Wenwen-D/HarnessDeltaAttribution
Skill Claude CodeCodex
Decompose an evolved harness's score gain S(E) − S(B) on ONE evaluation set into Overfitting (O), Test-Time Scaling (T), and a residual upper bound on Generalizable Improvement (G). Build two controlled variants — a compute-matched baseline Bcc and a neutralized Eneutral — and re-score them on the same set.…
not rated 13
+1 5d ago A 96 tokens