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 1 — Phase B smoke review (desk · complete)

Date: 2026-07-22 · Input: Phase A Isaac smokes (001A–C + 001D D0/D1)
Output: decisions for phase_c_proper_run_prereg_v1.0.md
Label: smoke / scaffold only — not confirmatory evidence


1. Smoke summary table

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.


2. What we know (smoke)

  1. Mode dominates at fixed S — CONTINUE ≪ REPLAN repeats across shift-only (001A) and occlusion (D0/D1).
  2. Timing cliff absent @ tested grid — 001B flat @ 3 cm; 001C no monotonic delay sensitivity; 6 cm @ d20 remains viable (001C recommend cell).
  3. Occlusion does not collapse REPLAN — D0/D1 REPLAN 5/5 @ 6 cm + gain_scale proxy.
  4. REOBSERVE viable but seed-sensitive — 4/5 both D0 and D1; seed 0 outlier (~0.80 m) matches 001A ep0 pattern.
  5. RESHAPE competitive with REPLAN — 4/5 (seed 2 timeout); environment-class path is plausible, not proven at scale.
  6. HANDOVER stub fails by design — 0/5 · handover_proxy terminal · not a recovery endpoint in v0.1.
  7. CF scaffold works — branch_replay_ok True on all promoted runs · pipeline RunPod-reproducible.

3. What we do not know (honest gaps)

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

4. Open questions → Phase C decisions

# 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

5. Hypothesis status (from method doc)

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

6. Phase B deliverable checklist

Phase B status: complete (desk). Next: freeze Phase C pre-reg → lit deep dive → execute proper run.