Auscult.ai combines language understanding with structured diagnostic reasoning.
Inspect competing diagnoses, the evidence used and what remains to be asked.
Both studies build on the reasoning architecture presented in our poster: explicit knowledge, structured evidence and inspectable differentials, adapted through dedicated profiles and mappings. They support adaptability within diagnosis. Different cohorts and scoring paths mean the percentages are not directly comparable.
Rare diseases · poster study
5,349 published cases.
HPO phenotype findings, 10,793 disease concepts, ontology-based native reasoning. ReasoningLayer and LIRICAL used the same case corpus and annotation release.
Earlier frozen configuration; retrospective published cases. In 74% of cases, the gold diagnosis has ontology annotations citing the case’s own publication. This measures recognition with shared source knowledge; clinical validation remains necessary.
On 256 synthetic cases, Auscult’s ranking component and Gemini received the same findings and diagnosis names. The primary Gemini comparator also received the same numerical profiles and scoring formula.
Top-1 accuracy on a fixed-evidence ranking task, using the recorded model google/gemini-3.7-flash. This does not test full interviews or isolate the native engine. Internal research on synthetic data; clinical validation is pending.
22 June 2026 · LIP6, Sorbonne Université, Paris. The official programme lists David Loiret’s presentation on structured, verifiable and actionable explanations with ReasoningLayer.
Conference participation is distinct from journal publication or endorsement. The original poster is accompanied by a current claim review and a correction to its June venue: LIP6, Sorbonne Université in Paris.
The October development revision strengthens evidence admission and source lineage. Targeted checks have passed; complete interviews through this revised stack and independent clinical validation remain ahead.
01
A differential you can inspect
Language models structure observations. Native inference links hypotheses to matched
and missing evidence; explicit application scoring produces the ranked differential.
02
Knowledge with source lineage
Ontology concepts give findings a shared vocabulary. Curated profiles now carry input
locations, document digests and reference lists through inference. Numerical support
in the cited publications remains to be verified.
03
Questions guided by gaps
Missing evidence and information-gain estimates guide follow-up questions. Explicit
unknowns, conflict checks and abstention rules make the system’s limits visible.
Clinical calibration is still needed.
04
Configurable patient context
Patient context, knowledge profiles and question policies are explicit inputs that can be adjusted without retraining the language model. A clinical benefit from personalization has not yet been measured.
05
Replay that can be checked
Recorded application scores were reconstructed for 18 artificial interview turns on an earlier build. This checks computation from captured state; it does not replay language extraction or the entire stack.
06
Evidence boundaries tested
Targeted native and application checks cover evidence admission, conflicting observations and source transport. Passing those checks is engineering evidence, with full-stack and clinical evaluation still to come.
Earlier product demo · 2 min 55
An illustrative consultation showing the interface. This recording predates the October engineering updates.
