What is ONC HTI-1?

Why It Matters

The HTI-1 final rule introduced transparency requirements for the AI and other predictive algorithms that are part of certified health IT. The reach is the point: ONC-certified health IT supports care at the large majority of US hospitals and office-based physicians, so a disclosure requirement attached to certification covers most of the clinical software market without anyone choosing to opt in.

For a builder, the practical effect is that the question “how does this model work” has a documented answer you have to produce, in a standard shape, on a schedule.

What It Requires

A predictive decision support intervention is in scope, and the criterion asks a certified module to expose the source attributes behind it: what the model was trained on, how it was validated, how it performs, how fairness was assessed, and how it is maintained. Developers attest to those attributes, and the attestation carries intervention risk management practices behind it rather than a statement of intent.

Two consequences follow. The disclosure has to stay true as the system changes, and the practices it attests to have to exist in the operation of the product, not only in the documentation.

Where It Breaks

The requirements were drafted for models with defined inputs and outputs. A feature backed by a language model is non-deterministic, its training data is a corpus rather than a table, and its behavior moves with prompt, retrieval, and tool changes. Teams end up overstating what they can attest to, or writing a disclosure once and letting it describe a system that no longer exists.

The second break is scope confusion inside the product. A retrieval feature, a coding suggestion, and a triage assistant may all sit behind one certified module, and each has different attributes worth disclosing. Treating the module as the unit of disclosure hides the differences that clinicians need to judge the output.

How Flytebit Handles It

Disclosure is a byproduct of how we build. Versioned prompts and models, an eval harness with clinician-graded sets, drift monitoring with a revalidation cadence, and a decision record per run together produce the evidence an attestation needs without a separate documentation exercise. The industry application is on our Healthcare & Life Sciences page, and the control design is our AI governance and risk work.