reward-lens

Do verifiers verify

Does the process reward model verify or style-read

refuted registered

The reading

  1. the PRM localizes ProcessBench error steps

    dense_localization_auc > 0.7 0.6 Does the process reward model verify or style-read refuted registered > 0.7 The frozen study "Does the process reward model verify or style-read", adjudicated on 19 July 2026 at the registered level of the project's evidence ladder. One frozen study, adjudicated once, on qwen-prm, over processbench-full. It does not establish that the same holds on another subject, another dataset or another run. Open the record › refuted

  2. the correctness preference is anchored, not style-carried

    style_share < 0.5 Does the process reward model verify or style-read refuted registered < 0.5 The frozen study "Does the process reward model verify or style-read", adjudicated on 19 July 2026 at the registered level of the project's evidence ladder. One frozen study, adjudicated once, on qwen-prm, over processbench-full. It does not establish that the same holds on another subject, another dataset or another run. Open the record › confirmed

methodological kill: the instrument failed on a real PRM, a works-ledger finding, not a scientific negative

dense_localization_auc < 0.55 Does the process reward model verify or style-read refuted registered > 0.7 The frozen study "Does the process reward model verify or style-read", adjudicated on 19 July 2026 at the registered level of the project's evidence ladder. One frozen study, adjudicated once, on qwen-prm, over processbench-full. It does not establish that the same holds on another subject, another dataset or another run. Open the record ›

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

Measurements

verification_score
Does the process reward model verify or style-read refuted registered none registered The frozen study "Does the process reward model verify or style-read", adjudicated on 19 July 2026 at the registered level of the project's evidence ladder. One frozen study, adjudicated once, on qwen-prm, over processbench-full. It does not establish that the same holds on another subject, another dataset or another run. Nothing was registered against this value in advance, so it is a supporting measurement rather than a tested one. Open the record ›

Provenance

card
VERIF-PRM
frozen_at
adjudicated_at
spec_hash
spec:616bec72ae5ab5714b01adaef378aba9
study_id
study:campaign-verif-prm@v1#616bec72
Evidence records
46e1500c4d39f9821c137ae0b00ac0da
spec_file
campaign-verif-prm.json