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 — Research questions v0.2

ARCHIVED · mock→physics direction · superseded 2026-07-29 · do not cite or extend Current Paper 002: WM expansion · archive index

Method: paper002_method_spec_v0.1.md
Pre-reg: paper002_prereg_v0.2.md
Supersedes: v0.1 (desk draft)
Date: 2026-07-28


Central question (sharp)

EN:

Under a frozen mock–Isaac occlusion contract, does mock-ranked generative curriculum selection identify perturbation specs whose same-state counterfactual informativeness (CONTINUE fail · REPLAN success @ S) replicates on Isaac — and does an LLM goal planner match or beat a rule planner on that validation metric?

KR:

Mock–Isaac occlusion contract 하에서, generative curriculum의 mock rank가 perturbation spec의 same-state CF informativeness를 Isaac에서 예측하는가 — LLM goal planner가 rule planner와 동등 이상의 validation을 달성하는가?


Measured object

Spec-level validation pair ((m_i, y_i)):

Counterfactual evaluation = Paper 001 method · generation = Paper 002 object.


Sub-questions (priority order)

ID Sub-RQ Hypothesis Tier
RQ-B1 Mock rank → Isaac rank? H1 · ρ ≥ 0.5 @ rule Confirmatory
RQ-B2 Mock top/bottom stratum separates Isaac yield? H2 · IR_bottom ≤ 0.5 · gap ≥ 0.4 Confirmatory
RQ-B3 LLM goal planner validates @ Isaac vs rule? H3a floor + H3b non-inferiority Confirmatory extension
RQ-B4 Coverage (Gaussian) vs diversity (diffusion) tradeoff? S2 · descriptive Exploratory
RQ-B5 Export robust to mock seed? S1 · median ρ across 5 seeds Sensitivity

Hypothesis map

RQ-B1 ──► H1  Spearman ρ (rule · n=40)
RQ-B2 ──► H2  Tier IR + gap (rule)
RQ-B3 ──► H3  LLM ρ floor + non-inferiority vs rule
            (gate: H1 ∧ H2 pass first)

What pilot established (not re-asked as primary)

Study 002 finding Paper 002 action
Diffusion > Gaussian yield Closed · exploratory S2 only
Occlusion misalignment Contract frozen in method spec §3
H3′ ρ=0.899 @ n=20 Replication target @ n=40 · threshold lowered to ρ≥0.5
LLM = rule on mock Isaac is RQ-B3 test

Success tiers (paper)

Tier Condition Title-level claim
A H1+H2+H3 CF-validated curriculum · mock filter · LLM goals
B H1+H2 · H3 fail CF-validated curriculum · rule planner
B− H1 · ¬H2 Rank correlation without reliable tier filter
C ¬H1 Honest negative · pilot did not replicate @ scale

Non-claims

Clinical deployment · foundation model · SOTA reach · latent mismatch (Paper 003) · RL menu (Paper 003) · classifier-guided diffusion