@sanity/ailf 2.0.0 → 2.0.1

package/dist/pipeline/compiler/scoring-bridge.js

@@ -18,7 +18,7 @@
  *
  * @see packages/core/src/services/scoring-engine.ts — the 4-tier engine
  * @see packages/eval/src/pipeline/calculate-scores.ts — the consumer
- * @see docs/exec-plans/architecture-overhaul/phase-7-migrate-literacy.md
+ * @see docs/archive/exec-plans/architecture-overhaul/phase-7-migrate-literacy.md
  */
 import { aggregateDimensions, computeTaskScore, normalizeScore, } from "../../_vendor/ailf-core/index.js";
 import { classifyRubric, parseRubricScore } from "../../_vendor/ailf-core/index.js";

package/dist/pipeline/compiler/task-bridge.d.ts

@@ -0,0 +1,41 @@
+/**
+ * task-bridge.ts — Bidirectional bridge between old TaskDefinition and new LiteracyTaskDefinition.
+ *
+ * Enables incremental migration: consumers can convert between the two types
+ * without changing their internal logic. Once all consumers use
+ * GeneralizedTaskDefinition, this module is deleted (Wave 3 task 6).
+ *
+ * Field mapping (TaskDefinition ↔ LiteracyTaskDefinition):
+ *   id              ↔ id
+ *   description     ↔ title
+ *   featureArea     ↔ area
+ *   taskPrompt      ↔ prompt.text (fallback: prompt.template)
+ *   canonicalDocs   ↔ context.docs
+ *   referenceSolution ↔ referenceSolution
+ *   docCoverage     ↔ docCoverage
+ *   assertions      ↔ assertions       (structurally identical)
+ *   baseline        ↔ baseline         (structurally identical)
+ *   tags            ↔ tags
+ *   status          ↔ status
+ *   extraVars       ↔ prompt.vars
+ *
+ * The assertion and doc-ref sub-types are structurally identical between
+ * the old and new type systems, so no field-level remapping is needed
+ * for those — only a TypeScript-level cast.
+ */
+import type { LiteracyTaskDefinition, TaskDefinition } from "../../_vendor/ailf-core/index.d.ts";
+/**
+ * Convert an old-style TaskDefinition to the new LiteracyTaskDefinition.
+ *
+ * Every field of TaskDefinition has a corresponding field in LiteracyTaskDefinition,
+ * so this conversion is lossless.
+ */
+export declare function toGeneralized(task: TaskDefinition): LiteracyTaskDefinition;
+/**
+ * Convert a new LiteracyTaskDefinition to the old TaskDefinition shape.
+ *
+ * Fields that only exist on LiteracyTaskDefinition (description, difficulty,
+ * metadata, rubric, providers, options, context.fixtures, prompt.systemMessage)
+ * are dropped — the old type has no place for them.
+ */
+export declare function toLiteracyTask(task: LiteracyTaskDefinition): TaskDefinition;

package/dist/pipeline/compiler/task-bridge.js

@@ -0,0 +1,92 @@
+/**
+ * task-bridge.ts — Bidirectional bridge between old TaskDefinition and new LiteracyTaskDefinition.
+ *
+ * Enables incremental migration: consumers can convert between the two types
+ * without changing their internal logic. Once all consumers use
+ * GeneralizedTaskDefinition, this module is deleted (Wave 3 task 6).
+ *
+ * Field mapping (TaskDefinition ↔ LiteracyTaskDefinition):
+ *   id              ↔ id
+ *   description     ↔ title
+ *   featureArea     ↔ area
+ *   taskPrompt      ↔ prompt.text (fallback: prompt.template)
+ *   canonicalDocs   ↔ context.docs
+ *   referenceSolution ↔ referenceSolution
+ *   docCoverage     ↔ docCoverage
+ *   assertions      ↔ assertions       (structurally identical)
+ *   baseline        ↔ baseline         (structurally identical)
+ *   tags            ↔ tags
+ *   status          ↔ status
+ *   extraVars       ↔ prompt.vars
+ *
+ * The assertion and doc-ref sub-types are structurally identical between
+ * the old and new type systems, so no field-level remapping is needed
+ * for those — only a TypeScript-level cast.
+ */
+// ---------------------------------------------------------------------------
+// toGeneralized — old TaskDefinition → LiteracyTaskDefinition
+// ---------------------------------------------------------------------------
+/**
+ * Convert an old-style TaskDefinition to the new LiteracyTaskDefinition.
+ *
+ * Every field of TaskDefinition has a corresponding field in LiteracyTaskDefinition,
+ * so this conversion is lossless.
+ */
+export function toGeneralized(task) {
+    const result = {
+        mode: "literacy",
+        id: task.id,
+        title: task.description,
+        area: task.featureArea,
+        prompt: {
+            text: task.taskPrompt,
+            ...(task.extraVars != null ? { vars: task.extraVars } : {}),
+        },
+        context: {
+            docs: task.canonicalDocs,
+        },
+        referenceSolution: task.referenceSolution,
+        docCoverage: task.docCoverage,
+        assertions: task.assertions,
+    };
+    // Only set optional fields when present to preserve round-trip identity
+    if (task.baseline != null)
+        result.baseline = task.baseline;
+    if (task.tags != null)
+        result.tags = task.tags;
+    if (task.status != null)
+        result.status = task.status;
+    return result;
+}
+// ---------------------------------------------------------------------------
+// toLiteracyTask — LiteracyTaskDefinition → old TaskDefinition
+// ---------------------------------------------------------------------------
+/**
+ * Convert a new LiteracyTaskDefinition to the old TaskDefinition shape.
+ *
+ * Fields that only exist on LiteracyTaskDefinition (description, difficulty,
+ * metadata, rubric, providers, options, context.fixtures, prompt.systemMessage)
+ * are dropped — the old type has no place for them.
+ */
+export function toLiteracyTask(task) {
+    const result = {
+        id: task.id,
+        description: task.title,
+        featureArea: task.area ?? "",
+        taskPrompt: task.prompt?.text ?? task.prompt?.template ?? "",
+        canonicalDocs: (task.context?.docs ?? []),
+        referenceSolution: task.referenceSolution ?? "",
+        docCoverage: task.docCoverage ?? false,
+        assertions: (task.assertions ?? []),
+    };
+    // Only set optional fields when present to preserve round-trip identity
+    if (task.baseline != null)
+        result.baseline = task.baseline;
+    if (task.tags != null)
+        result.tags = task.tags;
+    if (task.status != null)
+        result.status = task.status;
+    if (task.prompt?.vars != null)
+        result.extraVars = task.prompt.vars;
+    return result;
+}

package/dist/pipeline/compiler/task-graph-builder.d.ts

@@ -11,11 +11,8 @@
  * - Validate the graph is a DAG (reject cycles)
  * - Assign execution priority via topological sort
  *
- * This module exists alongside `generate-configs.ts` — it does NOT replace
- * the existing codegen path. Phase 7 will swap callers over to the compiler.
- *
  * @see packages/core/src/types/task-graph.ts — TaskGraph types
- * @see docs/exec-plans/architecture-overhaul/phase-2-config-compiler.md
+ * @see docs/archive/exec-plans/architecture-overhaul/phase-2-config-compiler.md
  */
 import type { DependencyEdge, FilterOptions, GeneralizedTaskDefinition, TaskGraph, TaskNode } from "../../_vendor/ailf-core/index.d.ts";
 /** Options for building a task graph */

package/dist/pipeline/compiler/task-graph-builder.js

@@ -11,11 +11,8 @@
  * - Validate the graph is a DAG (reject cycles)
  * - Assign execution priority via topological sort
  *
- * This module exists alongside `generate-configs.ts` — it does NOT replace
- * the existing codegen path. Phase 7 will swap callers over to the compiler.
- *
  * @see packages/core/src/types/task-graph.ts — TaskGraph types
- * @see docs/exec-plans/architecture-overhaul/phase-2-config-compiler.md
+ * @see docs/archive/exec-plans/architecture-overhaul/phase-2-config-compiler.md
  */
 // ---------------------------------------------------------------------------
 // Public API

package/dist/pipeline/compiler/telemetry/index.d.ts

@@ -4,7 +4,7 @@
  * Captures tool calls, token usage, cost, and timing for every evaluation.
  * Full traces go to blob storage; sanitized summaries to Content Lake.
  *
- * @see docs/exec-plans/architecture-overhaul/phase-6-observability.md
+ * @see docs/archive/exec-plans/architecture-overhaul/phase-6-observability.md
  * @see docs/design-docs/architecture-overhaul/observability-telemetry.md
  */
 export { collectTrace, mergeTraces, type ProviderResponse, type RawToolCall, type TraceCollectorOptions, } from "./trace-collector.js";

package/dist/pipeline/compiler/telemetry/index.js

@@ -4,7 +4,7 @@
  * Captures tool calls, token usage, cost, and timing for every evaluation.
  * Full traces go to blob storage; sanitized summaries to Content Lake.
  *
- * @see docs/exec-plans/architecture-overhaul/phase-6-observability.md
+ * @see docs/archive/exec-plans/architecture-overhaul/phase-6-observability.md
  * @see docs/design-docs/architecture-overhaul/observability-telemetry.md
  */
 // Trace collection

package/dist/pipeline/compiler/variable-resolver.d.ts

@@ -12,7 +12,7 @@
  * recorded in provenance for reproducibility tracking.
  *
  * @see docs/design-docs/architecture-overhaul/domain-model.md (VariableEnvelope)
- * @see docs/exec-plans/architecture-overhaul/phase-2-config-compiler.md
+ * @see docs/archive/exec-plans/architecture-overhaul/phase-2-config-compiler.md
  */
 import type { VariableDeclaration, VariableEnvelope } from "../../_vendor/ailf-core/index.d.ts";
 /** Options for variable resolution */

package/dist/pipeline/compiler/variable-resolver.js

@@ -12,7 +12,7 @@
  * recorded in provenance for reproducibility tracking.
  *
  * @see docs/design-docs/architecture-overhaul/domain-model.md (VariableEnvelope)
- * @see docs/exec-plans/architecture-overhaul/phase-2-config-compiler.md
+ * @see docs/archive/exec-plans/architecture-overhaul/phase-2-config-compiler.md
  */
 import { simpleHash } from "./hash.js";
 // ---------------------------------------------------------------------------

package/dist/pipeline/coverage-audit.d.ts

@@ -7,7 +7,7 @@
  *
  * Phase 3c of the Scenario Matrix implementation.
  *
- * @see docs/exec-plans/scenario-matrix-implementation/phase-3-gap-analysis.md
+ * @see docs/archive/exec-plans/scenario-matrix-implementation/phase-3-gap-analysis.md
  */
 import type { Logger, PluginRegistry } from "../_vendor/ailf-core/index.d.ts";
 import type { CoverageAuditReport, ProductFeature } from "./types.js";

package/dist/pipeline/coverage-audit.js

@@ -7,7 +7,7 @@
  *
  * Phase 3c of the Scenario Matrix implementation.
  *
- * @see docs/exec-plans/scenario-matrix-implementation/phase-3-gap-analysis.md
+ * @see docs/archive/exec-plans/scenario-matrix-implementation/phase-3-gap-analysis.md
  */
 import { ConsoleLogger } from "../adapters/loggers/index.js";
 import { tryLoadConfigFile } from "./compiler/config-loader.js";

package/dist/pipeline/degradations.d.ts

@@ -12,7 +12,7 @@
  *
  * These are deterministic, pure functions — no randomness, no side effects.
  *
- * @see docs/exec-plans/grader-reliability.md — Phase 4
+ * @see docs/archive/exec-plans/grader-reliability.md — Phase 4
  */
 /** A degradation targeting a specific scoring dimension */
 export interface Degradation {

package/dist/pipeline/degradations.js

@@ -12,7 +12,7 @@
  *
  * These are deterministic, pure functions — no randomness, no side effects.
  *
- * @see docs/exec-plans/grader-reliability.md — Phase 4
+ * @see docs/archive/exec-plans/grader-reliability.md — Phase 4
  */
 // ---------------------------------------------------------------------------
 // Task Completion degradations

package/dist/pipeline/failure-modes.d.ts

@@ -13,7 +13,7 @@
  * When both sources agree, confidence is boosted. When only ceiling
  * signals are available, they serve as a fallback for unclassified cases.
  *
- * @see docs/exec-plans/scenario-matrix-implementation/phase-3-gap-analysis.md
+ * @see docs/archive/exec-plans/scenario-matrix-implementation/phase-3-gap-analysis.md
  */
 import type { FailureMode, FailureModeReport, FeatureScore, GraderJudgment } from "./types.js";
 /**

package/dist/pipeline/failure-modes.js

@@ -13,7 +13,7 @@
  * When both sources agree, confidence is boosted. When only ceiling
  * signals are available, they serve as a fallback for unclassified cases.
  *
- * @see docs/exec-plans/scenario-matrix-implementation/phase-3-gap-analysis.md
+ * @see docs/archive/exec-plans/scenario-matrix-implementation/phase-3-gap-analysis.md
  */
 import { detectFeatureArea } from "../_vendor/ailf-core/index.js";
 // ---------------------------------------------------------------------------
@@ -23,6 +23,7 @@ import { detectFeatureArea } from "../_vendor/ailf-core/index.js";
 const CLASSIFICATION_THRESHOLD = 60;
 /** All failure mode types for initializing empty counts */
 const ALL_MODES = [
+    "api-error",
     "incorrect-docs",
     "missing-docs",
     "model-limitation",
@@ -33,6 +34,9 @@ const ALL_MODES = [
 // ---------------------------------------------------------------------------
 // Keyword patterns
 // ---------------------------------------------------------------------------
+/** API error pattern — checked FIRST to prevent timeout errors containing
+ *  "deprecated" from being misclassified as outdated-docs. */
+const API_ERROR_PATTERN = /\[api-error\]|timeout|timed out|rate limit|429|503|ECONNRESET|ETIMEDOUT|socket hang up|fetch failed/i;
 const OUTDATED_PATTERN = /deprecated|old api|v[0-9]+ syntax|no longer supported|legacy|previous version|outdated|superseded|replaced by/i;
 const MISSING_PATTERN = /no documentation|not covered|had to guess|not found|missing.*doc|no.*information|undocumented|couldn't find|without.*documentation/i;
 const INCORRECT_PATTERN = /contradicts|incorrect.*doc|doc.*incorrect|wrong.*doc|doc.*wrong|documentation says.*but|factual error|inaccurate|misleading.*doc/i;
@@ -226,6 +230,11 @@ function classifyByCeiling(score, ceilingScore, floorScore) {
 }
 /** Classify by keyword matching on the reason text */
 function classifyByKeyword(reason) {
+    // API errors checked first — prevents timeout messages containing
+    // "deprecated" from being misclassified as outdated-docs.
+    if (API_ERROR_PATTERN.test(reason)) {
+        return { confidence: "high", mode: "api-error", source: "keyword" };
+    }
     if (OUTDATED_PATTERN.test(reason)) {
         return { confidence: "high", mode: "outdated-docs", source: "keyword" };
     }
@@ -321,6 +330,7 @@ function findTopMode(modes) {
 /** Initialize mode counts to zero */
 function initModeCounts() {
     return {
+        "api-error": 0,
         "incorrect-docs": 0,
         "missing-docs": 0,
         "model-limitation": 0,
@@ -332,6 +342,8 @@ function initModeCounts() {
 /** Get icon for a failure mode */
 function modeIcon(mode) {
     switch (mode) {
+        case "api-error":
+            return "⚡";
         case "incorrect-docs":
             return "❌";
         case "missing-docs":

package/dist/pipeline/gap-analysis.d.ts

@@ -13,7 +13,7 @@
  * bottleneck dimension to the median of non-bottlenecked dimensions (not 100).
  * This produces realistic estimates rather than theoretical maximums.
  *
- * @see docs/exec-plans/scenario-matrix-implementation/phase-3-gap-analysis.md
+ * @see docs/archive/exec-plans/scenario-matrix-implementation/phase-3-gap-analysis.md
  */
 import type { FailureModeReport, FeatureScore, GapAnalysisReport, GapEstimate } from "./types.js";
 /**

package/dist/pipeline/gap-analysis.js

@@ -13,7 +13,7 @@
  * bottleneck dimension to the median of non-bottlenecked dimensions (not 100).
  * This produces realistic estimates rather than theoretical maximums.
  *
- * @see docs/exec-plans/scenario-matrix-implementation/phase-3-gap-analysis.md
+ * @see docs/archive/exec-plans/scenario-matrix-implementation/phase-3-gap-analysis.md
  */
 // ---------------------------------------------------------------------------
 // Constants
@@ -26,6 +26,7 @@ const DEFAULT_WEIGHTS = {
 };
 /** Map failure modes to the dimensions they typically bottleneck */
 const MODE_BOTTLENECKS = {
+    "api-error": [], // Infrastructure issue, not a docs problem
     "incorrect-docs": ["code-correctness", "task-completion"],
     "missing-docs": ["task-completion", "doc-coverage"],
     "model-limitation": [], // Not a docs problem
@@ -35,6 +36,7 @@ const MODE_BOTTLENECKS = {
 };
 /** Remediation descriptions by failure mode */
 const REMEDIATION_MAP = {
+    "api-error": "Check model provider config (timeoutMs, maxRetries), API quotas, and task complexity",
     "incorrect-docs": "Fix factual errors in existing documentation",
     "missing-docs": "Write new documentation for uncovered functionality",
     "model-limitation": "Not a documentation problem — track for model improvement",

package/dist/pipeline/generate-configs.d.ts

@@ -12,7 +12,7 @@
  * the new compiler has been validated in production.
  *
  * @see packages/eval/src/pipeline/compiler/ — the new compiler pipeline
- * @see docs/exec-plans/architecture-overhaul/phase-7-migrate-literacy.md
+ * @see docs/archive/exec-plans/architecture-overhaul/phase-7-migrate-literacy.md
  *
  * ---
  *
@@ -31,7 +31,7 @@
  * No process.argv parsing. No env var fallbacks. Callers provide typed options.
  *
  * @see config/models.yaml — the central model registry
- * @see docs/exec-plans/eliminate-lib-layer.md
+ * @see docs/archive/exec-plans/eliminate-lib-layer.md
  */
 import { type LiteracyTaskDefinition, type Logger } from "../_vendor/ailf-core/index.d.ts";
 import type { FilterOptions } from "./types.js";

package/dist/pipeline/generate-configs.js

@@ -12,7 +12,7 @@
  * the new compiler has been validated in production.
  *
  * @see packages/eval/src/pipeline/compiler/ — the new compiler pipeline
- * @see docs/exec-plans/architecture-overhaul/phase-7-migrate-literacy.md
+ * @see docs/archive/exec-plans/architecture-overhaul/phase-7-migrate-literacy.md
  *
  * ---
  *
@@ -31,14 +31,15 @@
  * No process.argv parsing. No env var fallbacks. Callers provide typed options.
  *
  * @see config/models.yaml — the central model registry
- * @see docs/exec-plans/eliminate-lib-layer.md
+ * @see docs/archive/exec-plans/eliminate-lib-layer.md
  */
-import { extractModelName, extractProvider, mergeConfig, modelMatchesMode, } from "../_vendor/ailf-core/index.js";
+import { extractModelName, extractProvider, mergeConfig, } from "../_vendor/ailf-core/index.js";
 import { existsSync, readdirSync, writeFileSync } from "fs";
 import { resolve } from "path";
 import { dump } from "js-yaml";
 import { ConsoleLogger } from "../adapters/loggers/index.js";
 import { loadConfigFile } from "./compiler/config-loader.js";
+import { modelMatchesLiteracyVariant } from "./compiler/mode-bases/literacy.js";
 import { LITERACY_PROMPT_TEMPLATES } from "./compiler/mode-handlers/literacy/index.js";
 import { expandTaskDefinitions, loadAndExpandTasks } from "./expand-tasks.js";
 import { validateModelsYaml } from "./validate.js";
@@ -135,8 +136,8 @@ const SOURCE_ISOLATION_ASSERT = {
 // Config generators
 // ---------------------------------------------------------------------------
 function generateAgenticConfig(models, tests, prompts, source, searchMode, allowedOrigins) {
-    const naiveModels = models.models.filter((m) => modelMatchesMode(m, "agentic-naive"));
-    const optimizedModels = models.models.filter((m) => modelMatchesMode(m, "agentic-optimized"));
+    const naiveModels = models.models.filter((m) => modelMatchesLiteracyVariant(m, "agentic-naive"));
+    const optimizedModels = models.models.filter((m) => modelMatchesLiteracyVariant(m, "agentic-optimized"));
     const providers = [];
     // Build doc source config to inject into providers
     const resolvedSearchMode = searchMode ?? "open";
@@ -170,6 +171,7 @@ function generateAgenticConfig(models, tests, prompts, source, searchMode, allow
                     model: modelName,
                     provider,
                 }),
+                ...(model.timeoutMs ? { timeoutMs: model.timeoutMs } : {}),
                 ...sourceConfig,
                 observe: true,
                 observerOptions: models.defaults.observerOptions ?? {},
@@ -189,6 +191,7 @@ function generateAgenticConfig(models, tests, prompts, source, searchMode, allow
                     model: modelName,
                     provider,
                 }),
+                ...(model.timeoutMs ? { timeoutMs: model.timeoutMs } : {}),
                 ...sourceConfig,
                 observe: true,
                 observerOptions: models.defaults.observerOptions ?? {},
@@ -220,9 +223,12 @@ function generateAgenticConfig(models, tests, prompts, source, searchMode, allow
     };
 }
 function generateBaselineConfig(models, tests, prompts) {
-    const baselineModels = models.models.filter((m) => modelMatchesMode(m, LiteracyVariant.STANDARD));
+    const baselineModels = models.models.filter((m) => modelMatchesLiteracyVariant(m, "baseline"));
     const providers = baselineModels.map((model) => ({
-        config: mergeConfig(models.defaults, model.config),
+        config: {
+            ...mergeConfig(models.defaults, model.config),
+            ...(model.timeoutMs ? { timeoutMs: model.timeoutMs } : {}),
+        },
         id: model.id,
         label: model.label,
     }));
@@ -246,12 +252,13 @@ function generateBaselineConfig(models, tests, prompts) {
     };
 }
 function generateObservedConfig(models, tests, prompts) {
-    const observedModels = models.models.filter((m) => modelMatchesMode(m, LiteracyVariant.OBSERVED));
+    const observedModels = models.models.filter((m) => modelMatchesLiteracyVariant(m, "observed"));
     const providers = observedModels.map((model) => {
         const modelName = extractModelName(model.id);
         return {
             config: {
                 ...mergeConfig(models.defaults, model.config),
+                ...(model.timeoutMs ? { timeoutMs: model.timeoutMs } : {}),
                 modelName,
                 observe: true,
                 recordOptions: models.defaults.observerOptions ?? {},

package/dist/pipeline/grader-compare-runner.d.ts

@@ -10,7 +10,7 @@
  * Migrated from lib/grader-compare.ts — no process.argv, no process.exit(),
  * no module-level constants. Accepts rootDir as parameter.
  *
- * @see docs/exec-plans/grader-reliability.md — Phase 3
+ * @see docs/archive/exec-plans/grader-reliability.md — Phase 3
  */
 import type { Logger } from "../_vendor/ailf-core/index.d.ts";
 import { type GraderComparison } from "./grader-comparison.js";

package/dist/pipeline/grader-compare-runner.js

@@ -10,7 +10,7 @@
  * Migrated from lib/grader-compare.ts — no process.argv, no process.exit(),
  * no module-level constants. Accepts rootDir as parameter.
  *
- * @see docs/exec-plans/grader-reliability.md — Phase 3
+ * @see docs/archive/exec-plans/grader-reliability.md — Phase 3
  */
 import { existsSync, mkdirSync, readFileSync, writeFileSync } from "fs";
 import { join } from "path";
@@ -41,6 +41,12 @@ function classifyDimension(component) {
 }
 function detectFeatureArea(description) {
     const desc = description.toLowerCase();
+    if (desc.includes("portable text"))
+        return "portable-text";
+    if (desc.includes("content lake"))
+        return "content-lake";
+    if (desc.includes("image handling") || desc.includes("image asset"))
+        return "image-handling";
     if (desc.includes("studio"))
         return "studio-setup";
     if (desc.includes("visual") ||

package/dist/pipeline/grader-comparison.d.ts

@@ -10,7 +10,7 @@
  * This module has NO side effects — no file I/O, no API calls.
  * It operates on pre-collected data only.
  *
- * @see docs/exec-plans/grader-reliability.md — Phase 3
+ * @see docs/archive/exec-plans/grader-reliability.md — Phase 3
  */
 import type { ComparisonReport, ScoreSummary } from "./types.js";
 /** Per-dimension comparison between two graders */

package/dist/pipeline/grader-comparison.js

@@ -10,7 +10,7 @@
  * This module has NO side effects — no file I/O, no API calls.
  * It operates on pre-collected data only.
  *
- * @see docs/exec-plans/grader-reliability.md — Phase 3
+ * @see docs/archive/exec-plans/grader-reliability.md — Phase 3
  */
 import { compare } from "./compare.js";
 import { pearsonCorrelation } from "./grader-validation.js";

package/dist/pipeline/grader-consistency-runner.d.ts

@@ -12,7 +12,7 @@
  * Migrated from lib/grader-consistency.ts — no process.argv, no process.exit(),
  * no module-level constants.
  *
- * @see docs/exec-plans/grader-reliability.md — Phase 1
+ * @see docs/archive/exec-plans/grader-reliability.md — Phase 1
  */
 import { type Logger } from "../_vendor/ailf-core/index.d.ts";
 import type { RawPromptfooFile } from "./calculate-scores.js";

package/dist/pipeline/grader-consistency-runner.js

@@ -12,7 +12,7 @@
  * Migrated from lib/grader-consistency.ts — no process.argv, no process.exit(),
  * no module-level constants.
  *
- * @see docs/exec-plans/grader-reliability.md — Phase 1
+ * @see docs/archive/exec-plans/grader-reliability.md — Phase 1
  */
 import { existsSync, mkdirSync, readFileSync, writeFileSync } from "fs";
 import { join } from "path";
@@ -44,6 +44,12 @@ function classifyDimension(component) {
 // ---------------------------------------------------------------------------
 function detectFeatureArea(description) {
     const desc = description.toLowerCase();
+    if (desc.includes("portable text"))
+        return "portable-text";
+    if (desc.includes("content lake"))
+        return "content-lake";
+    if (desc.includes("image handling") || desc.includes("image asset"))
+        return "image-handling";
     if (desc.includes("studio"))
         return "studio-setup";
     if (desc.includes("visual") ||

package/dist/pipeline/grader-consistency.d.ts

@@ -10,7 +10,7 @@
  * This module has NO side effects — no file I/O, no API calls.
  * It operates on pre-collected data only.
  *
- * @see docs/exec-plans/grader-reliability.md — Phase 1
+ * @see docs/archive/exec-plans/grader-reliability.md — Phase 1
  */
 /** Per-dimension consistency aggregates */
 export interface DimensionConsistency {

package/dist/pipeline/grader-consistency.js

@@ -10,7 +10,7 @@
  * This module has NO side effects — no file I/O, no API calls.
  * It operates on pre-collected data only.
  *
- * @see docs/exec-plans/grader-reliability.md — Phase 1
+ * @see docs/archive/exec-plans/grader-reliability.md — Phase 1
  */
 // ---------------------------------------------------------------------------
 // Pure computation

package/dist/pipeline/grader-sensitivity-runner.d.ts

@@ -11,7 +11,7 @@
  * Migrated from lib/grader-sensitivity.ts — no process.argv, no process.exit(),
  * no module-level constants. Accepts rootDir as parameter.
  *
- * @see docs/exec-plans/grader-reliability.md — Phase 4
+ * @see docs/archive/exec-plans/grader-reliability.md — Phase 4
  */
 import type { Logger } from "../_vendor/ailf-core/index.d.ts";
 import { type GraderSensitivityResult } from "./grader-sensitivity.js";

package/dist/pipeline/grader-sensitivity-runner.js

@@ -11,7 +11,7 @@
  * Migrated from lib/grader-sensitivity.ts — no process.argv, no process.exit(),
  * no module-level constants. Accepts rootDir as parameter.
  *
- * @see docs/exec-plans/grader-reliability.md — Phase 4
+ * @see docs/archive/exec-plans/grader-reliability.md — Phase 4
  */
 import { existsSync, mkdirSync, readdirSync, readFileSync, writeFileSync, } from "fs";
 import { basename, join } from "path";

package/dist/pipeline/grader-sensitivity.d.ts

@@ -11,7 +11,7 @@
  *
  * This module has NO side effects — no file I/O, no API calls.
  *
- * @see docs/exec-plans/grader-reliability.md — Phase 4
+ * @see docs/archive/exec-plans/grader-reliability.md — Phase 4
  */
 /** Sensitivity broken down by degradation type */
 export interface DegradationSensitivity {

package/dist/pipeline/grader-sensitivity.js

@@ -11,7 +11,7 @@
  *
  * This module has NO side effects — no file I/O, no API calls.
  *
- * @see docs/exec-plans/grader-reliability.md — Phase 4
+ * @see docs/archive/exec-plans/grader-reliability.md — Phase 4
  */
 // ---------------------------------------------------------------------------
 // Pure computation

package/dist/pipeline/grader-validate-runner.d.ts

@@ -11,7 +11,7 @@
  * Migrated from lib/grader-validate.ts — no process.argv, no process.exit(),
  * no module-level constants. Accepts rootDir as parameter.
  *
- * @see docs/exec-plans/grader-reliability.md — Phase 2
+ * @see docs/archive/exec-plans/grader-reliability.md — Phase 2
  */
 import type { Logger } from "../_vendor/ailf-core/index.d.ts";
 import { type GraderValidation } from "./grader-validation.js";

package/dist/pipeline/grader-validate-runner.js

@@ -11,7 +11,7 @@
  * Migrated from lib/grader-validate.ts — no process.argv, no process.exit(),
  * no module-level constants. Accepts rootDir as parameter.
  *
- * @see docs/exec-plans/grader-reliability.md — Phase 2
+ * @see docs/archive/exec-plans/grader-reliability.md — Phase 2
  */
 import { existsSync, mkdirSync, readFileSync, readdirSync, writeFileSync, } from "fs";
 import { join } from "path";
@@ -43,7 +43,7 @@ function loadReferenceGrades(rootDir) {
     if (!existsSync(refsDir)) {
         throw new Error(`Reference grades directory not found: ${refsDir}. ` +
             "Create canonical/grader-references/ with YAML reference files. " +
-            "See docs/exec-plans/grader-reliability.md — Phase 2.");
+            "See docs/archive/exec-plans/grader-reliability.md — Phase 2.");
     }
     const files = readdirSync(refsDir)
         .filter((f) => f.endsWith(".yaml") || f.endsWith(".yml"))

package/dist/pipeline/grader-validation.d.ts

@@ -11,7 +11,7 @@
  *
  * This module has NO side effects — no file I/O, no API calls.
  *
- * @see docs/exec-plans/grader-reliability.md — Phase 2
+ * @see docs/archive/exec-plans/grader-reliability.md — Phase 2
  */
 /** Quality label for a correlation value */
 export type CorrelationQuality = "excellent" | "good" | "moderate" | "poor" | "very-poor";