Notes
What we are building, and what we are learning.
Short, plain notes on decision models: how they are measured, how they are built and what we ship.
- Certified error rates, explainedHow a calibrated model decides what it keeps, and why the threshold is set before go-live, not after.Note
- Why a decision gets one forward passScoring every option directly, instead of generating text and parsing it, is faster, cheaper and the reason the probabilities can be trusted.Note
- What is in a lineage recordThe base model, the data, the licences and who made them: what we publish with every new Myles model, and why procurement should ask for it.Note
- The Decision API, version 1Three question types, a probability for every option, several questions about one situation in one request.Release
- Myles-400M: 91.8% on 1,400 held-out decisionsAhead of the medium tier of a leading hosted LLM family, on one CPU thread, at about one 450th of the cost.Evidence