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 002: Failure-Conditioned Model-Order Expansion

Status: confirmatory complete; venue-neutral submission package v1.1 built

Frozen preregistration: paper002-model-order-confirmatory-v1.0

Result artifact: 73a7e16

Headline Result

In a complete 400-cell Isaac Sim confirmatory grid, the gated constant-velocity model reduced H=10 prediction error by 10.806 mm and fixed-horizon final distance by 13.304 mm relative to repaired zero order. Both crossed-bootstrap 95% intervals cleared the preregistered 5 mm criterion. Static retention and all gate-specificity controls passed.

Start Here

Document Purpose
Manuscript v1.1 · LaTeX · PDF Submission-oriented review manuscript
Supplement v1.1 · LaTeX · PDF Conditions, validity audit, full tables, and statistical details
Overleaf main ZIP · supplement ZIP Root-main.tex upload packages for separate Overleaf projects
BibTeX Verified submission bibliography
Status Current phase, decisions, and next publication work
Frozen preregistration Fresh sample, endpoints, and locked decision rules
Confirmatory result Exact accounting, estimates, intervals, and provenance
Figures and CSV tables Reproducible manuscript panels and machine-readable tables
Process-isolated pilot Excluded engineering calibration
Model-order protocol v0.3 Isolation amendment that closed simulator carryover
Related work v0.2 Positioning and bibliography checklist
Physical validation roadmap Deferred hardware evidence track

Confirmatory Design

Static-first seed selection
  -> shared Ep1 persistent drift
  -> best allowed zero-order parameter repair
  -> persistent/directional/CV-explainable residual gate
  -> A frozen order 0 | B repaired order 0 | C gated order 1 | D oracle
  -> 10 fresh seeds x 10 fresh drift conditions x 4 arms
  -> static retention and gate controls

All arms use the same scripted Cartesian controller. The model's H=10 forecast is the controller target. Every seed-arm-condition cell runs in a fresh Isaac process.

Confirmatory Summary

Endpoint B repaired zero order C gated constant velocity C-B
H=10 prediction error 18.401 mm 7.595 mm -10.806 mm, 95% CI [-11.360, -10.331]
Fixed-horizon final distance 18.905 mm 5.601 mm -13.304 mm, 95% CI [-13.599, -12.982]
20 mm resolution rate 76% 100% +0.24, secondary
Static retention 10/10 10/10 non-inferior

Gate activation was 100/100 on persistent drift and 0/100 on each of static, observation-noise, and single-impulse controls. There were no missing cells, unexpected resets, or forbidden-region violations.

Reproduce Derived Materials

No simulator or GPU is required to recreate the paper figures and CSV tables:

python scripts/plot_paper002_model_order.py
python scripts/build_paper002_submission_tex.py
python scripts/build_paper002_overleaf_zip.py
python scripts/build_paper002_submission_pdf.py

The generated .tex files are standalone venue-neutral sources. From docs/paper002, a standard BibTeX build for the main manuscript is:

pdflatex paper002_manuscript_model_order_v1.1.tex
bibtex paper002_manuscript_model_order_v1.1
pdflatex paper002_manuscript_model_order_v1.1.tex
pdflatex paper002_manuscript_model_order_v1.1.tex

The supplement has no bibliography and requires two pdflatex passes. No TeX engine is bundled with this repository.

Upload To Overleaf

Use the ZIP packages rather than uploading the standalone .tex file from the project dashboard:

  1. In Overleaf, select New Project and then Upload Project.
  2. Upload paper002_overleaf_main_v1.1.zip for the manuscript.
  3. Confirm Compiler: pdfLaTeX and Main document: main.tex in Settings.
  4. Upload paper002_overleaf_supplement_v1.1.zip as a separate project when a separately compiled supplement is required.

Both archives place main.tex at the ZIP root. The manuscript archive also contains the BibTeX database and all five referenced PNG figures.

The Isaac experiment itself should run only on the documented VESSL image and frozen source/config. See the VESSL runbook.

Claim Boundary

Supported: in the specified Isaac target-drift family, structured failure after zero-order repair warrants a prepared velocity-state expansion that improves prediction-linked control without static regression.

Not supported: unconstrained model invention, arbitrary world-model expansion, tissue/contact validity, hardware transfer, clinical efficacy, or operating- room deployment.

Historical mock-to-physics documents remain under archive/mock_to_physics and must not be cited as Paper 002 evidence.