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 — Literature positioning v1 (public)

Date: 2026-07-22 · Status: post Lit Sprint v2 · smoke-informed · not a survey


Positioning (one paragraph)

Prior work strong on separate axes: uncertainty-gated handover in surgery, binary task–recovery switching, autonomous exploration recovery, tissue reshaping for exposure (MEDiC), and situation-driven replanning stacks (VAP-TAMP). None standardize same-state counterfactual comparison of a multi-mode intervention menu toward successful resolution with explicit timing. Our Stage 1 contribution is an evaluation protocol and empirical profiles ( \hat{R}_a(s,t) ) on an ORBIT-Surgical reach proxy — smoke evidence (n=5) shows mode separation under shift and an occlusion proxy; confirmatory runs are pre-registered separately (Phase C).


Closest prior art (honest)

Line They optimize Our wedge
Surgical UQ continue vs handover Multi-mode profile @ same onset
Recovery RL / RTA task ↔ backup policy Menu + reshape + reobserve
MEDiC tissue reshape for exposure Reshape as one timed mode among many
VAP-TAMP situation → perceive → replan Eval not new stack
VLA / FM task success No public multi-mode recovery profile tables

Kill test (Day-1 + smoke): full combination → 0 Kill · refine positioning, do not claim empty field.


What smoke supports (Tier B only)


What we do not claim