Seed 42
Clean stabilization: early accepted edits, then skips once the admission rule stops improving.
SkillOpt recipe drilldown
Three seeded runs learn the same causal-claim admission contract.
This second-level page exposes the experiment behind the UDL story: SkillOpt edits a Markdown admission rule, evaluates each candidate through validation gates, and keeps the best behavior when later edits drift.
Headline metrics
Reproducible validation convergence with a visible valid-unseen split.
The learned skills turn ambiguous news and scientific causal surfaces into TICKET-compatible admission bundles without promoting hypotheses into interventional proof. Seed 44 adds a useful stricter generalization signal.
Optimization trace
Different paths, same winning validation step.
Seed 42 gate scoreSeed 43 gate scoreSeed 44 gate score
Run comparison
The acceptance dynamics are the demo.
Seed 42 stabilizes with skips, seed 43 keeps proposing tie-score patches, and seed 44 rejects every post-step-5 candidate. All three preserve the same UDL validation winner.
Held-out examples
The learned skill fixes concrete field-level failures.
Seed drilldowns
Open the run trace only when you want the engine room.
Seed 43
Exploration pressure: many later candidate patches are rejected by the gate.
Seed 44
Sharper generalization check: validation converges, while valid-unseen exposes the harder edge.