ENXERTO
Generating the image is the half that got cheap. Deciding whether it is acceptable is the half that did not, because it depends on a standard somebody had to write down and then defend under a deadline.
A standard that lives in one head does not survive a schedule.
Every operation generating at volume hits the same wall. The tool is fast, the output is random, and the brand book that was supposed to prevent that is a PDF nobody opens before writing a prompt.
The rules usually exist. They just live in a format no machine can execute and no new hire can absorb, which means the person holding them becomes the bottleneck and the work stops when they go on holiday.
One registry. Everything answers to it.
Two files hold the system: the written laws, and the archive of named assets. Nothing generates from a description typed that afternoon. Everything derives from the registry, which is what makes the output reproducible rather than lucky.
A law is not a preference. “Cinematic and moody” is not a standard, because two people reading it produce two different images. “The only light in frame is daylight or a lamp you can see” is a standard, because it can be broken and you can point at where.
Writing that distinction is not engineering work. Knowing which sentence is which is the job of whoever runs the picture, and it is the reason the system cannot be built from the infrastructure side alone.
The thing that generates does not get to grade itself.
The judge runs on a model from a different family than the generator. A model does not reliably catch the mistake it is itself inclined to make, and a check that agrees with everything is not a check.
It does not return a score. It names the law that broke, which is the only output a person can act on, and it is also what makes a rejection teachable instead of a matter of taste. Rejected frames are kept beside the approved ones, because the rejection is where the reason is stored.
The check runs before the expensive call, not after it. Auditing a prompt against the laws costs nothing; auditing the render costs the render.
Geometry does not come from language.
There is no unit of measurement inside a sentence. Ask a model for a taller figure and it returns a different person, because “taller” is a comparison and the model has nothing to compare against.
So proportion left the prompt and moved to the mesh. The canon is a set of measurements, a race arrives as numbers before it arrives as an image, and the check on it is a measurement rather than an opinion about whether it looks right. The measuring runs locally, in Blender, on the actual geometry.
This is the part that is not a prompting problem. Matching something one to one, down to hardware and geometry, is a control and reproducibility problem, and it is solved with a file that can be measured rather than a sentence that can be reinterpreted.
Findings, in the wording that made them work.
Each of these was paid for in failed generations. They are in the system because the general version of the advice did not hold.
A negative goes in its own block at the very end of the prompt, because the model weights the end. The same prohibition buried in the middle of a style block loses the argument to the positives around it.
“No cast shadow” failed twice in a row, because a model draws a shadow under an object by reflex. What worked on the first try was describing what occupies the place of the shadow: the sky is a single uniform luminous dome, light arrives from every direction at once, so the ground holds exactly the same value everywhere, continuous and uninterrupted underneath each object. Forbidding a thing is weaker than filling the space where it would have been.
A reference locks likeness and drags everything toward photorealism at the same time. If the grammar is painted rather than photographed, the grammar has to be restated in the same prompt or the reference wins.
The last frame of one shot is fed in as the first frame of the next, so the camera never teleports between them. Continuity stops being something you hope for in post.
Judging is moving the file.
Approval is not a note or a memory, it is a location. A piece sits in a tray until it is judged, and judging moves it into approved or into rejected. A file still sitting in the tray is a generation nobody ever opened, and the system can say so.
The rules are checked by a command rather than remembered by discipline. One command verifies that every named asset resolves to a real file, that every name points inside approved, and that nothing approved is missing a name. It exits non-zero when it fails, so it can sit in front of a build.
The root of the archive resolves from the file that asks for it, never from the working directory, because the working directory depends on who called the script. That is the difference between a convention and a standard: a standard runs on somebody else’s machine without editing anything.
The laws are portable. The provider is not the moat.
The laws, the archive and the criterion run on any provider. Tying the argument to one vendor would narrow the audience and defend precisely the part that already turned into a commodity.
The system ships empty on purpose. The registry belongs to whoever writes it, and what is being handed over is the ability to defend a standard at volume, not a house style.
The public repository is the first version of this, the image-only judgment engine, open under Apache 2.0. What is written above is the system as it actually runs in production now: video, geometry and production state came after, and the repository has not caught up.
What it was built for