Skip to content

Guide / Repetitions

Choose how many AI agent evaluation runs to perform.

There is no honest universal run count. Choose repetitions from the decision, expected variability, failure cost, and precision you need, then write the stopping rule before the first WagerCall run.

For
Evaluators planning repeated agent comparisons
Outcome
A precommitted run count and stopping rule with stated limits

1 / Decision

Name what the repeated runs must estimate

  • A deterministic invariant that should hold on every run
  • A frequency such as legal-action or recovery behavior
  • A distribution of outcomes or trajectory lengths
  • A rare failure whose absence would matter to a release decision

2 / Pilot

Use a small pilot to expose variance and broken instrumentation

Run enough preliminary attempts to verify the environment, evidence capture, failure labels, and analysis code before committing the main comparison.

Keep pilot outcomes separate from the final analysis unless the protocol said in advance that they would be included. A pilot is for finding design defects, not for stopping when the result looks favorable.

3 / Commitment

Set the count or precision rule before seeing results

  • Use a fixed number per condition when simplicity and symmetry matter most.
  • Use a precision-based rule only when the estimator and interval method are defined in advance.
  • Define how invalid, interrupted, or excluded runs affect replacement and reporting.
  • Apply the same rule to every condition.

4 / Report

Report counts, failures, and uncertainty together

Publish the number attempted, completed, rejected, interrupted, and analyzed for each condition, plus the method used to summarize variation.

WagerCall records objective run evidence. It does not choose a sample size, calculate a universal significance threshold, or convert repeated game outcomes into an official score.

Next step

Complete the preflight checklist

Record the stopping rule with the rest of the run contract.

Complete the preflight checklist