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 — Description · world-model expansion v0.1

Program L0: When reality cannot be explained by the agent’s model class, how should an embodied system revise parameters and the architecture/composition of its world-modeling system?
Architecture note: paper002_wm_system_expansion_v0.1.md


Working title

When Parameter Update Is Not Enough: Structural World-Model Expansion from Unexplained Failures in Embodied Simulation


Central question

Can persistent, structured task failures provide evidence that the current dynamics model class is inadequate — and does selecting a prepared structural expansion operator improve prediction and control on novel related encounters beyond parameter-only repair, without nominal regression?

Paper 002 v0.1 tests one adequacy decision: L1 parameter repair insufficient → invoke L3 operator (add dynamics expert M₁ + gating) vs L1 vs no update.

Recoverability is a measurement window, not the program center.

Related work: paper002_related_work_v0.2.md · closest prior: TMoW · MuSix · Worldscape-MoE · LMC.


Conceptual frame

Parameter update Structural expansion
Failure meaning Wrong value in known state Missing variable, mode, or relation
Action Re-estimate position, friction, noise Add target_mode, dynamics mode, etc.
Test Residual shrinks under same representation Residual persists until representation changes

Minimal environment

True simulator: target_mode ∈ {static, drifting} (hidden from initial agent WM)

Initial agent representation:

state = target_position   (static target assumed)

Agent cannot represent drift → interprets as repeated position-estimation error.


Protocol (two-encounter)

Episode 1 — unexplained failure

Static assumption → target drifts → persistent prediction error → recovery fails

Evidence needed: parameter-only updates do not absorb structured residual.

Competing update arms (between Ep1 and Ep2)

Arm Expansion level Mechanism
No Update — frozen M₀
Parameter Update L1 θ on static dynamics only
Modular expansion L3 add expert M₁ (drift) + regime gating

Deferred arms (extension): monolithic fine-tune · diffusion future generator · M₂ external perturb.

Diffusion is an expansion operator candidate, not the whole world model — see taxonomy doc.

Change: start pose · drift direction · drift speed · mismatch onset
Hold fixed: hidden structure = drifting mode

Avoid pure memorization of Ep1.


Expansion gate (parsimony · model-adequacy test)

Do not expand on first failure. Expand only when evidence supports structural inadequacy (not distribution shift or uncertainty alone):

structured residual persists
after K parameter-update attempts
+ errors cluster by provisional mode / hidden condition
+ not absorbed by L1 repair on Ep1
→ candidate structural gap → invoke prepared L3 operator

Three-way distinction (method text): distribution shift · epistemic uncertainty · structural inadequacy (Paper 002 target).


Primary outcomes (Ep2 · vs Ep1 baseline)

Metric Role
Next-state prediction error Prediction
Mismatch detection latency Diagnosis
Correct response selection Decision
Task success / recoverability Outcome (window)
Repeated failure rate Utility
Static-condition regression Guardrail

Success pattern: Structural Expansion > Parameter Update > No Update on novel drift · no harm on static nominal.


Defensible claim (first paper)

Failure-conditioned model-adequacy testing: persistent structured task failures justify switching from parameter repair to a prepared model-class expansion operator, improving prediction and control on novel related encounters beyond parameter-only adaptation — with latent reorganization as observed mechanism, not primary claim.

Not claimed: adding experts alone (TMoW · Worldscape-MoE precedent) · latent representation learning SOTA · arbitrary z invention without behavior gain · unconstrained causal ontology invention.


Figure 1 (storyboard)

A. Reality — unobserved motion mode
B. Initial WM — position only
C. Failure — persistent mismatch
D. Parameter update vs structural expansion
E. Next encounter — mode predicted · action changed
F. Evaluation — improved prediction/recovery · no static regression

Layout: Primary results table (L1 repair failure on Ep1 → L3 advantage on Ep2 · Layer 3) · Reality | Agent belief · Latent before/after (Fig 4 · supporting · mechanistic secondary).


Program position

L0  Representation reconstruction under model-inadequate reality
Paper 001 (optional)  Recoverability @ S — credential · measurement window
Paper 002 (this)      Missing dynamic mode — detect · expand · validate
Next                  Missing relations / causal variables
Later                 Human-provided representation · surgical exceptions

Paper 001 not required as logical prerequisite.


Doc Path
Confirmatory spec paper002_confirmatory_spec_v0.1.md
Pre-reg skeleton paper002_prereg_wm_expansion_v0.1.md
Related work paper002_related_work_v0.2.md
Industry context paper002_industry_context_v0.1.md
Status status.md
Archive (mock→physics) archive/mock_to_physics/