Four Widely Used RLVR Verifiers Disagree by 41 Points on Their Own Ground Truth, and Whitespace Is 93% of the Damage

arXiv 2609.01354·high signalXBlueskyLinkedInCopy link

Applying metamorphic testing to the verifier rather than the model, the authors generate certified meaning-preserving answer rewrites so any rejection is a provable false negative, then measure 307,420 verdicts across four widely used verifiers. Self-validation ranges from 53.8% to 95.2% on identical inputs, a 41.3-point spread, and two configurations of the same library disagree on 49.9% of pairs. The residual is not spread across parsing categories but concentrated in whitespace and punctuation, which account for 93.0% of in-contract failures for the default LaTeX configuration, and a reference numeric cascade accepts off-by-one wrong answers as a step function of magnitude, from 0% below 10^4 to 100% at or above.