How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
Status 2026-08-05 — the design is CLOSED, and the result is negative
Paper 002's mode operator lands 0.957 to 1.000 of physical cells; the relational arm lands 0.174 to 0.583 and never beats the single-entity arm. 107 cells, three configurations. The preregistration is closed and no confirmatory run will follow, because the ordering is at ceiling and no
nreverses it.Why. The relation is made necessary by intermittency, and a real arm cannot pause: 22 steps at the median to come to rest, against 54 steps of grip on the block. Two candidate repairs were derived in advance and tried — a better-held object (the needle, 68 steps) and a pause long enough to be real (25 steps, from the measured settling time). Arm C moved 0.957 → 0.958. The measurement
How much of that account survived being tested. Settling time was then injected into the CPU carrier and swept, under four predictions fixed in writing. All four failed. What replaced them is better: arm C's score tracks the residual velocity ripple at r = −0.940, which is a measured mechanism rather than an inferred one. But arm D does not collapse under a smoothed carrier — 0.450 on CPU at the arm's own settling time, still winning the paired comparison 27 : 0 — so settling explains the mode operator's return and not the relational arm's physical failure. That gap is open and named. Rule · Result
The CPU result stands as what it always was — a comparison of arms under injected coupling, where the carrier is a scripted point that stops on command. It remains true of scripted carriers. It does not transfer to this manipulator.
What is not claimed. No confirmatory cell was ever run, so H2 was not rejected at α = 0.05; H1 was never testable, because the variant set was never constructed; H3 and H4 are structurally CPU-only — every control works by overriding the target's motion, and under contact physics decides it.
The experimental line is closed. No further runs are planned. What remains unexplained is recorded as a limitation in the manuscript, not as pending work.
Read this file first — the folder is not safe to read in alphabetical order, and several documents carry corrections in place.
Program context: builds on Paper 002 (missing dynamic mode); this is the next taxonomy cell, a missing causal relation — and the finding is that in this scene the cell is empty: the smallest sufficient change remains a mode expansion.
When repeated failures reveal that the agent's world model is missing not a mode but a relation between entities, does adding a prepared relation-module expansion open task capability that neither parameter repair nor a single mode-expert reaches — measured as growth of the achievable-task space, not only reduced prediction error?
paper003_rq_v0.1.md ·
paper003_description_v0.1.md
The question is unchanged. What changed is what "relation" means in the implementation. Three candidates were measured against each other, and the decision turned on a single-entity control — a model that sees no second body and learns the burst pattern of the target's own trajectory.
| Relation | Relation necessary | Effect exceeds tolerance |
|---|---|---|
| Collision — struck and released (what the Isaac line ran all day) | yes | no |
| Carriage — rides the reference throughout | no | yes |
| Capture — reference arrives at a still target, then carries it | yes | yes |
Collision cannot clear the placement tolerance. The push moves the target away, which reduces penetration, which reduces the push, so displacement per contact sits at the order of the interaction radius — below the 20 mm criterion. Measured five independent ways (the ceiling).
Carriage clears it easily and makes the relation unnecessary: the single-entity model matched the relational arm exactly, so H2 fails.
Capture is the design. Before the arrival the target is perfectly still, so its own history says nothing; afterwards it rides, so the effect accumulates without bound.
| Document | What it settles |
|---|---|
paper003_manuscript_negative_v1.0.md |
The paper, complete. Four candidate relations, the requirement each fails, and how much of the mechanism survived being tested |
paper003_manuscript_negative_v1.0.pdf |
The compiled paper, 11 pages. Still carries the draft banner: the bibliography is unverified and no confirmatory run was made |
paper003_overleaf_v1.0.zip |
The Overleaf project — .tex + .bib + figures/, compiles as-is. Set \draftnotefalse and verify the bibliography before submission |
figures/ |
Four panels, regenerated by scripts/plot_paper003_negative.py; --check verifies every transcribed number against its source |
paper003_settling_sweep_prereg_v1.0.md |
Four predictions, four failures. The rule that turned §4.3 from an assertion into a measurement, and the outcome recorded beside it |
paper003_related_work_v0.2.md |
Positioning for the negative result: shortcut learning, benchmark design, and the reality gap in actuator dynamics |
paper003_capture_design_v0.1.md |
The relation. Why capture, and why carriage was recommended and then rejected |
paper003_prereg_v1.0.md |
The preregistration, CLOSED. Locked in design; the pilot ran and the design's precondition failed. Read its closure section for what may and may not be reported |
paper003_self_arm_prereg_v1.0.md |
H2's decisive comparison, and its rule fixed before the arm existed |
paper003_onset_is_not_predictable_v0.1.md |
What the paper may not claim, and why the three-step band is the right answer |
paper003_rendezvous_v0.1.md |
Why a grasp needs an arrival, not a fly-by |
paper003_capture_pilot_runbook_v0.1.md |
How to run the Isaac calibration pilot, and what to read first |
paper003_derived_from_physics_v0.1.md |
The two numbers the scene decides, and the rule for reading them off it |
paper003_carry_is_not_slip_v0.1.md |
Three positions on what carrying is, and why only the gate's test was wrong |
paper003_cv_gain_horizon_v0.1.md |
The gate statistic was measuring the horizon as well as the motion |
paper003_where_collapse_is_defended_v0.1.md |
Why the collapse threat belongs to H2 and not to the gate |
paper003_servo_encounter_v0.1.md · paper003_rendezvous_v0.1.md |
Why the encounter arrives before it carries |
paper003_displacement_ceiling_v0.1.md |
Why collision cannot work |
paper003_branch_b_scene_v0.1.md |
The Isaac scene with a rigid object, and where 20 mm comes from |
paper003_related_work_v0.1.md · paper003_lit_positioning_v0.1.md |
Positioning |
| Document | Status |
|---|---|
paper003_sliding_problem_v0.1.md |
Reversed the same day. Concluded the gate collapses under real contact; it does not |
paper003_two_body_encounter_v0.1.md |
Its tolerance argument is wrong (header says so). The machinery is real; the collision coupling under it is superseded |
paper003_encounter_does_not_scale_v0.1.md |
The scaling diagnosis holds; its proposed fix was overtaken by the change of relation |
paper003_real_contact_design_v0.1.md |
Its Branch A/B question is settled — the scene exists |
| Document | Problem |
|---|---|
paper003_prereg_draft_v0.1.md |
Superseded entirely by paper003_prereg_v1.0.md. Collision throughout; provenance only, not to be cited for any design decision |
paper003_commitment_task_v0.1.md |
Describes carriage — "a bread slice carried by a tray" — and its arm-B band came from that rejected model |
paper003_pilot_runbook_v0.1.md |
Targets the reach task; the scene is now the lift task |
The scene. Isaac-Lift-Block-PSM-IK-Rel-Play-v0 reports
rigid_objects: ['object'] and command object_pose. The bootstrap used to
delete it as "incompatible"; the incompatibility was two lines setting a debug
marker's scale.
The placement tolerance: 20 mm, on two independent grounds — the task
family's own mdp/terminations.py threshold, and the block's half-height.
The gate, and it now has two forms of positive evidence. The claim that
one threshold set covered both relations was wrong, and corrected
2026-08-04: with an off-by-one fixed in capture_displacement, the
proximity contrast abstains on every capture cell — a carrier that never
leaves supplies no far field. The 1.00 previously recorded came from that
defect, which threw a captured target one body-step past its carrier so the
pauses sorted into the far field.
One gate still covers both relations, by admitting carriage as a second
form of positive evidence: the target's displacement is a body's
displacement, on ≥ 0.80 of moving steps with a run of ≥ 3. Collision still
enters through the contrast; capture enters here; every other clause applies
unchanged. The constant-velocity ceiling is what keeps drift out, which
agrees with its body on 0.71 of steps and would otherwise pass.
The measurement
The gate under real contact. Fires on every Isaac trace containing a
strike, with cv_gain between −0.20 and −0.75 — a constant-velocity model is
worse than zero-order, so Paper 002's operator does not absorb the residual.
The cell loop. wm_expansion/cell.py runs the same code the GPU runs with
the simulator behind one callback. It has caught six defects before GPU time.
Connect capture to the cell loop and the encounter. Done.
CellSpec(coupling="capture") drives a cell, paired with the burst
schedule — a body that arrives and carries the target off has no reason to
withdraw. Under probe the carried target is dragged back with it.
The commit window relative to the capture. Done. Within one
dispense-length of a body's arrival, on either side, anchored on
contact_arrivals. Fixed on the structure of the action: the dispense takes
dispense_latency steps, so a commit further out than that either completes
before anything has happened or measures a regime the arrival no longer
governs. Symmetric because there is no reason to prefer a side. What it
removes is an artefact, not a preference — under capture the eligibility
screen admits every step after the arrival, so the commit distribution was
being set by episode_steps and sat almost entirely in the riding tail where
a constant-velocity model absorbs the motion.
Score the arms under capture in a cell. Done. Under capture +
burst, where arm D can act it is right 0.78 of the time against 0.06
where it declines; by commit offset the relation pays 0.53 → 0.71 against
0.00 for parameter repair and ≤ 0.07 for the mode operator. All four
controls and the collision cells are unchanged.
The measurement
Decide how the gate admits a capture. Done — carriage, as a second
form of positive evidence in the same gate. See Settled, above.
The overlap between the commit window and arm D's readiness. Not a
defect, and settled. Arm D cannot act before +4 because there is nothing
to act on before the capture: 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 contact,
a capturing approach and a non-capturing one are the same observation. An arm
that declined there was correct.
Why
Measure the single-entity arm under capture + burst. Done. H2 stands on
CPU and fails under contact — see the status block. Below is the CPU result
as recorded, unedited.
Preregistered rule locked before the arm was implemented; 200 paired cells; arm D 0.650, SELF 0.000, 130 discordant pairs all in D's favour, one-sided exact McNemar p = 7.3 × 10⁻⁴⁰, margin +0.650. SELF acted on 0.675 of cells and was not broken — it holds a median of 4 steps of its own motion against a 14-step cycle and extrapolates through a pause it cannot see. Rule · Result
The limitation that came with it is now measured too. The protection is
bounded in time, and the bound is +30 — a little over two burst cycles
after the arrival, on fresh seeds. The protocol band is [−6, +6] and SELF
scores ≤ 0.02 inside it, so the catch-up sits five times further out than any
commit the protocol makes. What stops arm D first is the gate: cv_gain
climbs as the carry lengthens and arm D declines exactly where it crosses the
ceiling, so the clause written for drift lands unprompted on the boundary
of H2's validity. Past it the sufficient model is a single-entity periodic
one — neither B, nor C, nor D, and proposed by neither paper.
The bound
Redraw the two-body encounter for capture. 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 to apply to. A
decoy repair was designed and rejected for the same reason as item 5: its
strong form requires predicting the onset, which would drive arm D to zero on
static and noise and fail H4, and its weak form moves the usable band by
a step or two. The two-body encounter stays with collision.
Why
Rewrite the preregistration, which assumes collision throughout.
Done — paper003_prereg_v1.0.md, which
supersedes the draft entirely. Locked in design: arms including SELF, the
commit window, the confirmatory test, the sizing rule, the scope limits,
and a list of tuning that will not be done. Open in numbers, all marked
PENDING, because none of them may be inherited from injected coupling.
An Isaac calibration pilot for capture. The first thing that could end
this design. It does not: capture happens under real contact. Servoed
onto the block, the gripper closes with 0.25 mm of disturbance and the object
then rides for 10 consecutive steps at 0.93 agreement — against the gate's
own requirement of 3 steps at 0.80. But the capture radius is under 1 mm,
while the arm's steady-state error following a moving script is 7.3 mm, so
a script-following encounter cannot produce a capture here: it has to
servo to contact first, then carry.
The measurement
Three hypotheses died on the way, all of them explanations for an artefact:
the environment was resetting mid-run and the teleport was being read as a
36 mm ejection. Engagement, normal_alignment and the observation noise are
still unmeasured.
Completed 2026-08-05, through the protocol. The capture radius above was read off a sweep grid I had chosen — a measurement of my own settings — and is corrected to 2.5 mm by the rule "the largest separation at which capture holds in a majority". Everything the pilot was asked for:
| Asked for | Returned |
|---|---|
| capture happens at all | yes, 60/60 |
| engagement under contact | 0.27 → n ≈ 43 |
normal_alignment |
retired — it measures displacement against the contact normal, which a carried target does not move along |
| observation noise | 0.00 mm/step = no noise model, not low noise |
And the derived scene parameters: capture radius 2.5 mm, carry duration
54 steps, dispense_latency 9, arm D when it acts 0.375 against
0.78 on CPU.
Then the confirmatory sample. Not run, and will not be. The pilot
was not asked to score the arms; scoring them closed the paper. Arm C at
0.957–1.000 makes H2's success(D) − success(C) > C_MARGIN unreachable at
any n, and SELF above arm D makes the crossing condition unreachable too.
Closure
Not a larger sample and not a threshold: a carrier that can come to rest inside the time it holds the object. Settling time 22 steps against a 54-step carry is the entire obstruction, and it is a property of the manipulator, not of this protocol. A different arm, a conveyor, or a second object driven directly would restore the intermittency the design needs, and the locked design would then apply unchanged.
Isaac records from 2026-08-04 were written on the VESSL pod and never pushed —
the sweeps under results/paper003_probe_sweep, results/gate_v*,
results/lift_*. Their numbers are recorded in the documents above and under
exp_surg_004_relation_expansion/results/.
A rerun should regenerate them rather than hunt for the files, since the
encounter has changed since.
| Module | Role |
|---|---|
wm_expansion/relation_dynamics.py |
Couplings, the gate, the estimators, capture |
paper003_capture_arms.py |
Scores every arm under each relation, in the commit window |
wm_expansion/cell.py |
One commitment cell, simulator behind a callback |
wm_expansion/encounter.py |
Encounter geometry and schedules |
wm_expansion/stopping.py |
How long a struck object keeps moving |
wm_expansion/commitment_episode.py |
The episode driver and the arms |
orbit_lift_relation_cell.py |
Isaac adapter for the lift scene |
orbit_lift_stopping_probe.py |
Strike-and-measure probe |
| paper003_self_arm.py | The preregistered SELF arm test, with no flag to relax the rule |
| paper003_self_arm_bound.py | Off-protocol probe: where the single-entity arm catches up |
| wm_expansion/capture_verdict.py | Was that trace a capture, a collision, or nothing |
| orbit_lift_grasp_probe.py | What grasp the scene supports, and at what separation |
232 tests, all CPU.
General causal discovery · relation invention outside the prepared operator · capability emergence outside the tested task family · clinical or hardware deployment.