The Authority Set

A better-looking stranger is worth nothing

Image models beautify by default. They smooth the skin, slim the jaw, take five years off, and hand back someone who reads as you-ish at thumbnail size. That is fine for a mood board and fatal for a personal brand, because the entire point of putting your face on the content is that people recognise it — on the page, in the DMs, and then in the room. If your posts show a slightly different person every week, you are spending money to look inconsistent.

The router

Type what the business actually is. This is the production router — the same registry, the same weights, the same tie-breaks — scoring all ten environments against your own words.

Scoring: an exact niche match is worth 3 points, a multi-word signal 2, a single keyword 1.5, and a loose description overlap 0.5 — then ties break alphabetically so the same input always returns the same set. No model call, no randomness, no server.

The clause nobody is copying

The prompts circulating are all environment: the lens, the lighting, the mood. Environment is the part the model is already good at. Everything below is the part that keeps you in the frame, and it exists because of what we measured, not because it sounded good.

Honest result: the anti-retouch clause helped, but it did not solve it. On the beautifying model it improved only 2 of the 4 renders. What solved it was changing the model — see the receipts below. Prompt wording is a weak lever compared to picking an engine that does not quietly retouch people.

Three measurements, because there are three ways to fail

Hard caps, not soft advice: an unmeasurable face caps the score at 60, a below-threshold similarity at 55, a "that is a different person" verdict at 45, and a smoothed face at 80 however good the rest looks. 85 and above is LOCKED and may ship. 65–84 is DRIFT. Below 65 is BROKEN. A leg that cannot be measured returns UNSCORED and caps the score — it never invents a number to fill the gap.

The airbrush test

This is the third leg of the real gate, running locally on canvas: it measures how much skin detail an AI portrait kept compared with the original photograph. Drop your own before-and-after, or load one of ours.

Skin detail retentionLocal · no upload

1 · Original photograph

2 · AI render of the same person

Same maths as production — YCrCb skin mask, high-pass against a sigma-2 blur, RMS energy over skin pixels, ratio capped at 2.0, bands at 0.70 and 0.45. One honest difference: the production gate measures the face box a detector found, and a browser has no face detector, so this crops the middle of the frame where a portrait's face lives. Verified against the Python reference on the same crop rule on 2026-08-28: 6 pairs run through both the Python engine and the browser port on the same crop rule — largest disagreement 0.006 on a 0 to 2 scale, identical band on every pair.

Six rooms from one selfie

Every frame below came out of a single phone photograph on 2026-08-28, under a real consent record (self-granted, 2026-08-01, unrestricted scope, not revoked). The scores are the gate's own, printed whether they flatter us or not.

6 of 6 frames cleared the gate on the identity-preserving model, with similarity between 0.938 and 0.961 against a source photograph — higher than two real photographs of the same man score against each other (0.667 and 0.755). Skin detail came back above 1.0 on every frame, meaning the renders carry more visible skin detail than the compressed source file: the opposite of the airbrush failure. The vision leg still flags a light beautification on all six and says so in the table — it is describing a real effect that is small enough to stay inside the band, which is why judgement is a 25-point leg and not a veto. Costs about four hundredths of a cent per frame to adjudicate.

The measurement we had to throw away

We already owned a face-fidelity score for talking-head video, and the obvious move was to reuse it here. It does not work across a scene change, and the only way to know that was to test it against a stranger. Same score, run on four renders of the real subject and three frames of a completely different man:

The true positives scored lower than the impostor. There is no threshold you can draw through that table, which means anything gated on it would have been gated on nothing. The replacement was calibrated against the same controls before it was trusted with a client's face:

The gap between the highest stranger and the lowest true positive is about 0.78, so the thresholds sit in empty space with room on both sides rather than being tuned to make our own output look good. Locked at 0.55, drift at 0.40.

The point is not the portrait

A locked still is the cheapest thing to hand to a talking-head engine — you get a spokesperson video without a camera, a light, or a good hair day, and the person in it is verifiably the client. A set of ten covers a quarter of posting: the same human, credibly, in the rooms his customers associate with authority, on a schedule no filming day can match. The gate is what makes that safe to automate. Without it, "post daily without filming" means "publish a stranger with your name on him, unattended, at scale."

Straight answers