What's this about?
The paper studies formal statements that produce the expected solver verdict despite encoding the original problem incorrectly, then trains a verifier to score reference equivalence.
What exactly is verdict-preserving unfaithfulness?
It is a wrong translation that accidentally gives the right final answer—like entering the wrong equation but still getting “yes,” so an answer-only checker is fooled.
Why is it worth attention?
Agents need checks for semantic faithfulness, not merely executable output. The reported 0.961 AUROC and 11.3-point downstream gain are strong but remain author-evaluated.