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

Independent research | Physical AI as the current testbed

After the Spill

How far should an embodied system change after unexpected experience?

Robots and embodied agents do not only need to learn from experience. They need to decide whether an experience is noise, a recoverable disturbance, a sign of model inadequacy, or evidence that the system itself must change.

Program charter Paper 002 evidence Mismatch Lab Builder Lab

The Question

Most AI research asks:

How should a system learn?

This program asks a narrower and harder question:

How much should a system change after reality contradicts its current model?

The research object is the change decision:

Experience -> Evidence -> Operator -> Closure

The working principle is simple enough to be useful and dangerous enough to audit:

Intelligence is the ability to determine the smallest sufficient change after unexpected experience.

What Exists Here

Foundation

D000-D006 define the program's formal objects, terminology, novelty audit, literature matrix, evaluation protocol, and roadmap.

Papers

Paper 001-005 test different change operators: recover, repair, expand, dispose, and coordinate.

Products

Robot Diff and Mismatch Lab turn rollout comparison into evidence for model adequacy and recovery decisions.

Builder Lab

Small, reproducible Physical AI experiments on machines that already exist, used without modification. Current status: Planning; no hardware purchased.

Evidence So Far

Track What it contributes Status
Paper 001: Recoverability Measures whether a system can still recover from the same post-mismatch state Tier C complete
Paper 002: Model Adequacy Tests when parameter repair should give way to a prepared model-order expansion Tier C complete
Paper 003: Representation Expansion Tests whether adding a missing relation can open capability repair cannot reach Negative under real contact; written up
Mismatch Lab Public surface for diff, replay, explanation, and benchmark design Spec/demo
Builder Lab Public execution surface for small Physical AI experiments Planning; selection rule recorded, no hardware

Selected proof points:

Program Map

Paper Core decision Operator
001 Recoverability Can the system still recover from here? Recover / Replan
002 Model Adequacy Is repair inside the current model class still enough? Repair / Expand
003 Representation Expansion Is the missing thing a representation? Expand
004 Experience Disposition What should be done with the experience itself? Ignore / Preserve / Escalate
005 Recovery Orchestration How should multiple recovery capabilities be sequenced? Coordinate

Start Here

If you want... Go to
The whole program in one document After the Spill v2
The document that tries to break the idea D001 Claim & Novelty Audit
The strongest current evidence Paper 002 project page
The public product/lab surface Mismatch Lab
The public execution lab Builder Lab
The repository GitHub

Boundary

This is independent personal research by Seonhye Gu. The current evidence is controlled simulation evidence. The program does not claim clinical deployment, hardware transfer, autonomous structural self-improvement, universal resilience, arbitrary causal discovery, or a complete theory of intelligence.

Negative and design-stage results are kept in the record because a program gets stronger by rejecting weak claims early. The same applies to directions: the physical track carries an explicit exclusion rule for hardware selection, and what it rejected is recorded alongside what it kept.

Updated 2026-08-07