Deterministic, safe, and auditable systems for using LLMs in patient-facing healthcare.

August 2026

The problem

A general-purpose language model can hold a fluent medical conversation, but its behavior at the decision boundary is unreliable. The same case can be escalated in one run and dismissed in another, and the outcome depends on phrasing luck rather than on a defensible procedure. The cases where this flip-proneness appears, such as a possible rabies exposure or postpartum leg swelling, are exactly the cases where a wrong answer carries the highest danger if missed.

What we built

The architecture relocates every load-bearing decision from the model into deterministic machinery around it. Extracted clinical facts are accepted only with verbatim supporting quotes from the member's own words. A deterministic falsification pass tries to break the leading diagnosis before any conclusion is communicated. An independent danger vote adjudicates the care tier separately from the reasoning chain that produced the differential. A post-draft audit stack, with redraft and a deterministic floor fallback, checks every reply before it ships. Each turn emits a structured visit note and a complete decision trace, so every conclusion is auditable end to end.

The architecture is model-agnostic: the underlying models are off-the-shelf and unmodified, and each stage's contract is defined over text, not over any model's internals.

What we measured

On a 300-case red-flag evaluation suite, the final gate run scored safety 1.0 on a 60-case sample under a ship / no-ship verdict protocol. A targeted probe of previously flip-prone cases (rabies bat exposure, airway burn, postpartum deep vein thrombosis) produced 15 of 15 correct escalations, compared with 3/5, 3/5, and 2/5 before the independent danger vote was introduced. A 40-sample false-reassurance probe found zero under-escalation. The known residual failure mode is calm-side over-escalation, which fails in the safe direction.

Read the working paper

The full write-up, with the stage-by-stage architecture and the evaluation protocol (working paper, August 2026, not yet peer-reviewed). Download the PDF.

For methodology questions or access to the underlying evaluation data: hello@pymander.app