Did the training actually help?

The key comparison is the post-trained checkpoint minus its exact untouched base under the same evaluation contract. A larger model, a new agent wrapper, and a different decoding budget are separate changes. Without the base control, a good benchmark score does not establish training lift.

A rise from 40% to 44% is four percentage points and a ten-percent relative increase. An expert-reviewed average on a zero-to-four scale is neither of those. Retain raw counts, item coverage, and the original scale next to any normalized view.

Tune on a development split and reserve a sealed test for final comparisons. Report repeated-seed variation when available, and use paired item-level analysis for matched evaluations. A single aggregate without uncertainty should stay labeled as a point estimate.

Track completion, formatting, correctness, and deeper reasoning separately. Add out-of-domain tests and regression gates for behaviors that matter. If a treatment improves proof critique but not task completion, preserve both findings in the scorecard.

Name the exact base and treatment; pin training data and verifier versions; state training and inference compute; list held-out item IDs or their public manifest; publish denominators, failures, and uncertainty. Explain any missing field rather than silently substituting a proxy.