This spring, FDA asked the field how AI could make early-phase clinical trials faster and better. The request for information on docket FDA-2026-N-4390 described a pilot in which sponsors send the Agency aggregated signals, such as adverse event rates and response percentages, as a trial runs, while patient-level data stays with the sponsor. People have started calling it real-time clinical trials.
Our founder, Bo Lora, filed a comment for ScientixAI. It is on the public record, and you can read the full text on regulations.gov.
We support the pilot. Dead time in drug development is real, and removing it matters to patients. But we told FDA that two design choices will decide whether the pilot becomes something the whole field can use or stays a good demonstration. Can the underlying architecture actually carry real-time? And does each reported signal carry its own evidence?
This piece walks through what we said and why.
Why real-time submission matters now
Today a reviewer sees trial results after a long chain of locks, snapshots, and handoffs. By the time a number arrives, the decisions it could have informed are often already made.
A real-time model changes what the regulator receives. Instead of a dataset, FDA gets a number. That is the right call for privacy. It also moves the hard question. As the filing put it:
"When a reviewer sees a number without the data beneath it, the question is no longer whether the data is correct. It is how the number was derived, by whom or by what, under which protocol version, and on what basis."
And then the line we keep coming back to:
"An aggregate without that lineage is a dashboard reading. An aggregate that carries its lineage is a regulatable artifact."
Everything else in our comment follows from that difference.
Provenance travels with the value, from the moment it is created
One point was about where provenance lives. Most systems treat it as something you reconstruct later: an audit trail, a report, a reconciliation exercise before inspection.
We asked FDA to treat it differently:
"The pilot should treat provenance as a property the data carries from the moment it is generated, not as a report assembled after the fact."
If the who, the what, the basis, and the version are captured when the value is created, they never have to be rebuilt. If they are added later, someone is guessing, and that guess becomes part of the record.
Real-time cannot sit on serial batch projections
Another point was the one experienced data engineers nod at. The clinical data pipeline is a chain of transformations. Source data is projected into one model, then another, then into the tables a reviewer reads. Each step waits on the last. Each tends to strip away the context that came with the value.
"Real-time cannot sit on top of a chain of serial batch projections."
We were direct that this limit is structural:
"[Experienced data engineers] know the status quo architecture cannot bear real-time no matter how hard it is tuned, because the limit is structural, not a matter of effort or compute. Calling a pipeline real-time does not make it real-time."
The risk for the pilot is a demonstration that works in a small, hand-tuned setting and then falls apart at production scale. A pilot that keeps the old architecture and adds speed on top "will measure whether a hothouse can be made fast, not whether the architecture can bear real-time in the world." We asked FDA to test the architecture, not only the timeline.
AI error is generative, and it propagates
The pilot is about AI, so we spent time on what AI changes. Traditional analytics errors and AI agent errors behave differently.
"An analytics error is forensic. It sits in the record and can be traced. An agent error is generative."
In a real-time optimization loop, a wrong signal does not stay put. It becomes the input to the next decision, a dose adjustment or a cohort change, and "propagates forward wearing the appearance of evidence."
Speed makes this worse, not better:
"Real-time review does not reduce this risk. It accelerates it. The faster the loop, the more important it is that every signal can be reconstructed and attributed."
Human attestation belongs in the data flow
As models compute more of the signal, the human act of standing behind a result matters more. In most systems today, that act is a separate sign-off that lives outside the data. We argued it should be captured as part of the flow itself, as a recorded event tied to the values it covers.
The reason fits in two sentences:
"An AI model is not bound by perjury. The human in the loop still is."
A model's output can be recorded. Only a person can attest. A pilot that keeps "a clear, machine-readable record of who attested to what, and when" protects both the patient and the integrity of the submission. We asked FDA to define a signal as review-ready only when its derivation can be reconstructed and its human attestation is on the record.
Measure time to reconstruct, not only latency
Every real-time program will report how fast signals arrive. We asked FDA to measure something else alongside it: how long it takes to reconstruct a reported result from its source inputs.
"A signal that arrives fast but cannot be reconstructed is not review-ready."
And from our detailed responses:
"A trial that runs fast but whose results cannot be audited has not saved time; it has moved the cost downstream to the review and inspection phase."
Latency tells you how quickly a number showed up. Time to reconstruct tells you whether anyone can check it against the record when it counts.
Capture has to be a byproduct of the work
None of this works if it adds a new job for site staff. Manual steps get skipped under time pressure, and then the record has holes. We were blunt:
"The decisive usability requirement is that provenance capture be a byproduct of the existing workflow, generated automatically as work is performed."
We asked FDA to measure how much new manual effort any system puts on investigators, and to treat a meaningful increase "as a structural defect, not a workflow issue to be smoothed over later."
A shared model, owned by no one
We also asked the pilot to point at open, implementation-agnostic standards rather than any single vendor's pipe, so signals stay comparable across sponsors and no one gets locked in. We named two. W3C PROV-O describes where a value came from. NGSI-LD, a linked data standard built on JSON-LD, "provides a graph-native way to carry that context and temporal versioning alongside the value itself." The pre-competitive layer, we said, is the shared data and provenance model, not the algorithms. Participants can compete on models and inference while agreeing on how a signal records where it came from. That model should belong to the commons.
Where we stand
Our closing line to FDA is still the clearest way we know to say it:
"Build the evidence into the architecture now, and real-time clinical trials become not just faster, but more defensible than the system they replace."
This is the problem ScientixAI works on. We think evidence should be the starting point of clinical data, not something assembled at the end. Our platform is designed to capture who produced each value, on what basis, and under which version, at the moment it is created, and to record each human attestation as part of that same flow.
We have written before about why clinical trials are still in the switchboard era, and about what we mean by evidence infrastructure. This comment is that argument made to a regulator.
If you are working on real-time review, AI in early-phase trials, or simply trying to answer "how was this number derived?" faster, come see how we approach it at scientix.ai.