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 — Pre-registration · WM structural expansion v0.1 (DRAFT)

Superseded before confirmatory data by model-order confirmatory preregistration v1.0.

Status: design · not frozen · supersedes mock→physics pre-reg v0.3
Confirmatory spec: paper002_confirmatory_spec_v0.1.md
Analysis: paper002_analysis_plan_v0.3.md


0. Primary RQ

When a fixed dynamics model cannot explain persistent drift failure with parameter update alone, does adding a drift dynamics expert (L3) improve held-out drift prediction and Ep2 control, without nominal static regression?


1. Hypotheses

H1 — Prediction

L3 modular expansion shows lower held-out drift trajectory prediction error than L1 parameter update on Ep2 (multi-step horizon H=10).

H2 — Behavior

L3 shows higher Ep2 task success (or recoverability) than L1 on held-out drift conditions.

H3 — Retention

L3 does not degrade nominal static success beyond pre-specified margin δ (non-inferiority).

H4 — Gate validity

Rule-based expansion gate activates on persistent drift (M1) · rarely on observation noise (N1) · single impulse (N2) · nominal static (M0).


2. Design

Factor Levels
Update arm A No update · B Parameter (L1) · C Modular expansion (L3) · D Oracle (diagnostic)
True mode M0 static · M1 drift
Protocol Phase 0 pretrain → Ep1 → K repairs → gate → intervention → Ep2 → static retention

Hidden from initial WM: target_mode · drift velocity
L3 adds: F1 drift expert + gate G (not latent scalar alone)

Controller: MPC / model-based trajectory selection (v0.1).


3. Arms

Arm Intervention
A W_after = W0
B Fine-tune F0 only · θ0 → θ0'
C Add F1 + G · F0 frozen or low LR
D Oracle mode label to gate · upper bound only

Primary contrast: C vs B · equal data budget across arms.


4. Protocol phases

Phase Content
0 Pretrain W0 on M0 static only
1 Ep1 drift exposure · log full trajectory
2 K parameter repair attempts on F0
3 Rule-based expansion gate
4 Arm intervention
5 Ep2 novel drift (≥2 params differ from Ep1)
6 Static retention (M0)

Ep1 evidence identical across arms before split.


5. Expansion gate (rule-based · pre-specified)

Gate = 1 iff ALL:
  mean(residual_after_repair) > τ_error
  K repair attempts completed without absorbing residual
  residual_autocorrelation > τ_a
  ΔNLL = NLL(F0_repaired) - NLL(F1_candidate) > τ_nll   [held-out Ep1 slice]

Structural inadequacy · not distribution shift or epistemic uncertainty alone.

Latent cluster metrics do not trigger gate (Appendix A).


6. Ep2 novelty

≥2 differ from Ep1 among: initial target · drift direction · drift speed · drift onset · robot initial joints.
Same regime: M1 drift.


7. Gate negative controls (H4)

ID Condition
N1 ↑ target observation noise · M0 static
N2 Single target impulse then stop
— Nominal M0 · persistent M1

8. Outcomes

Presentation: Ep1 L1 failure → Ep2 O1/O2 → static retention → H4 → latent (supporting).

ID Metric Role
O1 Multi-step prediction error Ep2 · H=10 Primary
O2 Ep2 task success Primary
O6 Static retention success H3 · non-inferiority
O9 Gate activation rate by condition H4
O7 Ep1 residual after K L1 repairs H1 diagnostic
O2b Recoverability composite Secondary (optional)
O3–O5 Latency · tracking · safety · effort Secondary
O8 Latent / expert / residual mechanistic Exploratory

9. Sample size

Tier Design In confirmatory?
Engineering pilot 3 arms × 5 seeds × 10 Ep2 No
Confirmatory 3 arms × 10 seeds × 30 Ep2 Yes
Minimum start 10 seeds × 20 Ep2 conditions Yes

Paired: same (seed, condition) across arms.


10. Analysis

See paper002_analysis_plan_v0.3.md.


11. Environment


12. Open before freeze


Appendix A — Latent logging (mechanistic · exploratory)

Field Spec
When Every step Ep1/Ep2 · snapshots pre/post expansion
Log z_t · predicted z rollouts · residual · uncertainty · gate p · selected expert · action · success
Analysis static/drift probe · silhouette · expert entropy · CKA before/after
Rule Not primary endpoint · not gate trigger
Claim text Must link to O1/O2: separation associated with lower error and higher success

Diffusion WM: extension only · not confirmatory v0.1.


13. Tags

Freeze → paper002-prereg-wm-v0.1 · experiment exp_surg_003_wm_expansion

First milestone (engineering): L1 repair fails · L3 explains held-out drift on pilot seeds.