# Three-shot benchmark protocol

Protocol prepared 22 September 2026. **No model results are claimed.** This is a test design, not a ranked recommendation.

Use the same cleared fictional person, wardrobe, prop and location plates for every candidate that supports references. Keep native outputs, including failures. Record model/version, date, resolution, actual duration, reference mode, seed if available, prompt, elapsed queue/render time, quoted and billed cost separately. Do not silently compare text-only outputs with reference-conditioned outputs.

1. **Identity and geography:** fixed medium view; the subject turns toward a clearly located door. Accept only if face, wardrobe and screen direction stay consistent.
2. **Contact:** subject places one parcel on a bench and withdraws an empty hand. Inspect contact frame by frame: no duplication, floating, penetration or disappearance.
3. **Continuity:** start from the retained state; subject exits, leaving the parcel behind. Check before/after geography and the intended join to shot two.

Choose a shared supported duration before the run. If no common duration exists, publish separate groups. Use one baseline attempt per shot/model; keep retries separate, capped at one repair per failed shot within the approved budget. Stop two identical failures and record fallback rather than cherry-picking success.

Score each visible criterion pass/fail/unverified and record usable in/out interval, not a single aesthetic number. Independently record beat readability, identity, prop/hand, contact, extra subjects/flashes, joins, sound/silence and emotional intent. Native audio support is a separate track; do not penalize a silent-only model as though it promised audio.

Publish all attempted outputs with permission, exact settings, rejected results and actual billed costs. A three-shot run is a bounded comparison on these tasks, not a universal best-model claim. Retest a version change; old findings remain dated.

Candidate families requested in review: Kling, Veo, Sora, Luma, Pika, Firefly. Exact accessible versions and paid/free allowance must be checked before running. No subscription access, price or licence is inferred by this document.