# Limits of the facial symmetry coordinate audit

- All 52 examples are constructed coordinates. Human subjects, photos processed and detector runs: zero.
- These examples verify arithmetic and illustrate sensitivity to a chosen mirror axis. They do not calibrate or validate MediaPipe on photographs.
- Two-dimensional coordinates omit depth. Head turn, perspective, expression, occlusion and detection errors can change the measured geometry. None is quantified by this audit.
- Uniform coordinate scaling is not a camera-distance experiment. In-plane roll correction does not correct yaw or pitch.
- The published index is a normalized landmark gap, not a 0–100 symmetry percentage, population percentile, beauty probability or medical diagnosis.
- The result depends on 12 selected pairs, eight midline points, cheek width and iris-based alignment. Other metrics and landmark choices need their own definitions.
- The threefold difference between a single midline shift and a single mouth shift holds for this symmetric baseline and these isolated perturbations. It is not a universal detector-error multiplier.
- The coordinate rounding effect is measured for this finite test set, not proven as a maximum for all input sizes or poses.
- The site's quality gate and calibrated attractiveness scoring are separate from the raw measurement functions tested here. This audit does not bypass the gate or recommend scoring rejected photos.
- No real person images, calibration-set records or user measurements are distributed. External authors neither reviewed nor endorse this asset.