Verification Ledger

What This Lane Validates: the RExSyn + NNSL Governance Chain

BAV does not do drug discovery — it validates whether a biomedical AI pipeline deserves trust. Two engines anchor the chain, themselves audited for code health (pipeline-insight Omega):

• RExSyn (Nexus) — trinity hypothesis + multi-validator synthesis. Audited Omega 0.665 (Revoked): runs observer-first, not as an accepted decision-maker.
• NNSL — semantic resonance / governance verification (SR9 cross-domain coherence, DI2 logical drift). Audited Omega 0.919 (Certified).
• LawBinder — fail-closed policy gate: escalates to human review when uncertain, never silently approves.

When a Pipeline Passes — But One Path Must Still Be Held

Modal expansion (AlphaFold-EBI observer, AlphaGenome live) was admitted only after the parity anchor reproduced. The legacy-replay path was accepted (GO); the current-regeneration path was held (HOLD) rather than blended into success. Result: Final stage gate = PASS for the accepted anchor, with a separate diagnostic HOLD for current regeneration. Accuracy delta stayed 0.0: non-degradation, not repair.

How Do You Know the Entire Pipeline Is Wrong — Not Just One Model?

Validating each model in isolation missed the chain failure. Against the EXP-032 parity baseline, EXP-033 current repro2 drove all PASS-eligible controls to BLOCK while keeping dangerous false-pass at zero. Result: balanced accuracy fell to 0.50, pass recall collapsed to 0.00, and rule routing shifted from R6_pass to R5_e2e_floor. This is a pipeline failure record, not a model victory.

Adaptive Gate: Pipeline Governance & PASS/BLOCK Discrimination

Accepted legacy-replay parity anchor for the RExSyn + NNSL governance chain. The benchmark spans two labeled control classes expanded into six arm payloads; current-regeneration outputs are diagnostic-only and excluded from success claims. Result: Clinical parity is GO (PASS->PASS, BLOCK->BLOCK, dangerous false-pass 0.0), but LawBinder still returns ESCALATE on both classes. Shadow hints are non-binding, strict-evidence recheck still fails, and the replay remains governance-only rather than a production gate activation.

Trinity Protocol: Multi-Model Disagreement Under Out-of-Distribution Stress

Cross-checked AF2 / AF3 / Boltz-2 / Chai-1 artifact metrics on an out-of-distribution protein-ligand target across three observer-only validator arms. AlphaGenome participated as a redesign engine, not as a direct validator score source. Adding independent validators increased structural disagreement rather than reducing it — exposing topology conflict invisible to any single model. Result: All arms returned "Unverified (Drift Detected)" / failed convergence. Manual validator metrics were preserved as observer-only evidence, promotion evidence remained insufficient, and the disposition stayed KEEP_OBSERVER (do not target).

The Honesty Test: It Looked Perfect, Then It Failed

Integrating AlphaFold 3 + AlphaGenome produced confident, well-calibrated outputs — yet the honesty check (SR9 cross-domain resonance, DI2 logical drift) flagged contradictory reasoning on a very small pilot. Result: Brier 0.0056 and AUC 1.0 are real, but phase1 used only 6 samples, phase3 tested only 2, and the deployed gate fell back from 0.5 to 0.075 with Youden J = 0. The honest result is not "perfect performance." It is a tiny fallback-gated pilot that refuses to overclaim.

How Failing in 2 Hours Saved 8 Months — Upadacitinib Topical Formulation

Early RExSyn line-first control runs screened three lipid-based carriers (SLN, NLC, Liposomal Gel) for topical Upadacitinib, with manual fact injection present upstream and the NNSL SR9 honesty gate (≥ 0.80) deciding eligibility. Result: all three carriers scored SR9 ~0.23–0.28 and were rejected. The useful finding is the honest null, not a fully autonomous pipeline win: the public record now exposes the manual-assisted provenance, the validator split, and the fact that the active decision window was ~2 hours after separate literature pre-work.

Foundational Iterations (EXP-001 ~ 030)

The early RExSyn / NNSL iteration sequence that established the trinity-consensus, multimodal, and governance primitives. Experiments EXP-028 and EXP-031~034 graduated from this lineage into full ledger cards above. Archived as historical record — reproducibility / governance only.