decider: a language model that returns typed decisions in one forward pass

decider is a language model that does not generate text: it reads a state and typed questions and returns a probability distribution per question in a single forward pass. Weights are Apache 2.0; decider-2b takes 18 ms and decider-35b-a3b 41 ms per decision on a B300 in bf16, and the 35B ranks fourth on Decision Index at 54.3 behind Jev's 59.5.

• Model takes a JSON state and choice, score or noul questions, returning a distribution with no decoding or parsing
• decider-2b: 18 ms per decision; decider-35b-a3b (3B active): 41 ms on one B300 in bf16
• On Decision Index the 35B is fourth at 54.3 vs Jev's 59.5, with the best calibration of 32 entries (ECE 3.1)
• Limits: no chain of thought, weak 2B calibration on hard items (ECE 0.30), English only