FEATURED ESSAY · AUTUMN 2026
“The Kuleshov Effect, Measured on Machines”
Publishes with Edition 1.
All essaysAutumn 2026 Edition / Battery v1.1
AI video,
judged by film craft.
The industry-reference benchmark for AI-video realism. Blind human raters score every Take against anchored, published rubrics.
The instrumentThe Battery is 25 prompts, 3 takes each. Scores are reported on the Realism Index, a 1–10 scale.
25prompts
3takes each
8scored axes
Leaderboard
Edition 1 · pending grading
The Realism Index
A score for what survives the cut.
One number, built from 25 prompts and three independent Takes per prompt. Human raters score every Take across eight axes (COP, PHY, ANA, PHO, CAM, TEM, SPA, EMO).
METHOD / 01
Blind panel · published anchors · 95% confidence intervals
1–2 Broken · 3–4 Uncanny · 5–6 Passable in cuts · 7–8 Broadcast-adjacent · 9–10 Indistinguishable
Edition 1 · autumn 2026 · capture and blind scoring in progress · first six models: late October 2026 · all twelve: late November 2026
Row structure preview. Values render only from computed results after grading day — never hand-typed.
The Battery
25 prompts / 3 takes each
The prompt set
Tests that expose the edges.
The prompts are short. The craft questions are not. Each one isolates a familiar failure mode and gives raters a shared reference.
19 Core prompts6 Frontier probesClass S · 8sClass X · 12s
Two duration classes · eight scored axes · one published rubricThe full structure is documented on The Battery page.
Field notes / 04
The part where the room gets quiet.
A critical record of what the scores can show, and what they cannot. Published in the Print Journal alongside each edition.