What this is — a small imitation of
Jev, TypeSafe's "System One" model for runtime-defined semantic decisions: give it state, a question, and typed options, and it reads the option probabilities straight from one forward pass — no answer sentence, no JSON repair, no decoding loop. Here the base method is
laya, a 421M non-autoregressive decision model (sub-35ms, temperature-calibrated, 100+ languages), compared against four classic baselines (tfidf-cosine, naive-bayes, jaccard, knn) so you can see what the decision model actually buys you.