Runtime & Integrations · Evaluation Core
RunRecord
The complete, reproducible record of a run.
The
RunRecord holds everything needed to reproduce or inspect a run:
run_id, the scorecard, the per-example scores,
provenance, and the comparable fingerprint result. It round-trips
through to_dict / from_dict.
Where the Scorecard is the summary, the RunRecord is the full evidence trail. It is the primary persisted artifact and the snapshot target tests assert against, so its shape is stable and its parsing fails closed.
Fields
| Field | Holds |
|---|---|
run_id | The run identifier. |
scorecard | The summary artifact (verdict, metrics, authority, baseline, diagnostics). |
scores | Every per-example Score, with status, validity, diagnostics, and optional subject identity (example_id / unit_id). |
provenance | The RunFingerprint across ten dimensions. |
comparable | The comparability result against a baseline, if requested. |
Round-trips, fails closed
A present-but-malformed
comparable or scores entry raises on parse
rather than being dropped — a record either reconstructs faithfully or refuses, never silently
loses evidence.
Per-score subject identity
Each
Score may carry example_id and unit_id — additive,
optional provenance pointing at the subject it measured. They are emit-when-present, so old
runrecord.json artifacts still parse. This identity is what
/evalglass view --by-call groups by. See
RunRecord JSON for the exact per-score shape.