Donker Ensemble Tree Regression

Generated by scripts/run_donker_ensemble_tree_regression.py.

This tests whether each recovered Donker section's readings can recover held-out branch placements on the two cross-section ensemble consensus trees. Lower p-adic loss means the target tree geometry is more recoverable from that section's readings.

Target tree	Section	Witnesses	Features	Nodes	Best model	Best loss	Best exact p-adic	Exact loss	Exact - logistic
`direct_tree_distance`	`donker_ath_1cor`	21	40	19	`path_logistic_l2_C1`	0.1365	`padic_exact_s1_k20`	0.1650	+0.0285
`direct_tree_distance`	`donker_ath_2cor_titus`	21	69	19	`path_logistic_l1_C1`	0.1268	`padic_exact_s1_k20`	0.1543	+0.0275
`direct_tree_distance`	`donker_ath_acts`	24	44	22	`opaque_logistic_l1_C1`	0.0561	`padic_exact_s1_k20`	0.0672	+0.0111
`direct_tree_distance`	`donker_ath_acts_13_28`	24	25	22	`opaque_logistic_l2_C1`	0.0542	`padic_exact_s2_k20`	0.0670	+0.0128
`direct_tree_distance`	`donker_ath_acts_1_12`	22	19	20	`padic_exact_s2_k20`	0.0627	`padic_exact_s2_k20`	0.0627	-0.0350
`direct_tree_distance`	`donker_ath_catholicepistles`	14	12	12	`opaque_logistic_l2_C1`	0.1474	`padic_exact_s1_k20`	0.5291	+0.3817
`direct_tree_distance`	`donker_ath_heb`	18	35	16	`path_logistic_l2_C1`	0.1571	`padic_exact_s2_k20`	0.1652	+0.0081
`direct_tree_distance`	`donker_ath_paulineepistles`	21	163	19	`opaque_logistic_l2_C1`	0.1438	`padic_exact_s2_k20`	0.1450	+0.0011
`direct_tree_distance`	`donker_ath_rom`	19	19	17	`opaque_logistic_l2_C1`	0.1619	`padic_exact_s2_k20`	0.1677	+0.0057
`exemplar_kernel`	`donker_ath_1cor`	21	40	19	`path_logistic_l2_C1`	0.1264	`padic_exact_s2_k20`	0.1476	+0.0212
`exemplar_kernel`	`donker_ath_2cor_titus`	21	69	19	`dummy_training_padic_medoid`	0.1451	`padic_exact_s1_k20`	0.1543	-0.0097
`exemplar_kernel`	`donker_ath_acts`	24	44	22	`padic_exact_s2_k20`	0.0626	`padic_exact_s2_k20`	0.0626	-0.0402
`exemplar_kernel`	`donker_ath_acts_13_28`	24	25	22	`opaque_logistic_l2_C1`	0.0651	`padic_exact_s2_k20`	0.0685	+0.0035
`exemplar_kernel`	`donker_ath_acts_1_12`	22	19	20	`padic_exact_s2_k20`	0.0644	`padic_exact_s2_k20`	0.0644	-0.0048
`exemplar_kernel`	`donker_ath_catholicepistles`	14	12	12	`opaque_logistic_l2_C1`	0.1721	`padic_exact_s1_k20`	0.4642	+0.2921
`exemplar_kernel`	`donker_ath_heb`	18	35	16	`path_logistic_l2_C1`	0.1261	`padic_exact_s2_k20`	0.1387	+0.0126
`exemplar_kernel`	`donker_ath_paulineepistles`	21	163	19	`opaque_logistic_l1_C1`	0.1160	`padic_exact_s2_k20`	0.1413	+0.0254
`exemplar_kernel`	`donker_ath_rom`	19	19	17	`padic_exact_s2_k20`	0.1716	`padic_exact_s2_k20`	0.1716	-0.0232

Target-Tree Summary

Target tree	Mean best loss	Mean exact loss	Mean logistic loss	Exact beats logistic
`direct_tree_distance`	0.1163	0.1692	0.1202	1/9
`exemplar_kernel`	0.1166	0.1570	0.1263	4/9

The exact p-adic runs here use the same bounded one- and two-feature through-points ensemble setting as the recovered-section exact rerun. They are a recoverability screen for the ensemble trees, not yet the final large sweep.