The first expert only opens the drawer. The second one waits for that gap, then drops the driver in without clipping the rail. Those used to live in one long demo that died on step three. That is the note @axisrobotics put up this morning. They are filing solved motions as parts you can chain, then packing the chain into a bigger policy. A deployed run trained its own successor on the misses it made in the real room and moved from 22 to 52. A cheap specialist can nail one skill for five to ten dollars of compute. A small proxy even throws out part of the pile so the leftover set covers more than the full dump. https://t.co/pPT0B53lU9 Raw trajectories still matter. A shelf of reusable skills is what they are stacking now.