How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
Date: 2026-07-22 · Input: Phase A Isaac smokes (001A–C + 001D D0/D1)
Output: decisions forphase_c_proper_run_prereg_v1.0.md
Label: smoke / scaffold only — not confirmatory evidence
| Study | Perturbation | Modes / axis | n | Key result (success) | Role |
|---|---|---|---|---|---|
| 001A | Target shift 3 cm | CONTINUE vs REPLAN | 5 ep | CONTINUE 0/5 · REPLAN 4/5 (1 unsafe) | Mode separation @ mild shift |
| 001B | Shift 3 cm | REPLAN delay 0–20 | 3 ep | REPLAN 3/3 all delays · flat band | Timing insensitive @ 3 cm |
| 001C | Shift 1–9 cm × delay 0–60 | REPLAN grid + CONTINUE | 5/cell | Mostly 0.8–1.0 · no monotonic cliff | Severity×delay surface mostly flat |
| 001D D0 | Shift 6 cm + occlusion L1 | CONTINUE · REPLAN · REOBSERVE | 5 | REPLAN 5/5 · CONTINUE 0/5 · REOBSERVE 4/5 | Occlusion anchor @ 6 cm |
| 001D D1 | Same + 5-mode | + RESHAPE · HANDOVER | 5 | REPLAN 5/5 · RESHAPE 4/5 · REOBSERVE 4/5 · HANDOVER 0/5 | Multi-class menu smoke |
Cross-cutting: scripted IK-Rel · same-state CF fork · tol 2 cm · n=5 smoke scale.
handover_proxy terminal · not a recovery endpoint in v0.1.| Gap | Why it matters | Phase C action |
|---|---|---|
| Statistical power | n=5 · seed outliers drive rates | n↑ pre-registered |
| Timing headline | Flat bands may be true or under-powered | Keep delay fixed @ d20 for P1; optional extended delay arm |
| Occlusion realism | gain_scale v0.1 ≠ geometric occlusion | Decide v0.1 vs v0.2 before GPU |
| RESHAPE vs REPLAN mechanism | One seed failure · no path metrics | Report path length · violation · n↑ |
| HANDOVER semantics | Stub only | Separate collaboration log metric or defer to Stage 2 |
| Baselines | Surgical UQ · B-VAP not implemented | Pre-reg optional baseline arm |
| Learned selector | Not tested | Out of Phase C — Stage 2 |
| # | Question | Phase B decision |
|---|---|---|
| 1 | Paper 1 headline axis? | Intervention mode under 6 cm + occlusion — not golden-time |
| 2 | Keep gain_scale occlusion? | Yes for Phase C v1.0 · document as proxy · v0.2 geometry = optional arm |
| 3 | Include HANDOVER in success table? | Log only · primary endpoint = policy/info/environment modes that target resolution |
| 4 | Primary comparison cell? | 6 cm · delay 20 · occlusion L1 (001C recommend · D0/D1 validated) |
| 5 | Baselines in first proper run? | Tier 1: CONTINUE + REPLAN only profile · Tier 2 (stretch): binary UQ-handover rule |
| 6 | Claim label? | Smoke = direction · Phase C = confirmatory only after pre-reg freeze |
| ID | Hypothesis | Smoke verdict | Proper run needed? |
|---|---|---|---|
| H1 | Mode matters @ fixed S | Supported (direction) | n↑ to confirm rates |
| H2 | Delay matters | Weak / flat @ 3 cm | Not P1 headline; optional arm |
| H3 | Occlusion ≠ shift-only | Partial — REPLAN still 5/5; profile shape differs (REOBSERVE role) | Compare D1 vs 001C @ 6 cm no-occlusion cell |
| H4 | RESHAPE adds path | Plausible 4/5 | n↑ + ablation |
| H5 | Beats binary baseline | Not tested | Implement in Phase C stretch |
| H6 | Not replay artifact | Not tested | n≥20 · second seed block |
Phase B status: complete (desk). Next: freeze Phase C pre-reg → lit deep dive → execute proper run.
phase_c_proper_run_prereg_v1.0.mdresearch_question.mdroadmap.mdstatus.md