harness:certify

To verify LLM agent evaluation score stability after evolution, run repeated evaluations across the same dataset and codebase to compute summary statistics like mean and standard deviation. This confirms whether the reported improvement holds consistently before deployment.

Certifying LLM agent reliability uses mean, standard deviation, range, and a categorical verdict of STABLE, MARGINAL, or UNSTABLE. These summary statistics are aggregated from combined_scores across multiple independent evaluation runs to determine score variance.

Yes, you need the project's .evolver.json file to run agent stability evaluations. The certification process also requires the Evolver tools run_eval.py and read_results.py, along with a Python runtime to execute those scripts.

You should run certification for an evolved LLM agent after an evolution run merges an improved agent. Running multiple evaluations confirms that the improvement holds across independent runs before you tag and deploy the merged agent version.

If your agent evaluation score is unstable, you should add rubrics or rerun heavy evolutions. The certification process provides actionable guidance based on the stability verdict, advising deployment only when scores are stable across repeated runs.