Verification vs. Authoring vs. Orchestration: The Three Camps of Clinical-Trial AI

The authoring camp points AI at the blank page. Protocols, narratives, regulatory documents, submission sections. Feed it structured inputs, get back a draft in minutes instead of days. Companies whose core product points at the blank page — protocol design, medical and regulatory writing — include Faro Health, Narrativa, and Yseop.

The value is real. The blank page is expensive, and a competent first draft compresses weeks of writing into an afternoon.

The ceiling is also real. A faster draft is still a probabilistic draft. The model can produce a fluent eligibility section in seconds. What it cannot produce, in the same motion, is the record of which rule it applied, which source value it read, and what it did not check. Fluency is not verification. A document that reads correctly and a document that has been shown to be correct are two different objects, and inspection only cares about the second.

The camp has heard this, and it has answered. The newer entrants no longer sell speed alone. They sell speed plus defensibility: explainable generation, regulator-ready output, documentation you can carry into a review. Companies whose core product points at the blank page and now pairs generation with an explainability or accuracy claim include Biorce and AlphaLife Sciences. This is the right thing for them to be reaching for. The reach is worth taking seriously, and it is worth being precise about what it delivers.

Look at how that defensibility is established. In the common pattern, the system that generated the draft also scores the draft, and the score is reported as an accuracy figure. That is the model grading its own homework. The score inherits the blind spots of the thing it scores, because it comes from the same place. When the figure is high, the residue is the problem: a self-reported accuracy rate is silent about which outputs fall in the remainder, and silent about who is accountable for them at submission. An inspector does not ask how often the system is right. An inspector asks whether this determination, on this document, respected every rule that governed it, and who attested to that.

Explainability has the same shape. A model that narrates its reasoning has produced more text from the same model. The narration is an artifact of the generator, not an independent check on it. It is legible, which is worth something. Legible is not verified.

The orchestration camp points AI at the workflow. Route tasks, track status, coordinate sites, move a study through its stages. Companies operating in trial workflow and content management include Veeva and, at operations scale, IQVIA. Adjacent to them sit the systems that capture and carry the data itself: Medidata in electronic data capture, Medable in decentralized trials and patient-facing assessment.

A newer cohort has taken the word itself. Companies operating in agent-based trial coordination, where software plans and sequences steps rather than only tracking them, include Veridix from the Emmes Group in modular study build agents, ConcertAI in agent-assisted feasibility and study design, and Azra AI in care and enrollment coordination inside health systems. The layers differ. The shared property does not: this is the plumbing of modern trial operations, and it is genuinely hard to build well.

Agents raise the ceiling on orchestration. They do not change what orchestration is for. A system that plans its own next step is answering "what happens next" with more autonomy than a workflow engine. It is still answering "what happens next." The question of whether the determination inside any one of those steps was defensible sits outside the loop, no matter how capable the loop becomes. Autonomy is not evidence. A more sophisticated proposer is still a proposer.

Orchestration answers "where is this in the process." It does not answer "is the determination inside this step defensible." Knowing that a document reached the right desk on the right day is a logistics fact. Whether the decision recorded on that document respected every rule that governed it is a different fact, and workflow software is not built to establish it.

Both camps make the work move. Neither camp proves the work is right. Nearly the entire field sits in these two.

The verification camp points AI at a different target: the correctness of a specific decision, established at the moment the decision is made, in a form an inspector can reconstruct later.

Define it functionally, because the category is larger than any one company. A system is doing verification when it produces a deterministic, inspectable artifact that records the rule invoked, the values checked, the operation that returned pass or fail, and the boundary of what was not checked. Anyone who produces a conforming artifact is in the camp. This is a posture, not a product claim. The category is open by design.

It is also no longer empty. Companies whose core product points at checking rather than producing include Beaconcure in statistical output and clinical document review, Leucine in validated execution for regulated manufacturing environments, and Veriprajna in clinical safety checks. The lanes are different and the architectures are different. The shared instinct is the one that matters: the check is built as its own object, not as a feature of the generator.

A probabilistic model proposes. An independent check disposes. The model proposes; the proof disposes.

More members strengthen this camp rather than crowding it. A category with one occupant is a marketing position. A category with several, arguing openly about what a conforming artifact requires, is a standard forming. The second is the useful outcome, and it is the one worth wanting.

What makes verification a different animal is one architectural decision: the layer that proposes a decision is separated from the layer that verifies it.

This is why a confidence score does not qualify. A confidence score is generated by the same model whose output it scores. It inherits that output’s blind spots. It is the model grading its own work: correlated evidence wearing the label of a check. The same defect appears in any arrangement where one model checks another model trained on the same data, however the second model is labeled. Verification requires uncorrelated evidence — a check whose failures do not share a common cause with the thing being checked.

It also requires more than the word "deterministic." The field has learned to say deterministic, and a deterministic rule script is not the end of the argument. A script can be wrong in ways no one checked; determinism guarantees the same output every time, not that the output is correct. The higher bar is formal validation: a check whose own correctness is machine-verifiable, producing an artifact an inspector can reconstruct rather than take on trust. That is the boundary this camp is actually defined by — not that a check runs the same way twice, but that the check can be shown to be right.

In the three-gate form of this architecture, that means a deterministic rule check, a formal structural proof, and a mandatory human attestation. Three substrates that fail in different ways, converging on defensible evidence. The formal-proof gate is designed to emit a machine-checkable artifact; it is an architecture target, not a shipped production claim. The point that carries at category altitude is simpler: independence is structural, or it is not real.

You cannot bolt verification onto an authoring tool after the release. The same model that wrote the draft cannot be the independent check on the draft, whatever the second pass is called. Verification is a property of how the system is built, decided before the first line of output, not a feature added when a buyer asks for it.

That is the whole distinction. A confidence interval describes how often a system is right across a population. A proof artifact addresses whether one specific output respected every constraint that governed it. Both matter. Only the second survives an inspector asking, two years later, to see the decision reconstructed. In addition to statistical validation, never instead of it.

This map matters most in one window of the trial lifecycle, the period from study start-up to First-Patient-In, the category we call Trial Activation Intelligence. That is where a decision is made once, carried into a submission, and read back by a regulator long after the model that proposed it has moved on.

Authoring gives you a faster draft. Orchestration gives you a managed workflow. Verification gives you an artifact you can defend. The buyer who asked "will this survive inspection" was asking for the third thing the entire time.

Provably right, not probably right.