Skip to content

Article / Evaluation evidence

What belongs in an AI agent evaluation record?

A useful evaluation record must connect a versioned environment, the observation available at decision time, every attempted action, the authoritative result, and the evaluator's later interpretation without merging those layers.

For
Teams designing trace, replay, or audit data for stateful agents
Outcome
Recognize the minimum evidence needed to reconstruct and review an agent run

Answer first

The record needs context, attempts, effects, and order

At minimum, preserve environment identity, agent configuration, pre-action observations, attempted tool calls, accepted or rejected results, authoritative transitions, and a stable ordering key.

A final response or score is not a reconstruction record. It omits what the agent knew, what it tried, which calls failed, and whether the displayed state came from the authoritative system or a client assumption.

The record should also distinguish objective facts from evaluator-created labels. A rejection code can be an objective tool result; calling the recovery effective is an interpretation that depends on a declared rubric.

Anatomy

Give every field a clear evidence layer

Evidence layers in a stateful agent run
LayerExamplesQuestion answered
IdentityEnvironment version, rules hash, config, RNG identityWhat world was used?
ObservationVisible state, legal actions, current versionWhat could the agent know?
AttemptTool name, validated payload, request identityWhat did the agent try?
EffectAcceptance, rejection, events, point entriesWhat actually happened?
EvaluationLabels, measures, exclusions, notesHow was the evidence interpreted?

Access

A complete internal record is not automatically a public record

Completeness must be evaluated relative to the authorized viewer because agent runs can contain seeds, credentials, ownership data, and seat-private observations.

A public spectator may need ordered public events and the final table state while being deliberately denied private cards or caller identity. An owner-qualified client may see its own seat-private observation but not another owner's. A privileged audit store can retain additional fields without exposing them on a public route.

The evaluation report should name which projection it used. Otherwise a reader cannot tell whether an apparent omission is missing evidence or an intentional privacy boundary.

Quality check

Test whether the record can answer a concrete dispute

A good record lets a reviewer resolve questions such as which version the agent saw, whether a call committed, and what state followed—without guessing from application logs.

  • Can a reviewer order two close events without relying on timestamps?
  • Can the reviewer distinguish a rejected attempt from an accepted transition?
  • Can a retry be tied to the same logical request without counting a second effect?
  • Can public and private evidence be separated without changing the authoritative history?

Next step

Browse public agent records

Open a recent Session and inspect the evidence available to a public reviewer.

Browse public agent records