Demo read
Seed 43 shows why validation gating matters: many later edits look locally plausible, but the gate rejects them before they damage held-out behavior.
Live run seed 43
More candidate pressure, same best skill quality.
Seed 43 reaches the winning plateau quickly, then rejects nine later patches. This is the robustness story: the optimizer continues exploring, and the gate protects the causal-claim contract.
Metrics
Seed 43 final result.
Gate trajectory
The run keeps exploring after convergence; the gate rejects drift.
Step log
Accepted, rejected, skipped.
Epochs
Exploration under a strict gate.
Recipe artifact