@sanity/ailf 4.6.0 → 6.0.0

package/dist/_vendor/ailf-core/types/grader-judgment.d.ts

@@ -0,0 +1,125 @@
+/**
+ * GraderJudgment core domain types — canonical shapes for structured
+ * grader output (Doc 03, GRAD-02).
+ *
+ * Authored INDEPENDENTLY of any Zod schema (D0045 doctrine — Plan 02's
+ * `GraderJudgmentSchema` `satisfies` against this type, not the other
+ * way around). A tautological `satisfies z.ZodType<z.infer<typeof
+ * GraderJudgmentSchema>>` is forbidden.
+ *
+ * Phase 1 retained the existing pipeline core (`taskId`, `modelId`,
+ * `dimension`, `reason`, `score`, `outputFailure?`) as required for
+ * backward compat with Phase 0 callers (Doc 03 §"existing, unchanged")
+ * and added the GRAD-02 additive fields (`judgmentId`, `subJudgments`,
+ * `docCitations`, `failureMode`, `confidence`,
+ * `hallucinationCheckedAgainst`, `metadata`) as additive in Phase 1;
+ * required from Phase 3 GRAD-05.
+ *
+ * Phase 3 GRAD-05 has flipped the additive fields to required (this
+ * file) and the corresponding Zod schema in
+ * `packages/eval/src/adapters/grader-outputs/promptfoo-grader-output.ts`
+ * is `.strict()` with `graderJudgmentsVersion = "1.0.0"`. The
+ * read-only legacy parser at `…/legacy/promptfoo-grader-output-legacy.ts`
+ * (against `LegacyGraderJudgment`) is the named consumer for already-
+ * stored historical reports through GRAD-06 cutover.
+ *
+ * @see docs/decisions/D0049-shared-confidence-contract.md
+ * @see docs/decisions/D0052-judgment-ref-granularity.md
+ * @see docs/design-docs/actionability-ladder/03-structured-grader-judgments.md
+ */
+import type { JudgmentId } from "./branded-ids.js";
+import type { Confidence } from "./confidence.js";
+/**
+ * Role enum for doc citations attached to a grader judgment (GRAD-02).
+ * Closed string-literal union — Phase 3 may extend.
+ */
+export type DocCitationRole = "supports" | "contradicts" | "missing" | "irrelevant";
+/**
+ * A single doc the grader cited while reasoning. `documentId` is the
+ * canonical D0052 reference (id, not slug); `slug` is a human-readable
+ * annotation only. `hallucinated` is set true at adapter time when the
+ * `slug` does not resolve against the task's `contextDocs` set.
+ */
+export interface DocCitation {
+    /** Canonical D0052 document ref (id, not slug). */
+    documentId: string;
+    /** Optional human-readable annotation. Never the identity. */
+    slug?: string;
+    role: DocCitationRole;
+    /** True when `slug` is not in the resolvable-set. */
+    hallucinated?: boolean;
+}
+/**
+ * Per-criterion sub-judgment — one entry per task-criterion bullet
+ * (Doc 03 §"per-criterion sub-judgments"). The `criterionId` is the
+ * stable identifier declared on the task's `criteria` array (Phase 2
+ * GRAD-01 schema-sync), not synthesized at grade time.
+ */
+export interface CriterionSubJudgment {
+    /** Stable criterion identifier — matches `CriterionRef.id` from the task definition (D0052). */
+    criterionId: string;
+    met: boolean;
+    /** ≤280 chars — quote or paraphrase. */
+    evidence: string;
+    /** Grader self-confidence on this single criterion (D0049). */
+    confidence: Confidence;
+}
+/**
+ * The structured grader judgment — Phase 3 GRAD-05 shape.
+ *
+ * Existing pipeline core (Doc 03 §"existing, unchanged"): `taskId`,
+ * `modelId`, `dimension`, `reason`, `score`. The pre-existing
+ * `outputFailure?` remains optional. The
+ * `contextDocs? (legacy alias: canonicalDocs)` annotation (StoredJudgment
+ * extension) lives on the storage extension type, not here.
+ *
+ * Additive in Phase 1; required from Phase 3 GRAD-05: `judgmentId`,
+ * `subJudgments`, `docCitations`, `failureMode`, `confidence`,
+ * `hallucinationCheckedAgainst`, `metadata`. The corresponding Zod
+ * schema in `packages/eval/src/adapters/grader-outputs/promptfoo-grader-output.ts`
+ * is `.strict()` with `graderJudgmentsVersion = "1.0.0"`.
+ */
+export interface GraderJudgment {
+    /** Rubric template name (e.g. "task-completion", "code-correctness"). */
+    dimension: string;
+    /** The model that produced the response being graded. */
+    modelId: string;
+    /**
+     * True when the model failed to produce meaningful output (empty
+     * response, API error, or refusal). Distinguishes infrastructure
+     * failures from genuinely incorrect responses — a score of 0 from no
+     * output is fundamentally different from a score of 0 from wrong
+     * output.
+     */
+    outputFailure?: boolean;
+    /** The grader's natural-language reasoning. */
+    reason: string;
+    /** Numeric score in [0, 100] (normalized). */
+    score: number;
+    /** The task this judgment belongs to. */
+    taskId: string;
+    /**
+     * D0052 granular branded id. Required from Phase 3 GRAD-05 — every
+     * grader emission carries one.
+     */
+    judgmentId: JudgmentId;
+    /** Per-criterion sub-judgments. */
+    subJudgments: CriterionSubJudgment[];
+    /** Doc citations with role + hallucinated flag. */
+    docCitations: DocCitation[];
+    /**
+     * Per-dimension failure mode. Phase 3 GRAD-03 stamps the taxonomy
+     * literal at the runtime grader-prompt; the value is a free-form
+     * string for forward compat with future taxonomy extensions.
+     */
+    failureMode: string;
+    /** Grader self-confidence per D0049. */
+    confidence: Confidence;
+    /** Hallucination cross-check (Pitfall #11) — union of task.context.docs and run.documentManifest. */
+    hallucinationCheckedAgainst: string[];
+    /** Metadata about the grader run. */
+    metadata: {
+        graderModel: string;
+        graderJudgmentsVersion: string;
+    };
+}

package/dist/_vendor/ailf-core/types/grader-judgment.js

@@ -0,0 +1,30 @@
+/**
+ * GraderJudgment core domain types — canonical shapes for structured
+ * grader output (Doc 03, GRAD-02).
+ *
+ * Authored INDEPENDENTLY of any Zod schema (D0045 doctrine — Plan 02's
+ * `GraderJudgmentSchema` `satisfies` against this type, not the other
+ * way around). A tautological `satisfies z.ZodType<z.infer<typeof
+ * GraderJudgmentSchema>>` is forbidden.
+ *
+ * Phase 1 retained the existing pipeline core (`taskId`, `modelId`,
+ * `dimension`, `reason`, `score`, `outputFailure?`) as required for
+ * backward compat with Phase 0 callers (Doc 03 §"existing, unchanged")
+ * and added the GRAD-02 additive fields (`judgmentId`, `subJudgments`,
+ * `docCitations`, `failureMode`, `confidence`,
+ * `hallucinationCheckedAgainst`, `metadata`) as additive in Phase 1;
+ * required from Phase 3 GRAD-05.
+ *
+ * Phase 3 GRAD-05 has flipped the additive fields to required (this
+ * file) and the corresponding Zod schema in
+ * `packages/eval/src/adapters/grader-outputs/promptfoo-grader-output.ts`
+ * is `.strict()` with `graderJudgmentsVersion = "1.0.0"`. The
+ * read-only legacy parser at `…/legacy/promptfoo-grader-output-legacy.ts`
+ * (against `LegacyGraderJudgment`) is the named consumer for already-
+ * stored historical reports through GRAD-06 cutover.
+ *
+ * @see docs/decisions/D0049-shared-confidence-contract.md
+ * @see docs/decisions/D0052-judgment-ref-granularity.md
+ * @see docs/design-docs/actionability-ladder/03-structured-grader-judgments.md
+ */
+export {};

package/dist/_vendor/ailf-core/types/index.d.ts

@@ -13,6 +13,7 @@ import type { DocumentRef as _DocumentRef, EvalMode, RunContext } from "../../ai
 import type { ArtifactType } from "../artifact-registry.js";
 import type { SymbolPreflightReport } from "./symbol-preflight-report.js";
 import type { AssociationValues, RunId } from "./branded-ids.js";
+import type { GraderJudgment } from "./grader-judgment.js";
 export type { ActualScoreEntry, ComponentResult, TestResult, UrlMetadata, } from "./scoring-input.js";
 export type { DocumentRef, RunContext, RunTrigger } from "../../ailf-shared/index.d.ts";
 export type { StoredBaseline, StoredReport, StoredRun, StoredTaskResult, StoredTrace, SchemaVersioned, } from "./storage-schema.js";
@@ -32,9 +33,13 @@ export type { SymbolPreflightDeduction, SymbolPreflightFinding, SymbolPreflightR
 export { DEFAULT_PREFLIGHT_CODE_CORRECTNESS_WEIGHT, type PreflightRubricContext, type PreflightScoringConfig, } from "./preflight-scoring.js";
 export type { Confidence, ConfidenceDerivation } from "./confidence.js";
 export { CONVENTIONAL_DERIVATIONS, isConfidence } from "./confidence.js";
-export type { ArtifactId, AssociationAxis, AssociationValues, Brand, EntryKey, Err, FixtureId, IdValidationError, NewReportId, Ok, ProviderId, PromptId, Result, ResultId, RubricId, RunFingerprint, RunId, SuiteId, TaskId, TaskSlug, TraceId, } from "./branded-ids.js";
-export { err, fixtureId, generateRunId, ok, providerId, resultId, runId, suiteId, taskId, traceId, } from "./branded-ids.js";
-export type { AgentHarnessTaskDefinition, ContentLakeAuthorableMode, ContentLakeAuthorableTask, CustomTaskDefinition, GeneralizedAssertionDefinition, GeneralizedDocRef, GeneralizedTaskDefinition, GeneralizedTemplatedAssertion, GeneralizedValueAssertion, IdDocRef, KnowledgeProbeTaskDefinition, LiteracyTaskDefinition, MCPServerTaskDefinition, PromptVars, PathDocRef, PerspectiveDocRef, ReservedPromptVarKey, RubricRef, SlugDocRef, TaskCommonFields, TaskDifficulty, TaskOptions, TaskProviderConfig, TaskStatus, } from "./generalized-task.js";
+export type { ArtifactId, AssociationAxis, AssociationValues, Brand, EntryKey, Err, FixtureId, IdValidationError, JudgmentId, NewReportId, Ok, ProviderId, PromptId, Result, ResultId, RubricId, RunFingerprint, RunId, SuiteId, TaskId, TaskSlug, TraceId, } from "./branded-ids.js";
+export { err, fixtureId, generateJudgmentId, generateRunId, judgmentId, ok, providerId, resultId, runId, suiteId, taskId, traceId, } from "./branded-ids.js";
+export type { AgentHarnessTaskDefinition, ContentLakeAuthorableMode, ContentLakeAuthorableTask, CriterionRef, CustomTaskDefinition, GeneralizedAssertionDefinition, GeneralizedDocRef, GeneralizedTaskDefinition, GeneralizedTemplatedAssertion, GeneralizedValueAssertion, IdDocRef, KnowledgeProbeTaskDefinition, LiteracyTaskDefinition, MCPServerTaskDefinition, PromptVars, PathDocRef, PerspectiveDocRef, ReservedPromptVarKey, RubricRef, SlugDocRef, TaskCommonFields, TaskDifficulty, TaskOptions, TaskProviderConfig, TaskStatus, } from "./generalized-task.js";
+export type { ActionSuggestion, AreaSummaryBody, CardMeta, CardType, Diagnosis, DiagnosisCard, DocAttributionSpotlightBody, FailureModeSummaryBody, JudgmentRef, LowConfidenceAttributionBody, NoIssuesBody, RegressionVsBaselineBody, TopRecommendationsBody, VersionedInputs, WeakestAreaBody, } from "./diagnosis.js";
+export type { AttributionMeta, DocAttribution, JudgmentAttribution, } from "./attribution.js";
+export type { CriterionSubJudgment, DocCitation, DocCitationRole, GraderJudgment, } from "./grader-judgment.js";
+export type { LegacyGraderJudgment } from "./legacy-grader-judgment.js";
 type DocumentRef = _DocumentRef;
 /** Aggregated retrieval metrics for a feature area */
 export interface AreaRetrievalMetrics {
@@ -128,8 +133,31 @@ export interface FailureModeReport {
     /** Total judgments analyzed */
     totalJudgments: number;
 }
-/** Failure mode classification for a low-scoring judgment */
-export type FailureModeType = "api-error" | "incorrect-docs" | "missing-docs" | "model-limitation" | "outdated-docs" | "poor-structure" | "unclassified";
+/**
+ * Failure mode classification for a low-scoring judgment.
+ *
+ * Open-set string (Plan 03-02 per-dimension taxonomies introduce modes
+ * outside the original literacy enum: `false-floor`, `spec-mismatch`,
+ * `tool-misuse`, `factual-error`, `hallucination`, etc. — the grader
+ * is told these are legal answers via the rubric prompt). The legacy
+ * literacy enum survives as `LegacyFailureModeType` for the
+ * report-aggregation helpers that need stable bucket ordering and
+ * icon tables; consumers that only care about presence/absence treat
+ * `FailureModeType` as `string`.
+ */
+export type FailureModeType = string;
+/**
+ * Closed enum of the original literacy failure modes — used by the
+ * report formatters that iterate buckets in a stable order. Adding to
+ * this list is a deliberate extension; modes outside it still flow
+ * through the report (per-area `modes` record), just without a
+ * pre-allocated bucket in `summary`.
+ */
+export type LegacyFailureModeType = "api-error" | "incorrect-docs" | "missing-docs" | "model-limitation" | "outdated-docs" | "poor-structure" | "unclassified";
+/** Set of canonical legacy modes — exported for report-formatter use. */
+export declare const LEGACY_FAILURE_MODES: readonly LegacyFailureModeType[];
+/** Type guard for legacy modes. */
+export declare function isLegacyFailureMode(mode: string): mode is LegacyFailureModeType;
 /** Per-feature-area score breakdown */
 export interface FeatureScore {
     /**
@@ -261,30 +289,16 @@ export interface GapEstimate {
     /** Specific remediation description */
     remediation: string;
 }
-/** A single grader judgment — one per assertion per test */
-export interface GraderJudgment {
-    /** The rubric template used (task-completion, code-correctness, doc-coverage) */
-    dimension: string;
-    /** The model that produced the response being graded */
-    modelId: string;
+/** Enriched grader judgment with stored documentation context refs. */
+export interface StoredJudgment extends GraderJudgment {
     /**
-     * True when the model failed to produce meaningful output (empty response,
-     * API error, or refusal). Distinguishes infrastructure failures from
-     * genuinely incorrect responses — a score of 0 from no output is
-     * fundamentally different from a score of 0 from wrong output.
+     * Documentation context the task expected the model to use.
+     *
+     * Legacy alias `canonicalDocs` may appear on stored reports written
+     * before Phase 2 — readers should tolerate both. Writers (the pipeline)
+     * always emit `contextDocs`.
      */
-    outputFailure?: boolean;
-    /** The grader's natural language reasoning */
-    reason: string;
-    /** The numeric score (0–100) */
-    score: number;
-    /** The task this judgment belongs to */
-    taskId: string;
-}
-/** Enriched grader judgment with canonical doc references, stored in reports */
-export interface StoredJudgment extends GraderJudgment {
-    /** Canonical docs that the task expected the model to use */
-    canonicalDocs?: DocumentRef[];
+    contextDocs?: DocumentRef[];
 }
 /**
  * Per-test result stored in reports for drill-down and audit.
@@ -296,8 +310,11 @@ export interface StoredJudgment extends GraderJudgment {
 export interface StoredTestResult {
     /** Resolved feature area (from __featureArea or description) */
     area: string;
-    /** Canonical docs the task expected the model to use */
-    canonicalDocs?: DocumentRef[];
+    /**
+     * Documentation context the task expected the model to use.
+     * Legacy alias `canonicalDocs` may appear on pre-Phase-2 reports.
+     */
+    contextDocs?: DocumentRef[];
     /** Weighted composite score (gold variant only) */
     compositeScore?: number;
     /** Per-test cost (USD) */
@@ -349,6 +366,40 @@ export interface StoredTestResult {
 }
 /** Grader consistency diagnostics — does not affect scores, reported alongside */
 export interface GraderReliability {
+    /**
+     * Plan 03-03 — count of grader-emission vs ceiling cross-check disagreements.
+     *
+     * Incremented by the live pipeline when `validateFailureMode(...)` returns
+     * `level: "medium"` (the grader's emitted `failureMode` does not agree with
+     * the ceiling-decomposition mode). Surfaces calibration drift over time
+     * without affecting scores. Optional — undefined when the run did not
+     * exercise the failure-mode validator (e.g., grader-consistency-only paths).
+     *
+     * @see docs/decisions/D0049-shared-confidence-contract.md
+     */
+    failureModeCalibration?: number;
+    /**
+     * Plan 03-03 — count of strict-schema parse failures during grader-output
+     * extraction. Wired at the parse-fail branch in `extractGraderJudgments`;
+     * incremented when `GraderJudgmentSchema.safeParse` rejects a payload and
+     * the pipeline drops to the Phase 1 minimal-shape fallback.
+     *
+     * Plan 03-04 will tighten the strict schema (`.strict()` + GRAD-02 fields
+     * required) and graders will emit the structured wire format in earnest;
+     * this counter measures pre-hard-fail drift.
+     */
+    parseFailures?: number;
+    /**
+     * Phase 4 ATTR-01 — count of grader citations whose `slug` was not
+     * in the resolvable-set (`hallucinationCheckedAgainst`).
+     * Incremented by `computeJudgmentAttribution(...)` for every
+     * citation that fails the hallucination short-circuit (Success
+     * Criterion #5). A counter, not a ratio — consumers compute the
+     * rate by dividing by total-citations if needed.
+     *
+     * @see docs/decisions/D0049-shared-confidence-contract.md
+     */
+    hallucinationCount?: number;
     /** Inter-grader agreement (from multi-grader comparison) — Phase 3 */
     agreement?: {
         /** Models compared against the primary grader */

package/dist/_vendor/ailf-core/types/index.js

@@ -18,7 +18,21 @@ export { InMemoryPluginRegistry } from "./plugin-registry.js";
 export { evalModeType } from "./eval-mode-config.js";
 export { DEFAULT_PREFLIGHT_CODE_CORRECTNESS_WEIGHT, } from "./preflight-scoring.js";
 export { CONVENTIONAL_DERIVATIONS, isConfidence } from "./confidence.js";
-export { err, fixtureId, generateRunId, ok, providerId, resultId, runId, suiteId, taskId, traceId, } from "./branded-ids.js";
+export { err, fixtureId, generateJudgmentId, generateRunId, judgmentId, ok, providerId, resultId, runId, suiteId, taskId, traceId, } from "./branded-ids.js";
+/** Set of canonical legacy modes — exported for report-formatter use. */
+export const LEGACY_FAILURE_MODES = [
+    "api-error",
+    "incorrect-docs",
+    "missing-docs",
+    "model-limitation",
+    "outdated-docs",
+    "poor-structure",
+    "unclassified",
+];
+/** Type guard for legacy modes. */
+export function isLegacyFailureMode(mode) {
+    return LEGACY_FAILURE_MODES.includes(mode);
+}
 // ---------------------------------------------------------------------------
 // Comparison (Approach 2: structured comparison output)
 // ---------------------------------------------------------------------------

package/dist/_vendor/ailf-core/types/legacy-grader-judgment.d.ts

@@ -0,0 +1,55 @@
+/**
+ * LegacyGraderJudgment — Phase 1 superset core only, with NO GRAD-02
+ * additive surface. Used by the read-only legacy parser at
+ * `packages/eval/src/adapters/grader-outputs/legacy/` for historical
+ * pre-Phase-3 reports.
+ *
+ * Reports are immutable events — once a Report is written to Content
+ * Lake the structured grader-judgment shape it captures cannot be
+ * back-filled. The legacy parser exists so historical-report rendering
+ * paths can keep deserializing pre-Phase-3 output through Phase 7
+ * (GRAD-06 cutover removes Studio's `reason`-only fallback rendering
+ * paths and the legacy adapter alongside).
+ *
+ * Authored INDEPENDENTLY of any Zod schema (D0045 doctrine — the
+ * legacy schema in
+ * `packages/eval/src/adapters/grader-outputs/legacy/promptfoo-grader-output-legacy.ts`
+ * `satisfies z.ZodType<LegacyGraderJudgment>` against this type, not
+ * the other way around). A tautological
+ * `satisfies z.ZodType<z.infer<typeof Schema>>` is forbidden.
+ *
+ * Invariant — live grader output that fails the strict
+ * `GraderJudgmentSchema` MUST NOT fall back to this schema. Drop to
+ * `failureMode: "unclassified"` instead. The legacy parser is invoked
+ * ONLY by historical-report rendering paths.
+ *
+ * @see ./grader-judgment.ts — the Phase 1+ structured shape (live path)
+ * @see docs/decisions/D0045-type-architecture-and-contract-enforcement.md
+ * @see docs/design-docs/actionability-ladder/03-structured-grader-judgments.md
+ */
+/**
+ * The Phase 1 free-prose grader judgment as historical reports captured
+ * it. Mirrors the existing-pipeline-core surface of {@link GraderJudgment}
+ * (the required fields) and the pre-existing optional `outputFailure`
+ * flag. NO GRAD-02 additive fields (`subJudgments`, `docCitations`,
+ * `failureMode`, `confidence`, `hallucinationCheckedAgainst`,
+ * `metadata`) — those are by construction absent on pre-Phase-3 output.
+ */
+export interface LegacyGraderJudgment {
+    /** Rubric template name (e.g. "task-completion", "code-correctness"). */
+    dimension: string;
+    /** The model that produced the response being graded. */
+    modelId: string;
+    /**
+     * True when the model failed to produce meaningful output (empty
+     * response, API error, or refusal). Same semantics as
+     * {@link GraderJudgment.outputFailure}.
+     */
+    outputFailure?: boolean;
+    /** The grader's natural-language reasoning (free-prose Phase 1 shape). */
+    reason: string;
+    /** Numeric score in [0, 100] (normalized). */
+    score: number;
+    /** The task this judgment belongs to. */
+    taskId: string;
+}

package/dist/_vendor/ailf-core/types/legacy-grader-judgment.js

@@ -0,0 +1,30 @@
+/**
+ * LegacyGraderJudgment — Phase 1 superset core only, with NO GRAD-02
+ * additive surface. Used by the read-only legacy parser at
+ * `packages/eval/src/adapters/grader-outputs/legacy/` for historical
+ * pre-Phase-3 reports.
+ *
+ * Reports are immutable events — once a Report is written to Content
+ * Lake the structured grader-judgment shape it captures cannot be
+ * back-filled. The legacy parser exists so historical-report rendering
+ * paths can keep deserializing pre-Phase-3 output through Phase 7
+ * (GRAD-06 cutover removes Studio's `reason`-only fallback rendering
+ * paths and the legacy adapter alongside).
+ *
+ * Authored INDEPENDENTLY of any Zod schema (D0045 doctrine — the
+ * legacy schema in
+ * `packages/eval/src/adapters/grader-outputs/legacy/promptfoo-grader-output-legacy.ts`
+ * `satisfies z.ZodType<LegacyGraderJudgment>` against this type, not
+ * the other way around). A tautological
+ * `satisfies z.ZodType<z.infer<typeof Schema>>` is forbidden.
+ *
+ * Invariant — live grader output that fails the strict
+ * `GraderJudgmentSchema` MUST NOT fall back to this schema. Drop to
+ * `failureMode: "unclassified"` instead. The legacy parser is invoked
+ * ONLY by historical-report rendering paths.
+ *
+ * @see ./grader-judgment.ts — the Phase 1+ structured shape (live path)
+ * @see docs/decisions/D0045-type-architecture-and-contract-enforcement.md
+ * @see docs/design-docs/actionability-ladder/03-structured-grader-judgments.md
+ */
+export {};

package/dist/_vendor/ailf-core/types/pipeline-request.d.ts

@@ -84,6 +84,7 @@ export interface PipelineRequest {
     dataset?: string;
     debug?: PipelineRequestDebug | boolean;
     executor?: PipelineRequestCallerExecutor;
+    borderlineReplications?: number;
     gapAnalysis?: boolean;
     graderContext?: "rubric-only" | "with-docs";
     graderReplications?: number;

package/dist/_vendor/ailf-core/types/repo-config.d.ts

@@ -47,6 +47,14 @@ export interface RepoPublishConfig {
 /** Execution-tier knobs — replaces the retired `--concurrency` / `--api-url` flags. */
 export interface RepoExecutionConfig {
     apiUrl?: string;
+    /**
+     * Plan 03-04 GRAD-04 — replications per borderline judgment for the
+     * intra-grader consensus pass. Default 3 (set in composition-root).
+     * A judgment is "borderline" when its score lies within ±5 of any
+     * severity boundary (30/50/60). Non-borderline judgments are not
+     * re-graded.
+     */
+    borderlineReplications?: number;
     concurrency?: number;
     gapAnalysis?: boolean;
     graderReplications?: number;

package/dist/_vendor/ailf-shared/document-ref.d.ts

@@ -8,7 +8,7 @@
  * Attachable at every level of the report hierarchy:
  * - ScoreSummary.documentManifest — all docs used in the evaluation
  * - FeatureScore.documents — docs used for a specific area
- * - StoredJudgment.canonicalDocs — docs expected for a specific task
+ * - StoredJudgment.contextDocs (legacy alias: canonicalDocs) — docs expected for a specific task
  */
 export interface DocumentRef {
     /**

package/dist/adapters/api-client/build-request.d.ts

@@ -59,6 +59,7 @@ export interface RemoteConfigSlice {
     perspectiveOverride?: string;
     graderContext?: "rubric-only" | "with-docs";
     graderReplications?: number;
+    borderlineReplications?: number;
     gapAnalysisEnabled?: boolean;
     noRemoteCache?: boolean;
     /**

package/dist/adapters/api-client/build-request.js

@@ -130,6 +130,9 @@ export async function buildRemoteRequest(options) {
     if (config.graderReplications) {
         raw.graderReplications = config.graderReplications;
     }
+    if (config.borderlineReplications) {
+        raw.borderlineReplications = config.borderlineReplications;
+    }
     if (config.gapAnalysisEnabled)
         raw.gapAnalysis = true;
     if (config.noRemoteCache)

package/dist/adapters/attribution/attribution-meta-writer.d.ts

@@ -0,0 +1,35 @@
+/**
+ * attribution-meta-writer.ts — Zod schema for the run-scoped
+ * attribution metadata artifact (ATTR-01) emitted by Phase 4 and read
+ * back alongside the per-entry attribution objects.
+ *
+ * The schema asserts `satisfies z.ZodType<AttributionMeta>` against the
+ * canonical domain type in `packages/core/src/types/attribution.ts`
+ * (D0045 / W0187) — drift is a build error.
+ *
+ * `embeddingModel` is REQUIRED (Pitfall #6): silently downgrading to a
+ * default has caused regressions in adjacent codebases — model swaps
+ * MUST invalidate cached weights.
+ *
+ * Phase 1 lands the SHAPE only — no compute, no file I/O.
+ *
+ * @see docs/decisions/D0045-type-architecture-and-contract-enforcement.md
+ * @see docs/design-docs/actionability-ladder/04-per-document-attribution-ensemble.md
+ */
+import { z } from "zod";
+/**
+ * Canonical schema for {@link AttributionMeta}. Persisted at
+ * `runs/{runId}/attribution/_meta.json` (or whatever bulk path the
+ * Phase 4 descriptor pins) and parsed on read.
+ */
+export declare const AttributionMetaSchema: z.ZodObject<{
+    ensembleVersion: z.ZodString;
+    embeddingModel: z.ZodString;
+    calibrationSetVersion: z.ZodOptional<z.ZodString>;
+    weights: z.ZodObject<{
+        citation: z.ZodNumber;
+        canonical: z.ZodNumber;
+        retrieved: z.ZodNumber;
+    }, z.core.$strip>;
+}, z.core.$strip>;
+export type { AttributionMeta } from "../../_vendor/ailf-core/index.d.ts";

package/dist/adapters/attribution/attribution-meta-writer.js

@@ -0,0 +1,34 @@
+/**
+ * attribution-meta-writer.ts — Zod schema for the run-scoped
+ * attribution metadata artifact (ATTR-01) emitted by Phase 4 and read
+ * back alongside the per-entry attribution objects.
+ *
+ * The schema asserts `satisfies z.ZodType<AttributionMeta>` against the
+ * canonical domain type in `packages/core/src/types/attribution.ts`
+ * (D0045 / W0187) — drift is a build error.
+ *
+ * `embeddingModel` is REQUIRED (Pitfall #6): silently downgrading to a
+ * default has caused regressions in adjacent codebases — model swaps
+ * MUST invalidate cached weights.
+ *
+ * Phase 1 lands the SHAPE only — no compute, no file I/O.
+ *
+ * @see docs/decisions/D0045-type-architecture-and-contract-enforcement.md
+ * @see docs/design-docs/actionability-ladder/04-per-document-attribution-ensemble.md
+ */
+import { z } from "zod";
+/**
+ * Canonical schema for {@link AttributionMeta}. Persisted at
+ * `runs/{runId}/attribution/_meta.json` (or whatever bulk path the
+ * Phase 4 descriptor pins) and parsed on read.
+ */
+export const AttributionMetaSchema = z.object({
+    ensembleVersion: z.string().min(1),
+    embeddingModel: z.string().min(1),
+    calibrationSetVersion: z.string().optional(),
+    weights: z.object({
+        citation: z.number(),
+        canonical: z.number(),
+        retrieved: z.number(),
+    }),
+});

package/dist/adapters/attribution/index.d.ts

@@ -0,0 +1,9 @@
+/**
+ * attribution adapter barrel — named re-exports only (W0124 / D0045).
+ *
+ * The attribution schemas live here so they enter the D0045
+ * `pnpm check-trust-boundary-satisfies` SCAN_ROOTS gate.
+ */
+export { JudgmentAttributionSchema } from "./per-entry-attribution-writer.js";
+export { AttributionMetaSchema } from "./attribution-meta-writer.js";
+export type { AttributionMeta, DocAttribution, JudgmentAttribution, } from "../../_vendor/ailf-core/index.d.ts";

package/dist/adapters/attribution/index.js

@@ -0,0 +1,8 @@
+/**
+ * attribution adapter barrel — named re-exports only (W0124 / D0045).
+ *
+ * The attribution schemas live here so they enter the D0045
+ * `pnpm check-trust-boundary-satisfies` SCAN_ROOTS gate.
+ */
+export { JudgmentAttributionSchema } from "./per-entry-attribution-writer.js";
+export { AttributionMetaSchema } from "./attribution-meta-writer.js";

package/dist/adapters/attribution/per-entry-attribution-writer.d.ts

@@ -0,0 +1,56 @@
+/**
+ * per-entry-attribution-writer.ts — Zod schema for the per-judgment
+ * attribution artifact (ATTR-01) emitted by Phase 4's
+ * `ComputeAttributionStep` and read back by Phase 5's diagnosis runner.
+ *
+ * The schema asserts `satisfies z.ZodType<JudgmentAttribution>` against
+ * the canonical domain type in `packages/core/src/types/attribution.ts`
+ * (D0045 / W0187) — drift between schema and type is a build error.
+ *
+ * Phase 1 lands the SHAPE only — no compute, no file I/O. Phase 4 wires
+ * the writer; Phase 5 wires the reader. Both `satisfies` against this
+ * single source-of-truth schema.
+ *
+ * `hallucinationCheckedAgainst` is REQUIRED (Pitfall #11): consumers
+ * must be able to audit citation grounding without re-deriving the
+ * resolvable-set. The canonical task field is `contextDocs`; do NOT
+ * invent `expectedDocs` / `usedDocs` synonyms.
+ *
+ * @see docs/decisions/D0045-type-architecture-and-contract-enforcement.md
+ * @see docs/decisions/D0049-shared-confidence-contract.md
+ * @see docs/decisions/D0052-judgment-ref-granularity.md
+ * @see docs/design-docs/actionability-ladder/04-per-document-attribution-ensemble.md
+ */
+import { z } from "zod";
+/**
+ * Canonical schema for {@link JudgmentAttribution}. Persisted at
+ * `runs/{runId}/attribution/{entryKey}.json` (Phase 4) and parsed by
+ * the diagnosis runner on read (Phase 5).
+ */
+export declare const JudgmentAttributionSchema: z.ZodObject<{
+    judgmentRef: z.ZodString;
+    taskId: z.ZodString;
+    modelId: z.ZodString;
+    dimension: z.ZodString;
+    attributions: z.ZodArray<z.ZodObject<{
+        documentId: z.ZodString;
+        slug: z.ZodOptional<z.ZodString>;
+        score: z.ZodNumber;
+        signals: z.ZodObject<{
+            citation: z.ZodOptional<z.ZodNumber>;
+            canonical: z.ZodOptional<z.ZodNumber>;
+            retrieved: z.ZodOptional<z.ZodNumber>;
+        }, z.core.$strip>;
+        confidence: z.ZodObject<{
+            level: z.ZodEnum<{
+                low: "low";
+                medium: "medium";
+                high: "high";
+            }>;
+            signalsPresent: z.ZodNumber;
+            derivation: z.ZodString;
+        }, z.core.$strip>;
+    }, z.core.$strip>>;
+    hallucinationCheckedAgainst: z.ZodArray<z.ZodString>;
+}, z.core.$strip>;
+export type { DocAttribution, JudgmentAttribution } from "../../_vendor/ailf-core/index.d.ts";