# Face swap by identity-attribute inconsistency (2026-09-11)

Evidence for `182.C.07.03`, whose face-swap detection was left unmeasured on
2026-09-10 with the reason: no manipulation detector is staged, and an
uncalibrated deepfake score invites belief it has not earned.

**That is still true of a manipulation detector, and this is not one.** It is a
consistency check, and it works because a swap is not a change of person — it is
a change of _face_ on a person who is otherwise still there.

## Fixture

Subject A's face region composited into subject B's face box in the committed
frame, with a feathered elliptical mask so the edges do not read as a tear. Real
pixels of a real person on another real person's head.

## Result

| Frame            | face cos(A) | face cos(B) | **hair lightness** | skin lightness |
| ---------------- | ----------- | ----------- | ------------------ | -------------- |
| clean, subject A | 1.000       | 0.261       | 9.80               | 45.49          |
| clean, subject B | 0.261       | 1.000       | 26.67              | 70.59          |
| **swapped**      | **0.737**   | 0.381       | **27.06**          | 49.41          |

The face embeds as **A** at 0.737 while the hair around it is **B's** — 27.06
against A's own 9.80. Neither reading is wrong on its own; together they are
impossible. That is the whole detector.

**Both halves are required.** A face that matches nobody is an unknown face, not
a swap. Hair that matches nobody is a haircut. The fault is the two being
confidently _different people at once_.

## A third signal fell out of it

The swapped face's **skin** lightness measured 49.41, between A's 45.49 and B's
70.59, because a feathered composite mixes the two. A face whose skin matches
_neither_ enrolled subject cleanly is a face that was blended — and
`182.C.19.04` already reports that shape as merging. It is reported as a
separate signal, because a blend and a clean paste leave different traces and a
caller should be able to tell which happened.

## Blind spots, named

- a whole-head replacement, where the hair travels with the face and nothing
  disagrees
- a swap between two people who look alike, where the face cosine does not
  separate them
- a fully synthesised person, who has no reference to be inconsistent with
- any manipulation that leaves the face and its surround belonging to the same
  person

A clean result from a consistency check is not proof of no manipulation, and
`SWAP_BLIND_SPOTS` carries that rather than leaving it implied.

## Lighting

Hair and skin lightness are colours, and `182.C.19.02` measured a stop of
exposure moving skin ITA by 58°. The comparison is valid **within one frame**,
where the light is the same by construction, or across frames the caller has
already matched. `assertSameFrame` enforces it.

## Per-item status this evidence supports

| Item          | Was                                    | Is                                                                                     |
| ------------- | -------------------------------------- | -------------------------------------------------------------------------------------- |
| `182.C.07.03` | partial — face-swap and age unmeasured | done: face swap detected by consistency, age resolved as human review by `182.C.19.02` |
