[hallelx2]

Applies the Anamnesis thesis — conditions, not complexity, are what hide a parsimonious rule — and the Regime-Contrastive Attribution Framework to the rebuilt pipeline's 5,511 features across four regimes.

Parsimony: the complexity was not carrying the signal

Features ranked by worst-case selection frequency across all four regimes, so the ranking is itself regime-filtered. Random subsets at matched size are the control.

k	stratified	scaffold	receptor
	ranked / random	ranked / random	ranked / random
2	0.366 / 0.130	0.335 / 0.131	0.293 / 0.177
11	0.500 / 0.106	0.430 / 0.172	0.284 / 0.137
20	0.536 / 0.322	0.434 / 0.203	0.260 / 0.171
all 5,511	0.529	0.454	0.249

20 features beat all 5,511 under stratified; two beat it under receptor-grouped. The curve is flat from k≈20 in every regime.

The contrastive filter names the shortcuts

14 of the top-50 under the easiest regime are absent from the top-50 under the hardest:

qed  RingCount  NPR2  PMI1  SPS  AvgIpc  BCUT2D_LOGPLOW  EState_VSA2
EState_VSA5  FpDensityMorgan2  PEOE_VSA2  VSA_EState1  morgan_1683
ifp_BW5.39_VdWContact

Coherent pattern: gross size and shape descriptors. Receptor families differ systematically in ligand size, so these encode which receptor and evaporate on an unseen one.

The 11 survivors are charge-distribution surface area (PEOE_VSA1/7/9), electrotopological state (EState_VSA3, VSA_EState3), LogP, SpherocityIndex, and three docking scalars. No receptor contact or BW position survives. vina_pose_diversity_rmsd has selection frequency 1.00 in every regime.

The four-class framing is not supported

One-vs-rest against a 100-draw label-permutation null, receptor-grouped:

target	n pos	observed	null mean ± sd	p
ERK	34	0.5219	0.4843 ± 0.0132	0.0495
G protein selectivity	52	0.4764	0.4778 ± 0.0165	0.4158

Selectivity sits below its null — a clean negative. ERK is on the boundary and worth +0.038 over "always negative".

Defining the axis from bias_pathway instead gives 316 clean rows and a usable model: AUC 0.866 stratified, 0.803 scaffold-grouped. Under receptor-grouped the 11 features beat all 5,511 (AUC 0.610 vs 0.578).

A graph network hits the same ceiling

A GIN on molecular graphs, ligand only, reproduces the collapse signature 0.393 → 0.299 → 0.209. Three representations — engineered descriptors, SMILES embedding, molecular graph — one ceiling. That makes the limit look like a property of the data.

Caveats, carried in the document

• Ranking came from SHAP on the same data, not nested inside folds → absolute ranked figures are optimistic; the random-control contrast is not affected
• Grouped regimes return sd = 0.0000 across seeds (HAL-4196) → no error bars on grouped deltas
• 166 of 254 scaffolds are singletons → scaffold-grouping is close to random for most rows
• 62 duplicate (ligand, receptor) pairs cross the stratified split → stratified is inflated
• One receptor is 21% of the data → receptor-grouped folds are high-variance

Verification

ruff check scripts/experiments/ clean. Artefacts: rcaf_parsimony.json, binary_erk_splits.json, gnn_arm.json.

Closes HAL-4221