# Measurement verification notes

Checked: November 4, 2026. Engine: geometry-2.0.0.
Site: https://attractiveness-test.co

## What the downloadable evidence contains

- `/evidence/synthetic-landmarks-v2.json`: 478 synthetic normalized coordinate pairs,
  with a 1000 x 1000 reference canvas. No photo, person or demographic label.
- `/evidence/geometry-2.0.0.json`: twelve executed formula checks and their outcomes.
- Project reproduction command: `pnpm exec tsx scripts/verify-face-evidence.ts`.
  The script asserts invariances and expected rejection codes before writing data.

The baseline deliberately matches product references and scores 100 in each
component. This is a formula fixture, not evidence of a perfect human face.
The record checks identical input, double resolution, equivalent pixel geometry
on a wider canvas, translation, uniform scale and 10-degree roll. It also checks
empty input, a tiny face, excessive nose displacement and 25-degree roll rejection.
It also verifies that a brow shift changes the thirds score and total, and that
changing eye widths at fixed eye-centre spacing changes the fifths score and total.
The source formula and numeric references are explained at `/how-it-works`.

## Report verification

Isolated temporary-database tests mock both payment receipts and AI responses.
They verify payment amount/currency/order identity, cross-account denial, consent,
concurrent generation claims, no call for unpaid reports, preserved numerical
values, invalid-output rejection, bounded retries, refunds during generation,
report-data deletion and PDF generation. These tests do not call a third-party AI processing service or Stripe
and do not establish real provider latency, availability or report quality.

The public `/sample-report` and its PDF are editorial examples with invented
measurements and explanations. They are not a real person's report and are not
presented as a live AI output. Paid reports retain internal generation metadata. Historical deterministic reports may lack those fields.

## Privacy inspection and limits

Source review: the free detector reads image pixels locally, downloads same-origin
model/WASM files, and returns a local measurement result. The separate report flow
submits normalized coordinates only with consent, persists derived metrics and
sends those derived metrics to a third-party AI processing service for a paid explanation. No original image,
raw mesh, account identity or payment details enter that a third-party AI processing service request.

To inspect your own free test, open Network before selecting a permitted adult
photo and review Fetch/XHR request bodies. Browser Network cannot show the
server-side AI request. No real-photo network audit or independent security
certification is claimed by this evidence record. Host logs, browser extensions,
provider retention and downloaded files are outside the synthetic test scope.

There is no published human preference validation, representative demographic
fairness study, cross-camera photo study or supported attractiveness accuracy
percentage for this product. Read `/accuracy` and `/privacy-policy` for details.