=== DATA ABLATION (pendulum, N=40, K=8, shared cached oracle, eval noise seed=0, 3 training seeds=[42, 43, 44]) === - 25% data (650/2600 examples) params=3,823,883 time_to_best=261.8s median regret vs oracle: +0.29% +/- 0.27pp over seeds [s42=+0.61%, s43=+0.16%, s44=+0.12%] coverage (mean over seeds): certified>=1/plant 0.992 stable>=1/plant 1.000 per-sample stable 1.000 - 50% data (1300/2600 examples) params=3,823,883 time_to_best=355.6s median regret vs oracle: +1.00% +/- 0.23pp over seeds [s42=+1.22%, s43=+0.76%, s44=+1.02%] coverage (mean over seeds): certified>=1/plant 1.000 stable>=1/plant 1.000 per-sample stable 1.000 - 100% data (= size_ref, same 3 seeds, reused) params=3,823,883 time_to_best=421.4s median regret vs oracle: +0.92% +/- 0.22pp over seeds [s42=+1.06%, s43=+0.67%, s44=+1.03%] coverage (mean over seeds): certified>=1/plant 1.000 stable>=1/plant 1.000 per-sample stable 1.000 - hidden=256 (DEPLOYED checkpoint -- pre-existing train_pendulum.py run, not part of the 3-seed sweep) params=3,823,883 time_to_best=232.0s median regret vs oracle: +1.15% +/- 0.00pp over seeds [ssingle=+1.15%] coverage (mean over seeds): certified>=1/plant 1.000 stable>=1/plant 1.000 per-sample stable 1.000