@sanity/ailf 4.6.0 → 5.0.0

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";

package/dist/adapters/attribution/per-entry-attribution-writer.js

@@ -0,0 +1,49 @@
+/**
+ * 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";
+import { ConfidenceSchema } from "../../_vendor/ailf-core/schemas/index.js";
+const DocAttributionSchema = z.object({
+    documentId: z.string().min(1),
+    slug: z.string().optional(),
+    score: z.number().min(0).max(1),
+    signals: z.object({
+        citation: z.number().min(0).max(1).optional(),
+        canonical: z.number().min(0).max(1).optional(),
+        retrieved: z.number().min(0).max(1).optional(),
+    }),
+    confidence: ConfidenceSchema,
+});
+/**
+ * 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 const JudgmentAttributionSchema = z.object({
+    judgmentRef: z.string().min(1),
+    taskId: z.string().min(1),
+    modelId: z.string().min(1),
+    dimension: z.string().min(1),
+    attributions: z.array(DocAttributionSchema),
+    hallucinationCheckedAgainst: z.array(z.string()),
+});

package/dist/adapters/config-sources/file-config-adapter.js

@@ -125,6 +125,7 @@ function mapEvalConfigToResolvedConfig(config, rootDir) {
         noCache: config.noCache ?? false,
         noRemoteCache: config.noRemoteCache ?? false,
         graderReplications: config.execution?.graderReplications,
+        borderlineReplications: config.execution?.borderlineReplications,
         graderContext: config.grader?.context,
         urls: config.urls,
         headers: config.agentic?.headers,

package/dist/adapters/grader-outputs/index.d.ts

@@ -0,0 +1,10 @@
+/**
+ * grader-outputs adapter barrel — named re-exports only (W0124 / D0045).
+ *
+ * The grader-output schema lives here so it enters the D0045
+ * `pnpm check-trust-boundary-satisfies` SCAN_ROOTS gate.
+ */
+export { GraderJudgmentSchema, graderJudgmentsVersion, } from "./promptfoo-grader-output.js";
+export type { GraderJudgment } from "../../_vendor/ailf-core/index.d.ts";
+export { LegacyGraderJudgmentSchema } from "./legacy/index.js";
+export type { LegacyGraderJudgment } from "../../_vendor/ailf-core/index.d.ts";

package/dist/adapters/grader-outputs/index.js

@@ -0,0 +1,8 @@
+/**
+ * grader-outputs adapter barrel — named re-exports only (W0124 / D0045).
+ *
+ * The grader-output schema lives here so it enters the D0045
+ * `pnpm check-trust-boundary-satisfies` SCAN_ROOTS gate.
+ */
+export { GraderJudgmentSchema, graderJudgmentsVersion, } from "./promptfoo-grader-output.js";
+export { LegacyGraderJudgmentSchema } from "./legacy/index.js";

package/dist/adapters/grader-outputs/legacy/index.d.ts

@@ -0,0 +1,11 @@
+/**
+ * legacy grader-outputs adapter sub-barrel — named re-exports only
+ * (W0124 / D0045).
+ *
+ * Read-only schema for the Phase 1 free-prose grader-output shape,
+ * invoked only by historical-report rendering paths through Phase 7
+ * (GRAD-06 cutover). The schema lives here so it enters the D0045
+ * `pnpm check-trust-boundary-satisfies` SCAN_ROOTS gate.
+ */
+export { LegacyGraderJudgmentSchema } from "./promptfoo-grader-output-legacy.js";
+export type { LegacyGraderJudgment } from "../../../_vendor/ailf-core/index.d.ts";

package/dist/adapters/grader-outputs/legacy/index.js

@@ -0,0 +1,10 @@
+/**
+ * legacy grader-outputs adapter sub-barrel — named re-exports only
+ * (W0124 / D0045).
+ *
+ * Read-only schema for the Phase 1 free-prose grader-output shape,
+ * invoked only by historical-report rendering paths through Phase 7
+ * (GRAD-06 cutover). The schema lives here so it enters the D0045
+ * `pnpm check-trust-boundary-satisfies` SCAN_ROOTS gate.
+ */
+export { LegacyGraderJudgmentSchema } from "./promptfoo-grader-output-legacy.js";

package/dist/adapters/grader-outputs/legacy/promptfoo-grader-output-legacy.d.ts

@@ -0,0 +1,49 @@
+/**
+ * promptfoo-grader-output-legacy.ts — Zod schema for the Phase 1
+ * free-prose grader-output shape, used by historical-report rendering
+ * paths.
+ *
+ * READ-ONLY: invoked only by historical-report rendering paths through
+ * Phase 7 (GRAD-06 cutover). 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 schema exists so
+ * pre-Phase-3 reports continue to deserialize cleanly.
+ *
+ * Live grader output that fails the strict {@link GraderJudgmentSchema}
+ * parse must NOT fall back to this schema. Drop to
+ * `failureMode: "unclassified"` instead. Strict and legacy schemas are
+ * deliberate siblings, not a legacy/canonical pair to consolidate.
+ *
+ * The schema asserts `satisfies z.ZodType<LegacyGraderJudgment>` against
+ * the canonical domain type in
+ * `packages/core/src/types/legacy-grader-judgment.ts` (D0045 / W0187) —
+ * drift between schema and type is a build error. The domain type is
+ * authored independently in `@sanity/ailf-core`; this file authors ONLY
+ * the schema and never derives the domain type from the schema itself
+ * (no schema-derived self-reference allowed by D0045).
+ *
+ * @see docs/decisions/D0045-type-architecture-and-contract-enforcement.md
+ * @see ../promptfoo-grader-output.ts — the strict (live-path) sibling
+ * @see docs/design-docs/actionability-ladder/03-structured-grader-judgments.md
+ *      §"Backwards compatibility"
+ */
+import { z } from "zod";
+/**
+ * Canonical schema for {@link LegacyGraderJudgment}. Mirrors the Phase 1
+ * superset core (`taskId`, `modelId`, `dimension`, `reason`, `score`,
+ * optional `outputFailure`). NO GRAD-02 additive fields — those are by
+ * construction absent on pre-Phase-3 output.
+ *
+ * Intentionally NOT `.strict()` — pre-Phase-3 reports may carry stray
+ * keys; the legacy parser tolerates them so historical-report rendering
+ * keeps working through the GRAD-06 cutover.
+ */
+export declare const LegacyGraderJudgmentSchema: z.ZodObject<{
+    taskId: z.ZodString;
+    modelId: z.ZodString;
+    dimension: z.ZodString;
+    reason: z.ZodString;
+    score: z.ZodNumber;
+    outputFailure: z.ZodOptional<z.ZodBoolean>;
+}, z.core.$strip>;
+export type { LegacyGraderJudgment } from "../../../_vendor/ailf-core/index.d.ts";

package/dist/adapters/grader-outputs/legacy/promptfoo-grader-output-legacy.js

@@ -0,0 +1,48 @@
+/**
+ * promptfoo-grader-output-legacy.ts — Zod schema for the Phase 1
+ * free-prose grader-output shape, used by historical-report rendering
+ * paths.
+ *
+ * READ-ONLY: invoked only by historical-report rendering paths through
+ * Phase 7 (GRAD-06 cutover). 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 schema exists so
+ * pre-Phase-3 reports continue to deserialize cleanly.
+ *
+ * Live grader output that fails the strict {@link GraderJudgmentSchema}
+ * parse must NOT fall back to this schema. Drop to
+ * `failureMode: "unclassified"` instead. Strict and legacy schemas are
+ * deliberate siblings, not a legacy/canonical pair to consolidate.
+ *
+ * The schema asserts `satisfies z.ZodType<LegacyGraderJudgment>` against
+ * the canonical domain type in
+ * `packages/core/src/types/legacy-grader-judgment.ts` (D0045 / W0187) —
+ * drift between schema and type is a build error. The domain type is
+ * authored independently in `@sanity/ailf-core`; this file authors ONLY
+ * the schema and never derives the domain type from the schema itself
+ * (no schema-derived self-reference allowed by D0045).
+ *
+ * @see docs/decisions/D0045-type-architecture-and-contract-enforcement.md
+ * @see ../promptfoo-grader-output.ts — the strict (live-path) sibling
+ * @see docs/design-docs/actionability-ladder/03-structured-grader-judgments.md
+ *      §"Backwards compatibility"
+ */
+import { z } from "zod";
+/**
+ * Canonical schema for {@link LegacyGraderJudgment}. Mirrors the Phase 1
+ * superset core (`taskId`, `modelId`, `dimension`, `reason`, `score`,
+ * optional `outputFailure`). NO GRAD-02 additive fields — those are by
+ * construction absent on pre-Phase-3 output.
+ *
+ * Intentionally NOT `.strict()` — pre-Phase-3 reports may carry stray
+ * keys; the legacy parser tolerates them so historical-report rendering
+ * keeps working through the GRAD-06 cutover.
+ */
+export const LegacyGraderJudgmentSchema = z.object({
+    taskId: z.string().min(1),
+    modelId: z.string().min(1),
+    dimension: z.string().min(1),
+    reason: z.string(),
+    score: z.number(),
+    outputFailure: z.boolean().optional(),
+});

package/dist/adapters/grader-outputs/promptfoo-grader-output.d.ts

@@ -0,0 +1,102 @@
+/**
+ * promptfoo-grader-output.ts — Zod schema for the structured grader output
+ * (GRAD-02) emitted by the promptfoo grader process and consumed by the
+ * eval pipeline.
+ *
+ * The schema asserts `satisfies z.ZodType<GraderJudgment>` against the
+ * canonical domain type in `packages/core/src/types/grader-judgment.ts`
+ * (D0045 / W0187) — drift between schema and type is a build error.
+ * The domain type was authored independently in Plan 01-01; this file
+ * authors ONLY the schema and never derives the domain type from the
+ * schema itself (no schema-derived self-reference allowed by D0045).
+ *
+ * `graderJudgmentsVersion` is co-located with the schema (VER-01 D-02 —
+ * source-of-truth file owns its version constant). Bumped by hand when
+ * the grader rubric, prompt template, or judgment shape changes.
+ *
+ * Phase 3 will replace the inline `JSON.parse` at
+ * `pipeline/calculate-scores.ts:380-392` (Pitfall #4) so all grader
+ * output flows through this schema.
+ *
+ * @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/03-structured-grader-judgments.md
+ */
+import { z } from "zod";
+/**
+ * VER-01 D-02 — co-located version constant. Bumped by hand when the
+ * grader rubric, prompt template, or judgment shape changes in a way
+ * that should invalidate cached Diagnoses.
+ *
+ * Phase 3 GRAD-05 bumped this from `"0.1.0"` to `"1.0.0"` (semver
+ * major) — the additive GRAD-02 surface is now required + the schema
+ * is `.strict()`. AILF has no installed external base; the legacy
+ * parser at `./legacy/promptfoo-grader-output-legacy.ts` is the named
+ * consumer for already-stored historical reports.
+ */
+export declare const graderJudgmentsVersion = "1.0.0";
+/**
+ * Canonical schema for {@link GraderJudgment}. Required fields mirror
+ * the existing pipeline core (Doc 03 §"existing, unchanged"):
+ * `taskId`, `modelId`, `dimension`, `reason`, `score`. Phase 3 GRAD-05
+ * has tightened the additive surface to required and added `.strict()`
+ * — the schema rejects unknown fields (defense-in-depth against future
+ * prompt-injection attempts that try to smuggle keys through the
+ * grader emission).
+ *
+ * Branded `JudgmentId` is represented at runtime by a non-empty string;
+ * the schema routes the brand through `brandedString<"JudgmentId">()`
+ * — the project's single audited cast site for branded-string
+ * schemas (project typescript rule: no `as` on `unknown`).
+ */
+export declare const GraderJudgmentSchema: z.ZodObject<{
+    taskId: z.ZodString;
+    modelId: z.ZodString;
+    dimension: z.ZodString;
+    reason: z.ZodString;
+    score: z.ZodNumber;
+    outputFailure: z.ZodOptional<z.ZodBoolean>;
+    judgmentId: z.ZodType<import("@sanity/ailf-core").Brand<string, "JudgmentId">, unknown, z.core.$ZodTypeInternals<import("@sanity/ailf-core").Brand<string, "JudgmentId">, unknown>>;
+    subJudgments: z.ZodArray<z.ZodObject<{
+        criterionId: z.ZodString;
+        met: z.ZodBoolean;
+        evidence: z.ZodString;
+        confidence: z.ZodObject<{
+            level: z.ZodEnum<{
+                low: "low";
+                medium: "medium";
+                high: "high";
+            }>;
+            signalsPresent: z.ZodNumber;
+            derivation: z.ZodString;
+        }, z.core.$strip>;
+    }, z.core.$strip>>;
+    docCitations: z.ZodArray<z.ZodObject<{
+        documentId: z.ZodString;
+        slug: z.ZodOptional<z.ZodString>;
+        role: z.ZodEnum<{
+            supports: "supports";
+            contradicts: "contradicts";
+            missing: "missing";
+            irrelevant: "irrelevant";
+        }>;
+        hallucinated: z.ZodOptional<z.ZodBoolean>;
+    }, z.core.$strip>>;
+    failureMode: z.ZodString;
+    confidence: z.ZodObject<{
+        level: z.ZodEnum<{
+            low: "low";
+            medium: "medium";
+            high: "high";
+        }>;
+        signalsPresent: z.ZodNumber;
+        derivation: z.ZodString;
+    }, z.core.$strip>;
+    hallucinationCheckedAgainst: z.ZodArray<z.ZodString>;
+    metadata: z.ZodObject<{
+        graderModel: z.ZodString;
+        graderJudgmentsVersion: z.ZodString;
+    }, z.core.$strip>;
+}, z.core.$strict>;
+export type { GraderJudgment } from "../../_vendor/ailf-core/index.d.ts";