decider turns typed questions into probabilities

decider makes a strong case for using language models as decision engines rather than expensive text generators, though its narrow interface is also its main constraint.

• –Eliminates decoding and JSON parsing by reading probabilities directly from option-token logits.
• –Enables confidence-based routing, escalation, and selective human review.
• –Local CPU, Metal, CUDA, and GGUF support makes deployment practical beyond large GPU servers.
• –One-pass inference is a poor fit for multi-step reasoning, arithmetic, or knowledge-heavy questions.
• –Calibration and accuracy vary significantly by model size, so developers still need task-specific evaluation.