The question
When a language model judges a phrase pair in both directions, can incompatible answers help select worthwhile rechecks? The evaluation separates correctness, agreement, and cost.
The comparison
The study compares conflict-selected rechecking with matched random rechecking and simpler prediction strategies. It also asks how many initial errors the conflict rule misses and how often another attempt makes a correct answer wrong.
Scope and status
This is ongoing work on short English phrase relations. This page describes the question and design; it is not a completed research result. Wider claims about production systems or other tasks would require separate evidence.
Start here
Read: An AI can agree with itself and still be wrong
Download the evaluation checklist
Related work
Mitchell et al. (2022), ConCoRD is relevant prior work on consistency correction. This project does not claim to invent consistency-based reasoning.