Checkpoint
One last thing before we move on. Same surface as a write step — but the lesson doesn't complete until this passes.
Apply the seed-noise yardstick. Three logged runs share one config and differ only in seed; a candidate with a new lr scores 0.89. Compute the noise band (max - min of the three seed accs) and the candidate's gain over the BEST seed run, then print: f"seed-noise band: {lo} - {hi} (width {noise:.2f})", f"candidate gain over best: {gain:+.2f}", and "improvement proven:" with gain > noise. Chapter 38's sampling lesson, wired into the tracker.