@sanity/ailf 0.5.0 → 2.0.0

package/dist/pipeline/plan.js

@@ -9,13 +9,15 @@
  *
  * @see docs/exec-plans/execution-preview.md
  */
-import { existsSync, readFileSync, readdirSync, statSync } from "fs";
+import { existsSync, readdirSync, statSync } from "fs";
 import { resolve } from "path";
-import { load } from "js-yaml";
 import { lookupPricing } from "../agent-observer/pricing.js";
 import { RepoTaskSource } from "../adapters/task-sources/repo-task-source.js";
+import { loadAllTsTaskFiles } from "../adapters/task-sources/task-file-loader.js";
 import { lookupCache } from "./cache.js";
-import { loadAndExpandTasks } from "./expand-tasks.js";
+import { compileLiteracyTasks } from "./compiler/literacy-bridge.js";
+import { tryLoadConfigFile } from "./compiler/config-loader.js";
+import { LiteracyVariant } from "./normalize-mode.js";
 import { validateConfiguration } from "./validate.js";
 /**
  * Known promptfoo provider prefixes — stripped to get the raw model name.
@@ -39,38 +41,37 @@ function extractModelName(id) {
     return parts.length > 1 ? parts.slice(1).join(":") : id;
 }
 function loadModelsFile(rootDir) {
-    const modelsPath = resolve(rootDir, "config", "models.yaml");
-    if (!existsSync(modelsPath))
-        return null;
-    try {
-        const raw = readFileSync(modelsPath, "utf-8");
-        return load(raw);
-    }
-    catch {
-        return null;
-    }
+    const result = tryLoadConfigFile("models", rootDir);
+    return result?.data ?? null;
 }
 /**
- * Map eval mode to the model "modes" array values from models.yaml.
- * Baseline mode maps to "baseline"; agentic maps to both naive and optimized.
+ * Map eval mode + variant to the model "modes" array values from models config.
+ *
+ * Literacy mode uses the variant to determine which model sub-modes match.
+ * Non-literacy modes accept all models by default (filtering is done
+ * elsewhere for those modes).
  */
-function modeMatchesModelModes(mode, modelModes) {
+function modeMatchesModelModes(mode, modelModes, variant) {
     if (!modelModes || modelModes.length === 0)
         return true;
-    switch (mode) {
-        case "agentic":
-            return (modelModes.includes("agentic-naive") ||
-                modelModes.includes("agentic-optimized"));
-        case "baseline":
-            return modelModes.includes("baseline");
-        case "full":
-            // Full mode uses all models — a model matches if it's in any sub-mode
-            return (modelModes.includes("baseline") ||
-                modelModes.includes("agentic-naive") ||
-                modelModes.includes("agentic-optimized"));
-        case "observed":
-            return modelModes.includes("observed");
+    if (mode === "literacy") {
+        switch (variant) {
+            case LiteracyVariant.AGENTIC:
+                return (modelModes.includes("agentic-naive") ||
+                    modelModes.includes("agentic-optimized"));
+            case LiteracyVariant.OBSERVED:
+                return modelModes.includes(LiteracyVariant.OBSERVED);
+            case LiteracyVariant.FULL:
+                return (modelModes.includes(LiteracyVariant.STANDARD) ||
+                    modelModes.includes("agentic-naive") ||
+                    modelModes.includes("agentic-optimized"));
+            case LiteracyVariant.STANDARD:
+            default:
+                return modelModes.includes(LiteracyVariant.STANDARD);
+        }
     }
+    // Non-literacy modes accept all models by default
+    return true;
 }
 // ---------------------------------------------------------------------------
 // Cost estimation
@@ -131,46 +132,103 @@ export async function buildPipelinePlan(opts, rootDir) {
     let totalTests = 0;
     let tasks = [];
     let repoTaskCount;
+    // -----------------------------------------------------------------------
+    // Load and compile tasks — unified path for all modes
+    // -----------------------------------------------------------------------
     try {
-        const { entries } = loadAndExpandTasks(rootDir, filter);
-        totalTests = entries.length;
-        tasks = entries.map((entry) => {
-            const desc = typeof entry.description === "string" ? entry.description : "(unknown)";
-            const isBaseline = desc.includes("[Baseline]") ||
-                desc.endsWith("(baseline)") ||
-                (Array.isArray(entry.prompts) && entry.prompts.includes("without-docs"));
-            return {
-                description: desc,
-                variant: isBaseline ? "baseline" : "gold",
-            };
-        });
+        const modelsForCompile = loadModelsFile(rootDir);
+        const graderProvider = modelsForCompile?.grader?.id ?? "openai:chat:gpt-4o";
+        const modelEntries = (modelsForCompile?.models ?? []).map((m) => ({ id: m.id, label: m.label }));
+        // Load *.task.ts files from tasks/<mode>/
+        const modeTasksDir = resolve(rootDir, "tasks", opts.mode);
+        if (existsSync(modeTasksDir)) {
+            const rawTasks = await loadAllTsTaskFiles(modeTasksDir);
+            if (rawTasks.length > 0) {
+                // Dynamic import of the handler module
+                const handlerModulePath = `./compiler/mode-handlers/${opts.mode}/index.js`;
+                const mod = await import(handlerModulePath);
+                const handler = mod.handler;
+                for (const rawFile of rawTasks) {
+                    for (const taskDef of rawFile.tasks) {
+                        const task = taskDef;
+                        // Apply area/task/tag filter
+                        if (filter) {
+                            if (filter.areas?.length &&
+                                !filter.areas
+                                    .map((a) => a.toLowerCase())
+                                    .includes((task.area ?? "").toLowerCase()))
+                                continue;
+                            if (filter.taskIds?.length && !filter.taskIds.includes(task.id))
+                                continue;
+                            if (filter.tags?.length &&
+                                (!task.tags || !task.tags.some((t) => filter.tags.includes(t))))
+                                continue;
+                        }
+                        const result = handler.compileTask(task, {
+                            rootDir,
+                            graderProvider,
+                            models: modelEntries,
+                            // For literacy mode, pass the variant as evalMode
+                            ...(opts.mode === "literacy"
+                                ? {
+                                    evalMode: opts.variant === LiteracyVariant.AGENTIC
+                                        ? LiteracyVariant.AGENTIC
+                                        : LiteracyVariant.STANDARD,
+                                }
+                                : {}),
+                        });
+                        totalTests += result.tests.length;
+                        for (const test of result.tests) {
+                            const desc = typeof test.description === "string"
+                                ? test.description
+                                : (taskDef.id ?? "unknown");
+                            const isBaseline = desc.includes("[Baseline]") || desc.endsWith("(baseline)");
+                            tasks.push({
+                                description: desc,
+                                variant: isBaseline
+                                    ? LiteracyVariant.STANDARD
+                                    : "gold",
+                            });
+                        }
+                    }
+                }
+            }
+        }
     }
-    catch {
-        errors.push("Failed to expand tasks — check task YAML files");
+    catch (err) {
+        const detail = err instanceof Error ? err.message : String(err);
+        errors.push(`Failed to compile tasks: ${detail}`);
     }
     // Scan repo tasks path for additional task count (preview only)
     if (opts.repoTasksPath) {
         try {
             const repoSource = new RepoTaskSource(opts.repoTasksPath);
-            const repoTasks = await repoSource.loadTasks(filter);
+            // Type-narrow to literacy tasks — compileLiteracyTasks accepts LiteracyTaskDefinition[]
+            const repoTasks = (await repoSource.loadTasks(filter)).filter((t) => t.mode === "literacy");
             repoTaskCount = repoTasks.length;
             if (repoTaskCount > 0) {
-                // Expand repo tasks to estimate test entries
-                const { expandTaskDefinitions } = await import("./expand-tasks.js");
-                const { entries: repoEntries } = expandTaskDefinitions(repoTasks, rootDir, opts.mode === "agentic" ? "agentic" : "baseline");
-                totalTests += repoEntries.length;
-                for (const entry of repoEntries) {
-                    const desc = typeof entry.description === "string"
-                        ? entry.description
-                        : "(unknown)";
-                    const isBaseline = desc.includes("[Baseline]") ||
-                        desc.endsWith("(baseline)") ||
-                        (Array.isArray(entry.prompts) &&
-                            entry.prompts.includes("without-docs"));
-                    tasks.push({
-                        description: desc,
-                        variant: isBaseline ? "baseline" : "gold",
-                    });
+                const modelsForCompile = loadModelsFile(rootDir);
+                const graderProvider = modelsForCompile?.grader?.id ?? "openai:chat:gpt-4o";
+                const compileResult = compileLiteracyTasks(repoTasks, {
+                    rootDir,
+                    evalMode: opts.variant === LiteracyVariant.AGENTIC
+                        ? LiteracyVariant.AGENTIC
+                        : LiteracyVariant.STANDARD,
+                    graderProvider,
+                    models: (modelsForCompile?.models ?? []).map((m) => ({ id: m.id, label: m.label })),
+                });
+                totalTests += compileResult.totalTests;
+                for (const { taskId, result } of compileResult.tasks) {
+                    for (const test of result.tests) {
+                        const desc = typeof test.description === "string" ? test.description : taskId;
+                        const isBaseline = desc.includes("[Baseline]") || desc.endsWith("(baseline)");
+                        tasks.push({
+                            description: desc,
+                            variant: isBaseline
+                                ? LiteracyVariant.STANDARD
+                                : "gold",
+                        });
+                    }
                 }
             }
         }
@@ -186,11 +244,11 @@ export async function buildPipelinePlan(opts, rootDir) {
     const models = [];
     let graderModelName = "";
     if (modelsFile) {
-        const activeModels = modelsFile.models.filter((m) => modeMatchesModelModes(opts.mode, m.modes));
+        const activeModels = modelsFile.models.filter((m) => modeMatchesModelModes(opts.mode, m.modes, opts.variant));
         // For agentic mode, each model appears twice (naive + optimized)
         for (const m of activeModels) {
             const modelName = extractModelName(m.id);
-            if (opts.mode === "agentic") {
+            if (opts.variant === LiteracyVariant.AGENTIC) {
                 if (m.modes?.includes("agentic-naive")) {
                     models.push({
                         id: m.id,
@@ -518,16 +576,16 @@ function collectFilesCreated(opts) {
 // ---------------------------------------------------------------------------
 function collectFilesRead(rootDir, _mode) {
     const files = [
-        "config/models.yaml",
-        "config/rubrics.yaml",
-        "config/prompts.yaml",
-        "config/sources.yaml",
+        "config/models.ts",
+        "config/rubrics.ts",
+        "config/prompts.ts",
+        "config/sources.ts",
     ];
     // Task files
     const tasksDir = resolve(rootDir, "tasks");
     if (existsSync(tasksDir)) {
         const taskFiles = readdirSync(tasksDir)
-            .filter((f) => (f.endsWith(".yaml") || f.endsWith(".yml")) && !f.startsWith("."))
+            .filter((f) => /\.(yaml|yml|task\.ts|task\.js)$/.test(f) && !f.startsWith("."))
             .sort();
         for (const f of taskFiles)
             files.push(`tasks/${f}`);
@@ -551,11 +609,11 @@ function collectFilesRead(rootDir, _mode) {
             files.push(`canonical/reference-solutions/${f}`);
     }
     // Thresholds (if readiness is involved)
-    if (existsSync(resolve(rootDir, "config", "thresholds.yaml"))) {
-        files.push("config/thresholds.yaml");
+    if (existsSync(resolve(rootDir, "config", "thresholds.ts"))) {
+        files.push("config/thresholds.ts");
     }
-    if (existsSync(resolve(rootDir, "config", "features.yaml"))) {
-        files.push("config/features.yaml");
+    if (existsSync(resolve(rootDir, "config", "features.ts"))) {
+        files.push("config/features.ts");
     }
     return [...new Set(files)].sort();
 }
@@ -616,16 +674,14 @@ function estimateCost(testCount, models, graderModelName, rubricAssertionsPerTas
 // Used by the plan builder without importing the full type to avoid circular deps.
 // ---------------------------------------------------------------------------
 function estimateRubricAssertionsPerTask(rootDir) {
-    // Load rubrics.yaml and count the default template set.
+    // Load rubrics config and count the default template set.
     // In practice, most tasks have 2-4 rubric assertions.
-    const rubricsPath = resolve(rootDir, "config", "rubrics.yaml");
-    if (!existsSync(rubricsPath))
+    const result = tryLoadConfigFile("rubrics", rootDir);
+    if (!result)
         return 2; // conservative default
     try {
-        const raw = readFileSync(rubricsPath, "utf-8");
-        const data = load(raw);
-        const templateCount = data?.templates
-            ? Object.keys(data.templates).length
+        const templateCount = result.data?.templates
+            ? Object.keys(result.data.templates).length
             : 2;
         // Most tasks use 2-3 of the available templates
         return Math.min(templateCount, 3);

package/dist/pipeline/pr-comment.js

@@ -320,6 +320,8 @@ function generateComment(summary, options = {}) {
                         ? "📉"
                         : "➡️";
                 const d = (n) => n > 0 ? `+${Math.round(n)}` : String(Math.round(n));
+                // TODO(multi-mode): These dimension keys are literacy-specific.
+                // For other modes, iterate Object.entries(a.dimensions) dynamically.
                 lines.push(`| ${a.area} | ${a.baseline} | ${a.experiment} | ${icon} ${d(a.delta)} | ${d(a.dimensions.taskCompletion.delta)} | ${d(a.dimensions.codeCorrectness.delta)} | ${d(a.dimensions.docCoverage.delta)} |`);
             }
         }

package/dist/pipeline/profile-resolution.d.ts

@@ -1,13 +1,16 @@
 /**
  * pipeline/profile-resolution.ts
  *
- * Resolves the correct weight profile for a given (mode, variant) pair.
- * The scoring engine calls this to determine which dimensions and weights
- * apply to each test entry's composite score.
+ * Resolves the correct weight profile for a given (mode, perspective, variant)
+ * tuple. The scoring engine calls this to determine which dimensions and
+ * weights apply to each test entry's composite score.
  *
  * Resolution order:
- *   1. Explicit binding: mode-profiles.<mode>.<variant> → profile name
- *   2. Fallback: the "default" profile
+ *   1. Nested binding (variant provided):
+ *      mode-profiles.<mode>.<variant>.<perspective> → profile name
+ *   2. Flat binding (no variant):
+ *      mode-profiles.<mode>.<perspective> → profile name
+ *   3. Fallback: the "default" profile
  *
  * Supports both the new `profiles` format and the legacy flat `weights`
  * format (treated as a single profile named "default").
@@ -23,17 +26,22 @@ import type { RubricConfig, WeightProfile } from "../_vendor/ailf-core/index.d.t
  */
 export declare function resolveProfiles(config: RubricConfig): Record<string, WeightProfile>;
 /**
- * Resolve the weight profile for a specific (mode, variant) pair.
+ * Resolve the weight profile for a specific (mode, perspective, variant) tuple.
  *
- * @param mode    - Evaluation mode (e.g., "baseline", "agentic", "agent-task")
- * @param variant - Entry variant: "gold" (with docs) or "baseline" (without docs)
- * @param config  - Parsed rubrics.yaml config
+ * @param mode        - Canonical mode (e.g., "literacy", "mcp-server")
+ * @param perspective - Entry perspective: "gold" (with docs) or "baseline" (without docs)
+ * @param config      - Parsed rubrics config
+ * @param variant     - Optional variant within the mode (e.g., "baseline", "agentic" for literacy)
  * @returns The resolved weight profile (dimension → weight map)
  *
  * @example
- *   resolveProfile("baseline", "gold", config)     // → default profile
- *   resolveProfile("baseline", "baseline", config)  // → output-only profile
- *   resolveProfile("agentic", "gold", config)       // → default profile
- *   resolveProfile("unknown-mode", "gold", config)  // → default (fallback)
+ *   // Nested: literacy mode with variant sub-keys
+ *   resolveProfile("literacy", "gold", config, "baseline")     // → default profile
+ *   resolveProfile("literacy", "baseline", config, "baseline")  // → output-only profile
+ *   resolveProfile("literacy", "gold", config, "agentic")       // → default profile
+ *
+ *   // Flat: non-literacy modes
+ *   resolveProfile("mcp-server", "gold", config)                // → mcp-behavior profile
+ *   resolveProfile("unknown-mode", "gold", config)              // → default (fallback)
  */
-export declare function resolveProfile(mode: string, variant: string, config: RubricConfig): WeightProfile;
+export declare function resolveProfile(mode: string, perspective: string, config: RubricConfig, variant?: string): WeightProfile;

package/dist/pipeline/profile-resolution.js

@@ -1,13 +1,16 @@
 /**
  * pipeline/profile-resolution.ts
  *
- * Resolves the correct weight profile for a given (mode, variant) pair.
- * The scoring engine calls this to determine which dimensions and weights
- * apply to each test entry's composite score.
+ * Resolves the correct weight profile for a given (mode, perspective, variant)
+ * tuple. The scoring engine calls this to determine which dimensions and
+ * weights apply to each test entry's composite score.
  *
  * Resolution order:
- *   1. Explicit binding: mode-profiles.<mode>.<variant> → profile name
- *   2. Fallback: the "default" profile
+ *   1. Nested binding (variant provided):
+ *      mode-profiles.<mode>.<variant>.<perspective> → profile name
+ *   2. Flat binding (no variant):
+ *      mode-profiles.<mode>.<perspective> → profile name
+ *   3. Fallback: the "default" profile
  *
  * Supports both the new `profiles` format and the legacy flat `weights`
  * format (treated as a single profile named "default").
@@ -29,31 +32,50 @@ export function resolveProfiles(config) {
         return { default: config.weights };
     }
     // Schema validation should prevent this, but be defensive
-    throw new Error("rubrics.yaml has neither 'profiles' nor 'weights' — cannot resolve scoring profiles");
+    throw new Error("rubrics config has neither 'profiles' nor 'weights' — cannot resolve scoring profiles");
 }
 /**
- * Resolve the weight profile for a specific (mode, variant) pair.
+ * Resolve the weight profile for a specific (mode, perspective, variant) tuple.
  *
- * @param mode    - Evaluation mode (e.g., "baseline", "agentic", "agent-task")
- * @param variant - Entry variant: "gold" (with docs) or "baseline" (without docs)
- * @param config  - Parsed rubrics.yaml config
+ * @param mode        - Canonical mode (e.g., "literacy", "mcp-server")
+ * @param perspective - Entry perspective: "gold" (with docs) or "baseline" (without docs)
+ * @param config      - Parsed rubrics config
+ * @param variant     - Optional variant within the mode (e.g., "baseline", "agentic" for literacy)
  * @returns The resolved weight profile (dimension → weight map)
  *
  * @example
- *   resolveProfile("baseline", "gold", config)     // → default profile
- *   resolveProfile("baseline", "baseline", config)  // → output-only profile
- *   resolveProfile("agentic", "gold", config)       // → default profile
- *   resolveProfile("unknown-mode", "gold", config)  // → default (fallback)
+ *   // Nested: literacy mode with variant sub-keys
+ *   resolveProfile("literacy", "gold", config, "baseline")     // → default profile
+ *   resolveProfile("literacy", "baseline", config, "baseline")  // → output-only profile
+ *   resolveProfile("literacy", "gold", config, "agentic")       // → default profile
+ *
+ *   // Flat: non-literacy modes
+ *   resolveProfile("mcp-server", "gold", config)                // → mcp-behavior profile
+ *   resolveProfile("unknown-mode", "gold", config)              // → default (fallback)
  */
-export function resolveProfile(mode, variant, config) {
+export function resolveProfile(mode, perspective, config, variant) {
     const profiles = resolveProfiles(config);
     const modeProfiles = config["mode-profiles"];
-    // Look up explicit binding: mode-profiles.<mode>.<variant> → profile name
-    const profileName = modeProfiles?.[mode]?.[variant];
+    const modeEntry = modeProfiles?.[mode];
+    let profileName;
+    if (modeEntry && variant) {
+        // Nested lookup: mode-profiles.<mode>.<variant>.<perspective>
+        const variantEntry = modeEntry[variant];
+        if (typeof variantEntry === "object" && variantEntry !== null) {
+            profileName = variantEntry[perspective];
+        }
+    }
+    if (!profileName && modeEntry) {
+        // Flat lookup: mode-profiles.<mode>.<perspective>
+        const directEntry = modeEntry[perspective];
+        if (typeof directEntry === "string") {
+            profileName = directEntry;
+        }
+    }
     if (profileName) {
         const profile = profiles[profileName];
         if (!profile) {
-            throw new Error(`mode-profiles.${mode}.${variant} references profile "${profileName}" ` +
+            throw new Error(`mode-profiles.${mode}.${variant ? variant + "." : ""}${perspective} references profile "${profileName}" ` +
                 `which does not exist. Available profiles: ${Object.keys(profiles).join(", ")}`);
         }
         return profile;
@@ -61,7 +83,7 @@ export function resolveProfile(mode, variant, config) {
     // Fall back to "default" profile
     const defaultProfile = profiles["default"];
     if (!defaultProfile) {
-        throw new Error(`No scoring profile found for mode="${mode}" variant="${variant}" ` +
+        throw new Error(`No scoring profile found for mode="${mode}" perspective="${perspective}" ` +
             `and no "default" profile exists. ` +
             `Available profiles: ${Object.keys(profiles).join(", ")}`);
     }

package/dist/pipeline/provenance.d.ts

@@ -42,7 +42,7 @@ export interface ProvenanceInput {
     promptfooUrl?: string;
     /** Per-mode Promptfoo share URLs */
     promptfooUrls?: PromptfooUrlEntry[];
-    /** Path to the package root (for reading models.yaml) */
+    /** Path to the package root (for reading config/models) */
     rootDir: string;
     /** Report ID that triggered this re-run (becomes lineage.rerunOf) */
     sourceReportId?: string;
@@ -58,7 +58,7 @@ export interface ProvenanceInput {
  *
  * Assembles provenance from:
  * - Pipeline options (mode, source, areas, tasks)
- * - config/models.yaml (model list, grader)
+ * - config/models.ts (model list, grader)
  * - Environment variables (CI metadata, trigger detection)
  * - Optional metadata (context hash, Promptfoo URL)
  */

package/dist/pipeline/provenance.js

@@ -11,16 +11,14 @@
  * @see docs/design-docs/report-store/domain-model.md
  * @see docs/design-docs/report-store/architecture.md — Provenance collection
  */
-import { readFileSync } from "fs";
-import { resolve } from "path";
-import { load } from "js-yaml";
 import { ConsoleLogger } from "../adapters/loggers/index.js";
+import { tryLoadConfigFile } from "./compiler/config-loader.js";
 /**
  * Build a ReportProvenance object from pipeline context.
  *
  * Assembles provenance from:
  * - Pipeline options (mode, source, areas, tasks)
- * - config/models.yaml (model list, grader)
+ * - config/models.ts (model list, grader)
  * - Environment variables (CI metadata, trigger detection)
  * - Optional metadata (context hash, Promptfoo URL)
  */
@@ -168,20 +166,17 @@ function detectTrigger() {
 // Model config loading
 // ---------------------------------------------------------------------------
 /**
- * Load config/models.yaml to extract model list and grader info.
+ * Load config/models to extract model list and grader info.
  * Falls back to a minimal config if the file can't be read.
  */
 function loadModelsConfig(rootDir, log) {
-    try {
-        const content = readFileSync(resolve(rootDir, "config", "models.yaml"), "utf-8");
-        return load(content);
-    }
-    catch {
-        log.warn("Could not read config/models.yaml for provenance");
-        return {
-            defaults: {},
-            grader: { id: "unknown" },
-            models: [],
-        };
-    }
+    const result = tryLoadConfigFile("models", rootDir);
+    if (result)
+        return result.data;
+    log.warn("Could not read config/models for provenance");
+    return {
+        defaults: {},
+        grader: { id: "unknown" },
+        models: [],
+    };
 }

package/dist/pipeline/release-report.js

@@ -133,8 +133,8 @@ export function formatReleaseImpactConsole(report) {
                 const docs = task.attributedDocs.length > 0
                     ? task.attributedDocs.join(", ")
                     : "(unattributed)";
-                const deltaStr = task.delta >= 0 ? `+${task.delta.toFixed(1)}` : task.delta.toFixed(1);
-                lines.push(`  ${docs.padEnd(32)} | ${area.area.padEnd(16)} | ${task.taskId.padEnd(23)} | ${deltaStr}`);
+                const taskDeltaStr = task.delta >= 0 ? `+${task.delta.toFixed(1)}` : task.delta.toFixed(1);
+                lines.push(`  ${docs.padEnd(32)} | ${area.area.padEnd(16)} | ${task.taskId.padEnd(23)} | ${taskDeltaStr}`);
             }
         }
         lines.push("");
@@ -194,9 +194,9 @@ export function formatReleaseImpactMarkdown(report) {
                 const docs = task.attributedDocs.length > 0
                     ? task.attributedDocs.map((d) => `\`${d}\``).join(", ")
                     : "—";
-                const deltaStr = task.delta >= 0 ? `+${task.delta.toFixed(1)}` : task.delta.toFixed(1);
+                const taskDeltaStr = task.delta >= 0 ? `+${task.delta.toFixed(1)}` : task.delta.toFixed(1);
                 const regressIcon = area.regressed ? " ⚠️" : "";
-                lines.push(`| ${docs} | ${area.area} | ${task.taskId} | ${deltaStr}${regressIcon} |`);
+                lines.push(`| ${docs} | ${area.area} | ${task.taskId} | ${taskDeltaStr}${regressIcon} |`);
             }
         }
         lines.push("");

package/dist/pipeline/repo-threshold-evaluator.d.ts

@@ -5,7 +5,7 @@
  * definitions (`.ailf/tasks/*.yaml` → `execution.threshold`).
  *
  * This is distinct from the readiness-gate threshold system in
- * `config/thresholds.yaml`. Repo thresholds are per-task, defined by
+ * `config/thresholds`. Repo thresholds are per-task, defined by
  * the product team, and drive PR check pass/fail status. Framework
  * thresholds are per-area, defined by the AILF team, and drive
  * readiness reports.

package/dist/pipeline/repo-threshold-evaluator.js

@@ -5,7 +5,7 @@
  * definitions (`.ailf/tasks/*.yaml` → `execution.threshold`).
  *
  * This is distinct from the readiness-gate threshold system in
- * `config/thresholds.yaml`. Repo thresholds are per-task, defined by
+ * `config/thresholds`. Repo thresholds are per-task, defined by
  * the product team, and drive PR check pass/fail status. Framework
  * thresholds are per-area, defined by the AILF team, and drive
  * readiness reports.

package/dist/pipeline/rubric-loader.d.ts

@@ -0,0 +1,20 @@
+/**
+ * pipeline/rubric-loader.ts — Load and validate rubric config.
+ *
+ * Extracted from the legacy expand-tasks.ts so that callers (e.g.,
+ * calculate-scores.ts) can load rubric templates without pulling in
+ * the deprecated task expansion machinery.
+ *
+ * @see packages/eval/config/rubrics.ts — the rubric configuration
+ * @see packages/core/src/schemas/pipeline.ts — RubricConfigSchema
+ */
+import { type RubricConfig } from "../_vendor/ailf-core/index.d.ts";
+/**
+ * Load and validate config/rubrics from the given root directory.
+ * Caches the result for subsequent calls with the same rootDir.
+ */
+export declare function loadRubricTemplates(rootDir: string): RubricConfig;
+/**
+ * Reset the rubric config cache. Useful in tests.
+ */
+export declare function resetRubricCache(): void;

package/dist/pipeline/rubric-loader.js

@@ -0,0 +1,37 @@
+/**
+ * pipeline/rubric-loader.ts — Load and validate rubric config.
+ *
+ * Extracted from the legacy expand-tasks.ts so that callers (e.g.,
+ * calculate-scores.ts) can load rubric templates without pulling in
+ * the deprecated task expansion machinery.
+ *
+ * @see packages/eval/config/rubrics.ts — the rubric configuration
+ * @see packages/core/src/schemas/pipeline.ts — RubricConfigSchema
+ */
+import { RubricConfigSchema } from "../_vendor/ailf-core/index.js";
+import { loadConfigFile } from "./compiler/config-loader.js";
+let cachedRubricConfig = null;
+/**
+ * Load and validate config/rubrics from the given root directory.
+ * Caches the result for subsequent calls with the same rootDir.
+ */
+export function loadRubricTemplates(rootDir) {
+    if (cachedRubricConfig)
+        return cachedRubricConfig;
+    const { data } = loadConfigFile("rubrics", rootDir);
+    const result = RubricConfigSchema.safeParse(data);
+    if (!result.success) {
+        const messages = result.error.issues
+            .map((i) => `  [${i.path.join(".")}]: ${i.message}`)
+            .join("\n");
+        throw new Error(`Invalid config/rubrics:\n${messages}`);
+    }
+    cachedRubricConfig = result.data;
+    return result.data;
+}
+/**
+ * Reset the rubric config cache. Useful in tests.
+ */
+export function resetRubricCache() {
+    cachedRubricConfig = null;
+}