Add eee source record

components/eval-detail.tsx

@@ -2038,6 +2038,7 @@ export function EvalDetail({
                                       score={formatRawScore(modelResult.score, lb.metric_config.unit)}
                                       sourceUrl={modelResult.source_metadata.source_url}
                                       sourceRecordUrl={modelResult.source_record_url}
+                                      eeeRecordUrl={modelResult.eee_record_url}
                                     />
                                   </div>
                                 </div>
@@ -2850,6 +2851,7 @@ function MultiMetricLeaderboard({
                             score={formatRawScore(matchingResult.score, summary.metric_config.unit)}
                             sourceUrl={matchingResult.source_metadata.source_url}
                             sourceRecordUrl={matchingResult.source_record_url}
+                            eeeRecordUrl={matchingResult.eee_record_url}
                           />
                         </div>
                       </div>

lib/eval-processing.ts

@@ -39,6 +39,8 @@ export interface ModelResultForBenchmark {
   result: EvaluationResult
   /** URL to the underlying record JSON in the upstream HF dataset, when known. */
   source_record_url?: string
+  /** Deep-link to the raw EEE_datastore source record this score came from, when known. */
+  eee_record_url?: string
   aggregate_components?: Array<{
     evaluation_id: string
     composite_benchmark_key: string

lib/view-data.ts

@@ -143,6 +143,7 @@ const EVAL_CELL_JOIN_COLUMNS = `
   CAST(to_json(r.source_metadata) AS VARCHAR) AS source_metadata,
   CAST(to_json(r.source_data) AS VARCHAR) AS source_data,
   r.source_record_url,
+  r.eee_record_url,
   CAST(to_json(r.eval_library) AS VARCHAR) AS eval_library,
   CAST(to_json(r.aggregate_components) AS VARCHAR) AS aggregate_components,
   CAST(to_json(r.evalcards_annotations) AS VARCHAR) AS evalcards_annotations,
@@ -559,6 +560,7 @@ function reshapeCellToModelResult(row: Row): ModelResultForBenchmark {
     source_metadata: sourceMetadataFromRow(row),
     source_data: sourceDataFromRow(row),
     source_record_url: optionalString(row.source_record_url),
+    eee_record_url: optionalString(row.eee_record_url),
     aggregate_components: asArray<NonNullable<ModelResultForBenchmark["aggregate_components"]>[number]>(
       aggregateComponents
     ),

tests/view-data.test.ts

@@ -268,6 +268,7 @@ async function writeSyntheticStageJSnapshot(snapshotDir: string) {
           dataset_version := 'v1'
         ) AS source_data,
         'https://example.test/record.json' AS source_record_url,
+        'https://example.test/eee-record.json' AS eee_record_url,
         struct_pack(name := 'openai-evals', version := '1.0', fork := NULL::VARCHAR) AS eval_library,
         ['first_party']::VARCHAR[] AS evaluator_relationships,
         true AS has_first_party,