Structured outputs can hurt model quality.
When they have (silent) bugs. We show that both XGrammar and llguidance make masking errors that degrade accuracy. Sometimes below the unconstrained baseline.
dotjson? 0 errors.
Blog post 👉 blog.dottxt.ai/comparing-the-…
The same @jsonschema can mean different things depending on the inference provider. Different subsets, interpretations, or dialects.
It completely defeats the point of structured outputs. We're working closely with the community to fix this.
json-schema.org/blog/posts/dot…
