Claude overengineered ten chores, then had to prove each machine works.
Machines designed offline by Claude Opus 5.5 using the same tools and checks. Live generation needs an API key and isn't running here.
General notes
Each sheet is one chore from the project's 20-chore eval set. Each revision is one real simulate call, and the report under it is word for word what the model would read back.
A part joins the chain when a chain part touches it and the push changes where it goes. The simulator runs every machine twice in lockstep, with and without the first push. A part counts once it's at least 5 cm off its push-free path and has itself moved at least 5 cm.
A machine works only when a chain of at least 5 parts reaches the finale, and nothing reaches the finale when nobody pushes.
Your browser re-runs every revision you open from its blueprint and checks the run's trace hash against the one recorded in Node. The balloons number the chain in order; the red pencil marks where each link happened.
Revisions every simulate call, in order, failures included