Reference
RunRecord JSON
The complete, reproducible record.
runrecord.json is the full record of a run: run_id,
scorecard, per-example scores, provenance, and the
comparable result. It round-trips through to_dict /
from_dict and is the snapshot target for tests. The compact
Scorecard is its embedded summary.
Canonical snapshot
json
{
"run_id": "quickstart",
"scorecard": { "verdict": { "verdict": "informational", "ci_should_fail": false }, "...": "..." },
"scores": [
{ "metric": "exact_match", "status": "scored", "validity": "valid",
"value": 1.0, "evaluator_version": "exact_match@1",
"example_id": "datasets/grounded.jsonl#1", "unit_id": "datasets/grounded.jsonl#1" }
],
"provenance": {
"framework": "…", "metric_spec": "…", "evaluator": "…", "dataset": "…",
"example": "…", "evidence": "…", "config": "…", "policy": "…",
"authority": "…", "baseline": "…"
},
"comparable": null
}Fields
| Path | Type | Meaning |
|---|---|---|
run_id | string | The run identifier. |
scorecard | object | The full Scorecard. |
scores[] | array | Every Score: metric, status, validity, value (or null), evaluator_version, diagnostics, and optional subject identity. |
scores[].example_id · scores[].unit_id | string | null | Optional subject identity — which Example / EvalUnit the Score measured. Additive provenance; emit-when-present. |
provenance | object | The RunFingerprint — a hash per dimension (framework, metric_spec, evaluator, dataset, example, evidence, config, policy, authority, baseline). |
comparable | object | null | The comparability result against a baseline, when requested. |
Round-trips or refuses
A present-but-malformed
comparable or scores entry raises on parse
rather than being dropped — the record reconstructs faithfully or not at all.
Subject identity is additive and optional
example_id / unit_id are emit-when-present: a Score without them
serializes byte-for-byte as before, and older runrecord.json artifacts still parse.
They are what /evalglass view --by-call groups by — explicit identity, never list order.