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 Capability-Crossing Preregistration — v1.0 (capture), CLOSED

Status 2026-08-05 — CLOSED. No confirmatory run will follow.

The Isaac calibration pilot required below has been run. It answered its first question yes — capture happens under real contact, in 60 of 60 cells — and then returned a result that stops the design: the discriminating control saturates. Paper 002's mode operator lands 0.957 to 1.000 of physical cells across three configurations and 107 cells, while the relational arm lands 0.174 to 0.583 and never once beats the single-entity competitor.

H1's crossing condition and H2's first clause are therefore not merely unlikely to pass — they cannot pass at any n. The document is closed on that ground and not amended to make them passable.

What may be reported, and what may not, is written out in Closure. In short: the physical runs are calibration, this preregistration's own text excludes calibration from confirmatory estimates, and no confirmatory cell was ever run. So "H2 was rejected at α = 0.05" is not a sentence this record supports. What it supports is that the design's precondition failed in calibration.

Every locked section below is left exactly as it was locked on 2026-08-04. Nothing in the design was rewritten after seeing a number. The additions are the closure section, the outcome stamps inside H2 and the pilot checklist, and three corrections of fact marked as such.


Supersedes paper003_prereg_draft_v0.1.md entirely. That draft assumes the collision coupling throughout, grades variants by reference speed (measured non-functional), and its sample-size section predates the change of relation. It is kept for provenance and must not be cited for any design decision.

Status 2026-08-04 — superseded by the block above, kept verbatim. The design is locked. The physical parameters are not, and are marked PENDING — they come from an Isaac calibration pilot that has not been run for capture. This document is frozen against everything a result could tempt us to change, and open only where the world has not been measured yet.

Everything reported below as measured is CPU with injected coupling, which is a calibration device: the coupling is a formula, normal_alignment is 1.0 by construction, and there is no contact jitter. It is legitimate for powering and for arm comparisons, and it is not confirmatory evidence about the world.


Research question — unchanged

When repeated task failure survives both parameter repair and a mode-level structural expansion, and the residual is conditioned on a second entity's state, does adding a prepared relation-module expansion convert task variants that are unachievable into achievable ones — without regressing on variants that do not require the relation?

The claim is limited to the specified task family and coupling. It is not a claim about general causal discovery, relation invention outside the prepared operator, or capability emergence in untested task families.

The endpoint is deliberately not prediction error. The open-loop prediction advantage is around 1 mm against Paper 002's 10.8 mm, so a Paper-002-style contrast would understate or miss the effect this paper is about.


What changed from v0.1, and why

The relation. Three candidates were measured against each other and capture was chosen — a body arrives at a still target and carries it off. Collision cannot clear the placement tolerance: the push moves the target away, which reduces penetration, which reduces the push. Carriage clears it and makes the relation unnecessary, because a single-entity model matched the relational arm exactly. Capture has neither failure. Design · The ceiling

The encounter. One body, burst schedule. The two-body encounter is retired for this relation, not redrawn: under capture the first body to arrive consumes the target, so "demonstrate on one body, apply with another" has nothing left to apply to. Why

The commit policy, below.

A fifth arm, below.


Locked arms

Arm Target model Role
A Frozen, no update Secondary baseline
B Parameter repair within the independent-entity model Primary comparator
C Mode expansion — Paper 002's prepared operator Discriminating control
SELF The target's own trajectory, no second entity observed Decisive competitor
D Relation expansion — prediction depends on the second entity Primary intervention
D* Relation with the true landing supplied Diagnostic ceiling only

All arms share the same controller, task, tolerance, and commit policy, and differ only in the predicted landing point.

SELF is new in v1.0 and it is the arm that decides the paper. It sees no second body at all and learns the periodic pattern of the target's own trajectory. It already killed one design: under carriage it matched the relational arm exactly. Under capture it is the reason the relation can be called necessary rather than merely sufficient.

SELF is deliberately ungated. Arm D may act only when the relation-adequacy gate fires; SELF acts whenever its own pattern is identifiable. The asymmetry favours the competitor and is not to be corrected — gating SELF would be tuning it down.

Arm D must estimate everything it uses. The capture radius is recovered from observation (estimate_capture); being handed the generating model is arm D*, excluded from every primary estimate.


Task: commitment point

The agent commits to an irreversible placement onto a target that a reference body captures and carries. The action takes dispense_latency steps and lands wherever the target is at completion. There is no correction.

This structure is required, not incidental: a continuous reach-and-hold version produced no separation between any arms, because continuous re-aiming averages prediction error away.


Commit policy — LOCKED 2026-08-04, before the run that tests it

A commit candidate must lie within one dispense-length of a body's arrival, on either side. Among the candidates that qualify, one is drawn uniformly at random per cell from the cell's seed.

Arrivals are steps at which a body crosses into contact range, per body, from contact_arrivals.

Why the window, on grounds no arm appears in

Fixed on the structure of the action. The dispense takes dispense_latency steps and lands where the target then is, so a commit further out than that either completes before anything has happened to the target, or measures a regime the arrival no longer governs. Symmetric, because there is no reason to prefer a side and preferring one is exactly how "so that arm B fails" would enter.

What it removes is an artefact, not a preference. The eligibility screen admits any step where the target will move, and under capture that is every step after the arrival, because a carried target rides forever. So the eligible set — and with it the commit distribution — was being set by episode_steps, an arbitrary number, and almost all of its mass sat in the riding tail. The same cell scored differently for running longer.

Uniform within the window, for the reason v0.1 gave and which still holds: nobody places at the earliest physically possible instant, and the commit moment must not depend on when any particular arm's information becomes available.

Declared in advance: the direction of this policy's effect

Relative to v0.1's policy — uniform over every eligible step — the window removes the deep riding tail. In that tail a constant-velocity model does well, the single-entity arm eventually does better still, and arm D declines because cv_gain rises past the gate's ceiling. Removing it therefore favours arm D, and that is stated here rather than discovered afterwards.

The policy is chosen on the structural ground above. Had the reasoning pointed the other way the window would be the same.

Not permitted

The window may not be moved to improve the rate at which arm D engages, and the gate's min_carriage_run may not be lowered for the same purpose. The window is fixed on the action's structure; the run requirement is fixed on the collision equilibrium, where a struck target momentarily rides at the body's own speed. Either may be changed only through the encounter, on grounds stated before the run.

Recorded per cell

commit_policy, arrivals, commit_offset, committed_in_window, and the full eligible_steps list, so the realised distribution of commit moments can be audited rather than taken on trust.


Engagement, and why the endpoint is reported twice

Arm D cannot score worse than arm B. When the gate does not fire, or the relation cannot be fitted, arm D falls back to arm B's aim and is identical to it cell by cell. So D ≥ B holds by construction and the paired difference can never be negative.

Two consequences:

  1. "The interval includes zero" is nearly automatic and carries almost no information. The lower bound is structural, not evidential.
  2. The informative quantities are how often arm D engages and how much it wins by when it does. Both are preregistered endpoints.

The marginal rate and the conditional rate are reported together, always. On the CPU calibration they are 0.23 and 0.78 — a difference large enough that reporting either alone would misrepresent the design.


Which test is confirmatory — LOCKED

A one-sided paired sign test on the discordant cells, α = 0.05.

The paired bootstrap's lower bound cannot go negative, for the structural reason above, so "the interval clears zero" is a statement about the fallback rather than about the model. The sign test conditions on exactly the cells where the two arms differ, which is where the whole signal is. It is also the more conservative of the two.


Sample size — LOCKED as a rule, PENDING as a number

The v0.1 table does not transfer. It assumed arm D landing 1.00 against arm B's 0.80 when engaged. Under capture arm B lands 0.00 in the commit window — the target travels 15 mm per step against a 20 mm tolerance, so a zero-order aim cannot survive a carry — and the conditional effect is therefore far larger, while engagement is lower. Recomputed for that structure, with arm D landing 0.78 when it engages:

Committed cells eng 0.10 eng 0.15 eng 0.23 eng 0.35 eng 0.50
20 0.02 0.08 0.28 0.67 0.94
30 0.08 0.27 0.65 0.94 1.00
40 0.20 0.51 0.87 0.99 1.00
60 0.51 0.85 0.99 1.00 1.00
80 0.76 0.96 1.00 1.00 1.00
120 0.96 1.00 1.00 1.00 1.00
for 0.90 101 67 43 28 19

Engagement is not a nuisance parameter. Cells where arm D declines are ties, and a sign test discards ties, so halving engagement costs far more than half the power.

Sizing rule — LOCKED. The confirmatory n is read off this table using the engagement rate observed in the real-contact calibration pilot, not the injected-coupling one, and not the 0.23 column. Under real contact the gate's carriage evidence has to survive contact jitter, and nothing yet measures how often it does.

n = PENDING, to be filled from the pilot before any confirmatory cell is run, and recorded here with the pilot's engagement rate beside it.

2026-08-05. The rule was applied and it returned a number. Real-contact engagement is 0.27 (block, 60 cells) rising to 0.50 with the derived pause, so the rule reads n ≈ 43 at 0.27 and ≈ 19–28 at the higher rates — all of them smaller than the 0.23 column would have given, and all of them moot. The sizing rule was never the binding constraint; the closure is. Recorded because the rule was locked and it is owed its answer.


Primary endpoint: capability threshold crossing

For a preregistered variant set T and arms a:

crossed(t) :=  success(t, B)    <= NEAR_ZERO_BAND
           AND success(t, SELF) <= NEAR_ZERO_BAND
           AND success(t, D)    >= ACHIEVABLE_THRESHOLD

The primary estimand is the number of variants in T satisfying crossed, with per-variant success rates and intervals reported for every arm.

SELF joins the crossing condition in v1.0. A variant where the relational arm succeeds and the single-entity arm also succeeds is not a capability the relation opened, whatever arm B does.

NEAR_ZERO_BAND cannot be zero. Under a strictly periodic reference, timing irregularity lifts a near-zero arm to 0.05–0.09 because some dispense windows contain fewer moving steps than the periodic minimum. Writing == 0 would preregister an artefact of perfect periodicity.

How T is graded: PENDING. Reference speed was the v0.1 grading variable and it is measured non-functional — it does not move the arms apart. The replacement must be chosen from the calibration pilot on the grounds that it varies the difficulty of the prediction, and it must be fixed before the confirmatory run.


Confirmatory hypotheses

H1 — Capability crossing (primary)

At least K_CROSS variants in T satisfy crossed, with the lower endpoint of the bootstrap interval for success(t, D) above ACHIEVABLE_THRESHOLD, and the upper endpoints for success(t, B) and success(t, SELF) below NEAR_ZERO_BAND, in each counted variant.

K_CROSS, ACHIEVABLE_THRESHOLD, NEAR_ZERO_BAND: PENDING.

H2 — The relation is necessary, not merely some expansion

Two competitors, and both must be beaten in every variant counted under H1:

Status on CPU: passed. Under a rule locked before the arm was implemented — 200 paired cells, offsets [+4, +6] — arm D scored 0.650 against SELF's 0.000, 130 discordant pairs none of them SELF's, p = 7.3 × 10⁻⁴⁰. The competitor was not broken: it acted on 0.675 of cells, and its median miss when acting was 60.0 mm against 30.0 mm when it declined, because it extrapolates through a pause it cannot see. Rule · Result

This is a result about the arms, not about the world. It is restated here as a confirmatory hypothesis because the physical run must be able to fail it.

Status under real contact, 2026-08-05: both clauses fail. The sentence above was written so that this could happen, and it did.

arm C arm D SELF cells
block 1.000 0.200 0.167 60
needle, pause 4 0.957 0.174 0.348 23
needle, pause 25 (derived) 0.958 0.583 0.750 24

success(D) − success(C) is negative in every configuration, so no C_MARGIN above zero is reachable. And across the 47 needle cells arm D wins zero pairs that SELF loses, while SELF wins four that D loses, in each configuration — the McNemar clause fails in the direction of the competitor. Result

H3 — Gate specificity

The relation-adequacy gate must fire on at least 90% of coupled trials and on at most 10% of each control: persistent drift, static, observation noise, and post-contact slide.

Firing on motion a constant-velocity model explains is the failure that matters — it would mean the existing mode operator already accounts for the evidence.

Two admissible forms of positive evidence, one gate:

The constant-velocity ceiling applies to both and is not to be scoped to either. The drift control agrees with its body on 0.71 of moving steps and would pass a carriage test outright; cv_gain 0.99 is what rejects it. A change that scopes the ceiling to the proximity path alone must fail test_drift_would_pass_a_carriage_test_and_is_rejected_anyway before it reaches this document.

H4 — No regression without the relation

Arm D must not score below arm B on any control condition, within the preregistered equivalence margin NO_REGRESS_MARGIN (PENDING).

This is the hypothesis that ruled out the strong form of the decoy design, which would have driven arm D to zero on static and noise in exchange for onset prediction.

H3 and H4 are CPU-only, and this is structural — noted 2026-08-05

Both are defined against control conditions, and every control here works by overriding the target's motion: static holds it still, drift moves it independently, noise perturbs it. Under contact physics decides the target's motion, so the override has nothing to act on. Ten cells requested as static had the arm grasp the block and carry it 35 to 186 mm, on every step.

The runner now refuses --condition other than coupled under contact rather than producing such a cell. So H3 and H4 hold as CPU results and are untested — not failed — physically, and any report of them must carry that label. Building a contact control means building a scene where a body approaches and does not capture, which is a different scene and was never built.


Scope limits — declared, not discovered

The paper does not claim onset prediction

Arm D cannot commit before the target is captured, and no arm could. static and noise are worlds where a body arrives at the target and nothing happens, and it arrives closer than in the treatment — 12 and 14 mm against 42 mm, in 1.00 of cells. Up to the moment of contact, a capturing approach and a non-capturing one are the same observation.

The claim is therefore:

Once a relation has taken hold, a prepared relation module predicts where the target is going when no model of the target alone can. It does not predict that the relation is about to take hold, and in this world nothing could.

The measurement

The advantage is bounded in time, and the bound is outside the window

Measured on fresh seeds, the single-entity arm catches up from commit offset +30, a little over two burst cycles after the arrival. The commit window is [−6, +6], where SELF scores ≤ 0.02. The bound sits five times further out than any commit the protocol makes, and it is reported rather than relied on.

What stops arm D first is the gate, not SELF: cv_gain climbs as the carry lengthens and arm D declines exactly where it crosses the ceiling. The clause written to keep drift out lands unprompted on the boundary of H2's validity. The bound

The capability window is a three-step band

Arm D engages from offset +4, because the gate needs a run of three carriage steps and two consecutive crossings, so 3 of the window's 13 steps carry the effect. That is why the marginal rate is 0.23 against a conditional 0.78.

A reader is entitled to ask whether a three-step band is a capability opening or a curiosity. The answer must come from the confirmatory sample and from whether the band survives real contact — not from this document.

2026-08-05: the band survives, and it is not enough. Under contact the gate still fires — engagement 0.27 to 0.50 — so the band is real rather than an artefact of injected coupling. But arm C lands the same cells without it. The question was well posed and the answer is the second one: in this scene the band is a curiosity, because nothing needed it.


The tuning that will not be done

Fixed here so that a later change has to argue with this list:


Open parameters — must come from an Isaac calibration pilot

Every one is PENDING, and none may be set from the injected-coupling runs:

Parameter Why it cannot be inherited
n Read from the real-contact engagement rate, per the sizing rule
Grading variable for T Reference speed measured non-functional; replacement unchosen
ACHIEVABLE_THRESHOLD, NEAR_ZERO_BAND, K_CROSS Depend on the realised spread under contact
C_MARGIN, NO_REGRESS_MARGIN Same
Capture radius, approach speed, duty cycle Physical properties of the scene
Observation noise Never measured. The gate's statistics were re-derived once already to stop depending on it being small

Outcome 2026-08-05. Four of these were measured, one is unmeasurable in this scene, one is inapplicable to this relation, and the rest never became meaningful because the design closed first.

Parameter What the pilot returned
Capture radius 2.5 mm — 8 separations × 6 seeds, the largest at which capture holds in a majority
Carry duration 54 steps median (block), 68 (needle) — bounded by the gripper dropping it
dispense_latency 9 — p10 of the L-step displacement clears the 20 mm tolerance
Engagement 0.27 block · 0.39 needle · 0.50 needle with the derived pause → n ≈ 28–43
Arm D when it acts 0.375, against 0.78 on CPU
Observation noise 0.00 mm/step, which means there is no noise model in this scene — not that noise is small. It remains an unmeasured threat
normal_alignment Inapplicable to a capture. See the correction below
Grading variable, thresholds, margins Never chosen. T was never constructed, so H1 was never physically testable

Correction — normal_alignment does not apply to this relation. It measures displacement against the contact normal, the direction a struck target departs along. A carried target moves with its carrier and the sign inverts. It was listed below as "the dominant threat under realistic contact" on the strength of a CPU study of collision, and it did not survive the change of relation to capture. Pilot item 3 is retired rather than answered.

What the calibration pilot must produce

  1. That capture happens at all under real contact. Everything about the relation so far is arithmetic: the cell computes the target's motion and writes it into the command. Under ContactWorld the cell commands the bodies, steps physics, and reads where the object went. Whether a gripper or a pusher produces a capture rather than a collision in Isaac-Lift-Block-PSM-IK-Rel-Play-v0 is unmeasured, and it is the first thing that could end this design.
  2. The engagement rate under contact jitter, which sets n.
  3. normal_alignment under real contact. It is 1.0 by construction in every result above. A contact that pushes off-normal returns correct coefficients while arm D aims the wrong way, and a CPU study found this to be the dominant threat under realistic contact laws.
  4. Observation noise, so the gate's margin against it can be stated.

Answers, 2026-08-05. (1) Yes — 60 of 60. The scene produces a capture, and the item written as "the first thing that could end this design" did not end it. (2) 0.27 to 0.50, giving n ≈ 28–43. (3) Retired, not answered: inapplicable to a capture. (4) 0.00 mm/step = no noise model, so the margin cannot be stated from this scene.

The pilot was not asked to score the arms, and scoring them is what closed the paper. That is the honest sequence and it is recorded in that order.


Analysis, stopping, reporting


Closure — what the pilot returned, and what it licenses

The three conditions for freezing this document were:

  1. The Isaac calibration pilot above, for capture. — Run.
  2. Every PENDING filled, each with the measurement it came from. — Partly. Four measured, one unmeasurable here, one retired as inapplicable.
  3. The grading variable for T chosen and justified without reference to which arm it favours. — Never chosen.

Condition 3 was never reached, because condition 1 returned a result that makes T pointless to grade. The document is therefore closed, not frozen, and no cell run against it counts as confirmatory — which was true before and stays true.

The precondition that failed

The design rests on the mode operator being a partial explanation. In the prereg's own words: "Mode expansion partially helping is expected and acceptable; mode expansion matching the relation arm falsifies the contribution over Paper 002."

Under real contact the mode operator does not match the relation arm. It beats it, at ceiling, in every configuration tried:

Configuration arm C arm D SELF cells
block 1.000 0.200 0.167 60
needle, burst_off 4 0.957 0.174 0.348 23
needle, burst_off 25 (derived) 0.958 0.583 0.750 24

Why no n rescues this. H1 counts variants where success(D) clears ACHIEVABLE_THRESHOLD while success(SELF) sits under NEAR_ZERO_BAND, and H2 requires success(D) − success(C) to clear a positive C_MARGIN. Arm C is at 0.957–1.000 and arm D is below it everywhere; SELF is at 0.167–0.750 and above arm D in two configurations of three. These are not underpowered estimates near a boundary — they are the wrong sign at ceiling. Sampling more cells narrows intervals around the same ordering.

Why the obvious repair was tried and failed. The relation is made necessary by intermittency, and the arm's measured settling time (22 steps median, over 139 goal-pauses) exceeded the burst_off of 4, so the pause never began. The rule for fixing it — the pause must exceed the settling time plus min_ride_steps — was written before the settling time was measured, and it gave 25 steps. Run at 25 on the needle, whose 68-step carry has room for it, arm C moved from 0.957 to 0.958. The intermittency was restored and it changed nothing about whether a constant-velocity model suffices. Measurement

What this record does and does not support

It does not support — and no sentence in the paper may claim —

It does support

What would reopen it

Not a larger n, and not a threshold. A scene where the carrier can come to rest inside the time it holds the object. That is a manipulator property, not a parameter of this protocol: settling time 22 steps median against a 54-step carry is the whole obstruction. A carrier that stops promptly — a different arm, a conveyor, a second object driven directly — restores the intermittency the design needs, and this document's design would then apply unchanged.

Whoever reopens it should treat every number above as calibration and re-run the locked rule from the top.


Version history