Stop LLM syntax errors: make broken code and DSL impossible by construction

A language model samples each token from a distribution, and even a model that's seen your language fluently assigns some nonzero probability to a token that breaks it — a mismatched bracket, an undefined keyword, a missing separator. Over enough generations that token eventually gets drawn. Adding 'output only valid syntax' to the prompt lowers the odds; it doesn't remove the bad tokens from the distribution.

Retrying on a failed parse helps but is expensive and unbounded: nothing guarantees the second attempt parses either, and you've now doubled latency and cost on the path that was already going wrong. You're treating the symptom every time instead of the cause.

dsl.ai takes the bad tokens off the menu. It compiles your language's grammar into a decoding constraint, so at every step the model can only sample from tokens that keep the output inside the grammar. A bracket that wouldn't close, a keyword that doesn't exist, a statement that breaks the structure — those aren't unlikely, they're simply not available to sample. The output can't be syntactically wrong because the wrong paths were never reachable.

This is the same mechanism behind GBNF grammars in llama.cpp, Outlines, and XGrammar; dsl.ai's job is to compile your grammar into it from a file you paste, no GPU and no training set. And because syntax is guaranteed, the model spends its probability mass on the choices that actually matter — what the code should do — rather than on whether it parses.

Guaranteeing generation is half the story; you still want to confirm any given string is valid — including DSL a human wrote, not just what the model produced. dsl.ai generates a deterministic parser from the same grammar, so checking is exact and reproducible: valid, or invalid at a specific position with what it expected there. A parser names the failure; a model asked to grade itself can be confidently wrong.