How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
Status: EXECUTED · 2026-07-22 · Tier C confirmatory
Date frozen: 2026-07-22 · Results:study1_proper/summary.json
This document records the frozen pre-registered design for Phase C. The run completed 2026-07-22; sections below are unchanged design cells (not retro-edited).
Smoke atlas (Phase A) must not be cited as confirmatory n.
Successful resolution — terminal distance to shifted target ≤ 0.02 m, no forbidden violation, within max steps.
Primary contrast (pre-registered):
At fixed mismatch onset S (6 cm shift + occlusion L1), recoverability rate of REPLAN_d20 vs CONTINUE.
Secondary contrasts:
| Parameter | Value | Rationale (Phase B) |
|---|---|---|
| Task | Isaac-Reach-Dual-STAR-IK-Rel-Play-v0 |
Locked Stage 1 testbed |
| Proposer | Scripted 6D IK-Rel · body 13 | Same as smoke |
| Shift | +0.06 m Y @ step 20 | 001C recommend · D0/D1 anchor |
| REPLAN delay | 20 steps | D0/D1 locked |
| Occlusion | Level 1 · visibility_fraction=0.35 · gain_scale_flag v0.1 |
Smoke-validated proxy |
| Modes | CONTINUE · REPLAN_d20 · REOBSERVE · RESHAPE · HANDOVER(log) | D1 menu |
| Seeds | 0–19 (n=20 per mode) | Power vs smoke n=5 |
| Branches | 100 (20 × 5 modes) | Single primary cell |
| Horizon | max_steps 160 · episode_length_s 20 | Smoke default |
Optional extension arm (not required for v1.0 pass):
| Tier | Baseline | Implementation | Primary metric |
|---|---|---|---|
| 0 | CONTINUE | Built-in | Success rate @ S |
| 1 | REPLAN_d20 | Built-in | Success rate @ S |
| 2 (stretch) | Surgical UQ binary | Fail detector → handover/continue rule | Profile AUC vs multi-mode |
| 3 (stretch) | B-VAP-style rule | Situation → fixed mode map | Same |
v1.0 minimum: Tier 0–1 with n=20. Tier 2–3 may slip to v1.1 if implementation not ready — do not delay mode profile run.
terminal_category=handover_proxy rate · time-to-handover · visibility state.Phase C run earns confirmatory label if:
results/study1_proper/ with manifest + git commit| Phase | Branches | Est. time/branch | Total GPU |
|---|---|---|---|
| Primary cell (100) | 100 | ~2–3 min | ~3–5 h |
| Optional no-occlusion (40) | 40 | ~2–3 min | ~1.5 h |
| Bootstrap (cold pod) | — | 15–25 min | once |
| Merge + figures | — | 15 min | CPU |
RunPod RTX 4090 · Isaac 4.1.0: budget ~1 pod-day including bootstrap + one retry buffer.
results/study1_proper/
isaac_aggregate.json
summary.json
tables/aggregate_by_mode.csv
run_manifest.json # commit · pre-reg hash · timestamp
study1_proper_report.md
| Version | Date | Change |
|---|---|---|
| v1.0 | 2026-07-22 | Initial freeze from Phase B · post D1 smoke |
Bump to v1.1 only via: new PR + dated changelog · never retro-edit after GPU start.