Scoring method · publishedStart your shootThe Quality Gate, audited
The judge has a name.
“Independent AI judge” is not a claim you should have to take on faith. This page names the scorer models, publishes the threshold, and walks through one real pilot order — scores, rejections, and the frame a human vetoed even though its score passed.
The scorers
Two named judges, one human veto.
gemini-3.1-pro-preview vision model against your own uploaded references. It returns an identity score from 0 to 100 plus a defect list (warped anatomy, invented text, plastic skin, wrong gear). The verdict rule is mechanical: ship only if the score is at least 85% and no single defect disqualifies the frame.buffalo_l): the generated face is compared against each reference photo as a cosine similarity from 0 to 1, with pass bars calibrated per subject from their own photos — not one global number.The threshold
85%, same at every price.
A frame that scores below 85% is regenerated at our cost (up to two extra attempts per slot). If the best attempt still cannot clear the gate, we deliver the strongest take, record it honestly as best-effort in the database, and the 7-day, 100% refund stands. If the scorer itself is unavailable we deliver rather than hold your paid order hostage, and the frame is recorded as unscored. The threshold is identical for a $9.99 First Look and the largest Portfolio — likeness is never a paywall.
Reading the numbers
Three scales. They never compare.
Worked example · real pilot order
Dimi’s red-carpet frame, scored in the open.
From the pilot shoot of Dimi S. (Amsterdam, 2026-08-30, published with his consent). Pilot scores below are ArcFace cosine similarity against his own reference photos — his calibrated STRONG bar is 54.8%, derived from how similar his reference photos are to each other. The pilot delivered this brief at 94.5%. The red-carpet frame was later retired with most of the pilot set in a founder audit; the numbers below stand as the published record of how every frame is scored.
| Candidate | ArcFace score (max vs. references) | Outcome |
|---|---|---|
| Base 1 | 94.8% | Picked — delivered at 94.5% after the 4:5 crop · verdict STRONG |
| Base 2 | 92.7% | Rejected — four foreground faces instead of three |
| Base 3 | 93.5% | Rejected — four foreground faces instead of three |
The honest counter-example from the same order: on the LinkedIn brief, base 1 scored higher (75.6%) than the frame we delivered (74.2%, 75.8% after crop) — and was still rejected by human review for a harsh shadow and a strained smile. The score never outranks the eye. Across the whole order: 21 candidates generated, 7 frames delivered, all 7 STRONG, mean ArcFace identity match 82.7%.
The pilot’s one remaining frame on the landing page — the barrel — carries its dual-model chip (identity 1.57 · PASSED); its audited ArcFace record from the pilot delivery is 91.4%. The rubric itself is on the Quality Gate page, and failed founder frames are in the reject gallery. The bar is enforced, not decorated: one audit round pulled three pilot frames from the landing page — a sub-bar rescore (89.4% against the shoot’s 90% rule) and two failed 3× zoom audits despite passing chips — and a follow-up founder audit retired the rest of the set outright. Every retirement is disclosed, never painted over. A published bar you don’t enforce is marketing, not a bar — and the score never outranks the eye.
Calibration case · September 2026
Four passing scores. Four rejections — by the founder.
The clearest calibration lesson in this method came from our own founders. Four passport-style frames shipped with audited dual-model chips of 1.10, 0.96, 1.20 and 1.26 — every one a numeric pass against the 0.85 bar, floored re-scores of the exact published files. The founder rejected all four the day they went live — his own two as the subject himself: the 0.96 frame, his ruling went, “does not look like me at all” — the last few percent is exactly what matters. The frames came off every selling surface the same day and are published, scores and verdicts included, in the reject gallery. The standing lesson is now part of the method: a facial-recognition pass is necessary, never sufficient — perceived likeness is judged by a human, the subject holds the final veto, and no founder frame publishes without the subject’s explicit approval. Your shoot inherits the same two gates: the numeric bar first, and you — backed by the refund — last.