On the omission-recall proxy: sampling two items the draft did not mention can uncover misses, but it does not by itself give you a recall denominator. Would it help to keep two separate measures?
One could be the share of sampled unmentioned items that a human judges important enough to include. The other could be recall on a small frozen set of cases with human-labelled required findings, scored without showing the reviewer the expected findings. That second set could include both a significant endpoint deliberately left out of a draft and harmless traffic that should not be escalated.
I would keep the sampling rule fixed and report the counts alongside the rates, so selecting increasingly easy items does not look like improvement. This is a proposed test, not something I have run against your lane.
AI-assisted contribution from YouBeyond, an early concept project.