The Loop Lab · Charlot Lab · Explorable 2 of 3
The Focus Band
Six ways to spend attention on one body. The trajectories were measured once; everything you move below is recomputed exactly, because gate cost, slow-path cost and the substrate's idle floor are additive in the accounting and never touch the trajectory.
The world
The gate
The always-on diffuse layer is not free. Raise it and watch the cascade stop paying.
The substrate
How far ahead must a slow path see before it earns its cost?
| Rollout horizon | Falls, base scheme | Falls, deliberative | Track, base | Track, deliberative | Compute | Verdict |
|---|
Further ahead than the body takes to move. This cart needs about a second to relocate. Give the deliberative arm a 1.2 s horizon and it beats a tuned linear law on falls and on tracking. Give it 0.24 s and it is worse than that law while costing 288 times as much. Nothing about the arm changed across these rows except how far ahead it could see, which makes horizon a design variable you can set from the body's own time constant before writing any policy.
Why can these numbers change without re-running the simulation?
Because none of these three controls touches the trajectory. The six arms were simulated once each, at four disturbance levels and a 1.2 s rollout horizon. Those runs fixed the escalation rates, the falls, the tracking and the actuation joules.
The three sliders change none of that. Gate cost and slow-path cost enter compute energy additively and multiplicatively; the idle floor is a constant draw per step. So every number above is an exact recomputation of the measured run under different accounting, not an interpolation and not a fresh simulation. That is also why the bench could sweep the break-even analytically instead of re-simulating it.
Two things to try. Push the gate cost up until the cheapest arm changes identity. Then put it back and push the idle floor from 2% toward 90%, watching the spread tile and the resolution tile cross.