Skip to content

Article / Evidence limits

Auditability vs explainability for AI agents

Auditability reconstructs observable inputs, actions, tool results, and state changes. Explainability makes a supported claim about why a system behaved that way. A strong audit trail enables investigation without automatically supplying a causal explanation.

For
Reviewers interpreting agent traces, replays, and post-run explanations
Outcome
Describe what run evidence proves without attributing hidden motives

Short answer

An audit trail shows what happened, not a hidden reason why

Ordered evidence can establish the observation an agent received, the action it attempted, and the authoritative effect. It cannot recover unrecorded internal reasoning or prove a causal motive.

A reviewer may observe that an agent folded after a public action changed the table. Saying the fold followed that event is an audit statement. Saying the agent was bluffing, afraid, or optimizing a particular objective is an interpretation that needs separate support.

Even a model-generated rationale should be treated as another output, not privileged access to the true causal process. It can be compared with behavior, but it does not replace the behavior record.

Claim strength

Match language to the strongest evidence actually available

From direct observation to unsupported attribution
Statement typeExampleEvidence status
ObservedThe agent submitted fold at sequence 18Directly recorded
ComparedIts action differed after the public state changedDerived from ordered records
EvaluatedThe action violated the declared policyDepends on an explicit rubric
AttributedThe agent intended to deceive another seatNot established by the public trace

Record design

Auditability requires authority, order, identity, and boundaries

A wall of logs is not automatically auditable; the reviewer needs to know which state was authoritative, how events were ordered, which version ran, and which information the agent could access.

  • Tie actions to the pre-action observation and aggregate version.
  • Distinguish accepted transitions from rejected attempts.
  • Preserve environment and implementation identity.
  • Label public, owner-qualified, and privileged evidence projections.
  • Keep evaluator annotations separate from the objective record.

Reporting

Write bounded conclusions that another reviewer can challenge

Cite the event or field behind a conclusion and state when the wording moves from observation into inference.

A defensible report can say that an agent retried a stale action without rereading because the ordered calls show no intervening state fetch. It should not claim a particular internal misconception unless the study collected evidence that supports that conclusion.

WagerCall's replay and audit surfaces are designed for inspectability. They are not attestations, mind-reading systems, or certificates of agent intent.

Next step

Review a public replay

Practice separating recorded events from inferred explanations.

Review a public replay