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

Paper 003 — SELF arm decision rule

Preregistration. Written 2026-08-04, before the arm existed and before any capture cell was scored against it. Fixed in advance precisely so that the result cannot change it. Nothing in this document may be revised after the run except by recording the revision, the date, and the reason in place.

Locked 2026-08-04 by the co-PI, before implementation began: the arm definition, the ungated asymmetry, n = 200 in the [+4, +6] band, α = 0.05, and the 0.15 margin. Implementation and run follow this line in the history.

Why this arm decides the paper

H2 requires the ordering

relation arm  >  parameter repair  >  the target's own trajectory

The third term has never been measured in this repository. It is the hypothesis's most dangerous competitor, and it is the one that already killed a previous design: carriage was recommended, and rejected within the hour, because a single-entity model that learned the burst pattern of the target's own trajectory matched the relational arm exactly. Capture was chosen over it on the argument that a still target has no history to learn from before the arrival.

That argument is now in question, and not from any doubt about the relation. It is from a measured property of arm D: arm D cannot act before commit offset +4, because the gate requires a run of at least three carriage steps and two consecutive crossings. So every commit that scores anything is at least four steps into the carry — and by then the target's own trajectory has begun to carry the carrier's burst pattern. The condition capture was chosen to avoid is exactly the condition every scored cell now sits in.

Either the relation survives that or it does not. Both outcomes are results.

The SELF arm

A single-entity model. It observes only the target's own position history and never sees that a second body exists.

It is scored on every cell as a fifth arm, alongside A, B, C, D.

It is deliberately given the easier deal

Arm D may act only when the relation-adequacy gate fires. The SELF arm is not gated at all — it acts whenever its own pattern is identifiable.

That asymmetry is not an oversight and is not to be corrected. It favours SELF, so it makes the test harder for the hypothesis, and a relation that beats an ungated single-entity competitor has answered the objection in its strongest form. Gating SELF would be tuning the competitor down.

Population, fixed in advance

Cells capture coupling, burst schedule, 1 body, coupled condition
Loop run_cell with InjectedWorld, CPU, EpisodeSpec() defaults
Eligibility valid, resolved, and commit offset in [+4, +6]
Sample n = 200 eligible cells, seeds drawn from 300 upward until 200 are collected
Pairing SELF and D are scored on the same cell, so outcomes are paired

The offset band is not chosen here — it is where arm D can act at all, measured before this document. Cells outside it are reported but do not enter the test, because there SELF is compared against an arm that declined.

n = 200 is fixed now and does not move. The exploratory run put roughly a third of seeds in the band, so this needs about 600 seeds.

Decision rule

Let p_D and p_SELF be the success rates on the 200 paired cells, and let the discordant pairs be those where exactly one of the two landed.

H2 survives only if both hold:

  1. Superiority. One-sided exact McNemar test (binomial on discordant pairs), H_a: p_D > p_SELF, α = 0.05.
  2. Margin. p_D − p_SELF ≥ 0.15.

If either fails, H2 is rejected. Not "inconclusive", not "underpowered" — rejected. The burden is on the relation to show it is necessary; a relation that cannot be distinguished from the target's own trajectory is not doing the work the paper claims for it.

Where 0.15 comes from

Set from what the design has already demonstrated against the arm it was built to beat, not from anything measured against SELF. At offsets +4 to +6 arm D scores 0.53 to 0.71 against the mode operator's 0.00 to 0.07 — a margin of roughly 0.5. The floor here is under a third of that. It is deliberately low enough that it cannot be accused of having been set to whatever arm D turns out to achieve, and high enough that a marginal edge does not count as a capability opening.

What happens next, decided now

Case A — p_D > p_SELF at α = 0.05 with margin ≥ 0.15. H2 stands. The relation is necessary in the regime where it can act. Proceed to the two-body redraw, then the preregistration rewrite, then the confirmatory sample.

Case B — either condition fails. H2 is rejected for capture as currently defined, and it is written up as such. The relation definition reopens; the collision and carriage results, and the reason each was rejected, are already recorded and stand as the map of what has been ruled out. Paper 003 becomes a paper about what the missing-relation cell requires and why three candidate relations did not supply it — which is a result, and is the reason this rule can be fixed without fear.


Outcome, recorded 2026-08-04

Case A. H2 stands. 200 paired cells from 623 seeds.

p_D 0.650
p_SELF 0.000
Discordant: D only / SELF only 130 / 0
One-sided exact McNemar p = 7.3 × 10⁻⁴⁰
Margin +0.650

Both conditions passed. Nothing in this document was changed to reach that.

SELF acted on 0.675 of cells and landed on none of them, and it was not broken: its median miss when acting was 60.0 mm against 30.0 mm when it declined to arm B — it extrapolates through a pause it cannot see, and 60 mm is exactly one pause at 15 mm/step. It holds a median of 4 steps of its own motion at commitment against a 14-step cycle.

Full record

One limitation is recorded there and not resolved: SELF's disadvantage rests on how little of its own history it holds at +4 to +6, so the protection is bounded in time rather than absolute. Where the bound lies is a separate question on fresh seeds, and the clause below is why it was not answered by widening this run.

Amendment, 2026-08-05 — recorded before the re-run

Invoked under the clause below, which says a proposal is recorded here with the result that prompted it and applies to a future run on fresh seeds.

What prompted it. dispense_latency was derived from the physical scene and rose from 6 to 8 — the block carries at 2.86 mm/step at the tenth percentile against a 20 mm tolerance, so six steps leave the target inside tolerance and the cell poses no prediction problem. The derivation

The commit window is one dispense-length either side of the arrival, so it moved with it. The band this comparison runs in is defined as where arm D can act at all, and that is measured, not chosen.

What changes.

before after
dispense_latency 6 8
commit window ±6 ±8
offset band [+4, +6] [+4, +8]
seeds 300–922 from 3000

The band was measured on seeds 2000–2119, disjoint from both the original run and the re-run, before this amendment was written. Arm D acts on 0.00 of cells at every offset from −7 to +3 and on 0.75–0.94 from +4 to +8. The lower bound is unchanged, because what sets it is unchanged: three carriage steps plus two consecutive gate crossings.

What does not change. The arm's definition, its ungated asymmetry, α = 0.05, the margin of 0.15, n = 200, the pairing, and the test. The band moves because the window moved, and for no other reason.

The gate also changed since the original run — the constant-velocity ceiling is now scoped to the proximity path, with contact doing the work on the carriage path. That changes when arm D may act and is another reason the original result cannot simply be carried forward. Why

Case A stands as what it was: a comparison at dispense_latency 6, on seeds 300–922, of a task the physical scene cannot pose. It is not retracted and it is not evidence for the amended design. The re-run may fail, and if it does, that is the result.

Outcome of the re-run, recorded 2026-08-05

Case A again. 200 paired cells from 401 seeds, band [+4, +8], seeds from 3000.

p_D 0.735
p_SELF 0.045
arm C 0.145
arm B 0.000
Discordant: D only / SELF only 146 / 8
One-sided exact McNemar p = 3.0e-34
Margin +0.690

Both conditions passed, and nothing in the rule was changed to reach them.

This is better evidence than the first Case A, not merely a repeat. There the competitor scored 0.000 and won no discordant pair, which required a separate check that it was not simply broken. Here it scores 0.045 and takes 8 pairs, so it demonstrably can win and still loses by 146 to 8. The suspicion the first result had to answer does not arise.

Not permitted after the run

Any of these may be proposed after the run, and if so the proposal is recorded here with the result that prompted it, and it applies to a future run on fresh seeds — never to this one.