# Limitations: AttractivenessTest calibration benchmark (2026-10-21, v1)

- **One small, narrow sample.** About 100 scored adults aged 18–54, photographed under studio light in London; roughly two-thirds are white. Results may not carry over to selfies, other ages or other groups. No subgroup results by ethnicity are reported because the groups are too small.
- **Same faces fitted and tested.** The live reference was fitted on this set. The in-sample correlation is optimistic; the held-out split-half figure is the fairer estimate, and it is still based on about 50 test faces per split.
- **Geometry is not attractiveness.** The free score measures facial proportions only. Skin, hair, expression, grooming and style are left out, so a modest correlation with human ratings is expected and the score should not be read as a verdict on a person.
- **Different scales.** The site's PSL runs 1–8 and the human ratings 1–7. Level comparisons are made as a share of each scale; they show where each scale is anchored, not that one is right.
- **Expression and noise tests are partial.** Only one alternative photo per person (smiling, front view) is used. Ungated results describe photos the live site would refuse. Landmark noise is independent Gaussian noise, which is a simplification of real detector error, perspective and lens distortion.
- **Not a test of other websites or AI models.** No other rating site or language model was run here. Comparisons with published studies of other raters are context, not a head-to-head test.
- **Not medical, cosmetic or psychological advice.** Nothing here should guide surgery, treatment or self-worth.
- **Versioned.** These numbers belong to engine 0.2.0 / reference 0.2.0-frll / weights fixed-0.1.0. A change to any of them gets a new version of this benchmark; v1 files are never replaced in place.

Corrections: support@attractivenesstest.dev