A reproducibility-audited, gate-based framework for early translational pharmacology: native-scale evidence, identifiability, and a 12-domain reporting checklist

Background Early translational pharmacology often combines docking scores, machine-learning structure predictions, and physiologically based pharmacokinetic (PBPK) simulations as if they formed a single evidentiary scale. This can create precise-looking but uncalibrated biological and clinical claims. Methods We applied a four-class provenance framework to the deposited erinacine A package. Vina records were re-executed where complete inputs existed; DiffDock-L and Boltz-2 outputs were audited on their native scales; PBPK project identity, executable content, and exported profiles were inspected; and conditional free-CNS calculations were recomputed over joint fu × Kp , uu , brain assumptions. A prospective G0–G8 gate-based decision architecture was then specified, separating identity control, decoy topology, docking, replicated MD, direct biophysical binding to purified molecular targets under assay conditions, optional mechanistic QM/MM analysis, unbound exposure, and pharmacodynamics. Erinacine A serves as a worked example and stress test of the framework rather than as evidence of validated molecular efficacy. Results The archived 8TKL value of −3.221 was compatible with more than one scoring/receptor footing and therefore had no unique recoverable origin. The direct 31-heavy-atom RMSD between archived and re-executed poses was 3.974 Å. All three complete Vina re-executions failed the prespecified protocol-specific 5 Å geometric adequacy rule. DiffDock top poses retained steric conflicts, and Vina versus Boltz-2 showed no global rank concordance (ρ = 0.00, exact p = 1.00). The deposited PK-Sim files encoded a human individual, not porcine physiology, and lacked the observed calibration dataset and a populated compound-reaction block; their numerical exports were internally reproducible but were not qualified as a pig PBPK reconstruction. Conditional CNS-band attainment ranged from 0% to 80.2%, whereas lower-bound attainment ranged from 0% to 100% across versioned sensitivity anchors. Conclusion The transferable result is a gate-based governance protocol for translational computational evidence, combining provenance classification, native-scale constraints, structural-identifiability controls, and prespecified resource-allocation decisions. It preserves negative findings and missing provenance. The deposited outputs support prioritization and falsification tests, not proof of affinity, target engagement, porcine or human exposure, efficacy, formulation viability or clinical dosing. Progression through the prospective decision architecture requires a distinct evidentiary pass at each gate.