How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
Design decision, 2026-08-04, taken after measuring all three. Supersedes the collision coupling the Isaac line has been running, and rejects an intermediate proposal that looked right for an hour.
A single-entity model is added to the comparison: it sees no second body at all and learns the burst pattern of the target's own trajectory. If it matches the relational arm, the relation is not necessary and H2 fails, whatever else the design achieves.
| Relation | B zero-order | C constant velocity | SELF pattern | D relation |
|---|---|---|---|---|
| Carriage — target rides the reference | 0.00 | 0.25 | 1.00 | 1.00 |
| Capture — reference arrives, then carries | 0.33 | 0.36 | 0.33 | 0.50 |
Superseded for capture, 2026-08-04, by a preregistered run. The capture row above came from a deliberately crude proxy. Measured in the cell, on 200 paired cells with the rule locked in advance: SELF 0.000, D 0.650, 130 discordant pairs none of them SELF's, margin +0.650. The 0.50 was a floor, as this section says below, and the argument this table was built to make holds. Rule · Result
Under carriage the single-entity model is identical to the relational one. The intermittency that makes the task hard is written into the target's own trajectory, so nothing needs to look at a second entity. H2 fails outright.
Under capture the single-entity model is identical to zero-order. Before the capture the target is still, so its own history contains no information about what is about to happen to it, and only the approaching body does.
| Design | Relation necessary | Effect exceeds tolerance |
|---|---|---|
| Collision — struck and released | yes | no — self-limiting |
| Carriage — rides throughout | no | yes |
| Capture — arrives, then carries | yes | yes |
Collision fails on magnitude for a reason now measured five ways: the push moves the target away, which reduces penetration, which reduces the push, so displacement per contact is on the order of the interaction radius, and the radius is below the placement tolerance (the ceiling).
Carriage fails on necessity because an accumulating effect leaves its own trace.
Capture has neither failure: the onset is invisible in the target's history, and what follows accumulates without bound.
Carriage was recommended first, on the grounds that it restores the task's
original framing — commitment_task.py describes "a bread slice carried by an
intermittently nudged tray", and the 1-D proxy that put arm B in a 0.05–0.09
band was that model. The reasoning was that the Isaac line had silently
substituted collision for carriage and the ceiling was an artefact of the
substitution.
The first half of that is true. The conclusion was wrong, and the single-entity model is what showed it. A recommendation that survives an hour of reasoning can still fail the first measurement aimed at it, which is why the measurement came before the implementation.
A two-phase relational prediction. The quick proxy scored arm D at 0.50 against arm B's 0.33 by rolling the reference's pattern forward and applying it to the target immediately — which is wrong before the capture, when the target is not yet attached. A correct arm D predicts when the reference reaches the target, and applies carriage only from that step. So 0.50 is a floor produced by a deliberately crude model, not a ceiling on the design.
A gate that fires on capture — which turned out to need nothing.
Reversed later the same day. Everything from here to the end of this section is wrong. The proximity contrast fires on capture at 0.00, not 1.00, and finds no post-contact far-field steps at all. The prediction below was correct and the measurement that overturned it was reading an off-by-one in
capture_displacement: a captured target was thrown one body-step past its carrier, which put the riding separation outside the radius and sorted the pauses into the far field. The variant deleted for failing its own test was deleted against an artefact.What replaced it is neither the old claim nor that variant. The gate admits a second form of positive evidence — carriage, the target's displacement being a body's own — so one gate still covers both relations, but not by sharing this statistic. Capture fires at 1.00 through the new path. The measurement
The prediction here was that capture inverts the evidence. Under collision the discriminating question is whether the target stops once the body leaves, so evidence from before the first contact is discarded; under capture the body never leaves, so there should be no post-contact far-field period and the gate should abstain on every cell.
Measured, it does not. A carried target keeps a small separation from its carrier, and the pauses supply far-field steps in any case: 20 usable deltas rather than none. The same thresholds fire on capture at 1.00 and reject a target moving on its own at 0.00, which is exactly what a capture-specific variant gave. That variant was written and then deleted rather than shipped on a premise its own test contradicted.
One gate covers both relations.
The first control for capture had the target begin moving with the same pattern at the same step the body arrived. That is not a different world: the trajectories are identical and only the causal story differs, so no gate can separate them and none should be claimed to. It fired at 1.00 with statistics identical to capture, which is the correct behaviour and an uninformative test.
The controls now differ in onset and direction, and are rejected at 0.00.
The commit window relative to the capture, the approach speed, and the duty cycle are all unfixed. They must be set on grounds independent of which arm they favour before any confirmatory run, and the numbers above are exploration.