A model whose decider_config.json says "layout": "chat" (or "chat_template": true) is read with every prompt wrapped in its
tokenizer's chat template (thinking off, no system prompt, the letter read after "Answer: (" in the assistant turn), on every
path: Decider (decide, decide_batch, decide_json, system_one, schema()), CUDA graphs, MPS and CPU, decider.serve
(/decide, /v1/systemone, shared-prefix fork, schema cache), decider.evaluate, the probes, the benchmarks and the vLLM check.
Plain-layout models are unchanged: 359 pinned prompt sets from 1.1.3 match, and decider-2b v10 gives identical answers.
An unknown layout is refused with a ValueError; serve_v1 refuses chat models. The eager DecisionModel now also switches
off the cuDNN SDPA backend, as the engines have since 1.0.2. No weights change: Mapika/decider-2b stays v10. 152 tests.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>