For medicine · compliance, risk and medical-legal review
Every line proved to the chart, or held.
Apodicta sits beside any AI model and makes its output hallucination-resistant by architecture. On a medical record that means: it reads the whole file, finds the questions the file raises, reasons each one through to the source lines, and renders nothing it cannot ground. What it cannot prove, it names as held. The model stays free; what it says does not.
What it does on a medical file
Reads the whole record
Notes, orders, results, imaging reports, correspondence, device labeling, policy text. Any size, no question attached. It finds the issues on its own, the way a reviewer with unlimited hours would.
Proves each line to a source
Every statement in what it writes carries a citation to a line in the file, and the citation is checked by code before the line can render. A claim with no source does not get softer words; it gets held.
Names what it holds
The output is a memo with two kinds of lines: proved, with the cite, and held, with the reason. A reviewer reads the held lines first. That is the review, done in the order the risk runs.
Where it is used first
Medical uses · stated plainly
- Compliance and risk. A device's authorized use is a sentence. Every note, claim or report that cites the device can be read against that sentence, line by line, with the mismatches held.COMPLIANCE
- Medical-legal review. A chart read the way a chronology demands: what the record shows, what it does not, where the two are confused. The same engine already does this on criminal discovery.MED-LEGAL
- Documentation review. Summaries and letters checked against the chart before they leave the building; an unsupported line is held, not sent.DOCUMENTATION
- Literature and guideline grounding. A curated library seat exists in the engine; medical sources are not yet seated. Until they are, the engine grounds to the file in front of it, not to the literature.LITERATURELATER
- What it is not. Not a diagnostic device, not clinical decision support, not a substitute for a clinician's judgment. It checks text against text and says what it could not check.NOT
The record it keeps on medical text
The discipline is easiest to see on public records. We keep the record of FDA AI-enabled medical devices: 1614 devices, each with its authorization, its Indications for Use copied verbatim from FDA's own document, its regulation, and the adverse-event report counts MAUDE holds for its product code. Nothing on those pages is summarized or characterized, which is the same rule the engine applies to your file.
Numbers on this site are labeled: every figure on the proof page is an internal test, not independently audited, and we say so beside each one. There are no medical outcome claims here, because we have not run a medical pilot yet. The first one is the pilot below.
How a medical pilot runs
Your hardware, your records
The workstation version runs on a machine in your building with a local model. No record leaves it. The cloud version runs on a frontier model under your key; the enforcement boundary is the same either way.
A closed set, scored
You pick redacted files you already know. We score what it proved, what it held and what it got wrong against your reading, and you keep the scores. The same harness scores our legal work; the method is on the evaluation page.
See how the gauntlet works, why any model can sit behind it, or what a pilot includes.