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 settling sweep — rule fixed before the measurement, v1.0

Written 2026-08-05, before the sweep was implemented or run. The manuscript draft states a criterion in §4.3 that it does not measure: the carrier must come to rest inside the time it holds the target, and its settling time must be short relative to the action's commitment latency. That is an inference from two separate physical measurements — settling 22 steps, grip 54 steps — and an inference is not a result.

This turns it into one. Settling time is a property of the carrier, and on CPU the carrier is ours to specify. So we inject settling into the injected coupling and sweep it, over a configuration where the relation is known to be necessary, and watch whether Paper 002's operator returns.

Nothing here is confirmatory. It is a claim about which model suffices for a given carrier, and that question does not need the world in it — which is exactly why this is the one part of the physical story CPU can settle.

The mechanism under test

The relation is made necessary by intermittency. The design commands the carrier to advance for burst_on steps and hold for burst_off. A real arm does not hold when told: it takes s steps to come to rest, so a commanded pause of burst_off steps produces an observable pause of roughly burst_off − s steps, and none at all when s ≥ burst_off.

This predicts the physical result exactly, and the prediction is what makes it worth testing rather than merely restating:

commanded pause settling pause that survives
CPU, scripted carrier 4 0 4
physical, burst_off 4 4 22 0
physical, burst_off 25 25 22 3

Implementation

EncounterSpec.settling_steps, defaulting to 0, which reproduces the current carrier exactly. The commanded velocity is smoothed by a boxcar of width settling_steps + 1 before being integrated into the body's position.

A boxcar rather than an exponential lag, for one reason: after the command stops, a boxcar of width w keeps the body moving for exactly w − 1 further steps. So settling_steps is the measured quantity — "steps until the arm reads as stopped" — with no time-constant conversion in between, and the sweep's x-axis is directly comparable to the 22 measured on the arm.

Predictions — fixed before the sweep runs

Configuration: capture + burst, burst_on 10, burst_off 4 (period 14), dispense_latency 9, the commit window and gate exactly as preregistered.

P4 is the one that carries the most weight, because it tests whether the physical burst_off 25 row's anomaly — arm B jumping from 0.087 to 0.625 — is a property of the protocol or of Isaac. If CPU reproduces it, the protocol explains it and the physical row needs no further defence.

What a failure of each prediction would mean

Not permitted

Seeds and size

Seeds from 7000, distinct from every run so far (300 for the arm sweeps, 3000 for the amended SELF band), so no configuration is being re-read on cells it was tuned against. 60 seeds per settling value, s ∈ {0, 1, 2, 3, 4, 6, 9, 14, 22}.

The value 22 is included because it is the arm's measured settling time, so one row of this sweep is the CPU counterpart of the physical run.

Outcome, 2026-08-05

All four predictions failed. Recorded here beside the rule they were written against, and in full at settling_sweep_v1.0.

Prediction Measured
P1 C ≥ 0.90 by settling 4, monotone 0.100; not monotone FAIL
P2 D − B ≤ 0.05 by settling 4 0.367 FAIL
P3 C overtakes D at settling ∈ [2, 6] at 14 FAIL
P4 B ≥ 0.40 at settling 22, burst_off 25 0.333 FAIL

The document above states what each failure would mean, and those consequences were applied rather than renegotiated:

Nothing in this document was edited after the sweep ran except the addition of this section. The predictions stand as written, including the one that named 0.40 and got 0.333.

Version history