How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
Status: FROZEN BEFORE CONFIRMATORY DATA Freeze identity: immutable tag
paper002-model-order-confirmatory-v1.0Config:model_order_confirmatory_v1.0.jsonPilot data: excluded from every confirmatory estimate
When persistent target drift produces structured residuals that a zero-order target model cannot absorb, does a gated constant-velocity state expansion improve true open-loop prediction and fixed-horizon control without degrading static behavior or firing on matched negative controls?
The claim is limited to the specified Isaac surgical-reach task and target dynamics. It is not a claim about tissue interaction or clinical efficacy.
| 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 only |
All arms use the same Cartesian scripted controller. The model prediction is the policy target; the simulator evaluation target is restored before each step. Prediction horizon is 10 steps.
Each seed-arm-condition cell runs in a fresh Isaac process. Missing cells, failed prefixes, cross-policy branch-start gaps above 1 mm, command gaps above 1 micrometer, unexpected resets, incomplete drift exposure, or forbidden violations make the run invalid.
The 10 C01-C10 conditions in the config reuse no exact pilot condition. They balance positive and negative x/y axes and all four planar diagonals, with locked speeds from 1.7 to 2.0 mm per step, delays from 1 to 5 steps, and durations from 20 to 23 steps. The Ep1 adaptation trace remains +x at 1.5 mm per step. No condition may be changed after the freeze tag.
For C minus B, the upper endpoint of the 95% crossed-bootstrap confidence
interval for mean true open-loop H=10 prediction error must be at most
-0.005 m. Thus the entire interval must support at least 5 mm improvement.
For C minus B, the upper endpoint of the 95% crossed-bootstrap confidence
interval for mean final end-effector distance after full drift exposure must
be at most -0.005 m.
The 20 mm binary success rate is secondary because the pilot showed threshold saturation. It is reported with its crossed-bootstrap interval but cannot replace H2.
C static success must be at least 95%, and the lower endpoint of the paired seed-bootstrap interval for C minus B static success must exceed the locked non-inferiority margin of -5 percentage points.
Confirmatory support requires every H1-H4 and diagnostic gate to pass. This conjunctive rule is fixed before data and no successful subset is sufficient.
The primary estimand includes every valid crossed selected-seed/condition cell regardless of oracle outcome. For each endpoint, compute C minus B within every cell and resample seeds and conditions independently with replacement for 10,000 crossed-bootstrap draws. The RNG seed is 20260730. Report the mean, two-sided percentile 95% interval, and the fraction of pairs favoring C.
Oracle-feasible cells may be summarized as a descriptive sensitivity analysis only. They cannot replace the intention-to-treat primary result. A and D are secondary/diagnostic and are excluded from the primary contrast.
The structural gate remains locked at: window 8, minimum 4 deltas, speed floor 0.5 mm/step, active fraction 0.75, directional consistency 0.90, and constant- velocity error improvement 0.50. Noise sigma is 0.2 mm. Report Wilson 95% intervals for persistent drift, static, noise, and impulse controls.
Forbidden-region violations and simulator resets are always reported. No episode is silently removed. A technical failure requires a new versioned protocol; it cannot be repaired after inspecting confirmatory outcomes.
There is no interim efficacy analysis and no optional stopping. Run every locked cell or declare the run invalid. Preserve the config, selection manifest, orchestration manifest, combined records, preconditions, trajectories, checksums, source commit, and frozen tag.
Results will be reported even if one or all confirmatory gates fail.