marginal

Businesses generating video at scale pay premium rates for every clip, and a lot of what comes back has defects that only surface after the money is spent. Video models on fal span 6.7x in price on a 5-second shot. The obvious fix is generating cheap and escalating when quality is poor. I built that, and on 5 shots it cost $11.45, 129% more than premium alone ($5.00). marginal is what I built after measuring why. A VLM judge scores each clip against a binary defect checklist instead of holistic quality, which proved ceiling-compressed and unreliable. Price does not predict quality: one prompt had premium scoring worse than draft, and camera instruction-following failed identically at every tier. Some defects respond to escalation and some don't. marginal rerolls cheap on transient defects, escalates only where tier is measured to help, and halts on tier-independent defects instead of charging 6.7x again. Result: $4.00 versus $7.00 flat-premium, 43% under, on the seven shots demoed. On the original five-shot test, this same policy cost $3.55 against $11.45 for naive escalation (69% under), and reproduced that $3.55 exactly on rerun. Every decision is auditable from the rubric and evidence frames it was made on.