The fourth dimension: comparing the models that draw the people

Under the hood

• The unit is a cell: one generator, one people, one prompt register. Ten peoples across five generators makes 50 cells at four renders each. Every generator gets the same prompt template and a square frame, with any built-in prompt rewriting switched off, a seed recorded where it accepts one, and a hash of each cell's prompt.
• Scoring reuses the vision analysis that built the real-photo baselines, on the same bucketed scales, restricted to the dimensions that transfer from notable-people photographs to a generated portrait: skin tone, eye color, eye fold, hair texture and hair color.
• The design rule is enforced by the shape of the code. Comparison pages live under the people, at /ethnic/{slug}/generators; no URL nests a people beneath a model; no function computes a per-people score or sorts peoples by one; and the export sorts by people, then generator.
• The blind pair deals two renders of the same people and register from different generators, with sides assigned by a coin flip on every serving. The pair's identity is derived from its two render IDs, so a reader who meets the same pair dealt the other way updates a single vote, and the side each model sat on is recorded so position bias can be measured. Readers answer left, right, both or neither.
• Reader pair votes publish as preference per people and are never folded into the automated score. Renders are stored apart from the catalog, and the render and pair-vote tables join the open dataset.