Measure how much ROVER voting actually helps: on a small hand-verified ground-truth set (and the clear 3:1 cases like prytool>pyytool), compare consensus output CER/WER against each single model. Confirms voting beats best-single, not just agrees. See 2026-06-27-consensus-rover.md.
Measure how much ROVER voting actually helps: on a small hand-verified ground-truth set (and the clear 3:1 cases like prytool>pyytool), compare consensus output CER/WER against each single model. Confirms voting beats best-single, not just agrees. See 2026-06-27-consensus-rover.md.