/hyperpower:eval

A prompt change can feel like an improvement and still make results worse. This command checks.

It runs the eval cases and prints a verdict. A prompt change ships only when all three of these hold: there are no blockers, correctness and safety are within 0.1 of the baseline or better, and the weighted score beats the baseline.

--baseline saves the current scores as the baseline to beat.

The plugin comes with fifteen cases that work with any stack. Add your own after a few real runs, based on work you have read yourself.

Neither init nor this command creates cases from your commit history. A case taken from a commit nobody reviewed makes a baseline nobody can trust, and every later comparison would carry that problem.