A minimal certification protocol

For a checkpoint-specific mathematical claim, retain:

1. the exact baseline and early-stage checkpoints;

2. binary masks and integer counts for every layer;

3. the optimizer-reset convention and training schedule;

4. independent recorded training and evaluation streams;

5. common latent test examples for both task labels;

6. clean endpoint accuracies and unclamped cue differences;

7. midpoint accuracy for each reference task separately;

8. repeated stress estimates and their conditional Monte Carlo uncertainty;

9. margin distributions and the constants in Theorems thm:parameterstability, thm:globalstress, and thm:frozengates;

10. an independent validation or simultaneous-testing procedure for selecting a pruning peak.

The numerical table then becomes evidence to be certified, not a replacement for the theorem's assumptions.