Topic
SemIf reads the probability a language model assigns to each option instead of letting it write an answer. It runs on the model already installed on the server, so the data never leaves the building. These pages describe how we use it and what it gives on real cases.
Qwen3.8-Flash-Next locates the optic disc correctly in 120 decisions out of 120, then gets the eye wrong on 51 images out of 60. The protocol, the numbers and the fix we adopted.
Read moreA decision model that answers in a hundred milliseconds is not used the way a language model that answers in ten seconds is. Decision loops, comparing hypotheses, the generator-and-judge pairing, and the return of the small specialized model: the second part of our article on Jev and SemIf.
Read moreFor three years, we have been asking models built to write to make decisions. A new family of models does nothing but decide. First part: what a decision model is, with Jev and SemIf.
Read more
