Ollaya is a local model server for a category most people didn't know they needed: decision models. Instead of paying an LLM to hallucinate a JSON blob, you get a classifier that reads a state plus typed questions (choice, score, "noul" — the source doesn't elaborate on that one) and returns calibrated probabilities in one forward pass. Laya:en, a ModernBERT-large at 421M, does five questions in 8–10 ms on a 4090. It never generates text, which is arguably the most refreshing feature list entry I've read this year. The engineering is sane: ~3 MB ONNX graphs that reference pinned, sha256-verified safetensors on the authors' own HF repos, no re-hosted weights. It speaks a "TypeSafe-compatible" wire format so existing Jev clients work by flipping one env var, which suggests there's a TypeSafe-shaped hole in the market this is aimed at. fp32 exports match the PyTorch reference on 100% of 2,383 questions per checkpoint. One binary, Ollama-style CLI, MCP serving for agents, Apache-2.0. Sam Altman will announce lowercase support for noul questions next quarter, pending $7 trillion and a quick eyeball scan. Open question: with kev:9b hitting 0.722 on typed-decisions, is anyone actually benchmarking these against a plain LLM-with-a-schema, or just against vibes?