Does Tessary use an LLM to check for malformed output?

No. Checking a structured output against its declared schema is deterministic: either it parses and matches the shape, or it doesn’t, so there’s nothing left for a model to judge. No LLM call happens.

That matters for two reasons. Cost is the first: a deterministic check runs at negligible per-trace cost, which is what lets Tessary validate every trace instead of a sample, unlike the model judgment reserved for traces a classifier has already flagged. Confidence is the second: a judged check can be uncertain about how bad something is, but a schema either matches or it doesn’t, so every malformed_output finding carries the actual mismatch as evidence rather than a probability.

The tradeoff is scope. The classifier can tell you an output broke its contract; it can’t tell you whether the content inside a technically valid output is any good. That’s a different, harder question, and not this classifier’s job.