Why one dimension at a time
A prompt that tests everything at once tells a lab nothing it can act on. Isolating one control turns a failure into a fixable note.
Most video prompts ask for everything at once: a scene, a mood, a move, a lens, a lighting style, an aspect ratio. When the result misses, nobody can say which part failed. Was it the lens? The move? The light? An evaluator can say the shot "doesn't work," but a lab can't fix "doesn't work."
- Was it the lens?
- The move? The speed?
- The light, the focus, the ratio?
- One control: focal length
- One clear failure
- Measurable between releases
How notes work on a set
On a film set, nobody gives notes that way either. A director says "tighter," or "slower on the push," or "bring the key around." One note, one change. Everything else holds. If the next take is wrong, everyone knows which change caused it.
This is not a stylistic preference. It is how a crew debugs a shot under time pressure. Change two things at once and you can't tell which one helped.
How FinalGrade prompts are written
FinalGrade prompts follow the same rule. Each one targets a single dimension, like camera height or lighting ratio, and holds everything else as simple and fixed as possible. The scene is plain: one person, one place, one action. The direction is one line in the craft's own terms. The brief says exactly what should change and what should stay.
Even the scene description is written to avoid accidental tests. If a prompt is about focus, it doesn't also ask for an unusual lens or a complex move, because a failure there would contaminate the focus result.
Anatomy of a single-dimension prompt
Every FinalGrade prompt has the same parts:
- A plain scene. One subject, one place, one action: "A woman in her fifties swims laps in an indoor pool with lights set into the walls."
- One direction, in the craft's terms. "Tracking alongside her at water level."
- Fixed defaults for everything else. A standard camera body and a held, simple frame, so nothing else is being tested by accident.
- Three checks. What a professional would look for to confirm the direction was delivered: the camera travels alongside her at water level; the foreground water passes faster than the background; her distance to camera stays roughly constant.
The checks are written before any clip is generated, so they define "delivered" in advance rather than after seeing what the models produced.
When dimensions interact
Some dimensions can't be fully separated. A rack focus usually needs something in the scene to motivate it. A tracking shot implies a speed. We handle this by keeping the secondary element as simple and unremarkable as possible, and by writing the checks only about the dimension under test. If a model nails the rack focus but the subject's walk looks slightly unnatural, the rack focus still passes; the walk goes in a defect note instead.
What it costs
That design costs us something. It takes more prompts to cover a craft: 13 dimensions, each tested as a fresh shot and as a revision, each needing several questions to be reliable. Every prompt has to be written carefully, by people who know the craft, and checked so it really tests only one thing.
What it buys
- A failure points somewhere. "Model C misses focal length" is a note an engineer can work on. "Model C's shots don't look professional" isn't.
- Evaluators agree more. When the question is narrow, experts converge. When it's broad, taste takes over.
- Progress is measurable. A lab can see one dimension improve between releases, instead of an overall score that moves for reasons nobody can name.
What round 1 shows
Our first round tested whether the second claim holds. It mostly does, and the exceptions are instructive.
Agreement was highest on camera movement and lighting, where the directions are concrete and visible. It was lowest on focus and depth of field, where the clips often differed only slightly and a few directions, like "deep focus," turned out to be more ambiguous than we intended. Those questions are being rewritten.
Agreement was also higher on fresh shots than on revisions. Revisions ask evaluators to judge two things at once: whether the note was taken and whether everything else held. That is closer to a broad question, and agreement drops accordingly. It is one reason we report the two tests separately.