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 — Physical validation roadmap v0.2 (DRAFT)

Status: design · confirmatory experiments/ promotion after sim Go gates
Sim (Track A): EXP-SURG-003 · confirmatory spec
Physical anchor (Track B): SO-101 · observability cell · private until freeze


Purpose

Validate whether failure-conditioned model adequacy and repair vs structural expansion survive real sensors — not to claim clinical deployment.

Working name: Physical Testbed for Failure-Driven World-Model Reconstruction

Not: “Open Recoverability Robot” (Paper 001 credential · not program center)


Research layers (program)

Layer Question Primary venue
L1 Mechanism Does L3 expansion beat L1 repair when a dynamics mode is missing? Simulation (EXP-SURG-003)
L2 Observability Can inadequacy be detected without privileged state? SO-101 anchor (Track B)
L3 Task value Does expansion improve success/safety in task-relevant settings? Precision arm · surgical-like · later

Ultimate question is embodied. Paper 002 v0.1 isolates mechanism in L1 before L2/L3 claims.


Parallel tracks (2026-07 rev.)

Track Share Action
A — Paper 002 sim 70–80% EXP-SURG-003 · Go gates · pre-reg · confirmatory
B — Physical anchor 20–30% SO-101 · logging · repeatability · failure-signal observability

Anchor purchase: justified as 2–3 year program instrument — not as Paper 002 confirmatory on day one.

EXP-REAL-001 A/B/C: runs after sim Go gates G1–G5 · hardware confirmatory protocol not yet public (promote after pre-reg freeze).


Program center (unchanged)

When the current model class cannot explain persistent failure, does prepared structural expansion beat parameter repair on novel related conditions — without nominal regression?

Recoverability = measurement window on Ep2 · not the program title.


Planned physical cell (private detail)

Candidate platform: SO-101 follower arm (LeRobot ecosystem) — low-cost, reproducible, not final clinical evidence.

First task (draft): planar push (preferred over grasp) · controlled friction / mass perturbations · same A/B/C arms as sim.

Full protocol: EXP-REAL-001 · promote to experiments/ after pre-reg freeze · not linked from this repo until then.


Go / No-Go gates (EXP-REAL-001 confirmatory — not anchor purchase)

Run full A/B/C protocol on hardware only if sim pilot + confirmatory show:

Gate Criterion (initial · tune at pre-reg)
G1 Mechanism C > B on Ep2 success · ≥15–20 pp · multi-step prediction ≥15% lower · static retention ≤5 pp drop
G2 Robustness Holds across ≥2 of: direction · onset · pose · drift magnitude
G3 Identifiability Oracle separates expert vs gate vs observation limits
G4 Physical contract Runs on joint pose · commanded action · EE pose · RGB object pose only (no privileged physics in claim path)
G5 Gate validity False expansion on noise/impulse ≤10%

No-Go → revise perturbation taxonomy · repair baseline · or task — do not run confirmatory on hardware.

Track B (anchor) may proceed regardless — assembly · logging · repeatability · raw perturbation trajectories only.


Sim–real contract (sketch)

Paper002Environment
├── IsaacAdapter      ← EXP-SURG-003
└── SO101Adapter      ← EXP-REAL-001 (after Go)

Shared: world model · repair/expansion arms · gate · metrics · episode protocol
Different: physics backend · sensors · command interface · reset


Evidence ladder (honest claims)

L1  Sim (Paper 002)        mechanism · L1 vs L3 · Go gates
        ↓
L2  SO-101 (Track B)       observability · sensor contract · residuals
        ↓ confirmatory after Go
L2b EXP-REAL-001           frozen A/B/C on real robot
        ↓ selective
L3  Precision / task       partial reproduction · safety · intervention
        ↓ later
    Clinical / surgical    NOT claimed from SO-101 alone

Public promotion rule

Tier Location
Frozen protocol + results research-os/experiments/physical_world_model/ (future)
BOM · purchase · assembly Private working records · not in research-os
Thinking · daily Private vault · not in research-os