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

After the Spill Research Program v2.0

Core question: How should unexpected experience change an embodied system?

Program mission

After the Spill studies decision-making after unexpected experience in embodied systems. It asks when a system should preserve its current model, recover within it, repair parameters, expand representation, or coordinate multiple recovery operators.

The program is built to survive individual paper outcomes. A paper may fail, but the foundation documents should continue to clarify the question.

Practical question

Given an unexpected experience, what is the smallest change that restores competent action without hiding the evidence that made change necessary?

One compact decision form is:

c* = argmin_c [task loss + change cost + risk + future cost]

where c ranges over the available change operators.

Change decision framework

Object Working role
Experience A situated event that violates, stresses, or confirms the current model
Evidence Information extracted from experience that bears on adequacy
Operator A change action applied to state, policy, model, representation, or coordination
Closure A test that the chosen change is sufficient for the current task boundary
Competence The ability to act successfully under the relevant post-experience conditions

The decision is not "learn or do not learn." It is a choice among change operators with different costs and risks.

Change operators

Operator Meaning Typical failure if overused
Preserve Record the episode without permanent adaptation Ignores real mismatch
Recover Restore competent behavior without changing the model Treats inadequacy as disturbance
Repair Update parameters or local policy within the current representation Overfits inside the wrong model
Expand Add a prepared mode, relation, entity, constraint, or representation Invents structure too eagerly
Coordinate Select, order, or recruit multiple recovery capabilities Adds policy complexity without evidence

Paper roles

Paper Decision studied Program contribution
Paper 001: Recoverability Which response restores execution at the same mismatch state? Defines the post-mismatch recovery window
Paper 002: Model Adequacy When is parameter repair insufficient? Tests evidence for moving from repair to model-order expansion
Paper 003: Representation Expansion Which representational operator is missing? Tests mode/relation discrimination and capability expansion
Paper 004: Experience Disposition How much should an experience change the system? Classifies ignore, preserve, recover, repair, expand, and escalate decisions
Paper 005: Recovery Orchestration How should multi-relation consequences be contained and resolved? Studies ordered operator choice under uncertainty

Capability recruitment is a Paper 005 mechanism, not a separate paper. Ecosystem adaptation remains a long-term open problem.

Foundation documents

ID Title Immediate question
D000 Mathematical Foundation What are the program's stable objects?
D001 Claim & Novelty Audit Is this already Active Inference, Continual Learning, Meta Learning, Recovery Planning, Resilience Engineering, Self-healing Systems, Adaptive Control, or Decision Theory?
D002 Literature Matrix Can every paper be read through experience, change, decision, evidence, and closure?
D003 Terminology Which words are allowed, and what do they mean?
D004 Formal Definitions What must be defined before experiments can claim anything?
D005 Unified Evaluation Protocol What counts as evidence that an operator was sufficient?
D006 Research Roadmap How do papers and products accumulate into one program?

Shared evaluation philosophy

The program evaluates decisions, not only outcomes. A system that succeeds after unnecessary expansion may still be making a bad change decision. A system that fails after refusing expansion may have learned the boundary of repair.

Every study should measure whether the amount of change was justified:

  1. recovery closure,
  2. unnecessary and missed revision,
  3. revision latency,
  4. nominal regression,
  5. cumulative task regret,
  6. unresolved relation burden,
  7. claim tier supported by the result.

Claim boundary

Current evidence supports controlled simulation claims. It does not support clinical deployment, autonomous self-improvement, general intelligence, universal resilience, arbitrary causal discovery, unbounded world-model expansion, or a solution to continual learning.

The line "intelligence chooses the smallest sufficient change" is a governing principle and an audit target. It becomes a scientific claim only where evidence and closure tests make it measurable.

Governing principle

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

Build less. Break more. Keep what survives.

The program should become harder to fool over time.

Living document - 2026-08-04