Why LLMs return invalid JSON — a language model has no parser

A model predicts tokens, not documents — it has no parser. So "JSON mode" constrains sampling rather than guaranteeing your schema. Measured, plus the repair layer.

Read Original

Related