Six briefs, the same prompt word for word in every tool — one generation each, zero retries. Credits, wall-clock time and export specs logged; watermarks verified frame by frame. Below: the two story briefs, screened side by side.
The fidelity test: only one tool rendered the scripted beat in full.
▦Second figure standing in the CCTV grid — she turns — empty store. The full beat, exactly as scripted.
Only tool that fully rendered the beat. Catch: mirrored gibberish in the timestamp overlay — and the watermark ships in the standard export, frame-verified.
◉A distant figure on the aisle feed — not “a person behind her.” But the heroine's face is the best of the three.
Beat landed only in part. Cleanest screen of the wall — Veo was the only tool with zero text artifacts across all six briefs.
Beat not verifiable from stills. Gibberish signage on screen (“2P XSC”). The frame came back landscape.
Competent — but pointed the wrong way. Gen-4.5 text-to-video was locked to landscape in every run; vertical on Runway is a different workflow, not a setting.
The life scene: all three completed the beat — what separates them is mood, steadiness and the clock.
☂Hand-over complete under the awning — the steadiest composition of the three takes.
Steadiest frame on the wall. True vertical, 1080×1920, beat complete on the first take. The watermark is the price of the price.
✦Wave — then the hand-over. The fullest emotional arc, and the strongest neon in the rain.
The mood winner — the patience loser. 720p export with slight black bars, and this brief waited roughly half an hour in the queue.
Beat complete — and the fastest run of the day. Casting reads as a Western street (the prompt didn't specify; logged as an interpretation difference).
Fastest screen here — wrong shape. ~3.5 minutes door to door, but the minute you buy is landscape, worth less to a vertical workflow.
Best story fidelity at the lowest cost — if the watermark is acceptable on your channel. ~8 min average, true 9:16.
The clean professional: best faces, zero garbled text, no watermark — budget for the queue (~15 min average, up to ~30).
Fast and competent (~10 min average) — but text-to-video is locked to 16:9; enter through image-to-video for anything vertical.
PROVISIONAL RANKING — scored on 3 of 6 protocol dimensions (output quality, cost, speed; 70% of weight, renormalized). Languages, rights/terms and editing control are not scored until they are run and logged. Zero retries on every brief, all three tools.