fix(AIC-3576): record deferred judge scores under the evaluated config

packages/client/src/__tests__/judges.test.ts

@@ -386,44 +386,42 @@ describe('runJudge result trackData', () => {
     vi.clearAllMocks();
   });
 
-  it('does not inherit the parent modelKey/modelVersion when the judge trackData has none', async () => {
+  it('returns the evaluated (parent) config identity, identical to the inline track payload', async () => {
+    // The judge's own trackData differs on every identity key; none of it may leak into the result.
     mockExecuteAndTrack.mockResolvedValue({
       usage: { input: 1, output: 1, total: 2 },
       response: '{"score":0.9,"reasoning":"good"}',
       trackData: {
         runId: 'judge-run',
         configKey: 'judge-key',
         variationKey: 'j1',
-        version: 1,
+        version: 3,
         modelName: 'claude',
         providerName: 'Anthropic',
+        modelKey: 'judge-model',
+        modelVersion: 2,
       },
     });
-    const result = await runJudge(makeTask(parentTrackData), [makeHandler()]);
-    expect(result).not.toBeNull();
-    expect('modelKey' in result!.trackData).toBe(false);
-    expect('modelVersion' in result!.trackData).toBe(false);
-    expect(result!.trackData.judgeConfigKey).toBe('judge-key');
-    expect(result!.trackData.graphKey).toBe('g1');
-  });
+    mockExtractVariation.mockResolvedValue({ config: mockJudgeConfig, meta: mockJudgeMeta });
+    const track = vi.fn();
+    mockGetClient.mockReturnValue({ track });
 
-  it('keeps the judge own modelKey/modelVersion over the parent values', async () => {
-    mockExecuteAndTrack.mockResolvedValue({
-      usage: { input: 1, output: 1, total: 2 },
-      response: '{"score":0.9,"reasoning":"good"}',
-      trackData: {
-        runId: 'judge-run',
-        configKey: 'judge-key',
-        variationKey: 'j1',
-        version: 1,
-        modelName: 'claude',
-        providerName: 'Anthropic',
-        modelKey: 'judge-model',
-        modelVersion: 2,
+    await runJudges({
+      config: {
+        model: { name: 'gpt-4o' },
+        provider: { name: 'OpenAI' },
+        instructions: 'Be helpful.',
+        judgeConfiguration: { judges: [{ key: 'judge-key', samplingRate: 1 }] },
       },
+      userContext: mockContext,
+      handler: makeHandler(),
+      llmResponse: 'response',
+      baseTrackData: parentTrackData,
     });
+    const inlineTrackData = track.mock.calls[0][2];
+
     const result = await runJudge(makeTask(parentTrackData), [makeHandler()]);
-    expect(result!.trackData.modelKey).toBe('judge-model');
-    expect(result!.trackData.modelVersion).toBe(2);
+    expect(result!.trackData).toEqual({ ...parentTrackData, judgeConfigKey: 'judge-key' });
+    expect(result!.trackData).toEqual(inlineTrackData);
   });
 });

packages/client/src/judges.ts

@@ -14,12 +14,7 @@ import type {
   TrackData,
   VariationMeta,
 } from './types.js';
-import {
-  collapseMessagesToInstructions,
-  normalizeMode,
-  omitModelStamps,
-  parseJSONWithPossibleFences,
-} from './utils.js';
+import { collapseMessagesToInstructions, normalizeMode, parseJSONWithPossibleFences } from './utils.js';
 
 /**
  * Resolves a judge's own AI Config, returning `null` instead of throwing when it
@@ -341,11 +336,7 @@ export const runJudge = async (task: JudgeTask, handlers: ProviderHandler[]): Pr
   const messageHistory = [actualOutput, FORMATTING_INSTRUCTIONS].join('\n\n');
 
   return withJudgeEvaluation(configKey, async (recordEvaluation) => {
-    const {
-      usage,
-      response: rawResponse,
-      trackData,
-    } = await executeAndTrack({
+    const { usage, response: rawResponse } = await executeAndTrack({
       configKey,
       config: effectiveConfig,
       meta: judgeMeta,
@@ -371,9 +362,8 @@ export const runJudge = async (task: JudgeTask, handlers: ProviderHandler[]): Pr
       score,
       response: reasoning,
       usage,
-      // A judge without a pinned model config must not inherit the parent's
-      // modelKey / modelVersion; other parent-only keys (graphKey, ...) still carry over.
-      trackData: { ...omitModelStamps(parentTrackData), ...trackData, judgeConfigKey: configKey },
+      // The score belongs to the evaluated (parent) config, exactly as in the inline path.
+      trackData: { ...parentTrackData, judgeConfigKey: configKey },
     };
   });
 };