The model never saw the reference's patent number, title, assignee or dates, so it could not lean on anything it may have memorised. The control row is what makes the other two mean anything: recall alone is trivially gained by flagging everything, so the same screener was run over references the examiner did not cite, drawn from the same corpus and passing the same priority-date gate.

The denominator is examiner-applied references that are granted US patents inside the 171,695-patent corpus. 73% of examiner citations point at pre-grant publications and 6% at non-patent literature; both sit outside this corpus and are excluded from numerator and denominator alike.

This is recall against the examiner, not against ground truth. An examiner's own search runs 45 to 85% recall, so every reference Nightshift finds that the examiner missed is scored here as a miss. The number is a floor, not an estimate.