Methodology

This project is an argument about measurement, so it publishes its own. Every threshold below is the one the code actually uses.

How a variant is made

• Relight — AI Photo Lighting applied to the real file. A real operator on a real photograph; the skin did not change.
• Capture frames — one image-to-video call returns 81 frames. These are generated, so this axis measures variance across plausible captures of one image, not across genuinely re-shot photographs. Worse, the generator lets expression drift between frames — a raised brow, a changed mouth — so some of the movement we attribute to capture is really the face doing something different. That makes this axis an upper bound on capture noise, not a clean estimate of it. We say so because it is the weaker of the two axes and hiding it would make the whole project dishonest.
• Identical normalisation — every variant is cropped and scaled by the same fixed transform before measurement. Without this, part of the reported variance would be our own preprocessing rather than the capture.

What we would fix with more budget

• Sample frames closer together, or reject frames whose facial landmarks moved, so the capture axis stops absorbing expression drift.
• Re-shoot genuinely separate photographs of a consenting subject, which is the only way to measure real capture variance rather than a generated proxy for it.
• More than four variants per subject. With four readings the observed range is a coarse instrument, which is exactly why we report the range instead of a confidence interval.

Trust thresholds

Cosmetic efficacy is commonly advertised around a 20% improvement. We take 20 points on the engine's raw 1–100 scale as the reference effect of that size. A metric can only support such a claim if its own noise is comfortably smaller, so we compare each metric's observed range to that reference:

• range < 5 points → trustworthy
• range < 10 points → borderline
• otherwise → noise
• all readings pinned to the top or bottom of the scale → saturated

We report the observed range rather than a confidence interval: with a handful of captures, a tight interval would be a fiction.

Raw data

The complete dossiers — every variant's raw scores, provenance labels and unit ledgers — are the exact JSON files the site renders: Subject A, Subject B, Subject C. The code that produced them is at github.com/Biniyoyo/tolerance.

The canons

We report conformance to each canon's stated ideal. We do not report a percentile: we have no population distribution, and inventing one would be the exact sin this project exists to expose.