The standard of care
made executable.
Compiled Health is a clinical AI research company. We’re building a future where the standard of care is executable — where what medicine knows is what medicine does, at every visit, in every setting.
The accumulated, evidence-graded knowledge of medicine still lives in PDFs built to be read, not executed. Given this, it’s natural that AI has been looked to as a salvation. However, no matter how much model capabilities advance, the logic that reaches a clinical decision must still arrive in a form a clinician can inspect and reproduce — because liability runs through licensing, and that requirement does not relax as the models improve. In fact, as we argue in the essay, it binds harder.
The current attempts to insert AI into clinical recommendations are a chimera — for statutory reasons, to be sure, but also because no one, until now, has methodically measured the distance between what a model says and what the standard of care requires. To bridge this gap, we’re building Compiled Health to make clinical AI measurable against the standard, deterministic where it must be, and deployable where care actually happens. Bringing AI into clinical practice in its most meaningful, impactful, and feasible form.
We are physicians, engineers, and operators. Our team has practiced medicine on the prototypical twelve-minute clock, contributed to AI safety at frontier labs such as OpenAI, and shipped the first deterministic standards compiler into public health care — live today across Canadian provinces through PCare, our operating company, supported by partners and collaborators including Merck, Moderna, Roche Diagnostics, Mila, and Palantir Technologies.
01
Determinism where decisions are made
A clinical recommendation should read the same on Tuesday as it did on Monday.
Models are most adherent to standards at build time: extracting logic and projecting it onto a structure that is immediately useful for and reliable to clinicians. What executes in front of a patient is a compiled artifact — schema-enforced, formally verified for exhaustiveness, deterministic, and traceable to the source page and paragraph that justify it. We believe at base there can be no inference-time compute in the decision layer (although, to be fair, we believe statutory demands for all practical purposes dictate this as well).
02
Measurement against the standard of care
What can be measured can be governed.
Today’s benchmarks test whether a model sounds like medicine. Our internal benchmark and tooling systems test whether models follow the specific kind of medicine a clinician is compelled to follow based on their location through thousands of compiled artifacts and patient scenarios, each carrying the rules that govern them. The distance between a model and the standard of care is measurable, and we measure it.
03
The gap can be closed
Where the standard is defined, a model can be taught to meet it.
Our physicians and engineers tune models against our library and the over 108 specific, grounded scenarios and outcomes contained therein until behaviour converges: same facts, same answer, and a refusal where the standard refuses. We believe probabilistic systems per se cannot be executed at the point of care — but, likewise, that they’re essential to improve and leverage at build time to create ever more robust deterministic logic at the point of care.
04
Each layer where it belongs
Models at build time; artifacts at the point of care.
We design the line between build time and run time and build the bridges across it — surfacing, versioning, and deploying compiled artifacts into the software clinicians already use. Where hosting is required, we operate it on data-sovereignty first principles: the artifact travels; the data does not.
is where this began.#
Our operating company runs the first deterministic standards compiler inside Canadian public health care: clinical standards distilled, approved by the governing bodies that hold them, and executed at the point of care in both official languages.
PCARE.CA →Work on the standard of care.
We’re a small team of physicians and engineers building the execution layer medicine has been missing. If that’s your kind of problem, we’d like to hear from you.