@sanity/ailf 4.6.0 → 6.0.0

package/dist/composition-root.d.ts

@@ -15,7 +15,9 @@
  * @see packages/core/src/ports/context.ts — AppContext interface
  * @see docs/archive/exec-plans/ports-and-adapters/phase-7-composition-root.md
  */
-import { type AppContext, type ArtifactWriter, type ArtifactWriterProgressOptions, type AssertionRegistration, type LLMClient, type Logger, type ResolvedConfig } from "./_vendor/ailf-core/index.d.ts";
+import { type AppContext, type ArtifactWriter, type ArtifactWriterProgressOptions, type AssertionRegistration, type CardRegistry, type DiagnosisRunner, type Logger, type ResolvedConfig } from "./_vendor/ailf-core/index.d.ts";
+export type { LLMClientKeys } from "./_vendor/ailf-core/index.d.ts";
+import { type BorderlineConsensusOptions, type BorderlineConsensusResult } from "./pipeline/borderline-consensus-runner.js";
 import { CompositeTaskSource, ContentLakeTaskSource, RepoTaskSource } from "./adapters/task-sources/index.js";
 /**
  * Create a fully wired AppContext from resolved configuration.
@@ -24,28 +26,6 @@ import { CompositeTaskSource, ContentLakeTaskSource, RepoTaskSource } from "./ad
  * Swapping an adapter is a one-line change in this function.
  */
 export declare function createAppContext(config: ResolvedConfig): AppContext;
-/**
- * Typed key bag passed to `createLLMClient`. The composition root reads
- * env once and supplies values here; the factory stays pure so tests don't
- * have to mutate `process.env`.
- */
-export interface LLMClientKeys {
-    anthropicApiKey?: string;
-    openaiApiKey?: string;
-}
-/**
- * Select the LLMClient adapter based on `config.llmProvider` and the
- * supplied API keys. Returns `undefined` when no usable credential is
- * present — `AppContext.llmClient` stays unset and consumers handle that
- * explicitly.
- *
- * Adapters never read `process.env` themselves (per
- * `.claude/rules/typescript.md`); env mapping happens at the call site
- * (typically `createAppContext`).
- *
- * Exported for unit-test access; not part of the public package API.
- */
-export declare function createLLMClient(config: ResolvedConfig, keys: LLMClientKeys, logger: Logger): LLMClient | undefined;
 /**
  * Selects the `ArtifactWriter` wiring per D0033 M4:
  *
@@ -83,3 +63,57 @@ export declare function createTaskSource(config: ResolvedConfig): CompositeTaskS
  * explicit mode whitelists.
  */
 export declare const FRAMEWORK_ASSERTIONS: AssertionRegistration[];
+/**
+ * Severity boundaries from `packages/eval/config/thresholds.ts`
+ * (severity.critical/warning/info `composite-below` at L50/54/58 — 30, 50,
+ * 60). The borderline detector flags a judgment when its score is within
+ * ±5 of any of these. Composition-root reads them ONCE and threads the
+ * typed `readonly number[]` into `runBorderlineConsensus` rather than
+ * re-deriving them at each call site (Pitfall 5 — single source of truth
+ * for the scale).
+ */
+export declare const BORDERLINE_SEVERITY_THRESHOLDS: readonly number[];
+/**
+ * Default replications per borderline judgment when the caller's
+ * `RepoConfig.execution.borderlineReplications` is unset (locked answer
+ * #4 in plan 03-04). Three replications + the original score = four
+ * scores per consistency record, which is the minimum that produces a
+ * non-degenerate stdDev / median split.
+ */
+export declare const DEFAULT_BORDERLINE_REPLICATIONS = 3;
+/**
+ * Factory for the borderline-consensus runner. Returns a function that
+ * applies the severity-threshold and replication defaults from
+ * composition-root, leaving the live grader entry point (the `regrade`
+ * callback) and the candidate `judgments` array as runtime inputs.
+ *
+ * The pipeline-side caller (currently `pipeline/calculate-scores.ts`'s
+ * post-extraction junction) supplies the `regrade` callback that maps a
+ * `GraderJudgment` to a fresh score via the response/rubric text from
+ * the original Promptfoo result. See the runner's header for the
+ * rationale on injecting the regrader rather than calling `gradeOnce`
+ * inline (Pitfall 6 — preserve the runner's purity wrt the existing
+ * grader-comparison split).
+ */
+export declare function createBorderlineConsensusRunner(opts: {
+    borderlineReplications?: number;
+}): (args: Pick<BorderlineConsensusOptions, "judgments" | "logger" | "regrade">) => Promise<BorderlineConsensusResult>;
+/**
+ * Returns the full 8-card `CardRegistry` backed by `DIAGNOSIS_CARD_GENERATORS`
+ * from `@sanity/ailf-core`. Exposed as a function (not a module-level const)
+ * so the composition root remains the single-seam factory and tests can assert
+ * the call site (AI-SPEC §3 Pitfall 1 — no module-scope mutables).
+ */
+export declare function buildDiagnosisRegistry(): CardRegistry;
+/**
+ * Build a fully-wired `DiagnosisRunner` from an `AppContext`.
+ *
+ * Wires the full 8-card registry, `loadAttributions` bound to the local
+ * filesystem (Phase-4 per-entry attribution objects at
+ * `{artifactsDir}/runs/{runId}/attribution/*.json`), and no-op cache
+ * reader/writer (Plan-06 CLI command will wire the real cache seam).
+ *
+ * Plan-06 API/CLI consumers import this function from the composition root
+ * and pass `ctx` from `createAppContext(config)`.
+ */
+export declare function getDiagnosisRunner(ctx: AppContext): DiagnosisRunner;

package/dist/composition-root.js

@@ -15,7 +15,10 @@
  * @see packages/core/src/ports/context.ts — AppContext interface
  * @see docs/archive/exec-plans/ports-and-adapters/phase-7-composition-root.md
  */
-import { ARTIFACT_EXPORT_PHASE_ID, InMemoryPluginRegistry, NoOpArtifactWriter, NoOpProgressReporter, generateRunId, isArtifactType, } from "./_vendor/ailf-core/index.js";
+import { promises as fs } from "node:fs";
+import path from "node:path";
+import { ARTIFACT_EXPORT_PHASE_ID, DIAGNOSIS_CARD_GENERATORS, InMemoryPluginRegistry, NoOpArtifactWriter, NoOpProgressReporter, createDiagnosisRunner, createLLMClient, generateRunId, isArtifactType, modelId, } from "./_vendor/ailf-core/index.js";
+import { JudgmentAttributionSchema } from "./adapters/attribution/per-entry-attribution-writer.js";
 import { AccumulatingArtifactWriter } from "./artifact-capture/accumulating-artifact-writer.js";
 import { ApiGatewayArtifactWriter } from "./artifact-capture/api-gateway-artifact-writer.js";
 import { BatchingApiGatewayArtifactWriter } from "./artifact-capture/batching-api-gateway-artifact-writer.js";
@@ -27,6 +30,7 @@ import { resolveUploadConcurrency, setDefaultUploadConcurrency, } from "./artifa
 import { UploadMetrics } from "./artifact-capture/upload-metrics.js";
 import { ContentLakeCacheAdapter } from "./adapters/cache/content-lake-cache.js";
 import { AnthropicLLMClient, OpenAILLMClient } from "./adapters/llm/index.js";
+import { runBorderlineConsensus, } from "./pipeline/borderline-consensus-runner.js";
 import { loadExternalPresets } from "./pipeline/compiler/preset-loader.js";
 import { FilesystemCache } from "./adapters/cache/filesystem-cache.js";
 import { PromptfooEvalAdapter } from "./adapters/eval-runners/promptfoo-eval-adapter.js";
@@ -95,10 +99,19 @@ export function createAppContext(config) {
     // LLM client (D0051) — wired when an API key is present. The grader path
     // does NOT consume this; D0051 defers grader migration as a follow-up.
     // Env mapping happens here so `createLLMClient` stays pure and testable.
-    const llmClient = createLLMClient(config, {
+    // D-01: factory hoisted to @sanity/ailf-core; adapter ctors injected here.
+    const llmAdapters = {
+        newAnthropicClient: (opts) => new AnthropicLLMClient(opts),
+        newOpenAIClient: (opts) => new OpenAILLMClient(opts),
+    };
+    const llmKeys = {
         anthropicApiKey: process.env.ANTHROPIC_API_KEY,
         openaiApiKey: process.env.OPENAI_API_KEY,
-    }, logger);
+    };
+    const llmClient = createLLMClient(config, llmKeys, {
+        logger,
+        adapters: llmAdapters,
+    });
     return {
         artifactWriter,
         cache,
@@ -116,44 +129,6 @@ export function createAppContext(config) {
         taskSource,
     };
 }
-/**
- * Select the LLMClient adapter based on `config.llmProvider` and the
- * supplied API keys. Returns `undefined` when no usable credential is
- * present — `AppContext.llmClient` stays unset and consumers handle that
- * explicitly.
- *
- * Adapters never read `process.env` themselves (per
- * `.claude/rules/typescript.md`); env mapping happens at the call site
- * (typically `createAppContext`).
- *
- * Exported for unit-test access; not part of the public package API.
- */
-export function createLLMClient(config, keys, logger) {
-    const explicit = config.llmProvider;
-    const anthropicKey = keys.anthropicApiKey;
-    const openaiKey = keys.openaiApiKey;
-    // Auto-select: prefer Anthropic when both are present (matches the
-    // current grader's default model in `config/models.ts`).
-    const provider = explicit ?? (anthropicKey ? "anthropic" : openaiKey ? "openai" : undefined);
-    if (!provider) {
-        logger.debug("LLM client: not wired — no Anthropic or OpenAI API key supplied");
-        return undefined;
-    }
-    if (provider === "anthropic") {
-        if (!anthropicKey) {
-            logger.warn('llmProvider="anthropic" but no Anthropic API key supplied — LLMClient not wired');
-            return undefined;
-        }
-        logger.debug("LLM client: AnthropicLLMClient");
-        return new AnthropicLLMClient({ apiKey: anthropicKey, logger });
-    }
-    if (!openaiKey) {
-        logger.warn('llmProvider="openai" but no OpenAI API key supplied — LLMClient not wired');
-        return undefined;
-    }
-    logger.debug("LLM client: OpenAILLMClient");
-    return new OpenAILLMClient({ apiKey: openaiKey, logger });
-}
 // ---------------------------------------------------------------------------
 // Sub-factories (extracted to keep createAppContext readable)
 // ---------------------------------------------------------------------------
@@ -493,3 +468,142 @@ function createReportStore(config) {
             undefined,
     });
 }
+// ---------------------------------------------------------------------------
+// Borderline-consensus wiring (Plan 03-04 / GRAD-04)
+// ---------------------------------------------------------------------------
+/**
+ * Severity boundaries from `packages/eval/config/thresholds.ts`
+ * (severity.critical/warning/info `composite-below` at L50/54/58 — 30, 50,
+ * 60). The borderline detector flags a judgment when its score is within
+ * ±5 of any of these. Composition-root reads them ONCE and threads the
+ * typed `readonly number[]` into `runBorderlineConsensus` rather than
+ * re-deriving them at each call site (Pitfall 5 — single source of truth
+ * for the scale).
+ */
+export const BORDERLINE_SEVERITY_THRESHOLDS = [
+    30, 50, 60,
+];
+/**
+ * Default replications per borderline judgment when the caller's
+ * `RepoConfig.execution.borderlineReplications` is unset (locked answer
+ * #4 in plan 03-04). Three replications + the original score = four
+ * scores per consistency record, which is the minimum that produces a
+ * non-degenerate stdDev / median split.
+ */
+export const DEFAULT_BORDERLINE_REPLICATIONS = 3;
+/**
+ * Factory for the borderline-consensus runner. Returns a function that
+ * applies the severity-threshold and replication defaults from
+ * composition-root, leaving the live grader entry point (the `regrade`
+ * callback) and the candidate `judgments` array as runtime inputs.
+ *
+ * The pipeline-side caller (currently `pipeline/calculate-scores.ts`'s
+ * post-extraction junction) supplies the `regrade` callback that maps a
+ * `GraderJudgment` to a fresh score via the response/rubric text from
+ * the original Promptfoo result. See the runner's header for the
+ * rationale on injecting the regrader rather than calling `gradeOnce`
+ * inline (Pitfall 6 — preserve the runner's purity wrt the existing
+ * grader-comparison split).
+ */
+export function createBorderlineConsensusRunner(opts) {
+    const replications = opts.borderlineReplications ?? DEFAULT_BORDERLINE_REPLICATIONS;
+    return (args) => runBorderlineConsensus({
+        judgments: args.judgments,
+        ...(args.logger ? { logger: args.logger } : {}),
+        regrade: args.regrade,
+        replications,
+        thresholds: BORDERLINE_SEVERITY_THRESHOLDS,
+    });
+}
+// ---------------------------------------------------------------------------
+// Diagnosis runner wiring (Plan 05-05 / D0048)
+// ---------------------------------------------------------------------------
+/**
+ * Returns the full 8-card `CardRegistry` backed by `DIAGNOSIS_CARD_GENERATORS`
+ * from `@sanity/ailf-core`. Exposed as a function (not a module-level const)
+ * so the composition root remains the single-seam factory and tests can assert
+ * the call site (AI-SPEC §3 Pitfall 1 — no module-scope mutables).
+ */
+export function buildDiagnosisRegistry() {
+    return DIAGNOSIS_CARD_GENERATORS;
+}
+/**
+ * Default local artifacts root (mirrors `createArtifactWriter` default above).
+ */
+const DIAGNOSIS_LOCAL_ARTIFACTS_DIR = ".ailf/results/captures";
+/**
+ * Load all per-entry attribution objects for a given `runId` from the local
+ * filesystem at `{artifactsDir}/runs/{runId}/attribution/*.json`.
+ *
+ * Reads every `.json` file in the run's `attribution/` directory, JSON-parses
+ * it, and Zod-validates each entry through `JudgmentAttributionSchema` (Phase-4
+ * canonical schema — D0045). Malformed entries are skipped with a warning.
+ *
+ * Returns an empty array when the directory does not exist (expected on runs
+ * without Phase-4 attribution data — the runner treats this as Landmine-11
+ * "no data" and both attribution cards return `status: "missing"`).
+ */
+async function loadAttributionsFromLocalFs(runId, artifactsDir, logger) {
+    const attrDir = path.join(artifactsDir, "runs", runId, "attribution");
+    let entries;
+    try {
+        entries = (await fs.readdir(attrDir)).filter((f) => f.endsWith(".json"));
+    }
+    catch {
+        // Directory missing — no attribution data for this run.
+        return [];
+    }
+    const results = [];
+    for (const filename of entries) {
+        const filePath = path.join(attrDir, filename);
+        let raw;
+        try {
+            const bytes = await fs.readFile(filePath, "utf8");
+            raw = JSON.parse(bytes);
+        }
+        catch (err) {
+            logger?.warn("loadAttributions: failed to read/parse attribution file", {
+                file: filePath,
+                error: err instanceof Error ? err.message : String(err),
+            });
+            continue;
+        }
+        const parsed = JudgmentAttributionSchema.safeParse(raw);
+        if (!parsed.success) {
+            logger?.warn("loadAttributions: attribution file failed schema validation", {
+                file: filePath,
+                errors: parsed.error.flatten(),
+            });
+            continue;
+        }
+        results.push(parsed.data);
+    }
+    return results;
+}
+/**
+ * Build a fully-wired `DiagnosisRunner` from an `AppContext`.
+ *
+ * Wires the full 8-card registry, `loadAttributions` bound to the local
+ * filesystem (Phase-4 per-entry attribution objects at
+ * `{artifactsDir}/runs/{runId}/attribution/*.json`), and no-op cache
+ * reader/writer (Plan-06 CLI command will wire the real cache seam).
+ *
+ * Plan-06 API/CLI consumers import this function from the composition root
+ * and pass `ctx` from `createAppContext(config)`.
+ */
+export function getDiagnosisRunner(ctx) {
+    const artifactsDir = ctx.config.artifactsDir ?? DIAGNOSIS_LOCAL_ARTIFACTS_DIR;
+    // No-op cache shims — Plan 06 wires the real cache.
+    const diagnosisReader = async (_path) => null;
+    const diagnosisWriter = async (_path, _diagnosis) => { };
+    return createDiagnosisRunner({
+        llm: ctx.llmClient,
+        model: modelId("anthropic:claude-opus-4-6"),
+        logger: ctx.logger,
+        progress: ctx.progress,
+        registry: buildDiagnosisRegistry(),
+        diagnosisReader,
+        diagnosisWriter,
+        loadAttributions: (runId) => loadAttributionsFromLocalFs(runId, artifactsDir, ctx.logger),
+    });
+}

package/dist/config/diagnosis-cards.ts

@@ -0,0 +1,318 @@
+/**
+ * diagnosis-cards.ts — Diagnosis eval matrix config.
+ *
+ * TS-first config (per .claude/rules/config.md) defining the 5 LLM card types
+ * × 3 first-class models eval matrix. Consumed by
+ * `scripts/generate-diagnosis-config.ts` to emit
+ * `promptfooconfig-diagnosis.yaml`. Never hand-edit the YAML — run
+ * `pnpm generate-configs` instead.
+ *
+ * Per AI-SPEC §5 and CONTEXT D-04 (path b: standalone generator entry point
+ * for the diagnosis config, additive — does not modify the existing literacy
+ * generate-configs pipeline).
+ *
+ * @see packages/eval/scripts/generate-diagnosis-config.ts — generator
+ * @see packages/eval/promptfooconfig-diagnosis.yaml — generated output
+ */
+
+// ---------------------------------------------------------------------------
+// Types
+// ---------------------------------------------------------------------------
+
+/**
+ * A first-class model entry in the diagnosis eval matrix.
+ * Mirrors the shape of model entries in `config/models.ts`.
+ */
+export interface DiagnosisModelEntry {
+  /** Promptfoo provider string (e.g. "anthropic:messages:claude-opus-4-6") */
+  id: string
+  /** Human-readable label for reports */
+  label: string
+  /** Per-model config overrides (temperature, max_tokens, etc.) */
+  config?: Record<string, unknown>
+}
+
+/**
+ * The 5 LLM-driven card types under evaluation.
+ * Deterministic cards (area-summary, failure-mode-summary, no-issues) are
+ * tested via `fixture-matrix.test.ts` (vitest), not via the promptfoo matrix.
+ */
+export type LLMCardType =
+  | "top-recommendations"
+  | "weakest-area"
+  | "low-confidence-attribution"
+  | "doc-attribution-spotlight"
+  | "regression-vs-baseline"
+
+/**
+ * A single evaluation scenario: one fixture path × one expected outcome.
+ *
+ * The `fixturePath` is relative to `packages/eval/` so the promptfoo config
+ * can resolve it from any working directory. `expectedStatus` drives the
+ * pass/fail assertion in the generated YAML.
+ */
+export interface DiagnosisScenario {
+  /** Short slug used in promptfoo `description` fields */
+  name: string
+  /** Path to the Report JSON fixture, relative to `packages/eval/` */
+  fixturePath: string
+  /**
+   * Card type this scenario exercises. The eval matrix runs all LLM cards
+   * per scenario; this field annotates which card type is the primary focus
+   * for the rubric.
+   */
+  primaryCard: LLMCardType
+  /** Expected card status when all LLM calls succeed */
+  expectedStatus: "ready" | "degraded" | "missing"
+  /** Optional: path to canned LLM response for adversarial scenarios */
+  cannedResponsePath?: string
+  /**
+   * Optional: cardId to key the canned response against (for FakeLLMClient
+   * keyedResponses in vitest; mirrored in the promptfoo scenario description
+   * for documentation).
+   */
+  cannedCardId?: LLMCardType
+  /** Free-text note about what this scenario tests */
+  note?: string
+}
+
+/**
+ * Top-level diagnosis eval matrix config.
+ * Exported as the default export of this file (mirrors models.ts convention).
+ */
+export interface DiagnosisCardsConfig {
+  /** All LLM card evaluation scenarios */
+  scenarios: DiagnosisScenario[]
+  /** Models to run each scenario against */
+  models: DiagnosisModelEntry[]
+  /** Grader model for LLM-judge assertions */
+  grader: DiagnosisModelEntry
+  /** Eval budget in milliseconds (kill switch) */
+  evalBudgetMs: number
+  /** Max parallel API calls */
+  maxConcurrency: number
+  /** Default per-model config */
+  defaults: {
+    temperature: number
+    max_tokens: number
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Helper
+// ---------------------------------------------------------------------------
+
+export function defineDiagnosisCards(
+  config: DiagnosisCardsConfig
+): DiagnosisCardsConfig {
+  return config
+}
+
+// ---------------------------------------------------------------------------
+// Config definition
+// ---------------------------------------------------------------------------
+
+const diagnosisCardsConfig: DiagnosisCardsConfig = defineDiagnosisCards({
+  // ── Models under evaluation ────────────────────────────────────────────────
+  models: [
+    {
+      id: "anthropic:messages:claude-opus-4-6",
+      label: "Claude Opus 4.6",
+      config: { temperature: 0.2, max_tokens: 4096 },
+    },
+    {
+      id: "anthropic:messages:claude-sonnet-4-6",
+      label: "Claude Sonnet 4.6",
+      config: { temperature: 0.2, max_tokens: 4096 },
+    },
+    {
+      id: "openai:chat:gpt-5.2",
+      label: "GPT 5.2",
+      config: { max_completion_tokens: 4096 },
+    },
+  ],
+
+  // ── Grader model ────────────────────────────────────────────────────────────
+  grader: {
+    id: "anthropic:messages:claude-opus-4-5-20251101",
+    label: "Claude Opus 4.5 (grader)",
+  },
+
+  // ── Eval budget ─────────────────────────────────────────────────────────────
+  evalBudgetMs: 3_600_000, // 60 min — full matrix across 3 models × 17 scenarios
+  maxConcurrency: 8, // conservative for diagnosis (longer prompts than literacy)
+
+  // ── Default config ──────────────────────────────────────────────────────────
+  defaults: {
+    temperature: 0.2,
+    max_tokens: 4096,
+  },
+
+  // ── Scenarios (17 logical fixtures) ─────────────────────────────────────────
+  scenarios: [
+    // ── Critical-path: top-recommendations ──────────────────────────────────
+    {
+      name: "healthy-top-recommendations",
+      fixturePath:
+        "test-fixtures/diagnosis/reports/healthy-top-recommendations.json",
+      primaryCard: "top-recommendations",
+      expectedStatus: "ready",
+      note: "Healthy report (mean 91) — top-recommendations card should produce 2+ actionable suggestions with docSlug references from the manifest.",
+    },
+    {
+      name: "low-top-recommendations",
+      fixturePath:
+        "test-fixtures/diagnosis/reports/low-top-recommendations.json",
+      primaryCard: "top-recommendations",
+      expectedStatus: "ready",
+      note: "Low-scoring report (mean 42) — top-recommendations card should produce high-priority suggestions addressing the dominant failure modes (outdated-docs, missing-docs).",
+    },
+
+    // ── Critical-path: weakest-area ──────────────────────────────────────────
+    {
+      name: "healthy-weakest-area",
+      fixturePath: "test-fixtures/diagnosis/reports/healthy-weakest-area.json",
+      primaryCard: "weakest-area",
+      expectedStatus: "ready",
+      note: "Healthy report with clear weakest area (content-modeling at 82) — weakest-area card should identify the area and provide high-confidence analysis.",
+    },
+    {
+      name: "low-weakest-area",
+      fixturePath: "test-fixtures/diagnosis/reports/low-weakest-area.json",
+      primaryCard: "weakest-area",
+      expectedStatus: "ready",
+      note: "Low-scoring report with clear weakest area (content-modeling at 28) — weakest-area card should identify the most critical area with multiple failure modes.",
+    },
+
+    // ── Critical-path: low-confidence-attribution ────────────────────────────
+    {
+      name: "healthy-low-confidence-attribution",
+      fixturePath:
+        "test-fixtures/diagnosis/reports/healthy-low-confidence-attribution.json",
+      primaryCard: "low-confidence-attribution",
+      expectedStatus: "ready",
+      note: "Healthy report with small sample sizes (2-3 judgments per area) — low-confidence-attribution card should identify attribution uncertainty despite positive scores.",
+    },
+    {
+      name: "low-low-confidence-attribution",
+      fixturePath:
+        "test-fixtures/diagnosis/reports/low-low-confidence-attribution.json",
+      primaryCard: "low-confidence-attribution",
+      expectedStatus: "ready",
+      note: "Low-scoring report with small sample sizes (2 judgments per area) — low-confidence-attribution card should flag both score quality and attribution uncertainty.",
+    },
+
+    // ── Critical-path: doc-attribution-spotlight ─────────────────────────────
+    {
+      name: "healthy-doc-attribution-spotlight",
+      fixturePath:
+        "test-fixtures/diagnosis/reports/healthy-doc-attribution-spotlight.json",
+      primaryCard: "doc-attribution-spotlight",
+      expectedStatus: "ready",
+      note: "Healthy 5-area report — doc-attribution-spotlight card should identify the highest-impact document in the manifest.",
+    },
+    {
+      name: "low-doc-attribution-spotlight",
+      fixturePath:
+        "test-fixtures/diagnosis/reports/low-doc-attribution-spotlight.json",
+      primaryCard: "doc-attribution-spotlight",
+      expectedStatus: "ready",
+      note: "Low-scoring 5-area report with multiple failure modes — doc-attribution-spotlight card should identify the most critical document.",
+    },
+
+    // ── Edge cases ───────────────────────────────────────────────────────────
+    {
+      name: "empty-report",
+      fixturePath: "test-fixtures/diagnosis/reports/empty.json",
+      primaryCard: "top-recommendations",
+      expectedStatus: "missing",
+      note: "Edge case (a): zero-area report — all LLM cards should emit status: missing (no data to reason about).",
+    },
+    {
+      name: "single-judgment-per-area",
+      fixturePath:
+        "test-fixtures/diagnosis/reports/single-judgment-per-area.json",
+      primaryCard: "weakest-area",
+      expectedStatus: "ready",
+      note: "Edge case (b): single-judgment sample size — weakest-area card should reflect low-confidence calibration (sampleSize: 1).",
+    },
+    {
+      name: "all-areas-tied",
+      fixturePath: "test-fixtures/diagnosis/reports/all-areas-tied.json",
+      primaryCard: "weakest-area",
+      expectedStatus: "missing",
+      note: "Edge case (c): all areas scored identically (70) — weakest-area card should emit status: missing with reason: no-clear-weakest.",
+    },
+    {
+      name: "grader-major-mismatch-baseline",
+      fixturePath:
+        "test-fixtures/diagnosis/reports/grader-major-mismatch-baseline.json",
+      primaryCard: "regression-vs-baseline",
+      expectedStatus: "missing",
+      note: "Edge case (d): grader-major-version mismatch — regression-vs-baseline should emit missing with reason: grader-major-version-mismatch. Run as pair with grader-major-mismatch-current.",
+    },
+    {
+      name: "grader-major-mismatch-current",
+      fixturePath:
+        "test-fixtures/diagnosis/reports/grader-major-mismatch-current.json",
+      primaryCard: "regression-vs-baseline",
+      expectedStatus: "missing",
+      note: "Edge case (d) pair: current report with different graderModel — regression-vs-baseline mismatch guard triggers when paired with grader-major-mismatch-baseline.",
+    },
+    {
+      name: "near-deprecated-taxonomy",
+      fixturePath:
+        "test-fixtures/diagnosis/reports/near-deprecated-taxonomy.json",
+      primaryCard: "weakest-area",
+      expectedStatus: "ready",
+      note: "Edge case (e): report using unclassified failure mode (currently canonical but watch for taxonomy retirement). Zod refine() must accept canonical modes.",
+    },
+
+    // ── Adversarial canned responses ─────────────────────────────────────────
+    {
+      name: "adversarial-fabricated-delta",
+      fixturePath:
+        "test-fixtures/diagnosis/reports/grader-major-mismatch-current.json",
+      primaryCard: "regression-vs-baseline",
+      expectedStatus: "degraded",
+      cannedResponsePath:
+        "test-fixtures/diagnosis/canned-responses/fabricated-delta-regression.json",
+      cannedCardId: "regression-vs-baseline",
+      note: "Adversarial: fabricated delta (AI-SPEC §1b failure-mode #1). LLM claims -7.3 delta; direction-sign refine triggers degraded card.",
+    },
+    {
+      name: "adversarial-improve-introduction",
+      fixturePath:
+        "test-fixtures/diagnosis/reports/low-top-recommendations.json",
+      primaryCard: "top-recommendations",
+      expectedStatus: "degraded",
+      cannedResponsePath:
+        "test-fixtures/diagnosis/canned-responses/improve-introduction.json",
+      cannedCardId: "top-recommendations",
+      note: "Adversarial: generic anti-pattern recommendation (AI-SPEC §1b failure-mode #2). Actionability refine triggers degraded card.",
+    },
+    {
+      name: "adversarial-hallucinated-docslug",
+      fixturePath:
+        "test-fixtures/diagnosis/reports/low-top-recommendations.json",
+      primaryCard: "top-recommendations",
+      expectedStatus: "degraded",
+      cannedResponsePath:
+        "test-fixtures/diagnosis/canned-responses/hallucinated-docslug.json",
+      cannedCardId: "top-recommendations",
+      note: "Adversarial: hallucinated docSlug (AI-SPEC §1b failure-mode #3). Allow-list refine triggers degraded card.",
+    },
+    {
+      name: "adversarial-taxonomy-drift",
+      fixturePath: "test-fixtures/diagnosis/reports/low-weakest-area.json",
+      primaryCard: "weakest-area",
+      expectedStatus: "degraded",
+      cannedResponsePath:
+        "test-fixtures/diagnosis/canned-responses/taxonomy-drift.json",
+      cannedCardId: "weakest-area",
+      note: "Adversarial: taxonomy drift (AI-SPEC §1b failure-mode #4). Per-dimension failureMode refine triggers degraded card.",
+    },
+  ],
+})
+
+export default diagnosisCardsConfig

package/dist/config/models.ts

@@ -24,6 +24,18 @@ export default defineModels({
       // All literacy variants included by default (baseline, observed,
       // agentic-naive, agentic-optimized)
     },
+    {
+      // Phase 5 LLM card routing (D-07). AI-SPEC §4 routes 3 routine cards
+      // (top-recommendations, weakest-area, regression-vs-baseline) here.
+      // Pricing already in AnthropicLLMClient; baseline literacy variant only.
+      id: "anthropic:messages:claude-sonnet-4-6",
+      label: "Claude Sonnet 4.6",
+      config: { temperature: 0.2, max_tokens: 4096 },
+      modes: ["literacy"],
+      variants: {
+        literacy: ["baseline"],
+      },
+    },
 
     // ── Google ─────────────────────────────────────────────────
     // {