Laya is an open-source (Apache 2.0) non-autoregressive decision engine that classifies typed questions (choice, score, boolean) over any input state in a single forward pass at ~35 ms on GPU.

It outputs calibrated confidence scores without text generation and supports fine-tuning on custom data. Benchmarks report ~10x lower latency and 83.8% in-task accuracy versus TypeSafe Jev.