Yep, that’s a good way to frame it. The verification funnel is really the important part. We discard candidates if the failure can’t be reproduced, a fix is already in flight, no safe patch is found, verification fails or times out, or the patch/apply/runner step breaks.
So “the model produced a diff” is never enough to open a PR. If verification is inconclusive, we stop there.
Surfacing those reasons more explicitly is a good suggestion.
