Demo read
Seed 44 is the generalization-check story: validation convergence is reproducible, but valid-unseen performance shows where the next demo-worthy diagnostics should focus.
Live run seed 44
Step 5 still wins, but valid-unseen exposes the harder edge.
Seed 44 accepts three updates and rejects thirteen candidates, with no skips. It matches the validation best score at step 5, then lands at 83.33% valid-unseen hard accuracy.
Metrics
Seed 44 final result.
Gate trajectory
The gate accepts step 5, then rejects every later candidate.
Step log
Accepted, rejected, never skipped.
Epochs
Sharp selection pressure.
Recipe artifact