VERIF-PRM S9 · Do verifiers verify

Does the process reward model verify or style-read

Refuted K-verif fired

frozen 2026-07-18 · adjudicated 2026-07-19 · ev:46e1500c4d39f9821c137ae0b00ac0da · spec:616bec72ae5ab

The frozen predictions

  1. H-verif-loc the PRM localizes ProcessBench error steps
    dense_localization_auc > 0.7 registered p 0.6 dense_localization_auc Value 0.2821441611813727 Subject VERIF-PRM Threshold > 0.7 Trust Adjudicated Evidence ev:46e1500c4d39f9821c137ae0b00ac0da Refuted
  2. H-verif-style the correctness preference is anchored, not style-carried
    style_share < 0.5 style_share Value -0.42684735632111825 Subject VERIF-PRM Threshold < 0.5 Trust Adjudicated Evidence ev:46e1500c4d39f9821c137ae0b00ac0da Confirmed

K-verif fired on dense_localization_auc < 0.55 , observed 0.2821 : methodological kill: the instrument failed on a real PRM, a works-ledger finding, not a scientific negative. A fired kill criterion is a registered negative result, not a failure of the machinery.

Adjudication figure for VERIF-PRM
VERIF-PRM fired its kill criterion (K-verif); the registered negative result.

Registered companions

Subjects and artifacts