How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
Status: v1.0 · aligns with RQ v1.1 · pre-reg v2.0 in design
Nominal rollout (0 … onset−1) → record actions → mismatch @ S → replay → fork(response) → judge → profile
Same-state counterfactual evaluation: all branches share identical trajectory through mismatch onset S; only post-onset intervention differs. CF is method, not a learned recoverability estimator.
| Parameter | Anchor (D0–D3) | D1 control |
|---|---|---|
| Task | Isaac-Reach-Dual-STAR-IK-Rel-Play-v0 |
same |
| Onset step | 20 | 20 |
| Target shift (P2) | +0.06 m Y | same |
| Occlusion (P3) | L1 · visibility 0.35 · gain_scale v0.1 | none |
| REPLAN delay | 20 steps | 20 |
| Horizon | 160 steps | 160 |
| Proposer | Scripted 6D IK-Rel · body 13 | same |
Proxy contract: study1d_occlusion_proxy_v0.1.md.
[0, onset) · store pre_actions.frozen_xyz · apply shift → shifted_xyz.pre_actions · diverge response from step onset.Invariant: branch_replay_ok = true on all branches.
Forbidden AABB (EE proxy): center [0.45, 0, 0.15] · half-extent 0.04 m.
| Mode | Behavior after S |
|---|---|
| CONTINUE | Chase frozen target · occluded gain scale |
| REPLAN_d20 | Frozen until t+20 · then shifted target |
| REOBSERVE | 10-step hold · clear visibility · replan shifted |
| RESHAPE | 20-step reshape proxy · clear · replan |
| HANDOVER | Zero-action stub · early terminal |
Runners: orbit_reach_study1a_counterfactual.py · orbit_reach_study1d_counterfactual.py.
B2 — UQ-inspired binary: visibility & distance thresholds → HANDOVER_stub vs CONTINUE (rule proxy · not learned UQ).
B3 — Situation rule: occluded_shift → REOBSERVE path · shift_only → REPLAN_d20 · else CONTINUE.
Constants frozen in pre-reg v2.0. Desk spec: private baseline_spec_phase_c_v1.0.md.
Primary — successful resolution:
Categories: successful_resolution · unsafe_failure · timeout_failure · safe_unresolved · handover_proxy.
HANDOVER excluded from primary endpoint · reported as burden proxy.
| Tier | Label |
|---|---|
| A | Scaffold / protocol |
| B | Smoke n≤5 (001A–D) |
| C | Pre-reg n=20 · D0–D3 executed · v2.0 complete 2026-07-24 |
Scripted proposer · single sim task · occlusion proxy v0.1 · HANDOVER stub · B2 rule proxy · offline CF (not closed-loop deployment).
research_question.mdphase_c_proper_run_prereg_v1.0.mdstudy1_proper/summary.json