Prospective adjudicated-cohort audit datasets of BVCT clinical trial outcome predictions 2022-2026

BACKGROUND. Most medical-AI evaluations are retrospective. We report 1,431 pre-readout clinical-benefit predictions on industry trials, each sealed by a third-party signature certificate before its sponsor-disclosed readout, with the full cohort and adjudication trail deposited as a single immutable Zenodo version for independent reconstruction. The unit of analysis is the per-trial prediction of clinically meaningful benefit vs SoC at the registrational endpoint. METHODS. Inputs are restricted to drug specification and the public trial design; the model is patient-data-free and first-in-human-blinded for the asset under prediction. Output is a deterministic effect-size estimate vs SoC; binarisation SUCCESS→GO at HR<0.8, NO-GO at HR≥0.8, fixed before first prediction. Each prediction was sealed by a third-party signature certificate before its readout. Reference: per-(TA, entry-phase) BIO/QLS/Informa 2011-2020 multi-phase Likelihood of Approval (LoA), cohort-weighted. CIs on proportions: Agresti-Coull 99% (headline) with 95% companion intervals tabulated in Supplement §S5; Cohen's κ uses cluster-independent bootstrap 95% CI, B=2000. RESULTS. 1,431 prediction-readout events issued 21 Aug 2022 – 30 Apr 2026 (median ex-ante lead 156 d). Combined cohort (FINAL n=691 + SIMPLE n=740; n_GO=626): TP=561, FP=65, TN=794, FN=11. PPV = 89.6% (Agresti-Coull 99% CI 86.0–92.4). Primary inferential benchmark: vs cohort True-GO prevalence 40.0%, lift +49.6 pp, P<0.001 (construct-equivalent; pre-specified). Contextual benchmark: vs BIO 2011-2020 LoA 28.1%, lift +61.5 pp (not construct-equivalent). FINAL-only sensitivity 89.7% (TP=234, FP=27; n_GO=261). S4 deterministic envelope (n_GO_S4=247): PPV 89.9% (83.7–93.7) with adjudicator discretion algorithmically removed. 374/1,801 issued predictions (20.8%) remain unread-out at cutoff. A locked reviewer protocol (REVIEWER_PROTOCOL.md, v3, 2026-05-16) and full-FINAL-tier census audit package (BVCT_691_FINAL_reviewer_audit_*.csv, F001–F691) are deposited for Cohen's κ inter-rater analysis. CONCLUSIONS. On a cohort sealed before each readout and deposited in full, performance was stationary across five readout years (P=0.80) and an adjudicator-discretion-removed deterministic envelope reproduces PPV within 0.3 pp (89.9%). Adjudication was author-performed and both authors are co-founders of BioinvestGPT (comprehensive CoI in Declarations); independent academic re-adjudication is invited under the locked v3 reviewer protocol, with a full-FINAL-tier census audit package (n=691) deposited as the primary κ input.