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 C proper-run pre-registration v1.0

Status: EXECUTED · 2026-07-22 · Tier C confirmatory
Date frozen: 2026-07-22 · Results: study1_proper/summary.json


0. Scope boundary

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.


1. Primary endpoint

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:


2. Design (frozen v1.0)

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):


3. Baselines (tiered)

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.


4. HANDOVER handling


5. Analysis plan (pre-registered)

  1. Per-mode success rate + Wilson 95% CI (n=20).
  2. Mean final distance · completion steps · forbidden violation rate.
  3. Profile table ( \hat{R}_a(s,t) ) for fixed anchor cell.
  4. Seed sensitivity: flag outliers (ep0 pattern) · report with/without sensitivity.
  5. No p-hacking across modes — primary hypothesis = REPLAN > CONTINUE; others exploratory (FDR noted in text).

6. Success criteria (confirmatory label)

Phase C run earns confirmatory label if:


7. GPU budget estimate

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.


8. Artifacts (on completion)

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

9. Explicitly out of scope (Phase C v1.0)


10. Version history

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.