Your JSON is valid. But is it correct?

ValidJSON is an independent research-and-engineering shop focused on the reliability of structured output from large language models. Grammar-constrained decoding makes the JSON syntactically valid; we work on whether the values are correct, where they fail, and what to do about it.

What we ship

• valjson — open-source diagnostic tool. Per-grammar-role loss decomposition, per-field confidence, inference-time abstention via margin gating. pip install valjson.
• The paper — Valid JSON, Wrong Answer. Per-role regression detection and presupposition labeling. Citable DOI; code and data on GitHub.
• Consulting — structured-output reliability for teams shipping JSON, tool calls, and document extraction. Diagnosis first; prompt, schema, gating, and fine-tuning applied as the diagnosis warrants.

Start here

Most JSON-from-LLM problems can be diagnosed and fixed without us. How to work with validJSON walks through the order to try things, cheapest first, with a free 30-minute consult at the end if you’ve worked through the ladder and still have an open question.