Inspectors in any lab will tell you the same thing: defects are cheaper at the door than at the loading dock. AI generation is no different. Most frames that fail the scan were lost before the job ran — in a brief that contradicted itself, a scene with three light sources, a destination spec nobody checked. The pre-flight checklist is five lines long, takes a minute, and it runs before the first credit moves.
The five checks
1. Anchors present and unedited. The character block — age stated as a number over 25, hair, build, one or two face details, one signature — pasted verbatim from the filed protocol, not retyped from memory. Memory edits; files do not. If an anchor has genuinely changed, that is a protocol revision: save it as a new version, dated, so old sets stay reproducible.
2. One scene, one action, one light. Count the scene lines. A sample that asks for a marble bar, a rainstorm and a mirror selfie in a single frame will fail on noise no matter which model runs it. One setting, one action, one declared light source is the combination that scans cleanly — and the combination whose failures you can actually diagnose.
3. Hands have an assignment. Hands around small props are the most common defect on the report. Decide before the run what the hands are doing — holding one large simple object, resting, out of frame — or compose the crop at the waist and remove the variable entirely.
4. Destination spec known. Feed frame, cover, or keyframe for video? The answer changes the composition (a keyframe needs somewhere for the motion to go) and decides whether premium or an upscale pass belongs in the budget at all. Checking the destination after generation is how finishing credits get spent on frames that were never going to ship.
5. Admission in order, for uploads. If the job involves real media — a swap source, an undress sample, an edit base — the consent paperwork is part of the pre-flight, not an afterthought: whose image, what scope, documented where. A job that cannot pass check five does not run, whatever the other four say.
Why the checklist pays
None of this is talent; it is sequence. The bench’s own scan catches what slips through, but a minute of pre-flight routinely turns a three-attempt frame into a one-attempt frame — and at 6 credits a standard sample, that discipline is the difference between a protocol that produces a series and one that produces a pile of almosts.