# Segmentation consistency, probed and disqualified — 2026-09-12

Evidence for `182.C.19.05` (malformed teeth/lips, jaw discontinuity, asymmetry
spikes, face boundary tearing, extra limbs/digits, fused props, impossible
object contact).

The item named two remaining candidates after the face mesh was disqualified on
2026-09-10: **a trained anomaly/artifact detector** or **a
segmentation-consistency check**. This probes the second one, because it is the
one that can be run today — `mediapipe` is Apache-2.0 and
`selfie_multiclass_256x256.tflite` is already staged.

**It does not work either.**

## Setup

`tools/human-video/segmentation-consistency.py`, MediaPipe `ImageSegmenter` with
the multiclass selfie model on CPU, against the same three panels the mesh probe
used:
`../2026-09-10-face-mesh-artifact-probe/clean-vs-doubled-mouth-vs-tear.png`,
split at 280 px. Face box from YuNet.

Three metrics, each aimed at a different failure the item lists:

- **`faceBelowJawFraction`** — face-skin pixels more than 10% of the box height
  below the face box, over all face-skin. A mouth pasted on the chin lands here.
- **`faceComponents`** — connected components of face-skin above an area floor.
  One face is one component; a detached paste is two.
- **`silhouetteJumpFraction`** — largest row-to-row step in the person mask's
  left or right edge, over the image width. A torn band steps; a shoulder
  slopes.

## The result

| Case              | belowJaw | components | silhouetteJump | YuNet box height |
| ----------------- | -------- | ---------- | -------------- | ---------------- |
| clean             | 0.0      | 1          | 0.064          | 194.9            |
| **doubled mouth** | **0.0**  | **1**      | **0.079**      | **208.0**        |
| tear              | 0.0      | 1          | 0.079          | 188.8            |

The doubled mouth and the tear score **identically** on the only metric that
moved, and both sit 22% from clean — no separation, no threshold to put between
them. A fourth metric was tried and is not in the tool because it separated
nothing either: a row-to-row silhouette-shift estimate gave 71 / 71 / 73 rows
shifted by four pixels or more.

## Why it failed, which is the part worth keeping

**YuNet grew its face box to swallow the paste.** The doubled-mouth box is 208
px tall against the clean 194.9 — 6.7% taller, and the extra height is exactly
the pasted mouth. A metric defined relative to the face box therefore cannot see
an artifact the detector has already accepted as part of the face.

That is the same failure as the mesh, in a different instrument. The mesh fitted
its canonical topology to a broken face; the detector widened its box around
one. **Every learned model with a facial prior rationalises the artifact instead
of reporting it**, and both of the instruments that can run on this hardware
today have a facial prior.

## What this leaves

One candidate: **a trained anomaly/artifact detector**, which needs labelled
artifact examples — `182.C.33`'s corpus. `182.C.19.05` stays unchecked, and the
reason is now measured rather than assumed.

The tool stays in the tree as a probe. Nothing wires it into a gate, and its
docstring says so, because a detector that does not separate is worse than no
detector: it would answer "no artifact" on the doubled mouth every time.
