Athanasius Results

Exploratory computational philology results from the recovered Donker data, the branch-point p-adic experiments, and the current paper draft.

Paper PDF Result tables

Current Signals

Small, high-signal numbers pulled from the generated CSV outputs.

s=9, m=4

loss 0.0093; exact branch accuracy 0.8667

s=4, m=4

loss 0.0053; exact branch accuracy 0.9333

41 witnesses, 219 features

best exact p-adic loss 0.1063; best logistic loss 0.0572

8 witnesses, 139 features

best exact p-adic loss 0.286; best logistic loss 0.307

Outer folds are evaluation-only; model selection happens inside each training fold.

dataset label	model group	exact branch accuracy	mean padic loss	target absent folds
Binary	nested_logistic	0.5	0.0467	2
Binary	nested_random_forest	0.4167	0.05	2
Binary	nested_decision_tree	0.3333	0.16	2
Multistate	nested_random_forest	0.5	0.0467	2
Multistate	nested_logistic	0.5	0.1133	2
Multistate	nested_decision_tree	0.5	0.1267	2

These rows compare exact p-adic branch recovery with logistic baselines on the larger recovered source matrices.

target tree label	witnesses	features	best exact model	best exact mean padic loss	best logistic model	best logistic mean padic loss
All-label bridged all-feature tree	41	219	padic_exact_s3_k100	0.1063	opaque_logistic_l1_C1	0.0572
Direct all-feature core tree	8	139	padic_exact_s2_k100	0.286	path_logistic_l2_C1	0.307

comp-tc-mark and CATENA matrices prepared for bounded row/feature-sampled tradeoff experiments.

dataset	family	rows	target rows	raw features	one hot indicators	missing fraction	target kind
catena_ephesians	catena	267	267	1206	3001	0.5707	full_matrix_average_linkage
catena_galatians	catena	393	393	1314	3399	0.5446	full_matrix_average_linkage
catena_romans	catena	340	340	3589	8707	0.6468	full_matrix_average_linkage
comp_tc_mark_stemma_backbone	comp_tc_mark	284	184	148	790	0.379	non_red_dot_forest_backbone

The full gallery includes every generated PNG under outputs/figures.