Last released Sep 24, 2026
Typed, direct decisions from a pretrained language model. State + question + runtime label set in, a probability distribution over those labels (plus abstain) out, in one forward pass with nothing generated.