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

Evidence for `182.C.19`, and **the second correction of the day**. All ten
`182.C.19` rows carried the same blocker text — "No GPU host; needs a RunPod
worker per `182.C.17.13`" — applied uniformly. That blanket claim was wrong, and
I repeated it in conversation before testing it.

## The central premise was false

The stated reason `182.C.19` could not proceed was that no CPU model here
produces an ArcFace-class embedding. **SFace does**, it is in the same OpenCV
Zoo as the YuNet detector, it is **Apache-2.0**, and `cv2.FaceRecognizerSF` runs
it on CPU.

Measured, 128-d embeddings, cosine similarity:

| Pair                                     | Cosine    |
| ---------------------------------------- | --------- |
| Reference person vs **itself**, degraded | **0.802** |
| Reference person vs other woman #1       | 0.269     |
| Reference person vs other woman #2       | **0.422** |
| Other woman #1 vs other woman #2         | 0.261     |

"Degraded" is the reference put through `boxblur=3:1`, `brightness=+0.12`, and a
0.6× downscale-then-upscale — standing in for the generation-and-re-encode round
trip an identity gate actually has to survive. It survives it: 0.802 against
0.26–0.42 for strangers.

So identity similarity is computable on this machine today.

## And it immediately found a threshold defect

`182.C.03.05` C4 gates identity at **cosine ≥ 0.35**, a value inherited from the
ArcFace-oriented defaults in
`libs/nous/portrait-animation/src/portrait-animation-quality-metrics-pipeline.ts`
and carried into `quality-policy.ts` as `minIdentityCosine`.

**Under SFace, two different people scored 0.422** — a false accept, comfortably
above that gate. SFace's own published operating point is ~0.363, which this
pair also clears.

This is not a reason to distrust the embedding; it is a reason the inherited
number cannot be used. A threshold is only meaningful paired with the model that
produced it, and the "these are a starting point, not a calibration — `182.C.34`
re-fits them" caveat written into `quality-policy.ts` is now demonstrated rather
than asserted. Whatever model ships, C4's cutoff has to come from a measured ROC
on our own corpus, and it has to be per-model.

## Per-item status, replacing the blanket blocker

| Item                                                                              | Real status                                                                                                                                                                                                             |
| --------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `.01` identity similarity + temporal drift                                        | **Not blocked.** SFace supplies the embeddings; `libs/isis/workflows/src/consistency/identity-drift.ts` already computes cosine and drift over supplied vectors (`02.04`). The missing piece was exactly the extractor. |
| `.02` face shape, complexion, age, hair, wardrobe                                 | Needs attribute models plus human review. Not GPU-blocked; model-and-corpus work.                                                                                                                                       |
| `.03` track through profile turns, cuts, multiple people                          | **Not blocked.** Per-frame detect + embed + assignment is CPU-feasible; cost is time, not capability.                                                                                                                   |
| `.04` face merging, swaps, duplicates, flicker                                    | **Not blocked.** Derivable from per-frame embeddings and boxes.                                                                                                                                                         |
| `.05` malformed teeth/jaw, extra limbs, fused props                               | **Genuinely blocked on models** — no anatomy/artifact detector here, and this is not something a face model provides.                                                                                                   |
| `.06` motion smoothness, optical flow, freeze, judder                             | **Not blocked.** ffmpeg has `freezedetect`; cv2 has Farneback optical flow.                                                                                                                                             |
| `.07` camera/control adherence, gaze, emotion, action completion                  | Partially reachable, weakly measurable. Needs care to avoid VLM judgement dressed as measurement.                                                                                                                       |
| `.08` calibrate thresholds by visibility and shot size                            | **Blocked on a corpus**, not on hardware — and the 0.422 false accept above shows why it matters.                                                                                                                       |
| `.09` fairness slices                                                             | **Blocked on a labelled corpus** with demographic slices, gathered with consent. Compute does not solve this one, and it is the hardest of the ten.                                                                     |
| `.10` store embeddings with encryption, tenant isolation, deletion linkage, audit | **Never blocked at all.** Pure storage and access engineering.                                                                                                                                                          |

Four are unblocked, one was never blocked, three need corpora or attribute
models, and one (`.05`) is genuinely blocked on models this box does not have.
None of the ten is blocked for the stated reason.

## Why the wrong blocker persisted

It was inherited rather than tested. The phase constraints describe a 16 GB
MacBook Air on which "SyncNet-class evaluators, face tracking, and ASR are
hardware-blocked locally"; that sentence became a row-level status for ten items
without any of them being tried. This worktree is Linux with 15 GiB and 4 cores,
`onnxruntime` was already installed, and both models needed are Apache-2.0 and
under 40 MB.

The rule this suggests: **`blocked-local` must name what was attempted and what
failed.** A blocker copied across ten rows is a claim about none of them.

## Licensing

Both models are **Apache-2.0** — YuNet (227 KB) and SFace (38.7 MB), OpenCV Zoo.
That matters for `182.C.02.06`, which flagged the insightface weights in
`apps/psyche/avatar-engine` as carrying non-commercial terms. The commercially
usable path does not run through those weights.
