How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
Paper 002 manuscript v1.0
Status: complete confirmatory draft
Frozen design:
paper002-model-order-confirmatory-v1.0Confirmatory artifact:
73a7e16
Embodied agents commonly respond to prediction failure by updating parameters inside a fixed model class. This response is insufficient when the observed dynamics require state variables that the current class cannot represent. We study a restricted model-adequacy decision in Isaac Sim surgical reach: after persistent target drift remains unexplained by zero-order parameter repair, should the agent retain the repaired model or activate a prepared constant-velocity state expansion? A preregistered structural gate detects persistent, directional residuals that are better explained by the expanded class. We compare a frozen zero-order model, a repaired zero-order model, a gated constant-velocity model, and a velocity oracle. The confirmatory design uses the first 10 statically eligible seeds from 40 fresh candidates, 10 fresh balanced drift conditions, and a fresh Isaac process for each seed-arm-condition cell. All 400 primary cells and 20 static-retention cells passed execution-validity checks. Relative to repaired zero order, gated constant velocity reduced mean open-loop H=10 prediction error by 10.806 mm (crossed-bootstrap 95% CI [10.331, 11.360] mm improvement) and fixed-horizon final distance by 13.304 mm (95% CI [12.982, 13.599] mm improvement). It was better on both continuous endpoints in 100/100 paired seed-condition cells. The secondary 20 mm resolution rate was 100% versus 76%. Static retention was 10/10 for both models with identical 1.501 mm mean final distance. The gate fired on 100/100 persistent-drift cases and 0/100 static, observation-noise, and single-impulse controls. These results support a narrow claim: structured failure after parameter repair can justify a prepared increase in target-model order and improve prediction-linked control without nominal regression. They do not establish autonomous world-model invention or clinical validity.
Keywords: model adequacy, structural adaptation, world models, embodied control, model order, surgical simulation, preregistration
World models support prediction and action by representing how an environment changes. When their predictions fail, a natural first response is parameter repair: update a filter coefficient, fit a new rate, or fine-tune network weights without changing the set of represented state variables. This is a parsimonious and often correct response. It cannot succeed, however, when the data-generating process depends on a dynamic variable absent from the model.
This distinction separates parameter error from structural inadequacy. A zero-order target model can update its estimate of current position but cannot carry velocity as state. Under persistent drift, aggressive position updates reduce lag without yielding a correct multi-step forecast. A first-order model can represent the missing dynamic mode, but indiscriminate expansion risks complexity and false activation under noise or transient impulses. The central question is therefore not merely whether an expanded model can perform better. It is whether observed failure provides sufficient evidence to warrant that expansion.
We operationalize this question in an Isaac Sim surgical-reach task. A shared Episode 1 trace exposes a zero-order target model to persistent drift. Parameter repair searches the permitted zero-order coefficient. A locked adequacy gate then tests whether the residual is persistent, directionally consistent, and better explained by constant velocity. Episode 2 evaluates four target-model arms on unseen directions, speeds, delays, and durations while holding the Cartesian controller fixed. Static retention and matched negative controls constrain the expansion claim.
The confirmatory study was frozen before fresh-data execution. It contributes:
The contribution is deliberately bounded. We evaluate one missing variable, velocity, and one prepared operator, constant-velocity expansion. The study is an adequacy-test cell, not unconstrained architecture search or autonomous causal concept invention.

Figure 1. Static-first confirmatory protocol. After zero-order repair and the locked adequacy test, all four arms are evaluated on a complete fresh seed-condition grid. C versus B is the primary contrast.
Latent predictive world models learn compact transition representations for planning, policy learning, simulation, evaluation, and synthetic experience. Continual world-model methods primarily ask how to absorb new experience while retaining old dynamics. Modular approaches instead route among models or add experts. Test-Time Mixtures of World Models (TMoW), multi-scale mixtures such as MuSix, Worldscape-MoE, and Local Module Composition (LMC) establish that modularity and test-time structural adaptation can be useful.
Our question precedes the choice of update algorithm: when does failure after parameter repair count as evidence that the current class is inadequate? The comparison is therefore between no update, parameter repair within order zero, and a gated increase to order one. The gate is judged both by downstream prediction and control and by its refusal to activate on nominal, noise, and impulse controls. Closest modular work supplies important precedents for adding or routing components; our empirical wedge is the failure-conditioned repair versus expansion decision and its guardrails.
Related causal and compositional world-model work seeks invariant mechanisms, sparse transition changes, or symbolic repair. Those broader goals motivate the program but are not tested here. Our expansion operator and candidate variable are known in advance.
Experiments use Isaac Sim 4.1 with the ORBIT Surgical dual-arm reach task
Isaac-Reach-Dual-STAR-IK-Rel-Play-v0. The evaluated end effector follows a
target command in Cartesian space. The nominal target mode M0 is static. In M1,
the target begins persistent planar drift after a specified delay:
M0: x(t+1) = x(t)
M1: x(t+1) = x(t) + v, after the locked onset and delay
All arms share the same scripted Cartesian controller, gain 1.0, maximum action delta 0.08, and 20 mm resolution tolerance. The model forecast is installed as the policy target while the true simulator command is restored for evaluation at every step. Consequently, arm differences arise from target prediction and gate state rather than controller tuning.
The shared Episode 1 trace contains 40 steps of +x drift at 1.5 mm per step.
The first model class is zero order: it estimates target position but has no
velocity state. Four locked position-update candidates are fit on the common
trace (alpha = 0.25, 0.5, 0.75, 1.0), and the best held-out H=10 predictor is
selected. This is the permitted L1 parameter repair.
The prepared expanded class adds a velocity estimate and extrapolates position at H=10 under constant velocity. Its position and velocity coefficients are selected from the same evidence without using confirmatory Episode 2 outcomes.
The gate uses an eight-step window and requires at least four target deltas. It fires only when all locked conditions hold:
The same rule is evaluated on persistent drift and on three controls: static target, 0.2 mm observation noise, and a single impulse followed by rest.
| Arm | Target model | Role |
|---|---|---|
| A | Frozen zero order, position alpha 0.5 | Secondary baseline |
| B | Repaired zero order, position alpha 1.0 | Primary comparator |
| C | Gated constant velocity, position/velocity alpha 1.0 | Primary intervention |
| D | True injected velocity | Diagnostic oracle |
A and B cannot represent velocity. C remains parsimonious until the adequacy gate fires, after which it carries an estimated velocity state. D supplies the true velocity from onset and bounds model-side error; it is excluded from the primary comparison.
Candidate seeds were fixed as integers 300 through 339. Static control was run before treatment. The first 10 statically eligible seeds in numeric order were selected, preventing treatment performance from influencing inclusion.
Episode 2 uses 10 conditions, C01-C10, with no exact pilot-condition reuse. The set balances positive and negative x/y axes and all four planar diagonals. Speeds range from 1.7 to 2.0 mm per step, delays from one to five steps, and durations from 20 to 23 steps. The full primary grid is 10 selected seeds by 10 conditions by four arms, or 400 cells. Static retention adds 20 B/C cells.
An engineering pilot initially batched conditions inside one Isaac process. Although the same reset seed was supplied, preceding trajectories affected later reset outcomes. That pilot version was rejected. The corrected pilot and confirmatory study launch a fresh Isaac process for every seed-arm-condition cell.
The run is invalid if the grid is incomplete, a prefix is not statically ready, paired branch-start end-effector states differ by more than 1 mm, commands differ by more than 1 micrometer, branch-start target distance exceeds 21 mm, prediction windows are missing, drift exposure is incomplete, the simulator resets unexpectedly, or a forbidden region is entered. No failed cell is silently removed.
The primary estimand is intention-to-treat over all valid selected-seed and condition cells.
The 20 mm binary resolution rate is secondary because the process-isolated pilot showed threshold saturation. Diagnostic gates require at least 80% success for C and D, reproduction of the Episode 1 parameter lock, and a valid execution grid. Confirmatory support requires all criteria to pass.
For H1 and H2, seeds and conditions are independently resampled with replacement in a crossed bootstrap with 10,000 draws and RNG seed 20260730. We report the mean paired C-minus-B contrast, two-sided percentile 95% interval, and the fraction of the 100 paired cells favoring C. Gate rates use Wilson 95% intervals.
The design, conditions, selection rule, thresholds, estimands, and decision
criteria were frozen under immutable tag
paper002-model-order-confirmatory-v1.0 before seeds 300-339 were run. Pilot
seeds and observations are excluded from every confirmatory estimate. There
was no interim efficacy analysis or optional stopping.
Thirty-four of 40 fixed candidates passed static control. The first 10 eligible seeds were 300, 301, 302, 303, 304, 305, 307, 308, 310, and 311. All 400 primary cells and all 20 retention cells completed. The maximum paired branch-start end-effector and command gaps were both zero. Maximum branch-start target distance was 15.396 mm, below the 21 mm limit. Every prefix was ready, every prediction window was present, and every cell received full drift exposure. There were no unexpected resets or forbidden-region violations. The run was therefore valid.
The zero-order repair selected alpha 1.0 but retained 15.000 mm held-out H=10 prediction error. Constant velocity selected position and velocity alpha 1.0 and reduced the same diagnostic error to numerical zero. The gate fired with active fraction 1.0, directional consistency 1.0, and constant-velocity fit improvement 1.0. The preregistered parameter and gate locks were reproduced.
| Arm | n | H=10 prediction error | Final distance | Resolution rate |
|---|---|---|---|---|
| A frozen zero-order | 100 | 20.081 mm | 20.388 mm | 39% |
| B repaired zero-order | 100 | 18.401 mm | 18.905 mm | 76% |
| C gated constant velocity | 100 | 7.595 mm | 5.601 mm | 100% |
| D oracle velocity | 100 | 0.000 mm | 4.087 mm | 100% |
For H1, C minus B prediction error was -10.806 mm (95% CI [-11.360, -10.331] mm). For H2, C minus B final distance was -13.304 mm (95% CI [-13.599, -12.982] mm). Both interval upper endpoints exceed the locked 5 mm improvement requirement in the favorable direction. C had lower prediction error and lower final distance in all 100 paired seed-condition cells.
The secondary resolution-rate difference was +0.24 (95% CI [+0.08, +0.45]): C resolved 100/100 cells and B resolved 76/100. A resolved 39/100. D resolved 100/100 and had zero forecast error. C's remaining 1.514 mm mean final-distance gap to D is consistent with the finite evidence window before its gate fires.

Figure 2. Arm-level confirmatory outcomes. Dots show condition means for the continuous endpoints; success bars include Wilson intervals. The 20 mm line is shown for context but is not the H2 primary criterion.
Condition-level C-minus-B prediction differences ranged from -12.462 to -9.998 mm. Final-distance differences ranged from -13.888 to -12.716 mm. C resolved all 10 seeds in every condition, while B ranged from 5/10 to 10/10. No condition reversed either continuous primary effect.

Figure 3. Mean C-minus-B effects by fresh Episode 2 condition. Every condition lies beyond the -5 mm criterion for both endpoints; the formal decision uses the crossed-bootstrap interval over all cells.
In the process-isolated seed 300/C04 branch, B and C begin from the same state and initially share the same zero-order prediction. C's gate fires after enough directional evidence accumulates. Its H=10 prediction error then falls from 20 mm to numerical zero, and end-effector distance decreases to 6.4 mm; B remains at 20 mm forecast error and finishes at 20.2 mm. The oracle fires earlier and finishes at 4.5 mm. This trace is illustrative; all inferential results use the complete crossed sample.

Figure 4. Representative seed 300/C04 branch. The vertical line marks the first C gate activation. B, C, and D are process-isolated but start from exactly matched reset and branch states.
B and C each retained 10/10 static successes and identical 1.501 mm mean final distance. The paired success-rate difference was 0.00 with 95% interval [0.00, 0.00], passing the -5 percentage-point non-inferiority margin.
The gate fired in 100/100 persistent-drift evaluations (Wilson 95% CI [96.30%, 100%]) and 0/100 evaluations for each of static target, observation noise, and single impulse (each Wilson upper 95% bound 3.70%). H3 and H4 passed.

Figure 5. The gate separates persistent structured drift from matched negative controls, while the expanded arm is unchanged from repaired zero order on static retention.
Every conjunctive decision field passed: execution validity, Episode 1 parameter lock, oracle behavior, C success floor, H1 prediction interval, H2 final-distance interval, H3 retention, H4 drift activation, and H4 control specificity. The preregistered confirmatory result is therefore positive.
The experiment isolates a case in which a stronger parameter update is not a substitute for the missing state variable. Moving alpha from 0.5 to 1.0 helps the zero-order model follow current position, improving prediction by 1.680 mm and final distance by 1.483 mm relative to frozen zero order, but it cannot represent velocity. The gated order-one model produces a much larger change: 10.806 mm lower forecast error and 13.304 mm lower final distance than the best allowed zero-order repair.
This prediction-to-control alignment matters. The expanded model does not only fit a held-out trace; its H=10 forecast changes the target presented to an otherwise fixed controller. The continuous control outcome improves in every paired cell. The oracle shows that remaining C error comes from evidence collection and gate latency rather than the controller's inability to exploit a velocity forecast.
The negative controls delimit the trigger. A high residual by itself would not justify expansion because observation noise and a transient impulse can also create error. Requiring sustained activity, directional consistency, and superior alternative-class fit yielded 100% activation on persistent drift and 0% on each control. Static retention further shows that the added state was not forced into nominal behavior.
The result supports a conditional statement: for the tested persistent-drift family, failure evidence that survives zero-order repair and is predictably explained by constant velocity warrants activating that prepared order-one model. It does not show that every persistent failure is structural or that a system can discover the correct variable without a candidate class.
The next scientific step is not to broaden the present claim by wording. It is to test whether the same repair-versus-expansion decision survives new dynamic families, richer observations, learned controllers, and hardware measurement.
In a preregistered, process-isolated embodied-simulation experiment, a repaired zero-order target model remained structurally unable to forecast persistent drift. Activating a gated velocity-state expansion reduced H=10 prediction error and fixed-horizon final distance by more than the locked 5 mm margins, with consistent benefit across every paired cell, no static regression, and no false activation on three negative controls. Failure after parameter repair can therefore support a restricted model-adequacy decision when the alternative class is prepared and the claim is guarded by prediction, control, retention, and specificity tests.
paper002_model_order_confirmatory_prereg_v1.0.mdRESULTS.mdmodel_order_confirmatory_v1.0.jsonisaac_model_order_confirmatory_v1.0plot_paper002_model_order.pyfigures/manifest.jsonAll raw artifacts carry SHA-256 checksums. The figure manifest hashes both source JSON files and every derived panel/table. Pilot results are retained in the repository but excluded from confirmatory estimates.
This draft uses linked reference anchors pending venue-specific bibliography formatting.
The fuller positioning matrix and submission bibliography checklist are in
paper002_related_work_v0.2.md.