Name the benchmark, the metric and the result as three separate things

every_eval_ever/converters/README.md

@@ -46,7 +46,18 @@ The exact command for converting an example evaluation log is:
 uv run --extra inspect every_eval_ever convert inspect --log_path tests/data/inspect/2026-02-07T11-26-57+00-00_gaia_4V8zHbbRKpU5Yv2BMoBcjE.json
 ```
 
-Optional: pass `--supplemental_eval_details path/to/supplemental_eval_details.json` to enrich converted output. `additional_details` maps are extend-only (existing keys are preserved), while synthetic `metric_config` defaults can be overridden for these fields: `evaluation_description`, `lower_is_better`, `score_type`, `level_names`, `level_metadata`, `has_unknown_level`, `min_score`, `max_score`. Use top-level fields (`model_info`, `source_data`, `generation_config`, `agentic_eval_config`) for shared details and `evaluation_results` for per-result metric/score details keyed by `evaluation_name`.
+Optional: pass `--supplemental_eval_details path/to/supplemental_eval_details.json` to enrich converted output. `additional_details` maps are extend-only (existing keys are preserved), while synthetic `metric_config` defaults can be overridden for these fields: `evaluation_description`, `lower_is_better`, `score_type`, `level_names`, `level_metadata`, `has_unknown_level`, `min_score`, `max_score`. Use top-level fields (`model_info`, `source_data`, `generation_config`, `agentic_eval_config`) for shared details and `evaluation_results` for per-result metric/score details.
+
+An `evaluation_results` entry selects which results it applies to by either key:
+
+| Key | Selects |
+|---|---|
+| `evaluation_result_id` | one result — `"<scorer>:<metric>"`, e.g. `choice:accuracy` |
+| `evaluation_name` | every result of that evaluation, e.g. `inspect_evals/pubmedqa` |
+
+`evaluation_result_id` wins where both match. An entry with neither key applies
+positionally. Convert once without a supplement to see the available keys — a key
+that selects nothing is logged as a warning rather than applied.
 
 Example `supplemental_eval_details.json`:
 
@@ -74,7 +85,7 @@ Example `supplemental_eval_details.json`:
   },
   "evaluation_results": [
     {
-      "evaluation_name": "inspect_evals/pubmedqa - choice",
+      "evaluation_result_id": "choice:accuracy",
       "score_details": {
         "details": {
           "notes": [
@@ -100,7 +111,7 @@ Example `supplemental_eval_details.json`:
 Use it with:
 
 ```bash
-uv run python -m eval_converters.inspect \
+uv run --extra inspect every_eval_ever convert inspect \
   --log_path tests/data/inspect/data_pubmedqa_gpt4o_mini.json \
   --supplemental_eval_details path/to/supplemental_eval_details.json
 ```

every_eval_ever/converters/helm/adapter.py

@@ -52,6 +52,7 @@
     HELMInstanceLevelDataAdapter,
     _evaluation_result_id,
     _score_from_stat,
+    _stat_name_part,
 )
 from every_eval_ever.converters.helm.metrics import is_core_metric
 from every_eval_ever.converters.helm.utils import extract_reasoning
@@ -513,6 +514,7 @@ def _transform_single(
 
             metric_config = MetricConfig(
                 evaluation_description=metric_name,
+                metric_name=metric_name,
                 lower_is_better=False,  # TODO schema.json check
                 score_type=ScoreType.continuous,
                 min_score=0,
@@ -521,19 +523,12 @@ def _transform_single(
 
             split = getattr(stat.name, 'split', None)
             perturbation = getattr(stat.name, 'perturbation', None)
-            name_parts = [metric_name]
-            if split:
-                name_parts.append(str(split))
-            if perturbation:
-                name_parts.append(str(perturbation))
-            evaluation_name = (
-                f'{" ".join(name_parts)} on {source_data.dataset_name}'
-            )
+            perturbation_label = _stat_name_part(perturbation)
 
             evaluation_results.append(
                 EvaluationResult(
                     evaluation_result_id=evaluation_result_id,
-                    evaluation_name=evaluation_name,
+                    evaluation_name=source_data.dataset_name,
                     source_data=source_data,
                     evaluation_timestamp=evaluation_timestamp,
                     metric_config=metric_config,
@@ -553,10 +548,8 @@ def _transform_single(
                         ),
                         details={
                             'count': str(getattr(stat, 'count', '')),
-                            'split': str(split) if split else '',
-                            'perturbation': str(perturbation)
-                            if perturbation
-                            else '',
+                            'split': _stat_name_part(split) or '',
+                            'perturbation': perturbation_label or '',
                         },
                     ),
                     generation_config=GenerationConfig(

every_eval_ever/converters/inspect/adapter.py

@@ -56,6 +56,7 @@ def _require_inspect_dependencies() -> None:
 )
 from every_eval_ever.converters.inspect.instance_level_adapter import (
     InspectInstanceLevelDataAdapter,
+    evaluation_result_id,
 )
 from every_eval_ever.converters.inspect.utils import (
     apply_supplemental_eval_details,
@@ -147,11 +148,15 @@ def _build_evaluation_result(
         num_samples: int = 0,
     ) -> EvaluationResult:
         return EvaluationResult(
-            evaluation_name=f'{metric_info.name} on {evaluation_task_name} for scorer {scorer_name}',
+            evaluation_result_id=evaluation_result_id(
+                scorer_name, metric_info.name
+            ),
+            evaluation_name=evaluation_task_name,
             source_data=source_data,
             evaluation_timestamp=evaluation_timestamp,
             metric_config=MetricConfig(
-                evaluation_description=metric_info.name,
+                evaluation_description=f'{metric_info.name} from scorer {scorer_name}',
+                metric_name=metric_info.name,
                 lower_is_better=False,  # no metadata available
                 score_type=ScoreType.continuous,
                 min_score=0,
@@ -175,8 +180,15 @@ def _extract_evaluation_results(
         generation_config: GenerationConfig,
         num_samples: int,
         timestamp: str,
-    ) -> List[EvaluationResult]:
+    ) -> Tuple[List[EvaluationResult], Dict[str, List[str]]]:
+        """Convert Inspect's per-scorer metrics into aggregate results.
+
+        Returns the results plus a scorer name -> `evaluation_result_id` map,
+        which the instance-level converter needs to emit one row per aggregate
+        result a sample contributed to.
+        """
         results: List[EvaluationResult] = []
+        result_ids_by_scorer: Dict[str, List[str]] = {}
 
         for scorer in scores:
             llm_grader = None
@@ -220,22 +232,26 @@ def _extract_evaluation_results(
 
                 scorer_name = scorer.name or scorer.scorer
 
-                results.append(
-                    self._build_evaluation_result(
-                        evaluation_task_name=evaluation_task_name,
-                        scorer_name=scorer_name,
-                        metric_info=metric_info,
-                        llm_grader=llm_grader,
-                        source_data=source_data,
-                        evaluation_timestamp=timestamp,
-                        generation_config=generation_config,
-                        stderr_value=stderr_value,
-                        stddev_value=stddev_value,
-                        num_samples=num_samples,
-                    )
+                result = self._build_evaluation_result(
+                    evaluation_task_name=evaluation_task_name,
+                    scorer_name=scorer_name,
+                    metric_info=metric_info,
+                    llm_grader=llm_grader,
+                    source_data=source_data,
+                    evaluation_timestamp=timestamp,
+                    generation_config=generation_config,
+                    stderr_value=stderr_value,
+                    stddev_value=stddev_value,
+                    num_samples=num_samples,
                 )
+                results.append(result)
+
+                if scorer_name and result.evaluation_result_id:
+                    result_ids_by_scorer.setdefault(scorer_name, []).append(
+                        result.evaluation_result_id
+                    )
 
-        return results
+        return results, result_ids_by_scorer
 
     # A HuggingFace repo identifier: exactly `namespace/name` with no
     # extra path segments, schemes, or path-unsafe prefixes. We use an
@@ -634,7 +650,7 @@ def _transform_single(
 
         evaluation_task_name = eval_spec.task_display_name or eval_spec.task
 
-        evaluation_results = (
+        evaluation_results, result_ids_by_scorer = (
             self._extract_evaluation_results(
                 evaluation_task_name,
                 results.scores if results else [],
@@ -644,7 +660,7 @@ def _transform_single(
                 evaluation_unix_timestamp,
             )
             if results and results.scores
-            else []
+            else ([], {})
         )
 
         supplemental_eval_details = parse_supplemental_eval_details(
@@ -690,6 +706,7 @@ def _transform_single(
                     model_info.id,
                     raw_eval_log.samples,
                     getattr(raw_eval_log, 'reductions', None),
+                    result_ids_by_scorer,
                 )
             )