05/03/2026
I have been thinking a lot about a question that does not get asked in public enough.
🔍 An inspector walks into a clinical trial site. She picks a patient at random and asks the doctor, "Walk me through how you decided this patient was eligible for this trial."
That is the question every regulatory inspection in the world starts with.
If the doctor cannot answer it, source data, eligibility check, the logic that connected them, that is a finding. It can shut down enrollment. It can disqualify the data. It can end a career.
Now imagine an AI system made that eligibility decision.
The AI vendor shows the inspector a number. 94 percent confident.
That is not an answer. That is a marketing slide.
⚖️ This is the gap I have been working on for the past two years at NexTrial.ai
We are building AI for clinical trials that does not produce confidence scores. It produces what we call proof certificates. Documents the inspector can actually read. The rule that was applied. The patient values that were checked. The exact verification that ran. What the AI deliberately did NOT decide, and left for the human doctor.
Four properties. Reproducible. Inspectable. Defensible in court.
🌎 Brazil moved first on this. CFM Resolution 2.454 mandates explicability for any AI used in regulated medical decisions. The European Union AI Act now requires the same thing, quietly, in three different articles. The United States and India will follow.
The vendors who are building toward this architecture today will be ready. The vendors who are not will be re architecting under deadline pressure tomorrow, while their customers absorb the regulatory risk.
📖 I just published a long form essay on this. It is the longest piece I have written, almost 5,000 words, written for regulatory affairs leads, ethics board members, and the doctors who carry the liability when AI is wrong.
It also lays out six questions the framework leaves open. Honest, public, unresolved. The kind of questions a field has to answer in conversation, not in private.
In June, I will be presenting this work at the DIA 2026 Global Annual Meeting in Philadelphia.
Until then, the document is open. The arguments continue. The architecture evolves.
✅ Provably right, not probably right.
That is the standard.
The certificate is the proof.
A confidence score cannot be reconstructed under inspection. A proof certificate can. Before DIA 2026, the field has to decide which one counts.