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

The collapse threat belongs to H2, not to the gate

Written 2026-08-05, before the change was measured. Follows the finding that no constant-velocity ceiling can admit a capture while rejecting a sustained push. States where the defence actually lives, and closes the hole that removing the ceiling would otherwise open.

What forced this

The gate's negative evidence refuses a relational claim where a constant-velocity model already accounts for the motion. Measured across 40 seeds each:

case min median max
capture / burst admit 0.756 0.779 0.796
capture / probe admit 0.636 0.652 0.672
steady push reject 0.252 0.415 0.636
slide reject 0.985 0.987 1.000
drift reject 1.000 1.000 1.000

A captured target is more constant-velocity than a pushed one, and the order is inverted, so no ceiling separates them. The derivation rule anticipated this and said to report it rather than pick a number from the range.

The reason is structural. A carried target rides with its carrier, so during a burst its motion is exactly constant velocity. A pushed target sits at an equilibrium separation and jitters. The ceiling was calibrated when the treatment was collision, where contact is episodic and the target's motion jumps; capture moved the treatment to the other side of the clause.

The category error underneath

slide and the sustained push are described as controls for H3, gate specificity. That is the wrong hypothesis for them.

H3 asks whether the gate can tell the relation is present from the relation is absent. In static, noise and drift the relation is genuinely absent — the target's motion has nothing to do with any body. The gate should refuse, and it does.

In slide and in a sustained push the relation is present. Contact really does cause the motion. What is true of them is something else: a simpler operator already suffices. That is not a question about the gate's ability to detect a relation. It is a question about which operator to choose, and choosing the smallest sufficient operator is the program's own governing principle.

So they are H2's controls, not H3's. H2 already requires arm D to beat the mode operator by a margin in every counted variant; where a constant-velocity model suffices, arm C succeeds, the margin vanishes, and H2 fails. The defence is in the scoring, measured on outcomes, and it does not depend on a threshold being right.

The gate refusing them as well was belt-and-braces, and the braces are now demonstrably mis-calibrated for this relation.

Two changes

1. Carriage evidence requires the carrier to be in contact.

Removing the ceiling from the carriage path alone would let drift in: its target runs along the first body's own axis at its own speed and agrees with it on 0.71 of moving steps. But that body is 183 mm away and never comes within the interaction radius.

You cannot carry what you are not touching. Requiring the agreeing body to be within the interaction radius during the riding run is the definition of carrying rather than a tuned threshold, and drift fails it outright while a capture passes it trivially.

2. The constant-velocity ceiling applies to the proximity path only.

Where it was derived, where it works, and where the treatment is episodic. The carriage path keeps the contact requirement above as its negative evidence.

What this gives up, stated plainly

The gate will now fire on a sustained push and on some slides — cases where a relation is present and arm C is sufficient. Arm D will act where it does not need to. That is a real cost and it shows up as engagement without advantage, which the endpoint already reports as marginal-versus-conditional rates.

Measured after the change, and the worry was mostly unfounded: slide fires on 0.05 of trials, still under H3's 0.10 ceiling, so the preregistered control set is unaffected. drift, static and noise are at 0.00, and the treatment at 1.00.

The only case that now fires and did not before is the sustained push, which was never in H3's list - it could not be expressed until the steady schedule was added today. So no preregistered hypothesis has to move for this change, and the argument above stands as the reason the sustained push is not added to H3's controls now that it can be run.

What is untouched

min_proximity_contrast, min_post_contact_far_deltas, min_carriage_run, min_ride_steps, the tolerance, and the commit window. The one-step cv_gain replacement stands on its own grounds — a threshold should not change meaning because the action's length changed — independently of where the clause is applied.