# Face-mesh artifact probe — 2026-09-10

Evidence for `182.C.19.05` (malformed teeth/lips, jaw discontinuity, asymmetry
spikes, face boundary tearing). The question was whether a dense face mesh — the
obvious first instrument — can detect these.

**It cannot, and it fails in the worst possible direction.**

## Setup

`mediapipe` 1.0.1 on CPU. Version 1.x drops the legacy `mp.solutions.face_mesh`,
so this uses the Tasks API with the `face_landmarker.task` bundle (3.8 MB). The
mesh detector would not find the face in a full 1408×768 frame at all, so the
pipeline is YuNet detect → crop with margin → mesh, which yields the full 478
landmarks reliably.

Three geometric metrics over the canonical rings: midline-reflection asymmetry
(Chamfer residual normalised by inter-ocular distance), and max discrete turning
angle along the outer-lip and face-oval rings.

## The result

| Case                   | Asymmetry  | Lip curvature | Jaw curvature |
| ---------------------- | ---------- | ------------- | ------------- |
| clean                  | 0.0281     | 1.8326        | 0.3492        |
| tear, 8 px band shift  | 0.0423     | 1.7816        | 0.3551        |
| tear, 20 px band shift | 0.0387     | 1.6620        | 0.3663        |
| **doubled mouth**      | **0.0188** | **1.0252**    | 0.3579        |

The doubled-mouth image has a second mouth pasted below the chin. It is
unmistakable — a person glances at it and sees it instantly.

The mesh scored it as **more regular than the clean face**: asymmetry 33% lower,
lip curvature 44% lower. A malformation gate built on these metrics would rank
the broken face _above_ the good one.

## Why

A face-landmark model is a learned model with a strong facial prior. Handed a
broken face, it does not report a broken face — it fits its canonical topology
to whatever pixels are there and returns a plausible mesh. The artifact is
absorbed, and the resulting geometry is _smoother_ than a real face's, because a
real face has genuine asymmetry and irregularity that the fitted template does
not.

This is the trap already noted for hand keypoints — a 21-point model will fit 21
points to a six-fingered hand — now demonstrated rather than predicted, on the
face. **The prior that makes the model robust is exactly what makes it blind to
malformation.**

## A methodological correction inside this probe

The first attempt applied a smooth radial bulge to the mouth and found almost no
metric movement. Two errors:

1. The warp was centred at 62% of the **full frame** height while the face sat
   near the top, so it distorted the subject's torso. Fixed by centring on the
   detected face box.
2. More importantly, once fixed, visual inspection showed a 50% radial bulge
   produces **a perfectly plausible face** — the same woman with a slightly
   fuller lower face and a wider smile. It is a deformation, not a malformation,
   and there was nothing there to detect.

Smooth geometric warps are the wrong stimulus. Real generative artifacts are
**topological**: tearing, duplication, features that merge. That is why the test
was rebuilt around band displacement and duplication.

## What this does NOT establish

Two pixel-domain heuristics were tried and both failed — a row-gradient
discontinuity score (clean 5.18, doubled mouth 4.27) and a block copy-move
detector that returned **zero hits on a literal copy-paste**. The second is a
bug in the implementation, not a result about the technique: a copied region is
trivially detectable in principle, so a 20-line version returning nothing means
the thresholds are wrong. Nothing is claimed from either.

So: pixel-domain and segmentation approaches are **untested here**, not
disproved.

## What this means for `182.C.19.05`

- **Do not build the malformation gate on landmark geometry.** It is not merely
  weak, it inverts on the clearest case.
- Dense landmarks remain useful for what they are good at — head pose, framing,
  gaze direction — where the prior is a feature.
- The remaining candidates are **face parsing / segmentation** (a second
  mouth-shaped region below the chin is an anomalous segment) and a
  **purpose-trained artifact detector** trained on generated output, which is
  what the deepfake-detection literature actually uses. Both are untested here.
- Whatever is chosen must be validated against images like these before it gates
  anything, because "the metric moved in the right direction" is not the same as
  "the metric catches the artifact".

## Reproducing

`anatomy_probe.py`, `artifact_probe.py`, `pixel_probe.py` in the session
scratchpad, with `face_landmarker.task` and `yunet.onnx`. `mediapipe` is
Apache-2.0.
