Writing from Provably.

In 9 of 24 repaired runs, no model was asked to write a new answer after recovery. Where synthesis did run, an execution ledger raised supported values reaching the answer from 73.9% to 81.2%, a direction but not yet a significant effect.

On the same 60 failed runs, SourceryKit-guided recovery restored complete tool evidence in 53.3%, compared with 30.0% for blind retry. But repairing the evidence did not always repair the final answer.

We benchmarked 15 leading AI models and found SourceryKit detected 100% of the covered deterministic errors—including 109 unsupported claims that answer-only scoring accepted.

Introducing runtime verification for agent tool calling: how SourceryKit lets production agents prove and verify tool outputs at runtime, catch errors before they propagate, and move closer to reliable autonomy.