Methods, benchmarks and results from the team and the Frontier network.
How we measure whether an answer is actually usable in a consult: grounding, latency and clinical acceptability, scored by practising specialists.
Structured disagreement between reviewers and models, written up and sourced, outperforms simple preference labels on safety-critical categories.
Before-and-after timing across twelve clinics: where the hours went, and where they came back.
Our layered approach to removing identity from clinical text, evaluated against modern linkage attacks.
A fixed battery of clinical evaluations every release must clear before it ships, and how successive versions have moved.