Paper I is live on
ChemRxiv.

Title. Architecture-Specific Failure Modes in hERG QSAR Predictions for Iboga Alkaloids: A Within-Family Empirical Test of the Binding-vs-Blockade Limitation.

What it shows. Three production hERG QSAR models -- Pred-hERG (LightGBM on ECFP4), ADMET-AI (a graph neural net), and admetSAR (an SVM) -- are stress-tested on an iboga-alkaloid evaluation set with a pre-registered applicability-domain protocol. Two architecture-specific failure modes emerge: the GNN and the SVM over-call Blocker on designed safer-scaffold analogs (18-MC, tabernanthalog); the gradient-boosted fingerprint model under-callsBlocker on the natural potent blockers (voacangine, noribogaine = the API inside DMX-1001). Both failures trace to a shared root cause -- the binding-vs-blockade dissociation that hERG ligands actually live in -- expressing in opposite directions per architecture.

The white-box catch.The Pred-hERG noribogaine false-negative isn't a black-box mystery. The model's binary sub-model correctly calls Blocker; the deployed consensus rule overrides it via the multiclass and regression sub-models. That's a fixable specification decision, not a learned bias. The paper isolates the override mechanism.

Pre-registration discipline. 8 of 8 pre-registered priors landed within their registered probability ranges on the canonical-SMILES rerun. The OSF pre-registration at osf.io/uwvx4 (DOI 10.17605/OSF.IO/UWVX4, CC0) is Public and frozen; the registration timestamp anchors the priors so the within-family test reads as confirmatory rather than retrofitted. The pattern is portable: the next QSAR architecture audit can fork the template family and run the same pre-commit cascade on its own compound class.

Why this matters past iboga. Iboga is the load-bearing case study, not the point. The point is that architecture-specificfailure modes -- different model families failing on opposite tails of the same chemical-space gap -- are reproducible enough to characterize with a few well-chosen compounds, and that the audit protocol survives transplant to families that look nothing like iboga. Paper II already runs the protocol across tryptamines, phenethylamines, and cathinones; v9 is submission-ready.