Result Serialization
Definition
Evaluation access preserves selected public Run fields, Case content and exact references, actual outputs and rationales, frozen criteria/treatment context, native interpretation, coverage/completeness and captured telemetry.
The current access contract is evaluation_access.v1. Its owning service defines per-endpoint envelopes and pagination. Do not serialize every response as an earlier proposed universal Run/Policy/Rubric object.
Judgments and scores
Preserve binary judgment, unavailable interpretation and the configured point/gate calculation separately. Current scores use exact fractions and can be negative. A completed mandatory-gate failure contributes zero; missing judgments remain incomplete and keep the full planned population unrankable.
Manual corrections retain original AI judgment and attributable change metadata. They do not rewrite the criterion or submitted response.
Outputs and telemetry
External response items contain exactly one native JSON output or execution failure. Accepted text whitespace and JSON null remain meaningful submitted values. Optional trajectories describe captured model/tool/agent steps, not hidden reasoning.
Latency, usage and cost are nullable captured facts. An absent field does not mean zero. Preserve units and currency identity rather than inventing a normalized total.
Evaluation page fields
The data payload contains:
| Field | Interpretation |
|---|---|
schema_version | evaluation_access.v1 |
run | Scoped execution and pinned identity |
rows | Current page of observations |
total, limit, offset | Filtered population and current page |
grain | case_rubric_observation when supplied |
selection, sampling | Selection context when supplied |
links, next_page | Navigation when supplied |
evaluation-context also carries procedure/context information. See BenchmarkEvaluationPage, BenchmarkEvaluationContext and BenchmarkComparison in the OpenAPI. Native scores, completeness and HTTP success have separate meanings.