How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
Updated: 2026-07-31 · Independent personal research
When reality cannot be explained by the agent’s current model class, how should a system revise not only its parameters, but the architecture and composition of its world-modeling system?
Recoverability = measurement window · not the program center.
Embodiment = current testbed · not part of L0. The studies run in embodied simulation because it makes mismatch, timing, and recovery observable — the question itself is stated without it. See program · domain generality.
Full architecture · expansion levels · taxonomy: paper002_wm_system_expansion_v0.1.md
| Layer | Content | Changes |
|---|---|---|
| L0 · Program | The question above | Does not change |
| Testbed | Physical AI · embodied simulation (now) | Over years; other domains admissible under the application policy |
| Projects | Paper 001 · 002 · 003 · Mismatch Lab | Per project |
Naming by audience is intentional and does not indicate a change of question — the program name, the testbed name, and a project title describe the same work at three different levels.
| Label | Experiment ID | Role | Status |
|---|---|---|---|
| Paper 001 | EXP-SURG-001 | Recoverability profiles @ fixed S (optional credential) | ✅ Tier C complete |
| Study 002 | EXP-SURG-002 | Dream curriculum pilot (Tier B) | ✅ archived |
| Paper 002 (mock→physics) | EXP-SURG-003 | Mock→Isaac selection | ❌ archived · archive |
| Label | Experiment ID | Role | Status |
|---|---|---|---|
| Paper 002 | EXP-SURG-003 | Structural WM expansion · hidden mode minimal cell | ✅ confirmatory complete · submission v1.1 |
| Paper 003 | TBD | Missing causal relation · capability-threshold expansion (not just error reduction) | 🔄 design v0.1 · relation module + commitment task built · docs |
| Paper 004+ | TBD | Human representation · surgical exceptions | design |
| Component | In Paper 003 | Status |
|---|---|---|
| Missing relation between entities (vs Paper 002's mode) | ✅ | relation module implemented |
| Relation-adequacy gate · must stay silent on Paper 002's drift | ✅ | separates 10/10 vs 0/10 on the CPU proxy |
| Four arms: none / parameter / mode / relation | ✅ | arm C is the discriminating control |
| Capability threshold crossing as primary endpoint | ✅ | constructed on a commitment-point task |
| Arm D estimating the reference pattern online | — | blocking before any confirmatory run |
| Isaac implementation · prereg | — | not started |
| Preference / user modelling | — | out of scope (different question — see application policy) |
Paper 001 not required as logical prerequisite for Paper 002.
| Component | In Paper 002 | Deferred |
|---|---|---|
Hidden target_mode (static / drift) |
✅ | — |
| Three arms: none / parameter / structural | ✅ | — |
| Ep1 fail · Ep2 novel drift | ✅ | — |
| Expansion gate (parsimony) | ✅ | — |
| Reality | Belief figures | ✅ | — |
| Missing causal variables | — | Paper 003 |
| Human-in-the-loop expansion | — | Paper 004+ |
| Mock→physics curriculum | — | archived |
| Term | Means | Does not mean |
|---|---|---|
docs/stage2/ |
Study 002 public docs | Paper 002 body |
| Parameter update | Re-estimate within fixed representation | Structural expansion |
| Structural expansion | Add variable · mode · relation | Mock rank transfer |
| Mode gap (Paper 002) | Regime the model cannot represent | Relation gap |
| Relation gap (Paper 003) | Dependency on a second entity the model treats as independent | Mode gap · preference |
| Capability crossing | Task variants going from ~0% to achievable | Lower mean prediction error |
| Commitment point | An irreversible action; a wrong prediction fails outright | Any hard task |
| Physical AI | Current testbed | The program's identity |
When failures are structurally unexplained, expand representation · validate on novel encounters · recoverability measures whether it helped.