karenina-results

Use ResultsIOManager.load_result_set_from_json(path) for current v2.2 exports, which validates rows and preserves metadata. For legacy v2.0/v2.1 exports, use ResultsIOManager.load_legacy_result_set(path), which migrates old row shapes before validation.

Call result_set.get_template_results().to_dataframe() for one row per parsed field, or result_set.get_rubrics_results().to_dataframe() for one row per trait. Scenario runs offer to_dataframe(), to_turn_dataframe(), and to_outcome_dataframe() views.

The answering_models filter matches the full display string like 'langchain:claude-sonnet-4-6', not the bare model name, using exact string equality. Inspect actual keys first with sorted({r.metadata.answering_model for r in result_set}).

Check result.metadata.failure: it is None on success and a Failure object on failure, exposing category, stage, reason, and group. There is no error or completed_without_errors field on the result.

Yes, use evaluate_rubric_on_results from karenina.benchmark with a VerificationResultSet, a Rubric, and a parsing model. It returns PostHocJudgment objects carrying scores, labels, and errors per row.

ResultsIOManager.iter_from_json only supports the v2.2 {"results": [...]} shape and legacy bare arrays. Merged {"runs": {...}} exports from ResultsStore.export_to_file must be loaded with load_result_set_from_json instead.