# Direction adherence: the noise floor for gaze, pose and expression (2026-09-10)

Evidence for `182.C.19.07`: camera/control adherence, action completion,
blocking, gaze, emotion, gesture, framing, and scene/wardrobe continuity.

## Method

The **same real face** measured in five framings — full frame, 0.7×, 1.4×, a
tight crop and a 30-pixel shift — with MediaPipe `face_landmarker` at 478
landmarks (iris included), 52 blendshapes and the 4×4 facial transformation
matrix. Nothing about the face changed, so every difference between readings is
measurement noise and nothing else.

| Framing     | yaw   | pitch | roll  | gaze L | gaze R | smile | browDown  |
| ----------- | ----- | ----- | ----- | ------ | ------ | ----- | --------- |
| full        | −0.20 | −1.20 | −2.78 | 0.430  | 0.615  | 0.942 | 0.541     |
| scale 0.7   | −0.02 | +0.21 | −2.92 | 0.428  | 0.618  | 0.951 | 0.491     |
| scale 1.4   | −0.21 | −1.02 | −2.81 | 0.431  | 0.614  | 0.944 | 0.542     |
| crop tight  | −0.27 | −1.30 | −2.73 | 0.451  | 0.608  | 0.938 | 0.443     |
| shift right | −1.57 | −0.21 | −2.61 | 0.443  | 0.616  | 0.940 | 0.581     |
| **spread**  | 1.55  | 1.51  | 0.30  | 0.023  | 0.010  | 0.014 | **0.138** |

## Two findings, and both shape the API

### Blendshapes are not equally reliable

The smile blendshape moves **0.014** across pure framing changes; `browDown`
moves **0.138** — ten times as much. An emotion inference leaning on brow
position would be reading the framing.

So each blendshape carries its own measured stability, a claim smaller than that
blendshape's noise is refused, and a blendshape whose stability was never
measured is refused too: an unmeasured stability is not a good one.

### Gaze has a large per-face bias and small noise

On this face **looking straight at the camera** the two eyes read 0.430 and
0.615 — a **0.185 asymmetry** — and that asymmetry is stable across all five
framings (spreads of 0.023 and 0.010). It is a property of the face and the
perspective, not of where the person is looking.

So gaze measures **change** well, where the bias cancels between two readings of
the same face, and **absolute direction not at all** without a per-subject
baseline. `absoluteGazeDirection` refuses without one; an absolute verdict
without it would report this face as looking sideways forever.

## Head pose and camera moves

Head pose is repeatable to about **1.5°** in yaw and pitch and **0.3°** in roll.
Good enough to gate a profile turn, not good enough to gate a subtle glance.

A **pan or tilt** is decidable from the median flow's dominant axis, remembering
that the _image_ moves opposite to the camera — getting that sign wrong turns a
compliant take into a violation. A **push-in or pull-out is not**: a zoom moves
pixels outward from the centre and the median of that is near zero, the same as
a static shot. It needs the flow's divergence, which is a different measurement,
and `assertMoveDecidable` refuses rather than guessing.

## Three aspects are judgements, not gaps

| Aspect            | Why it is refused                                                                                                                                                  |
| ----------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| emotion           | 52 blendshapes are measured and reported; mapping them onto the contract's eight emotions is not. "Amused" against "warm" is not decidable from `mouthSmileRight`. |
| gesture           | Pose landmarks give arm and hand positions, which are reported. Naming the gesture needs a classifier that is not staged.                                          |
| action completion | Needs to know what was asked, which lives in the request. The judgement lives with a person.                                                                       |

All three travel in every report. A report showing framing and camera adherence
clean reads as "the direction was followed" unless it also says nobody checked
the emotion, the gesture, or whether the action finished.

## Per-item status this evidence supports

| Item          | Was           | Is                                                                                            |
| ------------- | ------------- | --------------------------------------------------------------------------------------------- |
| `182.C.19.07` | `not-started` | done; five aspects measured with their noise floors, three refused as judgements with reasons |
