The canon: what is fixed, is really fixed

A character dossier is not a block of prose but a set of separate, named traits. That is exactly what lets the application assemble an instruction from it and check the result against it afterwards.

What a character dossier holds

• Identity — name, role in the brand, short description, personality, how the character moves and looks.
• Mandatory traits — each trait on its own line, with a class (mandatory, desired, forbidden), a weight (heavy or light) and the check question that is put to the vision model.
• Forbidden deviations — the list that serves both as the negative prompt and as a check.
• Colour lock — the hex values that may not shift, and at the same time the basis for the automated palette check.
• Proportions — how large this character is next to the main character. This prevents the classic error where an animal ends up the size of a child.
• Reference images — divided into canonical, turnaround, expression and wrong. The wrong category is explicitly NOT sent to the model: that only confuses it. Those images exist for the human and for the check.
• Prompt fragments per model — a hand-tuned translation per model profile, without changing the canon itself.

Styles, worlds and props

Besides characters the application knows three other canon kinds with the same build-up: style profiles (the drawing style apart from the characters), worlds (recurring locations and settings) and props (recurring objects that have to be just as right as a character).

Changing the canon

A canon change is version controlled and has visible consequences. As long as a new version is a draft, nothing about work in progress changes. On publication every existing asset made against the previous version is flagged canon outdated. That is not a rejection but a signal: the re-check screen shows those assets grouped together, and an agent can run the re-check in bulk. A canon change can never silently invalidate an approved image.