# CPU identity-consistency probe — 2026-09-10

Evidence for `182.C.07.03`: detecting mixed identities in a reference pack
before any provider credit is spent. It extends
[`../2026-09-10-cpu-identity-embedding/`](../2026-09-10-cpu-identity-embedding/),
which established that SFace runs on this box; this one commits the tool and
measures the pairs from a source image that is in the repository, so the numbers
regenerate.

## Tool and models

`tools/human-video/face-embedding.py` — YuNet
(`face_detection_yunet_2023mar.onnx`, 227 KB) locates the face and supplies the
five landmarks SFace aligns on; SFace (`face_recognition_sface_2021dec.onnx`,
38.7 MB) produces the 128-d embedding. Both OpenCV Zoo, **Apache-2.0**.

## Measured cosines

Source: the committed two-subject office photo. "Degraded" is that image through
`boxblur=3:1`, `eq=brightness=0.12`, and a 0.6× downscale-then-upscale — a stand
in for the generation-and-re-encode round trip an identity check has to survive.

| Pair                    | Cosine    |
| ----------------------- | --------- |
| Subject A vs subject B  | **0.261** |
| Subject A vs A degraded | **0.739** |
| Subject B vs B degraded | **0.819** |
| Subject A vs B degraded | 0.177     |
| Subject B vs A degraded | 0.146     |

Raw: [`cosines.json`](./cosines.json). **The embeddings themselves are not
committed** — an embedding is derived from a real person's face and falls under
the same handling rules as the reference image (`182.C.19.10`). Cosines are
committed; vectors are not.

## Why no threshold is set here

On the same day, in the earlier probe, a _different_ pair of strangers scored
**0.422** — above the 0.35 identity gate inherited from ArcFace-oriented
defaults. So this table's clean 0.26-against-0.74 separation is not a licence to
pick a cutoff from it: five pairs from one photograph is not an ROC, and the one
pair that would have broken the rule is not even in this table.

`human-video-identity-consistency.ts` therefore takes the calibration as an
input, refuses one fitted for a different model, and returns `uncalibrated`
rather than a verdict when none is bound. It also uses a two-sided band — same
person above, different below, **uncertain in between routed to a human** —
because a single cutoff is exactly what turned the 0.422 pair into an accept.
`182.C.19.08` fits the real operating points on the corpus `182.C.33` gathers,
which is where fairness slices (`182.C.19.09`) get their say too; a cutoff
fitted on a generic public benchmark would be skewed in precisely the way that
item is about.

## What this does not measure

Two of the four detections `182.C.07.03` names are **not delivered**, and the
module returns them as `not-measured` findings with their reason attached rather
than leaving them silent:

- **Face swaps.** No manipulation detector is available on this stack. An
  uncalibrated deepfake score would invite belief it has not earned.
- **Age mismatches.** No age estimator is available. The character pack carries
  a declared `apparent-age-band` (`182.C.07.01`), but nothing here measures the
  age a reference reads as, so a declared band cannot be checked against one.

Reference _contradictions_, the fourth detection, need no model at all and are
delivered: references naming two different subjects, one image submitted under
two roles, a `face-angle` reference that is front-on, and a `full-body`
reference whose measured frame stops above the ankles.
