Every clip, cut in half

A model is shown the first half of a Rube Goldberg machine and has to produce the second. The second half is the answer key. All 260 cases are below — press play through on any of them to watch a chain run straight across its own cut.

260cases
130machines
2samples each
1shared prompt
prompt.txt — identical in all 260 cases

The input video shows the first half of a Rube Goldberg machine: a chain of physical events in which each step is caused by the one before it, under gravity and rigid-body contact alone. Nothing in the scene is actuated, remote-controlled or scripted, and no object appears, vanishes or passes through another.

Predict the second half. Continue the video from the frame it stops on, from the same fixed camera, for the same duration.

You are scored on whether the machine's causal chain proceeds the way physics requires — which objects move, in what order, what they strike, and where they come to rest — not on pixel-level similarity to the reference.

Why one prompt, and not each machine's own. Every machine ships a design specification — “the impact must split the mechanism into simultaneous left and right chains … the finish sphere must travel diagonally into its own rotated box.” That is a written-out description of the half the model is being asked to predict. Handing it over would score reading comprehension. It travels in each case's metadata.json, where a scorer can read it and a model cannot.