Posts tagged Json Mode — AI × software engineering, made practical.
Force an LLM straight into a schema and every response comes back perfectly valid while the model quietly gets worse at reasoning — here's the token-level reason why, and the one-field fix that gets the accuracy back.
Your schema validator has one job — checking shape, not truth — and constrained decoding quietly turns that blind spot into a hallucination wearing a green checkmark.