so you can verify it actually did what you asked.

Nah, just build a harness that validates the output of one model by running it through the same model again to check for hallucinations... And to make sure that second pass isn't hallucinating, uh.... run it through a model a third time to check the second isn't hallucinating.

/s