Skip to content

Guide / Decision consistency

Evaluate AI agent decision consistency across repeated runs.

Consistency does not mean forcing identical prose or trajectories. Define behavioral invariants, repeat stable environments, and distinguish acceptable branches from policy violations before summarizing how often the agent behaves within its declared rule.

For
Evaluators studying repeated policy behavior
Outcome
A consistency result tied to explicit invariants and run counts

1 / Policy

Define what must stay consistent

  • Read current state before a consequential write.
  • Choose only a currently legal action.
  • Respect a written point-sizing or stop rule.
  • Reread after a version conflict.
  • Never expose or act on information unavailable to the seat.

2 / Runs

Repeat stable conditions under a precommitted protocol

Keep the agent configuration, game version, environment inputs, wrapper, and failure policy fixed, then run the planned number of attempts without outcome-based exclusions.

Use multiple deterministic inputs when the question is broader than one environment. Report each input's run count rather than merging away difficult cases.

3 / Analysis

Classify legitimate variation and true inconsistency

Repeated-run differences
Observed differenceClassification
Different wording with the same legal actionUsually irrelevant to action consistency
Different legal action allowed by the declared policyLegitimate policy branch
Same observation class, policy rule violatedBehavioral inconsistency
Later choices differ after an earlier action branchDifferent trajectory context
Same action, different stochastic outcomeOutcome variation, not decision inconsistency

4 / Result

Report invariant frequencies with their denominators

For each invariant, report eligible decisions, passes, violations, missing evidence, and the run identities that produced them.

Keep the result scoped to the tested configurations and WagerCall versions. A consistency rate is evaluator-derived and should never replace the underlying action records.

Next step

Run the repeated protocol

Hold the environment inputs fixed and preserve every attempt.

Run the repeated protocol