After the Spill

How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.

View the Project on GitHub higuseonhye/research-os

One contact cannot displace the target past its own tolerance

Structural finding, 2026-08-04. Reached from five independent directions over one day, and it decides what Paper 003 can claim.

The measurement

A coherent encounter, expressed in radii so it scales with the scene: the first body advances to penetrate, withdraws clear, and stops; the second is launched once at the object and travels four radii, passing through. Both bodies contact in every seed and the target settles afterwards.

Script speed Gain Body 1 hits Body 2 hits Window displacement Over 20 mm Settles
4 mm/step 0.5 1.00 1.00 11.4 mm 0.00 1.00
6 mm/step 0.5 1.00 1.00 14.9 mm 0.00 1.00
8 mm/step 0.9 1.00 1.00 16.0 mm 0.00 1.00
10 mm/step 0.9 1.00 1.00 16.0 mm 0.00 1.00

Displacement over the dispense window never clears the 20 mm tolerance, at any speed or coupling strength tried.

The same wall, five times

Where Displacement Tolerance
Injected coupling, 50 mm radius, speeds 4–80 mm/step ~15 mm 20 mm
Injected coupling, gains 0.3–0.95 10.5–17.2 mm 20 mm
Two-body encounter on the CPU proxy 15.6–17.3 mm 20 mm
Real contact in the lift scene, coast after a strike 10.5–26.6 mm 20 mm
This encounter, 12 mm radius, speeds 4–10 mm/step 8–17 mm 20 mm

Every route arrives within a few millimetres of the same number, and it sits just below the criterion.

Why it is bounded

Three constraints multiply out to a ceiling, and none of them is a parameter that can be relaxed without giving something else up.

The target cannot be pushed faster than the body pushing it. So displacement over a window is at most the body's speed times the window.

The body's speed is capped by observability. A body travelling further per step than the interaction radius crosses the contact zone between observations, so nothing sees the contact. That is already enforced.

Contact is self-limiting. The push moves the target away, which reduces penetration, which reduces the push. The target escapes after roughly one radius of travel, and a body that keeps up instead becomes a bulldozer — measured, a pusher that never stops shoves the target 111 mm and the residual becomes constant-velocity, which is arm C's case, not arm D's.

So displacement per contact is on the order of the interaction radius, and the radius here is 12 mm against a 20 mm tolerance. Raising the radius above the tolerance would mean declaring contact at distances the scene's geometry does not support — real contact in this scene happens at 2 to 5 mm.

What this means for H1

The capability-crossing endpoint requires success(B) in a near-zero band. Arm B aims where the target is and fails only when the target moves further than the tolerance during the dispense. One contact does not move it that far.

So H1 is not reachable in this task family from a single contact, and this is a property of the mechanism rather than of a parameter that was not tuned hard enough.

Three honest routes remain, and the choice between them is the researcher's:

  1. Accumulate contacts. Several strikes inside one window would clear the tolerance — but a sustained push is what a constant-velocity model explains, so the residual moves into arm C's regime and the paper loses its discrimination. Measured directly: the bulldozer case.
  2. Change the endpoint. Drop capability crossing for a statement the evidence supports: the relational operator predicts a displacement that neither zero-order nor constant-velocity does, at a magnitude below the task tolerance. Weaker, and true.
  3. Report the ceiling as the finding. That a prepared relational operator cannot cross a capability threshold in this family because contact is self-limiting is itself a result, and it is preregisterable, testable, and informative about when relation-level expansion pays.

None may be chosen by which arm it favours. Route 1 is already measured to fail on its own terms, which leaves 2 and 3 as genuine alternatives.

What is not in doubt

The gate works. On real contact traces it fires on every trace containing a strike, with the constant-velocity clause comfortably satisfied — cv_gain between −0.20 and −0.75, meaning a constant-velocity model is worse than zero-order on this motion. The discrimination Paper 003 rests on survives real physics. What does not survive is the endpoint built on top of it.