Evidence for 182.C.19.04: face merging, identity swaps, duplicate people,
flicker, sudden age/skin-tone shifts, and a speaker inheriting another person's
mouth or voice.
Real face crops from the committed frame
../2026-09-10-openrouter-face-detection/two-foreground-subjects-plus-bystanders.png,
alpha-blended, duplicated, blurred and tone-shifted, all measured with YuNet
detection and SFace embeddings.
The finding that matters most is not one of the six#
Blur destroys identity long before it destroys detection.
| Gaussian σ | detected? | confidence | cosine against its own unblurred self |
|---|---|---|---|
| 0 | yes | 0.938 | 0.866 |
| 4 | yes | 0.922 | 0.453 |
| 8 | yes | 0.879 | 0.033 — random |
| 12 | yes | 0.792 | −0.103 — negatively correlated |
| 16 | no | — | — |
At σ=8 the detector is 0.879 confident there is a face, and that face's
embedding has collapsed to noise. A pipeline that gates on detection score will
feed noise into identity comparison and get a confident wrong answer.
assertEmbeddingUsable runs before any of the six detections, because every one
of them reads an embedding.
The signatures#
Face merging#
| alpha | cos(A) | cos(B) | margin |
|---|---|---|---|
| 0.00 | 0.882 | 0.251 | 0.631 |
| 0.25 | 0.688 | 0.238 | 0.450 |
| 0.50 | 0.341 | 0.320 | 0.021 |
| 0.75 | 0.257 | 0.703 | 0.446 |
| 1.00 | 0.210 | 0.828 | 0.618 |
A 50/50 blend of two real faces embeds as neither person, and equally unlike both. The signature is a low best cosine and a near-zero margin — both, in conjunction. A low cosine alone is an unknown face; a small margin alone is two strangers who look alike.
Duplicate people#
The same crop twice in one frame gives two simultaneous detections at mutual cosine 0.9565; two different people side by side give 0.3243. Simultaneity is what makes it a fault: the same person across two frames is the point of tracking, and the same person twice in one frame is impossible.
Flicker#
Reproduced at the detection boundary — blur alternating σ 12 (detected, 0.792 confident) and σ 16 (not detected) flipped presence on 11 of 11 transitions.
It did not reproduce by size: YuNet held a detection down to a 16-pixel face and never dropped one oscillating between 56- and 40-pixel patches. So flicker here is a marginal-quality phenomenon, not a marginal-size one.
Sudden skin-tone shift#
An ITA move of +40.9° under constant lighting left the identity cosine at
0.954. SFace is essentially blind to skin tone, so an identity check will
not catch a tone change. They are independent detections and both are needed
— and a tone comparison across mismatched lighting is refused, because
182.C.19.02 measured a one-stop exposure change moving a subject 58.2°.
Identity swaps#
Measured under 182.C.19.03: occlusion swapped a track onto the other person
twice. See ../2026-09-10-identity-tracking/.
A speaker inheriting another person's mouth or voice#
Split, honestly. The mouth half is measurable and already built —
human-video-speaker-binding.ts (182.C.17) binds speech runs to mouth motion
and detects a face moving its mouth while somebody else speaks. The voice
half needs speaker embeddings or diarisation, which is not staged on this box,
so it is reported as not-measured rather than approximated.
Not measured at all#
| Fault | Why |
|---|---|
| age shift | No age model is staged; 182.C.19.02 records why a number would be worse than none. |
| inherited voice | Needs speaker embeddings or diarisation. The mouth half is covered; the voice half is not. |
Both travel in the report rather than in a comment, because a report listing clean detections reads as clean whether or not two more were never attempted.
Per-item status this evidence supports#
| Item | Was | Is |
|---|---|---|
182.C.19.04 |
not-started |
done for five of six; age shift and the voice half of the sixth are not measured |
182.C.19.08 |
pending | a fourth reason a universal cutoff cannot work: image quality moves the cosine to noise |