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

Vision narrative v0.2 — Understanding-first (public-safe)

Long-term research program framing · not a company or product roadmap in this repo
See PUBLIC_SCOPE.md · PUBLIC_BOUNDARY.md


Vision (research program · long-term)

Physical AI that keeps learning when reality breaks its assumptions.

Public scope today (Mismatch Lab spec)

Robot Diff — understand what changed between runs, and when the model may no longer be adequate.

Value proposition (teams using the tools)

Reduce wasted retraining and identify the right next experiment after deployment surprises.

한국어:

구조적으로 잘못된 모델을 계속 재튜닝하느라 발생하는 시간과 실험 비용을 줄인다.

Long-term research direction (not current repo scope)

An experience lifecycle framing: real-world surprises → hypotheses → counterfactual tests → model revision → safer redeployment.

Documented here for orientation only. Not a commitment to ship a platform from research-os.

Domain generality (orientation only)

While the present studies use embodied robotic systems as the testbed, the underlying question — when should an autonomous decision-making system revise its model class rather than repair within it — is domain-general and may extend beyond physical robotics.

This framing does not change current scope or claims (see PUBLIC_BOUNDARY.md). It is documented here only to record the origin of the research question.


Three sentences (public-safe)

  1. Physical AI breaks when the world changes faster than the model.
  2. We detect when parameter repair isn't enough.
  3. We study how field experience should inform world model revision.

Two public surfaces (this repository)

Research OS (this repo)
  Evidence · protocols · Paper 001/002 · promoted results

Mismatch Lab (docs/mismatch_lab/)
  Robot Diff · benchmark · SDK design · pilot invitation

Commercial hosting · customer contracts · operational SLAs are out of scope for research-os.


Program structure (research · not org chart)

Program question
  Physical AI that improves when reality violates its assumptions

Public wedge (now)
  Robot Diff · Replay · Explain

Mechanism chain (Paper 002)
  Difference → Mismatch → Adequacy → Revision

Honest capability boundary (now)

Help teams understand robot behavior differences — and flag when the current model may no longer be adequate, suggesting repair, replan, or structural review.

Not yet:

determine what is missing

(hypothesis generation milestone)


Roadmap by trust level (research · aspirational)

Stage Research question Timeline
Detect Sustained prediction–observation divergence 2026–28
Diagnose Sensor · policy · parameter · structure candidates 2026–28 partial
Test Counterfactual that best separates hypotheses 2026–28 partial
Revise What to retrain · add · replace 2028–30
Validate Novel + static regression 2030–33
Deploy Safe redeploy + new experience capture 2033+

Later stages are not documented as products in this repo.


Paper 002 ↔ public lab spec

Paper 002     Scientific evidence for adequacy decision rule
Mismatch Lab  Explainer demos · benchmark · open SDK sketch

Decision-interface framing (research positioning)

Field telemetry
        ↓
Model adequacy layer (research + open spec)
        ↓
Collect / test / retune / expand / escalate
        ↓
Training and deployment stack (external)

Not tied to a single foundation model or simulator.


Evidence (split)

Preliminary (Tier B+ · mock pilot)

Before stronger public claims


Updated 2026-07-31