How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
ARCHIVED · mock→physics direction · superseded 2026-07-29 · do not cite or extend Current Paper 002: WM expansion · archive index
Positioning: paper002_description_v0.1.md
Pre-reg: paper002_prereg_v0.3.md
Under a frozen mock–Isaac occlusion contract, does a continuous cheap proxy score (five-seed median) rank exported occlusion mismatches by their same-state counterfactual value in Isaac simulation — and does a constrained LLM proposer meet prespecified adequacy vs a rule proposer?
This is not “does LLM generate a good training curriculum?” It is “does the proxy predict which physics experiments are worth running?”
| ID | Question | Hypothesis |
|---|---|---|
| RQ-B1 | Proxy validity on export set | H1 |
| RQ-B2 | Selection utility (top vs bottom) | H2 |
| RQ-B3 | LLM proposer adequacy | H3 |
| RQ-B4 | Coverage vs diversity (G vs D) | S2 · exploratory |
| RQ-B5 | Consensus vs single-seed export | S1 · sensitivity |
Paper 001 asks how to measure response quality after mismatch.
Paper 002 asks which mismatches are worth measuring before spending GPU on physics.
POET/PLR/GenSim optimize learning from generated environments; Paper 002 validates experiment selection via mock-to-physics rank transfer @ same-state CF.