Research
AI can now draft and check regulated clinical trial work. A human still has to decide if it is right, and a patient lives with that decision. Today that decision usually rests on a confidence score, and a probability has never been something you can take to an inspection. Our research asks a more demanding question: how do you produce evidence an AI-assisted output is sound, evidence a named human can stand behind and an inspector can independently re-verify?
Proof, not probability. Evidence, not substitution. The human decides.
This page collects the work behind that standard. It is a contribution to a field that is still being defined, including the parts we have not solved yet. We would rather state the open question plainly than paper over it.
The DIA 2026 Poster
Our anchor work was presented at the DIA 2026 Global Annual Meeting in Philadelphia: Formal Verification as Regulatory Validation for AI Across the Clinical Trial Lifecycle (Abstract 116114, Thompson, 2026). It argues that the way to govern AI in regulated research is not to make the model more confident, but to wrap its output in verifiable evidence and mandatory human accountability. It introduces the three-gate architecture below and maps it against the January 2026 FDA and EMA joint guiding principles for AI.
Formal Verification as Regulatory Validation for AI
Abstract 116114 · Thompson · DIA 2026 Global Annual Meeting
The Three-Gate Framework
Every AI-assisted output passes through three independent checks before it counts, on one principle: nothing approves its own work, and the system that generates an output never verifies it. A risk classification runs first and decides how hard each gate works, so the effort is always proportionate to what is at stake.
Gate 1 — Jurisdiction
The output is read against the actual rules in force for the specific site it belongs to: regulation, the locked protocol, site SOPs, and ethics conditions. One site, one rulebook, no cross-border bleed. The result is a clear verdict with the exact rule cited by name and version, never a black box.
Gate 2 — Formal Proof
The architecture is designed to carry a formal, mathematical proof that the output's structure is sound, one anyone can independently re-verify without trusting us. It proves the document is well-formed. It does not claim to prove the document is clinically correct, and it states that boundary plainly. This gate is designed, not in production.
Gate 3 — Human Oversight
A named human attests to the output, on the record, at a depth set by risk before the trial starts. The certificate is evidence that person reads. It never replaces their judgment, and liability stays with people and organizations, never the model.
A set of architecture briefs covering each gate is available on request.
Extending GxP Validation to Non-Deterministic Systems
Computerized system validation was built for deterministic software, where the same input reliably produces the same validated output. AI breaks that assumption. The same prompt can yield different text, which is exactly why traditional validation strains against it.
The three-gate framework is a way to bring AI-assisted work back under the GxP expectations regulated teams already hold, rather than inventing a new standard to replace them. It is designed to speak the established language:
- Risk-based quality management (ICH E6(R3)) maps to the risk classification that runs before any gate and governs how much verification and human judgment each output requires.
- Rule-based verification (Gate 1) maps to the documented, traceable checks at the center of computerized system validation.
- Data integrity and ALCOA+ map to Gate 2's structural evidence and to a lineage record where nothing is silently deleted or backdated.
- 21 CFR Part 11 and human-in-the-loop control map to Gate 3's named, attributable electronic attestation.
The throughline is the direction regulators are already moving: assurance effort proportionate to risk, evidence over documentation for its own sake. We are building so that AI-assisted regulatory work can meet that bar, not ask for an exception to it.
Multi-Agency Regulatory Engagement
We engage with regulators as a contributor to the conversation, not a claimant. We filed the framework to the relevant public FDA docket as a contribution to the open record. We are now in active dialogue across six regulatory environments.
FDA
Framework filed to public docket. Ongoing engagement with CDER and CDRH on AI credibility standards.
EMA
Three-gate architecture mapped against the January 2026 FDA-EMA joint guiding principles for AI. EU AI Act high-risk classification alignment.
ANVISA
Blueprint encoded for Lei 14.874/2024 and RDC 945/2024. Framework presented to the Brazilian regulatory community.
CDSCO
CDSCO adapter live. Framework engagement through VeriTrial India partnership across 25 sites.
PMDA
Engagement initiated. PMDA alignment with ICH E6(R3) risk-based validation approach.
NMPA / CDE
Blueprint in development. Zero-egress architecture aligned with PIPL data localization requirements.
Engagement, not endorsement.
The Open Question
Here is the part the field has not settled. When a regulation becomes a computable rule, someone had to translate it. Who certifies that the encoding is faithful to the regulation it came from? A verification system can prove an output matches an encoded rule perfectly and still be checking against a flawed translation. The question is not who built the checker. It is who is accountable for the encoding.
Nine of the ten FDA and EMA joint guiding principles are addressed by the architecture itself. The tenth is this one, and we do not claim to have solved it. We put it to the field instead.
We are running a public poll on who should hold that responsibility. Add your view and see where the field lands.
Complementary and Open Work
This framework does not stand alone, and it is stronger for that. The ARCH implementation standard, authored independently by Jessica Stuyvenberg and published openly with a DOI, operationalizes the architecture as a CDISC USDM-native data schema, so a verification certificate travels with the submission instead of becoming another proprietary silo. Two complementary works, co-developed.
We filed the framework to the relevant public FDA docket as a contribution to the open record. We engage with regulators as a contributor to the conversation, and we are careful to claim nothing more than that.
ARCH Framework v6.0
Jessica Stuyvenberg
Adaptive Regulatory Compliance and Human Oversight — a field-level implementation specification for anchoring AI proof certificates natively within the CDISC USDM.
Engage with the Framework
This is a space to think, and we built it to be pressure-tested. If you work in regulatory affairs, clinical operations, data standards, or you sit on the agency side, we would value your read, including where you think it breaks.
Celina orchestrates. Humans decide. Patients get access.