or after it. */
-  readonly lastOccurredAt: string | null;
-  readonly audit: SearchAudit;
-  readonly completion: SearchCompletion;
-  /**
-   * The admission rule's room: `headroomUsd` is the largest reservation a new
-   * non-claim cell or operation may hold (committed spend, open holds and the
-   * unspent claim reserve all count against the cap), or null when the search
-   * declares no cap. `claimReserveUsd` is what claim cells may still draw.
-   */
-  readonly budget: {
-    headroomUsd: number | null;
-    claimReserveUsd: number;
-  };
-  private readonly state;
-  private readonly version;
-  constructor(state: SearchState, version: number);
-  /** The node the search started from: the first registered. */
-  get rootNodeId(): string | null;
-  node(nodeId: string): SearchNode | undefined;
-  hasNode(nodeId: string): boolean;
-  /** Node ids in registration order. */
-  nodeIds(): string[];
-  /** Nodes in registration order. */
-  nodes(): SearchNode[];
-  nodeIdForDigest(artifactDigest: string): string | undefined;
-  edge(edgeId: string): SearchEdgeRecordedEvent | undefined;
-  /** Edges in ledger order. */
-  edges(): SearchEdgeRecordedEvent[];
-  cell(cellId: string): SearchCell | undefined;
-  /** Cells in allocation order, optionally of one node. */
-  cells(filter?: {
-    nodeId?: string;
-  }): SearchCell[];
-  operation(operationId: string): SearchOperation | undefined;
-  /** A node's scored cells on one split, in allocation order. Unscored cells
-   * (errored, cancelled or in flight) are absent, never zero. */
-  scoredCells(nodeId: string, split: SearchSplit): SearchScoredCell[];
-  /** Per-unit score means of a node on one split, in unitId order. */
-  unitScores(nodeId: string, split: SearchSplit): SearchUnitScore[];
-  /** The lineage a RunRecord mint needs for one cell attempt. */
-  lineage(coordinates: RunSearchCoordinates): RolloutSearchLineage;
-  private live;
-}
 //#endregion
 //#region src/campaign/search-summary.d.ts
 /** The proposer's own read of a search: its node's cells on the train split,
@@ -1014,5 +1020,5 @@ interface CampaignResult<TArtifact = unknown, TScenario extends Scenario = Scena
   scenarios: Array<CampaignScenarioIdentity & Pick<TScenario, 'id' | 'kind'>>;
 }
 //#endregion
-export { SearchNode as $, LabeledScenarioWrite as A, ScoredSurfaceOutcome as B, JudgeDimension as C, LabeledScenarioSampleArgs as D, LabeledScenarioRecord as E, ProposeContext as F, labelTrustRank as G, SurfaceProposer as H, ProposedCandidate as I, renderSearchSummary as J, SearchProposerView as K, RedactionStatus as L, OptimizerConfig as M, ParetoParent as N, LabeledScenarioSource as O, ProposalTrackContext as P, SearchDecisionRecord as Q, Scenario as R, JudgeConfig as S, LabelTrust as T, TraceSpan as U, SessionScript as V, isProposedCandidate as W, SearchCell as X, searchProposerView as Y, SearchCompletion as Z, GateDecision as _, CampaignResult as a, SearchUnitScore as at, GenerationRecord as b, CampaignTraceWriter as c, searchTaskSetDigest as ct, DispatchContext as d, SearchOperation as et, DispatchFn as f, GateContribution as g, GateContext as h, CampaignCostMeter as i, SearchStateView as it, MutableSurface as j, LabeledScenarioStore as k, CodeSurface as l, searchUnitScores as lt, GateCheckStatus as m, CampaignArtifactWriter as n, SearchSpend as nt, CampaignScenarioIdentity as o, searchCellId as ot, Gate as p, SearchSummaryOptions as q, CampaignCellResult as r, SearchState as rt, CampaignTokenUsage as s, searchNodeId as st, CampaignAggregates as t, SearchScoredCell as tt, ComponentSurface as u, GateResult as v, JudgeScore as w, JudgeAggregate as x, GenerationCandidate as y, ScenarioAggregate as z };
-//# sourceMappingURL=types-CGEmVYOk.d.ts.map
+export { SearchCell as $, LabeledScenarioWrite as A, ScoredSurfaceOutcome as B, JudgeDimension as C, LabeledScenarioSampleArgs as D, LabeledScenarioRecord as E, ProposeContext as F, labelTrustRank as G, SurfaceProposer as H, ProposedCandidate as I, renderSearchSummary as J, SearchProposerView as K, RedactionStatus as L, OptimizerConfig as M, ParetoParent as N, LabeledScenarioSource as O, ProposalTrackContext as P, searchUnitScores as Q, Scenario as R, JudgeConfig as S, LabelTrust as T, TraceSpan as U, SessionScript as V, isProposedCandidate as W, SearchState as X, searchProposerView as Y, SearchStateView as Z, GateDecision as _, CampaignResult as a, SearchSpend as at, GenerationRecord as b, CampaignTraceWriter as c, searchNodeId as ct, DispatchContext as d, SearchCompletion as et, DispatchFn as f, GateContribution as g, GateContext as h, CampaignCostMeter as i, SearchScoredCell as it, MutableSurface as j, LabeledScenarioStore as k, CodeSurface as l, searchTaskSetDigest as lt, GateCheckStatus as m, CampaignArtifactWriter as n, SearchNode as nt, CampaignScenarioIdentity as o, SearchUnitScore as ot, Gate as p, SearchSummaryOptions as q, CampaignCellResult as r, SearchOperation as rt, CampaignTokenUsage as s, searchCellId as st, CampaignAggregates as t, SearchDecisionRecord as tt, ComponentSurface as u, GateResult as v, JudgeScore as w, JudgeAggregate as x, GenerationCandidate as y, ScenarioAggregate as z };
+//# sourceMappingURL=types-Cq8mSbq8.d.ts.map

package/dist/types-Cq8mSbq8.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"types-Cq8mSbq8.d.ts","names":[],"sources":["../src/campaign/search-state-identity.ts","../src/campaign/search-state-types.ts","../src/campaign/search-state-view.ts","../src/campaign/search-state-projection.ts","../src/campaign/search-summary.ts","../src/campaign/types.ts"],"mappings":";;;;;;;;;;iBAOgB,aAAa,kBAAkB;;iBAK/B,aACd,kBACA,gBACA,gBACA,OAAO,aACP;;iBAMc,oBAAoB,gBAAgB,eAAe;;;UCHlD;EACf;EACA,UAAU;EACV,OAAO;EACP;EACA;;UAGe;EACf;EACA;EACA;EACA;;;;UAKe;EACf;EACA;EACA;EACA;;;UAIe;EACf;EACA;;EAEA;EACA;;UAGe;EACf;;EAEA;EACA,gBAAgB;EAChB,UAAU;EACV,mBAAmB;;;EAGnB;;EAEA;;;;EAIA,kBAAkB;EAClB;EACA;EACA,QAAQ;EACR;;EAEA;EACA,oBAAoB;EACpB;EACA,OAAO;EACP;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA,OAAO;EACP;EACA,OAAO;EACP;EACA,aAAa;EACb;EACA;EACA,SAAS;;EAET;;EAEA;EACA,WAAW;;EAEX;;EAEA;EACA;EACA;EACA;;UAGe;EACf;EACA,eAAe;EACf,aAAa;EACb;;EAEA,SAAS;;;EAGT,WAAW;;;EAGX,oBAAoB;;EAEpB;;EAEA;;UAGe;EACf;EACA;;;;;;;;;;iBCnGc,iBAAiB,gBAAgB,qBAAqB;;;;;;cAyBzD;WACF;WACA,QAAQ;WACR,QAAQ;WACR;IAAQ;IAAkB,WAAW;;;WAErC;WACA,OAAO;WACP,YAAY;;;;;;;WAOZ;IAAU;IAA4B;;mBAC9B;mBACA;EAEjB,YAAY,OAAO,aAAa;;MAgB5B;EAIJ,KAAK,iBAAiB;EAItB,QAAQ;;EAKR;;EAKA,SAAS;EAKT,gBAAgB;EAIhB,KAAK,iBAAiB;;EAKtB,SAAS;EAKT,KAAK,iBAAiB;;EAKtB,MAAM;IAAU;MAAyB;EAKzC,UAAU,sBAAsB;;;EAMhC,YAAY,gBAAgB,OAAO,cAAc;;EAKjD,WAAW,gBAAgB,OAAO,cAAc;;EAKhD,QAAQ,aAAa,uBAAuB;UAmCpC;;;;;;;;;cCzIG,uBAAuB,gBAAgB,mBAAmB;WAC5D;UACD;UACA;UACA;UACA;UACA;UACA;mBACS;mBACA;mBACA;mBACA;mBACA;mBACA;mBACA;mBACA;UACT;UACA;;mBAES;;UAET;mBACS;EAEjB,YAAY;EAgCZ,MAAM,OAAO,mBAAmB;EA6BhC,YAAY;;EAKZ,KAAK;UAIG;UAgCA;UA6CA;UAsBA;UA8BA;UA8CA;UAsHA;UAmEA;UA4EA;UAoBA;UAmDA;UAaA;;;UA2EA;;EASR;;;UAmBQ;;;UAkBA;;;;;;;;;;;UAoBA;;;EAsBR;IAAgB;IAA4B;;;EAa5C,cAAc;;EAKd,cAAc;;EAKd;IAAc;IAAkB,WAAW;;;EAK3C;;EAKA,aAAa;;EAcb,SAAS,iBAAiB;;EAM1B;;EAKA,YAAY;;EAKZ,oBAAoB;;EAKpB,SAAS,iBAAiB;;EAK1B;;EAKA,SAAS,iBAAiB;;EAM1B,YAAY;;EAMZ,cAAc,sBAAsB;;EAMpC,gBAAgB,gBAAgB,OAAO,cAAc;;;;;;UCj4BtC;WACN;WACA;;;EAGT,YAAY,0BAA0B;;EAEtC,WAAW,0BAA0B;;;iBAIvB,mBAAmB,OAAO,kBAAkB;UAW3C;;;;;;;;EAQf,OAAO;;EAEP;;;;;;;;;;;iBAYc,oBAAoB,OAAO,iBAAiB,SAAS;;;;;UCtCpD;EACf;EACA;EACA;;;;;EAKA;;;UAIe,iCAAiC,KAAK;EACrD;;;;;UAMe;EACf;;EAEA;EACA;EACA;EACA;EACA,QAAQ;EACR,OAAO;EACP,WAAW;EACX,MAAM;;EAEN;;EAEA;;;;;;;;EAQA;;;;KAKU,WAAW,kBAAkB,UAAU,cACjD,UAAU,WACV,KAAK,oBACF,QAAQ;;;;UAOI,cAAc,WAAW;EACxC;EACA;EACA;;EAEA;;;EAGA,sBAAsB,UAAU,WAAW,sBAAsB,UAAU,cAAc;;UAK1E;;EAEf;;EAEA;;;;;;;;;UAUe,YAAY,WAAW,kBAAkB,WAAW;EACnE;EACA,YAAY;;;;EAIZ;;;EAGA,MAAM;IACJ,UAAU;IACV,UAAU;IACV,QAAQ;;IAER,aAAa;IACb;IACA,WAAW;MACT,aAAa,QAAQ;EACzB,aAAa,UAAU;;;;;;;;;;;UAYR;EACf,YAAY;EACZ;EACA;;EAEA,UAAU;;;;;;;EAOV;;;;EAIA,eAAe,eAAe;IAAQ;IAAe;;;;;EAIrD;;;EAGA;;EAEA;;EAEA,WAAW,eAAe;;;;;;;UAUX;WACN;;;WAGA;;WAEA;;WAEA;;WAEA;;WAEA;;WAEA;;;WAGA;aACE;aACA;aACA;;;WAGF;;;UAIM;WACN;WACA,YAAY,SAAS;;;;;;;;;;;KAYpB,0BAA0B,mBAAmB;;;;;;;UAQxC;EACf,SAAS;;EAET;;;;EAIA;;;;;;EAMA,cAAc,SAAS;;;;iBAKT,oBACd,OAAO,iBAAiB,oBACvB,SAAS;;;;;;;;;UAkBK;EACf,SAAS;EACT;;;EAGA,YAAY;;;EAGZ;;EAEA;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;EACA;;;;;UAMe;;;EAGf;;EAEA;EACA;EACA;EACA,YAAY;;;;;EAKZ,WAAW;IAAQ;IAAoB;IAAmB;IAAgB;;EAC1E;IACE;IACA;;;;;UAMa,eAAe,YAAY;;;;WAIjC,gBAAgB;;;;;WAKhB;;;;WAIA,UAAU;IAAgB;IAAgB,UAAU;;;;WAGpD,QAAQ;;;;WAIR;;;WAGA,SAAS,cAAc;WACvB,UAAU,cAAc;;WAExB;WACA;WACA,QAAQ;;WAER,QAAQ;;;WAGR,kBAAkB;;;WAGlB,mBAAmB;;;;WAInB,gBAAgB;;;;WAIhB;;;;;;;WAOA,gBAAgB,cAAc;;WAE9B,aAAa;WACb;;;;;;;;;;;;;;UAeM,gBAAgB,YAAY;EAC3C;;;;;EAKA,QAAQ,KAAK,eAAe,aAAa,QAAQ,MAAM,iBAAiB;;;EAGxE,QAAQ;IAAQ,SAAS,cAAc;;IAAwB;IAAe;;;UAG/D;EACf;EACA;EACA,mBAAmB,qBAAqB;;UAGzB,wBAAwB;EACvC,UAAU;;;KAMA;;KAGA;UAEK;EACf;EACA,QAAQ;EACR;;UAGe,YAAY,WAAW,kBAAkB;EACxD,oBAAoB,YAAY;EAChC,oBAAoB,YAAY;;EAEhC,aAAa,YAAY,eAAe;;;;;EAKxC,sBAAsB,YAAY,eAAe;;;;;;;EAOjD,yBAAyB,YAAY,eAAe;;;EAGpD,uBAAuB,YAAY;EACnC,WAAW;EACX;IAAQ;IAAmB;;;EAE3B,aAAa;EACb;EACA,QAAQ;;UAGO;EACf,UAAU;EACV;EACA,mBAAmB;EACnB;;;UAIe,KAAK,qBAAqB,kBAAkB,WAAW;EACtE;EACA,OAAO,KAAK,YAAY,WAAW,aAAa,QAAQ;;;;UAOzC;EACf,KAAK,cAAc,aAAa,0BAA0B;EAC1D,SAAS;;UAGM;EACf,IAAI,aAAa;EACjB,aAAa,aAAa;;;;;UAMX;EACf,MAAM,cAAc,kBAAkB,aAAa;EACnD,UAAU,cAAc,iBAAiB;;;;;;KAO/B,qBAAqB;;;;;UAMhB;;EAEf,YAAY,GACV,OAAO,KAAK,iBAAiB;IAC3B,UAAU;MAEX,QAAQ,eAAe;;;;;KAQhB;KAOA;;;;;;;;;;;;;;;KAgBA;;iBASI,eAAe,OAAO;;;;UAOrB,qBAAqB,kBAAkB,WAAW,UAAU;EAC3E,UAAU;EACV,UAAU;EACV,aAAa,eAAe;EAC5B,QAAQ;EACR;EACA;EACA,iBAAiB;;;;;EAKjB,aAAa;;EAEb;;UAGe,sBAAsB,kBAAkB,WAAW,UAAU,6BACpE,qBAAqB,WAAW;;EAExC;;;EAGA;;UAGe;EACf;;EAEA;;;;EAIA;EACA;IACE;IACA,SAAS,wBAAwB;IACjC;IACA;;;;;IAKA,WAAW;;;UAIE;EACf,QAAQ,OAAO,uBAAuB;EACtC,OAAO,MAAM,4BAA4B,QAAQ;EACjD,QAAQ;IACN;IACA;IACA,UAAU;;;IAGV,SAAS,OAAO;;;UAMH,mBAAmB;;;EAGlC;EACA;;EAEA;IAAY;IAAsB;;EAClC;EACA;EACA;EACA,UAAU;EACV,aAAa,eAAe;;;EAG5B;;EAEA,gBAAgB;;EAEhB;;;EAGA,YAAY;;;EAGZ;;;EAGA;EACA;EACA;EACA;;;;EAIA;;EAEA;;EAEA;EACA;;UAGe;EACf;EACA;EACA;EACA;;;;;;EAMA,cAAc;;UAGC;EACf;EACA;EACA;;;EAGA,cAAc;;UAGC;EACf;EACA,YAAY;EACZ;;;;;;UAOe;EACf;;;;EAIA;;;;EAIA;;;;EAIA;IACE;IACA;IACA,iBAAiB;MAAQ;MAAgB;;;;;EAI3C,YAAY;;;;;;;;;;EAUZ,WAAW;IAAQ;IAAoB;IAAmB;IAAgB;;;;EAG1E;;;;EAIA;;;;EAIA,cAAc,SAAS;;UAGR;EACf,SAAS,eAAe;EACxB,YAAY,eAAe;;EAE3B,MAAM;;EAEN;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;UAGe,eAAe,qBAAqB,kBAAkB,WAAW;;EAEhF;;EAEA;EACA;;EAEA;EACA;EACA;EACA;EACA,OAAO,MAAM,mBAAmB;EAChC,YAAY;EACZ;IACE,aAAa;IACb;;EAEF,OAAO;EACP;EACA;EACA,iBAAiB;;;EAGjB,WAAW,MAAM,2BAA2B,KAAK"}

package/dist/{types-CW5b90JH.js → types-IXD2HJzu.js}

@@ -16,6 +16,8 @@ import { createHash } from "node:crypto";
 * a RunRecord, a JudgeInput, or `custom`), and the registry routes
 * the matching input from `AnalystRunInputs`.
 */
+/** Every finding schema this package reads. Writers emit 1.1.0 only when a 1.1.0 field is set. */
+const ANALYST_FINDING_SCHEMA_VERSIONS = ["1.0.0", "1.1.0"];
 /** Finding severity. `AnalystSeverity` derives from this array, so the type
 *  and the schema that validates a finding cannot name different levels. */
 const ANALYST_SEVERITIES = [
@@ -45,6 +47,17 @@ function computeFindingId(input) {
 function normalizeClaim(c) {
 	return c.toLowerCase().replace(/\s+/g, " ").replace(/[.!?;:,]+$/g, "").trim();
 }
+/** Fields that exist only in schema 1.1.0. */
+const FINDING_SCOPE_FIELDS = [
+	"layer",
+	"scope",
+	"method",
+	"measure",
+	"confounders",
+	"valid_for",
+	"claim_id",
+	"evidence_units"
+];
 /**
 * Convenience factory: produce a fully-formed AnalystFinding with the
 * id computed automatically. Analyst code stays terse.
@@ -52,7 +65,7 @@ function normalizeClaim(c) {
 function makeFinding(init) {
 	const { id_basis, produced_at, ...rest } = init;
 	return {
-		schema_version: "1.0.0",
+		schema_version: FINDING_SCOPE_FIELDS.some((field) => rest[field] !== void 0) ? "1.1.0" : "1.0.0",
 		finding_id: computeFindingId({
 			analyst_id: rest.analyst_id,
 			area: rest.area,
@@ -73,6 +86,6 @@ function makeProposalFinding(init) {
 	};
 }
 //#endregion
-export { makeProposalFinding as i, computeFindingId as n, makeFinding as r, ANALYST_SEVERITIES as t };
+export { makeFinding as a, computeFindingId as i, ANALYST_SEVERITIES as n, makeProposalFinding as o, FINDING_SCOPE_FIELDS as r, ANALYST_FINDING_SCHEMA_VERSIONS as t };
 
-//# sourceMappingURL=types-CW5b90JH.js.map
+//# sourceMappingURL=types-IXD2HJzu.js.map

package/dist/types-IXD2HJzu.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"types-IXD2HJzu.js","names":[],"sources":["../src/analyst/types.ts"],"sourcesContent":["/**\n * Analyst contract — the missing orchestration layer over agent-eval's\n * existing analyzers (analyzeTraces, MultiLayerVerifier,\n * SemanticConceptJudge, JudgeFn, ...).\n *\n * Each existing primitive returns its own output shape. The Analyst\n * contract is the single envelope every primitive lifts into, so a\n * registry can run N analysts against a run and a single renderer can\n * compose findings without knowing which analyzer produced them.\n *\n * The contract is intentionally domain-agnostic: nothing here knows\n * about code, voice, RAG, or any particular agent stack. Analysts\n * declare what INPUT KIND they need (a trace store, an artifact dir,\n * a RunRecord, a JudgeInput, or `custom`), and the registry routes\n * the matching input from `AnalystRunInputs`.\n */\n\nimport { createHash } from 'node:crypto'\nimport type { CostLedgerHandle, CostProvenance } from '../cost-ledger'\nimport type { RunRecord, RunTokenUsage } from '../run-record'\nimport type { TraceAnalysisStore } from '../trace-analyst/store'\nimport type { JudgeInput } from '../types'\nimport type { ChatClient } from './chat-client'\nimport type {\n  FindingEvidenceUnit,\n  FindingLayer,\n  FindingMeasure,\n  FindingMethod,\n  FindingScope,\n} from './finding-scope'\n\n/**\n * Unified envelope every analyst emits. Schema-versioned so renderers\n * and time-series diffs survive future field additions.\n *\n * 1.1.0 adds what a finding is about and what it was measured on: `layer`,\n * `scope`, `method`, `measure`, `confounders`, `valid_for`, `claim_id` and\n * `evidence_units`. They are all optional, and a 1.0.0 finding carries none of\n * them; `readFindingScope` reads such a finding as layer and n unknown.\n * `parseAnalystFinding` (finding-scope.ts) is the one validator for both.\n */\nexport interface AnalystFinding {\n  schema_version: AnalystFindingSchemaVersion\n  /**\n   * Stable hash over identity-defining fields (analyst_id + canonical\n   * claim + area + optional subject). Two findings from two runs that\n   * \"are the same finding\" share this id — that's what `diffFindings`\n   * uses to compute appeared/disappeared sets across runs.\n   */\n  finding_id: string\n  analyst_id: string\n  produced_at: string\n  severity: AnalystSeverity\n  /**\n   * Coarse classification. Renderers group by this. Free-form so\n   * domain-specific analysts can introduce categories without a\n   * schema change ('agent-reasoning', 'verification', 'cost',\n   * 'tool-use', 'safety', 'latency', 'data-quality', ...).\n   */\n  area: string\n  claim: string\n  rationale?: string\n  evidence_refs: EvidenceRef[]\n  recommended_action?: string\n  validation_plan?: string\n  /** 0..1 — the analyst's own confidence. Not calibrated across analysts. */\n  confidence: number\n  /**\n   * Optional subject the finding is about — leaf id, agent id, request\n   * id. Included in finding_id when present so per-subject findings\n   * diff cleanly across runs.\n   */\n  subject?: string\n  /** True when this finding was lifted from a judge result rather than observed\n   *  directly in a trace or artifact. Descriptive only: proposal access is\n   *  controlled by `ProposalFinding.proposal_origin`. */\n  derived_from_judge?: boolean\n  /** Analyst-private extras; renderers ignore unless they know the analyst. */\n  metadata?: Record<string, unknown>\n  /** 1.1.0: the level the claim is about, from one event up to the whole lab. */\n  layer?: FindingLayer\n  /** 1.1.0: the population the claim was measured on. */\n  scope?: FindingScope\n  /** 1.1.0: how the claim was produced, and the judge's calibration when judged. */\n  method?: FindingMethod\n  /** 1.1.0: the number behind the claim, when it has one. */\n  measure?: FindingMeasure\n  /** 1.1.0: known differences between the compared units besides the treatment. */\n  confounders?: string[]\n  /** 1.1.0: the conditions the claim was measured under, in words. */\n  valid_for?: string\n  /**\n   * 1.1.0: stable key for the claim across subjects and layers. `promote`\n   * groups children by it; a finding without one never lifts.\n   */\n  claim_id?: string\n  /**\n   * 1.1.0: the measured units with their lineage. When present, `scope.n` and\n   * `scope.population` must match them and no unit may sit above `layer`.\n   */\n  evidence_units?: FindingEvidenceUnit[]\n}\n\n/** Every finding schema this package reads. Writers emit 1.1.0 only when a 1.1.0 field is set. */\nexport const ANALYST_FINDING_SCHEMA_VERSIONS = ['1.0.0', '1.1.0'] as const\n\nexport type AnalystFindingSchemaVersion = (typeof ANALYST_FINDING_SCHEMA_VERSIONS)[number]\n\n/** Finding severity. `AnalystSeverity` derives from this array, so the type\n *  and the schema that validates a finding cannot name different levels. */\nexport const ANALYST_SEVERITIES = ['critical', 'high', 'medium', 'low', 'info'] as const\n\nexport type AnalystSeverity = (typeof ANALYST_SEVERITIES)[number]\n\n/** Data sources that candidate generation may intentionally learn from. */\nexport type ProposalFindingOrigin = 'search' | 'production'\n\n/** A finding explicitly admitted as candidate-generation input. */\nexport type ProposalFinding = AnalystFinding & {\n  readonly proposal_origin: ProposalFindingOrigin\n}\n\nexport interface EvidenceRef {\n  /**\n   * Where the evidence lives. `span` and `event` refer to OTLP trace\n   * elements; `artifact` to a file inside the run's artifact tree;\n   * `finding` to another AnalystFinding (cross-analyst chaining);\n   * `metric` to a named scalar reading the renderer knows how to read.\n   */\n  kind: 'span' | 'event' | 'artifact' | 'finding' | 'metric'\n  uri: string\n  excerpt?: string\n}\n\n// ── Analyst contract ─────────────────────────────────────────────────\n\n/**\n * The discriminator the registry uses to pass the right input.\n * `custom` is the escape hatch — analysts that need something else\n * (e.g. an embedding cache, a partner SDK handle) read it from\n * `AnalystRunInputs.custom[<analyst id>]`.\n */\nexport type AnalystInputKind =\n  | 'trace-store'\n  | 'artifact-dir'\n  | 'run-record'\n  | 'judge-input'\n  | 'custom'\n\nexport interface AnalystCost {\n  /** `deterministic` analysts MUST NOT call the LLM. */\n  kind: 'deterministic' | 'llm'\n  /** Optional declared upper bound; the registry can enforce a budget. */\n  est_usd_per_run?: number\n  /** Models the analyst expects to use (informational). */\n  models?: string[]\n  /** Maximum post-cancellation wait for provider usage. Model analysts default to 5 seconds. */\n  settlement_timeout_ms?: number\n}\n\nexport interface AnalystRequirements {\n  /** Min number of shots / samples the analyst needs to produce signal. */\n  min_shots?: number\n  /** Capabilities the runtime must supply (e.g. ['network', 'gpu']). */\n  capabilities?: string[]\n}\n\n/**\n * What's passed to every analyst call. The registry resolves which\n * field the analyst's `inputKind` selects and asserts it's present.\n */\nexport interface AnalystRunInputs {\n  traceStore?: TraceAnalysisStore\n  artifactDir?: string\n  runRecord?: RunRecord\n  judgeInput?: JudgeInput\n  /** Keyed by analyst id; populated by callers that registered custom analysts. */\n  custom?: Record<string, unknown>\n}\n\nexport interface AnalystContext {\n  runId: string\n  /** Stable correlation id so logs from a single registry.run() share a tag. */\n  correlationId: string\n  /** Enforced wall-clock deadline (epoch ms). */\n  deadlineMs?: number\n  /** Per-analyst USD budget. Analysts MAY check before issuing LLM calls. */\n  budgetUsd?: number\n  /** Shared paid-call account when the analyst runs inside a larger campaign. */\n  costLedger?: CostLedgerHandle\n  /** Attribution phase used when writing to the shared paid-call account. */\n  costPhase?: string\n  /**\n   * Shared chat client. Analysts that call an LLM go through this so\n   * the operator picks transport (sandbox-sdk | router | cli-bridge |\n   * direct-provider | mock) at the registry boundary without touching\n   * analyst code.\n   */\n  chat?: ChatClient\n  /**\n   * Findings from a prior run the operator wants the analyst to see as\n   * retrieval context. Kinds that take advantage of cross-run memory\n   * (failure-mode \"I saw this cluster last run\", knowledge-gap \"the wiki\n   * page I asked for is still missing\") render these into the actor's\n   * working set. Filtering is the operator's job: pass the slice that\n   * matches the analyst's id, or pass everything and let the kind\n   * filter. Empty / absent means no cross-run context.\n   */\n  priorFindings?: ReadonlyArray<AnalystFinding>\n  /**\n   * Findings emitted by analysts that completed earlier in this registry run.\n   * This is separate from `priorFindings`: upstream findings are dependency\n   * context for the current pass, while prior findings are cross-run memory.\n   * The registry populates this only when `RegistryRunOpts.chainFindings` is on.\n   */\n  upstreamFindings?: ReadonlyArray<AnalystFinding>\n  /**\n   * Report metered work independently of findings. This keeps an empty finding\n   * set from erasing token/cost telemetry. Multiple receipts are accumulated.\n   */\n  recordUsage?: (receipt: AnalystUsageReceipt) => void\n  /** Free-form runtime tags (env, host, op). Findings can echo these into metadata. */\n  tags?: Record<string, string>\n  /** Logger callback — analysts SHOULD prefer this over console.* for testability. */\n  log?: (msg: string, fields?: Record<string, unknown>) => void\n  /** Optional abort signal. Analysts SHOULD pass it through to LLM calls. */\n  signal?: AbortSignal\n  /**\n   * Optional live-execution port. A runtime that owns a sandbox or checkout\n   * fills it so an analyst can execute a bounded probe against the run's\n   * produced state instead of reasoning about it from the trace alone. This\n   * package defines only the port: no field here reaches for an agent loop,\n   * and an absent probe means the analyst works from recorded evidence.\n   */\n  probe?: ExecutionProbe\n}\n\n// ── Live-execution port ─────────────────────────────────────────────\n\n/** One bounded command an analyst asks the probe to run. */\nexport interface ExecutionProbeRequest {\n  command: string\n  /** Working directory inside the probed environment. */\n  cwd?: string\n  /** Hard wall-clock deadline for this one execution. */\n  timeoutMs: number\n  /** Bytes of combined output retained; the prober truncates beyond it. */\n  maxOutputBytes?: number\n  signal?: AbortSignal\n}\n\n/**\n * Typed outcome of one probe execution. `succeeded: false` is a PROBE failure\n * (the environment could not run the command); a command that ran and exited\n * non-zero is a successful observation with a non-zero `exitCode`.\n */\nexport type ExecutionProbeOutcome =\n  | {\n      succeeded: true\n      exitCode: number\n      stdout: string\n      stderr: string\n      durationMs: number\n      /** True when output was cut at `maxOutputBytes`. */\n      truncated: boolean\n    }\n  | { succeeded: false; error: { class: string; message: string } }\n\n/**\n * The seam a runtime fills to let analysts observe produced state live.\n * Implementations own sandboxing, credentials, and cleanup; analysts only\n * submit bounded requests and read typed outcomes.\n */\nexport interface ExecutionProbe {\n  /** One plain sentence naming what is being probed (e.g. a sandbox id). */\n  readonly description: string\n  execute(request: ExecutionProbeRequest): Promise<ExecutionProbeOutcome>\n}\n\n/**\n * The minimal contract. Concrete analysts can refine `TInput` so\n * implementations stay type-safe (e.g. a trace analyst's `TInput` is\n * `TraceAnalysisStore`); the registry passes the right field from\n * `AnalystRunInputs` based on `inputKind`.\n */\nexport interface Analyst<TInput = unknown> {\n  /** Stable identifier — appears in finding_id, telemetry, and registry exclusion lists. */\n  readonly id: string\n  /** Human-readable. One sentence. */\n  readonly description: string\n  readonly inputKind: AnalystInputKind\n  readonly cost: AnalystCost\n  readonly requires?: AnalystRequirements\n  /** Bump on breaking changes to claim wording or area so old finding_ids don't collide. */\n  readonly version: string\n  analyze(input: TInput, ctx: AnalystContext): Promise<AnalystFinding[]>\n}\n\n/** Metered work performed by one analyst call. */\nexport interface AnalystUsageReceipt {\n  /** Number of model-usage records observed at the provider boundary. */\n  calls: number | null\n  /** Null when the provider did not return token accounting. */\n  tokens: RunTokenUsage | null\n  /** Observed, estimated, or explicitly uncaptured dollar cost. */\n  cost: CostProvenance\n  /** Known lower bound when one or more calls have uncaptured cost. */\n  knownCostUsd?: number\n  /**\n   * Token counts the provider reported on only one side. Present exactly when\n   * `tokens` is null and at least one side WAS reported: `RunTokenUsage` has no\n   * nullable side, so a one-sided count cannot live in `tokens` without writing\n   * a zero nobody measured. Read it as a lower bound, never as a total — the\n   * field exists so a null `tokens` cannot hide a real count.\n   */\n  partialTokens?: { input: number | null; output: number | null }\n  /**\n   * True when the token counts were DERIVED by the transport (from character\n   * lengths, say) rather than measured by the model provider. `cost.kind` is\n   * `estimated` both for a rate estimate over exact tokens and for one over\n   * derived tokens; this is the field that separates them.\n   */\n  tokensEstimated?: boolean\n}\n\n// ── finding_id stability ─────────────────────────────────────────────\n\n/**\n * Compute the stable finding_id from the identity-defining fields.\n * Default implementation hashes {analyst_id, area, subject, normalized claim}.\n * Analysts that emit findings whose claim text varies per run (timestamps,\n * counts) SHOULD either: (a) pass an explicit `id_basis` to fix the hash,\n * or (b) move the variable part into `rationale`/`metadata` and keep the\n * `claim` static.\n */\nexport function computeFindingId(input: {\n  analyst_id: string\n  area: string\n  subject?: string\n  claim: string\n  /** Override the claim for hashing — use when the displayed claim has run-specific bits. */\n  id_basis?: string\n}): string {\n  const basis = JSON.stringify({\n    a: input.analyst_id,\n    r: input.area,\n    s: input.subject ?? '',\n    c: normalizeClaim(input.id_basis ?? input.claim),\n  })\n  return `f_${createHash('sha256').update(basis).digest('hex').slice(0, 20)}`\n}\n\nfunction normalizeClaim(c: string): string {\n  // Lowercase, collapse whitespace, strip trailing punctuation. Goal:\n  // \"Leaf X failed install\" and \"Leaf X failed install.\" hash the same.\n  return c\n    .toLowerCase()\n    .replace(/\\s+/g, ' ')\n    .replace(/[.!?;:,]+$/g, '')\n    .trim()\n}\n\n/** Fields that exist only in schema 1.1.0. */\nexport const FINDING_SCOPE_FIELDS = [\n  'layer',\n  'scope',\n  'method',\n  'measure',\n  'confounders',\n  'valid_for',\n  'claim_id',\n  'evidence_units',\n] as const satisfies ReadonlyArray<keyof AnalystFinding>\n\n/**\n * Convenience factory: produce a fully-formed AnalystFinding with the\n * id computed automatically. Analyst code stays terse.\n */\nexport function makeFinding(\n  init: Omit<AnalystFinding, 'schema_version' | 'finding_id' | 'produced_at'> & {\n    id_basis?: string\n    produced_at?: string\n  },\n): AnalystFinding {\n  const { id_basis, produced_at, ...rest } = init\n  return {\n    schema_version: FINDING_SCOPE_FIELDS.some((field) => rest[field] !== undefined)\n      ? '1.1.0'\n      : '1.0.0',\n    finding_id: computeFindingId({\n      analyst_id: rest.analyst_id,\n      area: rest.area,\n      subject: rest.subject,\n      claim: rest.claim,\n      id_basis,\n    }),\n    produced_at: produced_at ?? new Date().toISOString(),\n    ...rest,\n  }\n}\n\n/** Build a finding whose source is explicitly allowed during candidate generation. */\nexport function makeProposalFinding(\n  init: Omit<ProposalFinding, 'schema_version' | 'finding_id' | 'produced_at'> & {\n    id_basis?: string\n    produced_at?: string\n  },\n): ProposalFinding {\n  const { proposal_origin, ...finding } = init\n  return { ...makeFinding(finding), proposal_origin }\n}\n\n// ── Registry result envelope ────────────────────────────────────────\n\nexport interface AnalystRunSummary {\n  analyst_id: string\n  status: 'ok' | 'skipped' | 'failed'\n  /** Why skipped — missing input, budget exceeded, capability unmet. */\n  reason?: string\n  findings_count: number\n  latency_ms: number\n  /** Additive model usage and cost provenance for this analyst. */\n  usage: AnalystUsageReceipt\n  /** When `status='failed'`: the error class + message, never the full stack. */\n  error?: { class: string; message: string }\n}\n\nexport interface AnalystRunResult {\n  run_id: string\n  correlation_id: string\n  started_at: string\n  ended_at: string\n  findings: AnalystFinding[]\n  per_analyst: AnalystRunSummary[]\n  /** Total LLM cost in USD across all analysts in this registry.run(). */\n  total_cost_usd: number\n  /**\n   * Provenance for `total_cost_usd`. When uncaptured, the numeric field is only\n   * the known subtotal and must not be treated as the run's total spend.\n   */\n  total_cost_provenance?: CostProvenance\n}\n\n// ── Streaming event envelope ────────────────────────────────────────\n\n/**\n * Events emitted by `AnalystRegistry.runStream(...)` in real time as\n * the registry executes. UIs subscribe via `for await (const ev of\n * registry.runStream(...))`; `registry.run(...)` is a thin collector\n * over the same stream, so the two surfaces share their invariants.\n *\n * Per-finding events are intentionally omitted — analyzers are batch\n * operations (a recursive engine returns the full `findings:json[]` at the\n * end of the responder), so streaming inside one analyst would only\n * emit partial JSON consumers can't render. The kind-completion event\n * is the right granularity; subscribers wanting per-finding rendering\n * iterate `event.findings` themselves.\n */\nexport type AnalystRunEvent =\n  | {\n      type: 'run-started'\n      run_id: string\n      correlation_id: string\n      started_at: string\n      /** The ordered list of analyst ids the registry will run. */\n      analyst_ids: ReadonlyArray<string>\n    }\n  | {\n      type: 'analyst-skipped'\n      summary: AnalystRunSummary\n    }\n  | {\n      type: 'analyst-started'\n      analyst_id: string\n      started_at: string\n    }\n  | {\n      type: 'analyst-completed'\n      /** `summary.status` is `'ok'` for clean completion or `'failed'` for thrown analysts. */\n      summary: AnalystRunSummary\n      findings: ReadonlyArray<AnalystFinding>\n    }\n  | {\n      type: 'run-completed'\n      result: AnalystRunResult\n    }\n"],"mappings":";;;;;;;;;;;;;;;;;;;AAwGA,MAAa,kCAAkC,CAAC,SAAS,OAAO;;;AAMhE,MAAa,qBAAqB;CAAC;CAAY;CAAQ;CAAU;CAAO;AAAM;;;;;;;;;AAiO9E,SAAgB,iBAAiB,OAOtB;CACT,MAAM,QAAQ,KAAK,UAAU;EAC3B,GAAG,MAAM;EACT,GAAG,MAAM;EACT,GAAG,MAAM,WAAW;EACpB,GAAG,eAAe,MAAM,YAAY,MAAM,KAAK;CACjD,CAAC;CACD,OAAO,KAAK,WAAW,QAAQ,CAAC,CAAC,OAAO,KAAK,CAAC,CAAC,OAAO,KAAK,CAAC,CAAC,MAAM,GAAG,EAAE;AAC1E;AAEA,SAAS,eAAe,GAAmB;CAGzC,OAAO,EACJ,YAAY,CAAC,CACb,QAAQ,QAAQ,GAAG,CAAC,CACpB,QAAQ,eAAe,EAAE,CAAC,CAC1B,KAAK;AACV;;AAGA,MAAa,uBAAuB;CAClC;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF;;;;;AAMA,SAAgB,YACd,MAIgB;CAChB,MAAM,EAAE,UAAU,aAAa,GAAG,SAAS;CAC3C,OAAO;EACL,gBAAgB,qBAAqB,MAAM,UAAU,KAAK,WAAW,KAAA,CAAS,IAC1E,UACA;EACJ,YAAY,iBAAiB;GAC3B,YAAY,KAAK;GACjB,MAAM,KAAK;GACX,SAAS,KAAK;GACd,OAAO,KAAK;GACZ;EACF,CAAC;EACD,aAAa,gCAAe,IAAI,KAAK,EAAA,CAAE,YAAY;EACnD,GAAG;CACL;AACF;;AAGA,SAAgB,oBACd,MAIiB;CACjB,MAAM,EAAE,iBAAiB,GAAG,YAAY;CACxC,OAAO;EAAE,GAAG,YAAY,OAAO;EAAG;CAAgB;AACpD"}

package/dist/{verdict-DhPoXApk.js → verdict-Dl0tiT1X.js}

@@ -1,4 +1,4 @@
-import { n as contentHash } from "./verdict-cache-BQEsVy9n.js";
+import { n as contentHash } from "./verdict-cache-Bm8EKL2s.js";
 //#region src/verdict.ts
 /**
 * Validator-output verdict — substrate primitive for "did this output pass,
@@ -58,4 +58,4 @@ function equivalenceVerdict(record) {
 //#endregion
 export { equivalenceVerdict as n, certificationEvidenceDigest as t };
 
-//# sourceMappingURL=verdict-DhPoXApk.js.map
+//# sourceMappingURL=verdict-Dl0tiT1X.js.map

package/dist/{verdict-DhPoXApk.js.map → verdict-Dl0tiT1X.js.map}

@@ -1 +1 @@
-{"version":3,"file":"verdict-DhPoXApk.js","names":[],"sources":["../src/verdict.ts"],"sourcesContent":["/**\n * Validator-output verdict — substrate primitive for \"did this output pass,\n * and how well?\"\n *\n * Used by:\n *   - `@tangle-network/agent-eval/matrix` — verdict per cell in the cartesian.\n *   - `@tangle-network/agent-runtime` — Validator<Output, Verdict = DefaultVerdict>.\n *     Runtime keeps `Validator` because it's coupled to runtime-shaped\n *     `ValidationCtx` (iteration, signal, traceEmitter); the verdict TYPE\n *     itself is a substrate concept and lives here.\n *\n * Repo layering: agent-eval is the substrate (no upward deps). Both\n * agent-runtime and agent-knowledge consume this type FROM agent-eval —\n * never the other way around. See CLAUDE.md \"Repo layering\" for the rule.\n */\n\nimport { contentHash } from './verdict-cache'\nimport type {\n  CheckerIdentity,\n  EquivalenceRecord,\n  VerificationStrategySource,\n} from './verification-strategy'\n\n/**\n * What certified a verdict — the epistemics a bare `valid` + `score` pair\n * cannot carry. A kernel-checked proof and an LLM judge can produce the\n * same `{ valid: true, score: 1 }`; this record is what tells them apart.\n *\n * Every field is required on purpose. A certification without a checker\n * identity or an evidence digest is exactly the unverifiable claim this\n * record exists to kill.\n */\nexport interface VerdictCertification {\n  /**\n   * Which verification-strategy member produced the certificate. Each\n   * member's failure mode is documented on the family type and in\n   * `VERIFICATION_STRATEGIES` (src/verification-strategy.ts).\n   */\n  strategy: VerificationStrategySource\n  /**\n   * Exact checker identity, e.g. `{ name: 'lean4', version: '4.33.0',\n   * pins: { mathlib: 'db584cd6d46c' } }` for a kernel, or a judge model\n   * id + snapshot for a judge.\n   */\n  checker: CheckerIdentity\n  /**\n   * Every step the certificate rests on that the checker did NOT verify,\n   * e.g. 'diagonal embedding argued in prose' or 'agreement arms share\n   * training-corpus blind spots'. An empty array is the producer's\n   * explicit, auditable claim that there are none — it is not a default.\n   */\n  assumptions: string[]\n  /** Digest (e.g. sha256) of the evidence artifact the checker consumed or produced. */\n  evidenceDigest: string\n}\n\n/**\n * Minimal verdict shape — `valid` + `score` are required; `scores` +\n * `notes` are optional surface. Validators that need richer shapes\n * parameterise `Validator<Output, MyVerdict>` with their own type.\n *\n * Need structured extras? Extend DefaultVerdict with typed fields — never\n * serialize extras into `notes`.\n */\nexport interface DefaultVerdict {\n  /** Whether the output meets the validator's pass criteria. */\n  valid: boolean\n  /** Aggregate score in [0, 1]. Drivers use this for winner selection. */\n  score: number\n  /** Per-dimension scores. Free-form; weighted into `score` by the validator. */\n  scores?: Record<string, number>\n  /** Human-readable rationale; surfaces in trace + final-result `winner.verdict`. */\n  notes?: string\n  /**\n   * What certified this verdict, when anything did. Absent = uncertified:\n   * scored, but no strategy vouches for it. A consumer that cares reads\n   * `certification.strategy` to tell a kernel-checked verdict from a\n   * judge-scored one; a consumer that does not read it sees the exact\n   * verdict it always saw.\n   */\n  certification?: VerdictCertification\n}\n\n/**\n * Digest a certification's evidence artifact: sha-256 over the canonical\n * JSON of the evidence AS SERIALIZED. The JSON round-trip drops\n * `undefined`-valued fields on purpose — certified verdicts persist as\n * JSON, so the digest must cover exactly what a reader of the persisted\n * artifact can re-digest, nothing the wire cannot carry. Throws when the\n * evidence has no JSON form at all.\n */\nexport function certificationEvidenceDigest(evidence: unknown): string {\n  const serialized = JSON.stringify(evidence)\n  if (serialized === undefined) {\n    throw new Error(\n      'certificationEvidenceDigest: evidence has no JSON form — a certificate over nothing is the unverifiable claim this record exists to kill',\n    )\n  }\n  return contentHash(JSON.parse(serialized))\n}\n\n/**\n * Land a finished statement-equivalence record in the shared verdict\n * spine. `valid` answers the protocol's question — are the two blind\n * statements equivalent — so a refutation is `valid: false` even though a\n * separating witness in hand is a successful protocol outcome.\n *\n * Certification is present exactly when the obligation was discharged\n * (proved or refuted): the checker vouches for the discharge either way.\n * An unresolved obligation certifies nothing and keeps its reason in\n * `notes`.\n */\nexport function equivalenceVerdict(record: EquivalenceRecord): DefaultVerdict {\n  const { obligation, spec, arms } = record\n  const discharged = obligation.status !== 'unresolved'\n  const notes =\n    obligation.status === 'proved'\n      ? `statements equivalent: ${spec.artifact}`\n      : obligation.status === 'refuted-with-separating-witness'\n        ? `formalization gap witnessed: ${obligation.separatingWitness}`\n        : `obligation undischarged: ${obligation.unresolvedReason}`\n  return {\n    valid: obligation.status === 'proved',\n    score: obligation.status === 'proved' ? 1 : 0,\n    notes,\n    ...(discharged && obligation.evidenceDigest !== undefined\n      ? {\n          certification: {\n            strategy: spec.source,\n            checker: obligation.checker,\n            // The arms' blindness declarations are attested by the arms,\n            // never verified by the checker — the one step every\n            // equivalence certificate rests on.\n            assumptions: [\n              `arm '${arms[0].armId}' and arm '${arms[1].armId}' self-declare blind derivation`,\n            ],\n            evidenceDigest: obligation.evidenceDigest,\n          } satisfies VerdictCertification,\n        }\n      : {}),\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;AA2FA,SAAgB,4BAA4B,UAA2B;CACrE,MAAM,aAAa,KAAK,UAAU,QAAQ;CAC1C,IAAI,eAAe,KAAA,GACjB,MAAM,IAAI,MACR,0IACF;CAEF,OAAO,YAAY,KAAK,MAAM,UAAU,CAAC;AAC3C;;;;;;;;;;;;AAaA,SAAgB,mBAAmB,QAA2C;CAC5E,MAAM,EAAE,YAAY,MAAM,SAAS;CACnC,MAAM,aAAa,WAAW,WAAW;CACzC,MAAM,QACJ,WAAW,WAAW,WAClB,0BAA0B,KAAK,aAC/B,WAAW,WAAW,oCACpB,gCAAgC,WAAW,sBAC3C,4BAA4B,WAAW;CAC/C,OAAO;EACL,OAAO,WAAW,WAAW;EAC7B,OAAO,WAAW,WAAW,WAAW,IAAI;EAC5C;EACA,GAAI,cAAc,WAAW,mBAAmB,KAAA,IAC5C,EACE,eAAe;GACb,UAAU,KAAK;GACf,SAAS,WAAW;GAIpB,aAAa,CACX,QAAQ,KAAK,EAAE,CAAC,MAAM,aAAa,KAAK,EAAE,CAAC,MAAM,gCACnD;GACA,gBAAgB,WAAW;EAC7B,EACF,IACA,CAAC;CACP;AACF"}
+{"version":3,"file":"verdict-Dl0tiT1X.js","names":[],"sources":["../src/verdict.ts"],"sourcesContent":["/**\n * Validator-output verdict — substrate primitive for \"did this output pass,\n * and how well?\"\n *\n * Used by:\n *   - `@tangle-network/agent-eval/matrix` — verdict per cell in the cartesian.\n *   - `@tangle-network/agent-runtime` — Validator<Output, Verdict = DefaultVerdict>.\n *     Runtime keeps `Validator` because it's coupled to runtime-shaped\n *     `ValidationCtx` (iteration, signal, traceEmitter); the verdict TYPE\n *     itself is a substrate concept and lives here.\n *\n * Repo layering: agent-eval is the substrate (no upward deps). Both\n * agent-runtime and agent-knowledge consume this type FROM agent-eval —\n * never the other way around. See CLAUDE.md \"Repo layering\" for the rule.\n */\n\nimport { contentHash } from './verdict-cache'\nimport type {\n  CheckerIdentity,\n  EquivalenceRecord,\n  VerificationStrategySource,\n} from './verification-strategy'\n\n/**\n * What certified a verdict — the epistemics a bare `valid` + `score` pair\n * cannot carry. A kernel-checked proof and an LLM judge can produce the\n * same `{ valid: true, score: 1 }`; this record is what tells them apart.\n *\n * Every field is required on purpose. A certification without a checker\n * identity or an evidence digest is exactly the unverifiable claim this\n * record exists to kill.\n */\nexport interface VerdictCertification {\n  /**\n   * Which verification-strategy member produced the certificate. Each\n   * member's failure mode is documented on the family type and in\n   * `VERIFICATION_STRATEGIES` (src/verification-strategy.ts).\n   */\n  strategy: VerificationStrategySource\n  /**\n   * Exact checker identity, e.g. `{ name: 'lean4', version: '4.33.0',\n   * pins: { mathlib: 'db584cd6d46c' } }` for a kernel, or a judge model\n   * id + snapshot for a judge.\n   */\n  checker: CheckerIdentity\n  /**\n   * Every step the certificate rests on that the checker did NOT verify,\n   * e.g. 'diagonal embedding argued in prose' or 'agreement arms share\n   * training-corpus blind spots'. An empty array is the producer's\n   * explicit, auditable claim that there are none — it is not a default.\n   */\n  assumptions: string[]\n  /** Digest (e.g. sha256) of the evidence artifact the checker consumed or produced. */\n  evidenceDigest: string\n}\n\n/**\n * Minimal verdict shape — `valid` + `score` are required; `scores` +\n * `notes` are optional surface. Validators that need richer shapes\n * parameterise `Validator<Output, MyVerdict>` with their own type.\n *\n * Need structured extras? Extend DefaultVerdict with typed fields — never\n * serialize extras into `notes`.\n */\nexport interface DefaultVerdict {\n  /** Whether the output meets the validator's pass criteria. */\n  valid: boolean\n  /** Aggregate score in [0, 1]. Drivers use this for winner selection. */\n  score: number\n  /** Per-dimension scores. Free-form; weighted into `score` by the validator. */\n  scores?: Record<string, number>\n  /** Human-readable rationale; surfaces in trace + final-result `winner.verdict`. */\n  notes?: string\n  /**\n   * What certified this verdict, when anything did. Absent = uncertified:\n   * scored, but no strategy vouches for it. A consumer that cares reads\n   * `certification.strategy` to tell a kernel-checked verdict from a\n   * judge-scored one; a consumer that does not read it sees the exact\n   * verdict it always saw.\n   */\n  certification?: VerdictCertification\n}\n\n/**\n * Digest a certification's evidence artifact: sha-256 over the canonical\n * JSON of the evidence AS SERIALIZED. The JSON round-trip drops\n * `undefined`-valued fields on purpose — certified verdicts persist as\n * JSON, so the digest must cover exactly what a reader of the persisted\n * artifact can re-digest, nothing the wire cannot carry. Throws when the\n * evidence has no JSON form at all.\n */\nexport function certificationEvidenceDigest(evidence: unknown): string {\n  const serialized = JSON.stringify(evidence)\n  if (serialized === undefined) {\n    throw new Error(\n      'certificationEvidenceDigest: evidence has no JSON form — a certificate over nothing is the unverifiable claim this record exists to kill',\n    )\n  }\n  return contentHash(JSON.parse(serialized))\n}\n\n/**\n * Land a finished statement-equivalence record in the shared verdict\n * spine. `valid` answers the protocol's question — are the two blind\n * statements equivalent — so a refutation is `valid: false` even though a\n * separating witness in hand is a successful protocol outcome.\n *\n * Certification is present exactly when the obligation was discharged\n * (proved or refuted): the checker vouches for the discharge either way.\n * An unresolved obligation certifies nothing and keeps its reason in\n * `notes`.\n */\nexport function equivalenceVerdict(record: EquivalenceRecord): DefaultVerdict {\n  const { obligation, spec, arms } = record\n  const discharged = obligation.status !== 'unresolved'\n  const notes =\n    obligation.status === 'proved'\n      ? `statements equivalent: ${spec.artifact}`\n      : obligation.status === 'refuted-with-separating-witness'\n        ? `formalization gap witnessed: ${obligation.separatingWitness}`\n        : `obligation undischarged: ${obligation.unresolvedReason}`\n  return {\n    valid: obligation.status === 'proved',\n    score: obligation.status === 'proved' ? 1 : 0,\n    notes,\n    ...(discharged && obligation.evidenceDigest !== undefined\n      ? {\n          certification: {\n            strategy: spec.source,\n            checker: obligation.checker,\n            // The arms' blindness declarations are attested by the arms,\n            // never verified by the checker — the one step every\n            // equivalence certificate rests on.\n            assumptions: [\n              `arm '${arms[0].armId}' and arm '${arms[1].armId}' self-declare blind derivation`,\n            ],\n            evidenceDigest: obligation.evidenceDigest,\n          } satisfies VerdictCertification,\n        }\n      : {}),\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;AA2FA,SAAgB,4BAA4B,UAA2B;CACrE,MAAM,aAAa,KAAK,UAAU,QAAQ;CAC1C,IAAI,eAAe,KAAA,GACjB,MAAM,IAAI,MACR,0IACF;CAEF,OAAO,YAAY,KAAK,MAAM,UAAU,CAAC;AAC3C;;;;;;;;;;;;AAaA,SAAgB,mBAAmB,QAA2C;CAC5E,MAAM,EAAE,YAAY,MAAM,SAAS;CACnC,MAAM,aAAa,WAAW,WAAW;CACzC,MAAM,QACJ,WAAW,WAAW,WAClB,0BAA0B,KAAK,aAC/B,WAAW,WAAW,oCACpB,gCAAgC,WAAW,sBAC3C,4BAA4B,WAAW;CAC/C,OAAO;EACL,OAAO,WAAW,WAAW;EAC7B,OAAO,WAAW,WAAW,WAAW,IAAI;EAC5C;EACA,GAAI,cAAc,WAAW,mBAAmB,KAAA,IAC5C,EACE,eAAe;GACb,UAAU,KAAK;GACf,SAAS,WAAW;GAIpB,aAAa,CACX,QAAQ,KAAK,EAAE,CAAC,MAAM,aAAa,KAAK,EAAE,CAAC,MAAM,gCACnD;GACA,gBAAgB,WAAW;EAC7B,EACF,IACA,CAAC;CACP;AACF"}

package/dist/{verdict-cache-BQEsVy9n.js → verdict-cache-Bm8EKL2s.js}

@@ -1,4 +1,4 @@
-import { a as hashCanonical, r as canonicalString } from "./canonical-BsZVgBhW.js";
+import { a as hashCanonical, r as canonicalString } from "./canonical-B2RqLofv.js";
 import { appendFileSync, existsSync, readFileSync } from "node:fs";
 //#region src/verdict-cache.ts
 /**
@@ -85,4 +85,4 @@ function fileVerdictCache(path) {
 //#endregion
 export { contentHash as n, fileVerdictCache as r, canonicalJson as t };
 
-//# sourceMappingURL=verdict-cache-BQEsVy9n.js.map
+//# sourceMappingURL=verdict-cache-Bm8EKL2s.js.map

package/dist/{verdict-cache-BQEsVy9n.js.map → verdict-cache-Bm8EKL2s.js.map}

@@ -1 +1 @@
-{"version":3,"file":"verdict-cache-BQEsVy9n.js","names":[],"sources":["../src/verdict-cache.ts"],"sourcesContent":["/**\n * Content-addressed judge-verdict caching.\n *\n * LAW: cache JUDGE VERDICTS only — judging the same artifact with the same\n * judge+rubric is pure. NEVER cache agent rollouts. (A router that cached\n * identical fanout prompts silently destroyed best-of-N diversity; rollout\n * caching reintroduces that failure class. Judging has no diversity to\n * destroy — same artifact + same rubric ⇒ same verdict is the desired\n * property, not a bug.)\n *\n * The cache key is `v2:` + a sha-256 over the canonical JSON of everything\n * that can change a verdict: the artifact content, the scenario id, the judge\n * name, the full dimension list (key + description — the description IS the\n * rubric text shown to the judge), and a caller-supplied `judgeVersion`.\n * `judgeVersion` is REQUIRED: a judge whose prompt/model/ensemble changes\n * without a version bump would otherwise silently serve stale verdicts.\n * The `v2:` prefix names the key scheme; a store written under another scheme\n * misses once and repopulates.\n *\n * Canonical JSON is RFC 8785 from `ledger-core/canonical` — the package's one\n * identity scheme. It refuses `undefined`, functions, symbols, non-finite\n * numbers, class instances (including `Date`), `Map`, `Set`, cycles, and\n * unpaired surrogates: an artifact that cannot be unambiguously serialized\n * cannot be content-addressed, and coercing it would let two different\n * artifacts collide on one key. Pass timestamps as ISO strings.\n */\n\nimport { appendFileSync, existsSync, readFileSync } from 'node:fs'\nimport type { JudgeConfig, JudgeScore, Scenario } from './campaign/types'\nimport { canonicalString, hashCanonical } from './ledger-core/canonical'\n\n// ── canonical JSON + content hash ─────────────────────────────────────────\n\n/** Key scheme of every entry `cachedJudge` writes. */\nexport const VERDICT_CACHE_KEY_SCHEME = 'v2'\n\n/**\n * Canonical JSON (RFC 8785): object keys sorted recursively, so two\n * semantically equal values produce byte-identical output regardless of key\n * insertion order. Throws `LedgerCanonicalizationError` on any value with no\n * faithful JSON form — see `ledger-core/canonical`.\n */\nexport function canonicalJson(value: unknown): string {\n  return canonicalString(value)\n}\n\n/** Hex sha-256 over `canonicalJson(value)`. The content address used by the\n *  verdict cache and report attestation. Equals `hashCanonical(value)` with\n *  the `sha256:` prefix removed. */\nexport function contentHash(value: unknown): string {\n  return hashCanonical(value).slice('sha256:'.length)\n}\n\n// ── store contract ─────────────────────────────────────────────────────────\n\n/** Pluggable verdict store. Sync or async on both legs — `cachedJudge`\n *  awaits the results either way. */\nexport interface VerdictCacheStore {\n  get(key: string): Promise<JudgeScore | undefined> | JudgeScore | undefined\n  set(key: string, score: JudgeScore): Promise<void> | void\n}\n\n/** Process-local Map-backed store. */\nexport function inMemoryVerdictCache(): VerdictCacheStore {\n  const entries = new Map<string, JudgeScore>()\n  return {\n    get: (key) => entries.get(key),\n    set: (key, score) => {\n      entries.set(key, score)\n    },\n  }\n}\n\ninterface VerdictCacheLine {\n  key: string\n  score: JudgeScore\n}\n\nfunction parseCacheLine(line: string, path: string, lineNo: number): VerdictCacheLine {\n  let parsed: unknown\n  try {\n    parsed = JSON.parse(line)\n  } catch (err) {\n    throw new Error(\n      `fileVerdictCache: corrupt JSONL at ${path}:${lineNo} — ${err instanceof Error ? err.message : String(err)}`,\n    )\n  }\n  const rec = parsed as Partial<VerdictCacheLine>\n  if (\n    typeof rec !== 'object' ||\n    rec === null ||\n    typeof rec.key !== 'string' ||\n    typeof rec.score !== 'object' ||\n    rec.score === null ||\n    typeof rec.score.composite !== 'number' ||\n    typeof rec.score.dimensions !== 'object'\n  ) {\n    throw new Error(\n      `fileVerdictCache: invalid record shape at ${path}:${lineNo} — expected {key, score:{dimensions, composite, notes}}`,\n    )\n  }\n  return rec as VerdictCacheLine\n}\n\n/**\n * JSONL-file-backed store: the full file is loaded into an in-memory index at\n * construction; every `set` appends one line synchronously (durable before\n * the verdict is returned). A corrupt or malformed line throws at load with\n * file:line — a skipped line would silently re-judge (cost) or, worse, mask\n * a half-written file that needs operator attention.\n */\nexport function fileVerdictCache(path: string): VerdictCacheStore {\n  const entries = new Map<string, JudgeScore>()\n  if (existsSync(path)) {\n    const lines = readFileSync(path, 'utf8').split('\\n')\n    for (let i = 0; i < lines.length; i++) {\n      const line = lines[i]\n      if (line === undefined || line.trim() === '') continue\n      const rec = parseCacheLine(line, path, i + 1)\n      entries.set(rec.key, rec.score)\n    }\n  }\n  return {\n    get: (key) => entries.get(key),\n    set: (key, score) => {\n      appendFileSync(path, `${JSON.stringify({ key, score })}\\n`, 'utf8')\n      entries.set(key, score)\n    },\n  }\n}\n\n// ── cached judge wrapper ───────────────────────────────────────────────────\n\nexport interface VerdictCacheStats {\n  hits: number\n  misses: number\n}\n\nexport interface CachedJudgeOptions {\n  /** REQUIRED — part of the cache key. Bump on any change to the judge's\n   *  prompt, model, ensemble, or scoring logic; silent judge upgrades must\n   *  never serve stale verdicts. */\n  judgeVersion: string\n}\n\n/** The wrapped judge: same `JudgeConfig` seam, plus hit/miss observability. */\nexport type CachedJudge<TArtifact, TScenario extends Scenario = Scenario> = JudgeConfig<\n  TArtifact,\n  TScenario\n> & {\n  stats(): VerdictCacheStats\n}\n\n/**\n * Wrap a `JudgeConfig` so repeat judgments of the same artifact are served\n * from the store instead of re-invoking `score()`. The wrapper is generic\n * over the judge's own type parameters and preserves `appliesTo` — it is a\n * drop-in replacement anywhere a `JudgeConfig` is accepted.\n *\n * A judge that throws is NOT cached: the error propagates and the next\n * attempt re-judges (caching a failure would pin a transient outage forever).\n */\nexport function cachedJudge<TArtifact, TScenario extends Scenario = Scenario>(\n  judge: JudgeConfig<TArtifact, TScenario>,\n  store: VerdictCacheStore,\n  options: CachedJudgeOptions,\n): CachedJudge<TArtifact, TScenario> {\n  if (typeof options.judgeVersion !== 'string' || options.judgeVersion.trim() === '') {\n    throw new Error('cachedJudge: judgeVersion is required and must be a non-empty string')\n  }\n  const stats: VerdictCacheStats = { hits: 0, misses: 0 }\n  const wrapped: CachedJudge<TArtifact, TScenario> = {\n    name: judge.name,\n    dimensions: judge.dimensions,\n    judgeVersion: options.judgeVersion,\n    async score(input) {\n      const key = `${VERDICT_CACHE_KEY_SCHEME}:${contentHash({\n        artifact: canonicalJson(input.artifact),\n        scenarioId: input.scenario.id,\n        judgeName: judge.name,\n        dimensions: judge.dimensions,\n        judgeVersion: options.judgeVersion,\n      })}`\n      const cached = await store.get(key)\n      if (cached !== undefined) {\n        stats.hits += 1\n        return cached\n      }\n      const score = await judge.score(input)\n      await store.set(key, score)\n      stats.misses += 1\n      return score\n    },\n    stats: () => ({ ...stats }),\n  }\n  if (judge.appliesTo) wrapped.appliesTo = judge.appliesTo\n  return wrapped\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA0CA,SAAgB,cAAc,OAAwB;CACpD,OAAO,gBAAgB,KAAK;AAC9B;;;;AAKA,SAAgB,YAAY,OAAwB;CAClD,OAAO,cAAc,KAAK,CAAC,CAAC,MAAM,CAAgB;AACpD;AA2BA,SAAS,eAAe,MAAc,MAAc,QAAkC;CACpF,IAAI;CACJ,IAAI;EACF,SAAS,KAAK,MAAM,IAAI;CAC1B,SAAS,KAAK;EACZ,MAAM,IAAI,MACR,sCAAsC,KAAK,GAAG,OAAO,KAAK,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,GAC3G;CACF;CACA,MAAM,MAAM;CACZ,IACE,OAAO,QAAQ,YACf,QAAQ,QACR,OAAO,IAAI,QAAQ,YACnB,OAAO,IAAI,UAAU,YACrB,IAAI,UAAU,QACd,OAAO,IAAI,MAAM,cAAc,YAC/B,OAAO,IAAI,MAAM,eAAe,UAEhC,MAAM,IAAI,MACR,6CAA6C,KAAK,GAAG,OAAO,wDAC9D;CAEF,OAAO;AACT;;;;;;;;AASA,SAAgB,iBAAiB,MAAiC;CAChE,MAAM,0BAAU,IAAI,IAAwB;CAC5C,IAAI,WAAW,IAAI,GAAG;EACpB,MAAM,QAAQ,aAAa,MAAM,MAAM,CAAC,CAAC,MAAM,IAAI;EACnD,KAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;GACrC,MAAM,OAAO,MAAM;GACnB,IAAI,SAAS,KAAA,KAAa,KAAK,KAAK,MAAM,IAAI;GAC9C,MAAM,MAAM,eAAe,MAAM,MAAM,IAAI,CAAC;GAC5C,QAAQ,IAAI,IAAI,KAAK,IAAI,KAAK;EAChC;CACF;CACA,OAAO;EACL,MAAM,QAAQ,QAAQ,IAAI,GAAG;EAC7B,MAAM,KAAK,UAAU;GACnB,eAAe,MAAM,GAAG,KAAK,UAAU;IAAE;IAAK;GAAM,CAAC,EAAE,KAAK,MAAM;GAClE,QAAQ,IAAI,KAAK,KAAK;EACxB;CACF;AACF"}
+{"version":3,"file":"verdict-cache-Bm8EKL2s.js","names":[],"sources":["../src/verdict-cache.ts"],"sourcesContent":["/**\n * Content-addressed judge-verdict caching.\n *\n * LAW: cache JUDGE VERDICTS only — judging the same artifact with the same\n * judge+rubric is pure. NEVER cache agent rollouts. (A router that cached\n * identical fanout prompts silently destroyed best-of-N diversity; rollout\n * caching reintroduces that failure class. Judging has no diversity to\n * destroy — same artifact + same rubric ⇒ same verdict is the desired\n * property, not a bug.)\n *\n * The cache key is `v2:` + a sha-256 over the canonical JSON of everything\n * that can change a verdict: the artifact content, the scenario id, the judge\n * name, the full dimension list (key + description — the description IS the\n * rubric text shown to the judge), and a caller-supplied `judgeVersion`.\n * `judgeVersion` is REQUIRED: a judge whose prompt/model/ensemble changes\n * without a version bump would otherwise silently serve stale verdicts.\n * The `v2:` prefix names the key scheme; a store written under another scheme\n * misses once and repopulates.\n *\n * Canonical JSON is RFC 8785 from `ledger-core/canonical` — the package's one\n * identity scheme. It refuses `undefined`, functions, symbols, non-finite\n * numbers, class instances (including `Date`), `Map`, `Set`, cycles, and\n * unpaired surrogates: an artifact that cannot be unambiguously serialized\n * cannot be content-addressed, and coercing it would let two different\n * artifacts collide on one key. Pass timestamps as ISO strings.\n */\n\nimport { appendFileSync, existsSync, readFileSync } from 'node:fs'\nimport type { JudgeConfig, JudgeScore, Scenario } from './campaign/types'\nimport { canonicalString, hashCanonical } from './ledger-core/canonical'\n\n// ── canonical JSON + content hash ─────────────────────────────────────────\n\n/** Key scheme of every entry `cachedJudge` writes. */\nexport const VERDICT_CACHE_KEY_SCHEME = 'v2'\n\n/**\n * Canonical JSON (RFC 8785): object keys sorted recursively, so two\n * semantically equal values produce byte-identical output regardless of key\n * insertion order. Throws `LedgerCanonicalizationError` on any value with no\n * faithful JSON form — see `ledger-core/canonical`.\n */\nexport function canonicalJson(value: unknown): string {\n  return canonicalString(value)\n}\n\n/** Hex sha-256 over `canonicalJson(value)`. The content address used by the\n *  verdict cache and report attestation. Equals `hashCanonical(value)` with\n *  the `sha256:` prefix removed. */\nexport function contentHash(value: unknown): string {\n  return hashCanonical(value).slice('sha256:'.length)\n}\n\n// ── store contract ─────────────────────────────────────────────────────────\n\n/** Pluggable verdict store. Sync or async on both legs — `cachedJudge`\n *  awaits the results either way. */\nexport interface VerdictCacheStore {\n  get(key: string): Promise<JudgeScore | undefined> | JudgeScore | undefined\n  set(key: string, score: JudgeScore): Promise<void> | void\n}\n\n/** Process-local Map-backed store. */\nexport function inMemoryVerdictCache(): VerdictCacheStore {\n  const entries = new Map<string, JudgeScore>()\n  return {\n    get: (key) => entries.get(key),\n    set: (key, score) => {\n      entries.set(key, score)\n    },\n  }\n}\n\ninterface VerdictCacheLine {\n  key: string\n  score: JudgeScore\n}\n\nfunction parseCacheLine(line: string, path: string, lineNo: number): VerdictCacheLine {\n  let parsed: unknown\n  try {\n    parsed = JSON.parse(line)\n  } catch (err) {\n    throw new Error(\n      `fileVerdictCache: corrupt JSONL at ${path}:${lineNo} — ${err instanceof Error ? err.message : String(err)}`,\n    )\n  }\n  const rec = parsed as Partial<VerdictCacheLine>\n  if (\n    typeof rec !== 'object' ||\n    rec === null ||\n    typeof rec.key !== 'string' ||\n    typeof rec.score !== 'object' ||\n    rec.score === null ||\n    typeof rec.score.composite !== 'number' ||\n    typeof rec.score.dimensions !== 'object'\n  ) {\n    throw new Error(\n      `fileVerdictCache: invalid record shape at ${path}:${lineNo} — expected {key, score:{dimensions, composite, notes}}`,\n    )\n  }\n  return rec as VerdictCacheLine\n}\n\n/**\n * JSONL-file-backed store: the full file is loaded into an in-memory index at\n * construction; every `set` appends one line synchronously (durable before\n * the verdict is returned). A corrupt or malformed line throws at load with\n * file:line — a skipped line would silently re-judge (cost) or, worse, mask\n * a half-written file that needs operator attention.\n */\nexport function fileVerdictCache(path: string): VerdictCacheStore {\n  const entries = new Map<string, JudgeScore>()\n  if (existsSync(path)) {\n    const lines = readFileSync(path, 'utf8').split('\\n')\n    for (let i = 0; i < lines.length; i++) {\n      const line = lines[i]\n      if (line === undefined || line.trim() === '') continue\n      const rec = parseCacheLine(line, path, i + 1)\n      entries.set(rec.key, rec.score)\n    }\n  }\n  return {\n    get: (key) => entries.get(key),\n    set: (key, score) => {\n      appendFileSync(path, `${JSON.stringify({ key, score })}\\n`, 'utf8')\n      entries.set(key, score)\n    },\n  }\n}\n\n// ── cached judge wrapper ───────────────────────────────────────────────────\n\nexport interface VerdictCacheStats {\n  hits: number\n  misses: number\n}\n\nexport interface CachedJudgeOptions {\n  /** REQUIRED — part of the cache key. Bump on any change to the judge's\n   *  prompt, model, ensemble, or scoring logic; silent judge upgrades must\n   *  never serve stale verdicts. */\n  judgeVersion: string\n}\n\n/** The wrapped judge: same `JudgeConfig` seam, plus hit/miss observability. */\nexport type CachedJudge<TArtifact, TScenario extends Scenario = Scenario> = JudgeConfig<\n  TArtifact,\n  TScenario\n> & {\n  stats(): VerdictCacheStats\n}\n\n/**\n * Wrap a `JudgeConfig` so repeat judgments of the same artifact are served\n * from the store instead of re-invoking `score()`. The wrapper is generic\n * over the judge's own type parameters and preserves `appliesTo` — it is a\n * drop-in replacement anywhere a `JudgeConfig` is accepted.\n *\n * A judge that throws is NOT cached: the error propagates and the next\n * attempt re-judges (caching a failure would pin a transient outage forever).\n */\nexport function cachedJudge<TArtifact, TScenario extends Scenario = Scenario>(\n  judge: JudgeConfig<TArtifact, TScenario>,\n  store: VerdictCacheStore,\n  options: CachedJudgeOptions,\n): CachedJudge<TArtifact, TScenario> {\n  if (typeof options.judgeVersion !== 'string' || options.judgeVersion.trim() === '') {\n    throw new Error('cachedJudge: judgeVersion is required and must be a non-empty string')\n  }\n  const stats: VerdictCacheStats = { hits: 0, misses: 0 }\n  const wrapped: CachedJudge<TArtifact, TScenario> = {\n    name: judge.name,\n    dimensions: judge.dimensions,\n    judgeVersion: options.judgeVersion,\n    async score(input) {\n      const key = `${VERDICT_CACHE_KEY_SCHEME}:${contentHash({\n        artifact: canonicalJson(input.artifact),\n        scenarioId: input.scenario.id,\n        judgeName: judge.name,\n        dimensions: judge.dimensions,\n        judgeVersion: options.judgeVersion,\n      })}`\n      const cached = await store.get(key)\n      if (cached !== undefined) {\n        stats.hits += 1\n        return cached\n      }\n      const score = await judge.score(input)\n      await store.set(key, score)\n      stats.misses += 1\n      return score\n    },\n    stats: () => ({ ...stats }),\n  }\n  if (judge.appliesTo) wrapped.appliesTo = judge.appliesTo\n  return wrapped\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA0CA,SAAgB,cAAc,OAAwB;CACpD,OAAO,gBAAgB,KAAK;AAC9B;;;;AAKA,SAAgB,YAAY,OAAwB;CAClD,OAAO,cAAc,KAAK,CAAC,CAAC,MAAM,CAAgB;AACpD;AA2BA,SAAS,eAAe,MAAc,MAAc,QAAkC;CACpF,IAAI;CACJ,IAAI;EACF,SAAS,KAAK,MAAM,IAAI;CAC1B,SAAS,KAAK;EACZ,MAAM,IAAI,MACR,sCAAsC,KAAK,GAAG,OAAO,KAAK,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,GAC3G;CACF;CACA,MAAM,MAAM;CACZ,IACE,OAAO,QAAQ,YACf,QAAQ,QACR,OAAO,IAAI,QAAQ,YACnB,OAAO,IAAI,UAAU,YACrB,IAAI,UAAU,QACd,OAAO,IAAI,MAAM,cAAc,YAC/B,OAAO,IAAI,MAAM,eAAe,UAEhC,MAAM,IAAI,MACR,6CAA6C,KAAK,GAAG,OAAO,wDAC9D;CAEF,OAAO;AACT;;;;;;;;AASA,SAAgB,iBAAiB,MAAiC;CAChE,MAAM,0BAAU,IAAI,IAAwB;CAC5C,IAAI,WAAW,IAAI,GAAG;EACpB,MAAM,QAAQ,aAAa,MAAM,MAAM,CAAC,CAAC,MAAM,IAAI;EACnD,KAAK,IAAI,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;GACrC,MAAM,OAAO,MAAM;GACnB,IAAI,SAAS,KAAA,KAAa,KAAK,KAAK,MAAM,IAAI;GAC9C,MAAM,MAAM,eAAe,MAAM,MAAM,IAAI,CAAC;GAC5C,QAAQ,IAAI,IAAI,KAAK,IAAI,KAAK;EAChC;CACF;CACA,OAAO;EACL,MAAM,QAAQ,QAAQ,IAAI,GAAG;EAC7B,MAAM,KAAK,UAAU;GACnB,eAAe,MAAM,GAAG,KAAK,UAAU;IAAE;IAAK;GAAM,CAAC,EAAE,KAAK,MAAM;GAClE,QAAQ,IAAI,KAAK,KAAK;EACxB;CACF;AACF"}

package/dist/wire/index.d.ts

@@ -1,7 +1,7 @@
-import { b as CustomTokenPricing, c as CostLedgerHandle } from "../cost-ledger-CO7UyEcQ.js";
-import { x as ChatClient } from "../types-DBZ7keha.js";
+import { a as CostLedgerHandle, p as CustomTokenPricing } from "../cost-ledger-Jk62OBfF.js";
+import { x as ChatClient } from "../types-C02qtnG8.js";
 import { s as TraceStore } from "../store-C4avFjIi.js";
-import { m as FeedbackTrajectoryStore } from "../feedback-trajectory-BLN_6fEs.js";
+import { m as FeedbackTrajectoryStore } from "../feedback-trajectory-C4EG0nNH.js";
 import { z } from "zod";
 import { ServerType } from "@hono/node-server";
 import { Hono } from "hono";

package/dist/wire/index.js

@@ -1,2 +1,2 @@
-import { A as WIRE_VERSION, C as RUBRIC_VERSION_SCHEME, D as TracesIngestRequestSchema, E as TraceEventSchema, O as TracesIngestResponseSchema, S as ListRubricsResponseSchema, T as RubricSchema, _ as FeedbackIngestResponseSchema, a as runRpcBatch, b as JudgeRequestSchema, c as WireError, d as handleListRubrics, f as handleTracesIngest, g as ErrorResponseSchema, h as listBuiltinRubrics, i as dispatchRpc, j as hashRubric, k as VersionResponseSchema, l as handleFeedbackIngest, m as getBuiltinRubric, n as startServer, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi, t as createApp, u as handleJudge, v as FeedbackTrajectorySchema, w as RubricDimensionSchema, x as JudgeResultSchema, y as HealthResponseSchema } from "../server-DbcC7xMl.js";
+import { A as WIRE_VERSION, C as RUBRIC_VERSION_SCHEME, D as TracesIngestRequestSchema, E as TraceEventSchema, O as TracesIngestResponseSchema, S as ListRubricsResponseSchema, T as RubricSchema, _ as FeedbackIngestResponseSchema, a as runRpcBatch, b as JudgeRequestSchema, c as WireError, d as handleListRubrics, f as handleTracesIngest, g as ErrorResponseSchema, h as listBuiltinRubrics, i as dispatchRpc, j as hashRubric, k as VersionResponseSchema, l as handleFeedbackIngest, m as getBuiltinRubric, n as startServer, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi, t as createApp, u as handleJudge, v as FeedbackTrajectorySchema, w as RubricDimensionSchema, x as JudgeResultSchema, y as HealthResponseSchema } from "../server-e9oWO_bN.js";
 export { ErrorResponseSchema, FeedbackIngestResponseSchema, FeedbackTrajectorySchema, HealthResponseSchema, JudgeRequestSchema, JudgeResultSchema, ListRubricsResponseSchema, RUBRIC_VERSION_SCHEME, RubricDimensionSchema, RubricSchema, TraceEventSchema, TracesIngestRequestSchema, TracesIngestResponseSchema, VersionResponseSchema, WIRE_VERSION, WireError, buildOpenApi, createApp, dispatchRpc, getBuiltinRubric, handleFeedbackIngest, handleJudge, handleListRubrics, handleTracesIngest, handleVersion, hashRubric, listBuiltinRubrics, runRpcBatch, runRpcOnce, startServer, startServerAsync };

package/docs/concepts.md

@@ -418,6 +418,7 @@ release decision.
 - **Need the feature map?** Read [feature-guide.md](./feature-guide.md) for integration patterns and operational limits.
 - **Scoring a string from another language?** Read [wire-protocol.md](./wire-protocol.md) for the HTTP/RPC interface.
 - **Diagnosing failures across a batch of runs?** Read [trace-analysis.md](./trace-analysis.md) for the recursive DSPy-RLM analyst and the public benchmark it's calibrated against.
+- **Stating what a finding holds for?** Read [finding-layers.md](./finding-layers.md) for layers, scope, lineage, promotion and the findings ledger.
 - **Building a driver and worker loop?** Read [control-runtime.md](./control-runtime.md) for coding, browser, computer-use, and research patterns.
 - **Turning review feedback into reusable data?** Read [feedback-trajectories.md](./feedback-trajectories.md) for dataset, optimization, and preference-memory examples.
 - **Building a code-generator eval?** → Start with `BuilderSession`, `SandboxHarness`, and `MultiLayerVerifier`.

package/docs/design/mlbenchmarks-book-review.md

@@ -111,7 +111,7 @@ These are implementation findings, not an assessment of adoption in every consum
 | Concern | Checked implementation | Assessment and remaining gap |
 | --- | --- | --- |
 | Statistical comparisons | [statistics](../../src/statistics/index.ts), [paired decisions](../../src/paired-promotion-decision.ts), [heldout pairing](../../src/campaign/gates/statistical-heldout.ts) | Present. Paired tests, uncertainty, exact binary methods, multiplicity, and power do not need replacement. Claim scope and independent sampling units need stronger binding. |
-| Cluster-aware design | [power](../../src/experiment/power.ts), [registered rule AST](../../src/experiment/ast.ts) | Present. The high-level campaign path does not automatically select these methods from a declared generalization target. |
+| Cluster-aware design | [power](../../src/experiment/power.ts), [registered rule AST](../../src/experiment/ast/) | Present. The high-level campaign path does not automatically select these methods from a declared generalization target. |
 | Executable preregistration | [define/seal/open](../../src/experiment/define.ts), [acceptance tests](../../tests/experiment/preregistration-acceptance.test.ts) | Present. Extend this rule representation instead of creating a second experiment language. |
 | Sequential testing | [sequential gate](../../src/campaign/gates/sequential.ts), [e-process](../../src/statistics/sequential-eprocess.ts) | Present. Optional stopping within a valid stream differs from adapting hypotheses across streams or counting dependent replicas as independent evidence. |
 | Search/final separation | [selfImprove](../../src/contract/self-improve.ts), [method comparison](../../src/campaign/presets/compare-optimization-methods.ts) | Present within calls. Final cases are withheld from method inputs. A persistent cross-campaign exposure policy is missing from these paths. |

package/docs/design/mlbenchmarks-review/probes.mts

@@ -8,7 +8,8 @@ import { sequentialPairedGate } from '../../../src/campaign/gates/sequential.ts'
 import type { JudgeScore } from '../../../src/campaign/types.ts'
 import { HoldoutAuditor } from '../../../src/contamination-guard.ts'
 import { selfImprove } from '../../../src/contract/self-improve.ts'
-import { evaluatePowerFloorGate } from '../../../src/experiment/ast.ts'
+import { ValidationError } from '../../../src/errors.ts'
+import { evaluatePowerFloorGate } from '../../../src/experiment/ast/validity.ts'
 import { compareCodeUnits, hashCanonical } from '../../../src/ledger-core/canonical.ts'
 import { correlationStudy } from '../../../src/meta-eval/correlation-study.ts'
 import { InMemoryOutcomeStore } from '../../../src/meta-eval/outcome-store.ts'
@@ -21,7 +22,8 @@ import { TraceEmitter } from '../../../src/trace/emitter.ts'
 import { InMemoryTraceStore } from '../../../src/trace/store.ts'
 
 // These diagnostics record behavior without asserting that a defect must remain.
-const reviewedBaseRevision = 'fe1cc5111aab5d588bf7db3a3785325635937a91'
+const reviewedObservations = JSON.parse(await readFile(new URL('./observations.json', import.meta.url), 'utf8'))
+const reviewedBaseRevision = reviewedObservations.reviewedBaseRevision
 const command = 'pnpm exec tsx docs/design/mlbenchmarks-review/probes.mts'
 const repositoryRoot = fileURLToPath(new URL('../../../', import.meta.url))
 const sourcePaths = ['src', 'package.json', 'pnpm-lock.yaml', 'tsconfig.json']
@@ -223,6 +225,13 @@ async function adaptationPairing() {
   const b = await runAdaptationCurve<{ scenarioId: string; score: number }>({
     scenarios: scenariosB, ks, reps, runner,
   })
+  let comparison: ReturnType<typeof compareAdaptationCurves> | { refused: true; code: string; message: string }
+  try {
+    comparison = compareAdaptationCurves(a, b, { seed: 1 })
+  } catch (error) {
+    if (!(error instanceof ValidationError) || !error.message.includes('scenario pairs differ')) throw error
+    comparison = { refused: true, code: error.code, message: error.message }
+  }
   return {
     inputs: {
       scenariosA,
@@ -234,7 +243,7 @@ async function adaptationPairing() {
       sameRunnerForBothArms: true,
       bootstrapSeed: 1,
     },
-    results: compareAdaptationCurves(a, b, { seed: 1 }),
+    results: comparison,
     limitations: [
       'The two arms differ in task difficulty; zero task identities overlap.',
       'This probes the adaptation helper, not the separately implemented campaign paired comparison.',
@@ -278,6 +287,7 @@ async function negativeOutcomeDirection() {
   }))
   const runs: RunRecord[] = []
   const outcomes = new InMemoryOutcomeStore()
+  const targetOutcome = { id: 'success_rate', direction: 'higher-is-better' as const }
   for (const [i, row] of rows.entries()) {
     const runId = `direction-fixture-${i}`
     runs.push({
@@ -307,14 +317,14 @@ async function negativeOutcomeDirection() {
   const report = await rubricPredictiveValidity({
     runs,
     outcomes,
-    outcomeMetrics: ['success_rate'],
+    outcomeMetrics: [targetOutcome],
     rubrics: ['quality'],
     seed: 1,
     bootstrapResamples: 100,
   })
   const researcher = new PredictiveValidityResearcher({
     outcomes,
-    outcomeMetrics: ['success_rate'],
+    targetOutcome,
     rubrics: ['quality'],
   })
   const researcherReport = await researcher.runValidityCheck(runs)
@@ -325,8 +335,7 @@ async function negativeOutcomeDirection() {
       n: rows.length,
       rows,
       rubric: 'quality',
-      outcome: 'success_rate',
-      desiredOutcomeDirection: 'increase',
+      outcome: targetOutcome,
       seed: 1,
       bootstrapResamples: 100,
       researcherBootstrapResamples: 500,
@@ -347,9 +356,8 @@ async function negativeOutcomeDirection() {
       },
     },
     limitations: [
-      'Magnitude-based bucketing is intentional in existing tests, despite contradictory interface prose.',
-      'A negative association can be desirable for an outcome such as failure rate; direction needs explicit interpretation.',
-      'The researcher recommends increasing rubric weight despite its negative association with desired success rate; it does not execute or deploy that recommendation.',
+      'The desired outcome direction is explicit; association remains descriptive and does not authorize a rubric change.',
+      'The researcher reports recommendations; this diagnostic does not execute or deploy them.',
       'Constructed perfect correlation establishes neither causal validity nor held-out predictive performance.',
     ],
   }
@@ -411,6 +419,7 @@ function powerFloor() {
   const gate = {
     kind: 'power-floor' as const,
     target: 0.8,
+    minimumEffect: 0.01,
     effectGrid: [0.01, 1],
     sim: { trials: 1, resamples: 1, seed: 1 },
   }
@@ -419,14 +428,13 @@ function powerFloor() {
     inputs: {
       gate,
       curve,
-      practicalEffectForInterpretation: 0.01,
+      practicalEffectForInterpretation: gate.minimumEffect,
       curveSource: 'Supplied deterministic fixture; no power simulation is run.',
     },
     results: evaluatePowerFloorGate('floor', gate, curve),
     limitations: [
-      'The inspected gate documents a maximum-over-grid structural feasibility check; this output matches that contract.',
-      'Passing this gate does not establish target power at the practical effect of 0.01.',
-      'The fixture powers are inputs, not measured or simulated power estimates.',
+      'The gate compares power at minimumEffect with target; power 0.1 at effect 0.01 fails despite maximum power 1.',
+      'The fixture powers are supplied diagnostic inputs, not measured or simulated estimates.',
     ],
   }
 }
@@ -454,23 +462,7 @@ const observations = {
     sequentialDependence: await sequentialDependence(),
     powerFloor: powerFloor(),
   },
-  separateVerificationAtReviewedBase: {
-    provenance: 'Historical checks at reviewedBaseRevision; this diagnostic does not rerun them.',
-    sourceRevision: reviewedBaseRevision,
-    checks: [
-      { command: 'pnpm typecheck', result: 'passed' },
-      { command: 'pnpm build', result: 'passed' },
-      { command: 'pnpm verify:package', result: 'passed' },
-    ],
-    tests: {
-      command: 'pnpm test -- tests/experiment/preregistration-acceptance.test.ts tests/experiment/power.test.ts tests/contamination-guard.test.ts tests/rl-predictive-validity-researcher.test.ts tests/rubric-predictive-validity.test.ts tests/meta-eval.test.ts',
-      observedScope: 'The package command expanded to the full Vitest suite.',
-      files: { passed: 399, skipped: 2 },
-      tests: { passed: 5876, skipped: 3 },
-      result: 'passed',
-    },
-    limitation: 'Passing repository checks do not establish correctness of the counterexample behaviors recorded above.',
-  },
+  separateVerificationAtReviewedBase: reviewedObservations.separateVerificationAtReviewedBase,
 }
 
 console.log(JSON.stringify(observations, null, 2))

package/docs/experiment.md

@@ -21,7 +21,7 @@ Calling `registered.decide()` does not automatically run admission, power, budge
 
 | Object | Entry point | Purpose |
 | --- | --- | --- |
-| Registered rules | [Rule types](../src/experiment/ast.ts) and [`ExperimentSpec`](../src/experiment/define.ts) | Describe admission, estimation, intervals, and decisions as data. |
+| Registered rules | [Rule types](../src/experiment/ast/) and [`ExperimentSpec`](../src/experiment/define.ts) | Describe admission, estimation, intervals, and decisions as data. |
 | Define / seal / execute | `defineExperiment`, `sealExperiment`, `amendExperiment`, `openSealedExperiment` | Validate, identify, and execute the registered rules. |
 | Cluster-aware power | `clusteredPower`, `assertDesignAdequate` | Assess a declared effect under a simulated outcome model and cluster-count policy. |
 | Denominator chain | `buildFunnel`, `executeAdmissionRule`, `composeFunnels`, `renderFunnelTable` | Reconcile retained and excluded evidence. |