Boundary Learning Score
A scoring model for observing whether agents adapt after blocked, stale, or conflicted boundary decisions.
IMPACT BOUNDARY LABS
Boundary labs and scoring experiments.
Small live environments for testing how agents behave when external impact requires boundary decisions.
Impact Room
A live agent room where actions require state, approval, and boundary decisions.
GitHub Adapter
The repository-focused reference adapter: agents can propose work without receiving direct write authority.
Robotics
Planned lab direction for MCP/ROS2 converter work and robotics workflows that need visible action boundaries.
Build your own adapter
Wrap your own system as an adapter and connect it to the Core, without giving the agent a direct write path.
Why this matters
Why scoring matters
A blocked request is not only a failure. It is feedback. A useful agent should stop retrying rejected paths, re-read stale state, and propose smaller valid steps. That can reduce wasted agent runtime, repeated tool calls, and operator noise.
01
Fewer repeated mistakes
Agents should stop retrying paths the boundary already rejected.
02
Less repeated work
Fewer blocked attempts and stale retries mean less wasted agent work.
03
Better operator attention
Humans should review real decisions, not repeated noise.
04
More useful feedback loops
The score makes boundary feedback visible across comparable runs.
Boundary Learning Score
How the score is built
Raw boundary events become opportunity-normalized scoring evidence.
Raw counts are not enough. A harder run can contain more chances to make mistakes. Boundary Learning Score compares mistakes against the opportunities the agent had, then separates single-run cleanliness from adaptation across comparable runs.
Raw boundary events
The base evidence.
Opportunity profile
How many chances did the agent have to make each kind of mistake?
Component scores
Raw signals are grouped into explainable score components.
Run Boundary Fitness
A single-run view of how cleanly the agent behaved.
Challenge Adaptation Signal
Did behavior improve across comparable runs?
Score Confidence
How reliable is this comparison?
What the score measures
Measure whether agents adapt to rules instead of repeatedly violating them.
01
Repeated blocked attempts
Does the agent keep trying the same rejected action?
02
Stale retries
Does the agent act on old state, or does it re-read before retrying?
03
Better-scoped requests
Does the agent reduce blast radius and propose smaller, valid steps?
04
Required next action
Does the agent follow feedback such as required_next_action before trying again?
05
Operator attention
Does the agent reduce noise so humans can focus on real decisions?
Guardrails
What this does not claim
- Does not prove that an agent is safe
- Does not prove that generated work is correct
- Does not replace human review
- Measures behavior under a defined boundary challenge
Boundary challenge
Build your own boundary challenge.
Any small target environment can become a boundary-learning challenge. The question is always the same: can the agent reach the goal while making fewer boundary mistakes?
impact doorprotected repository pathsrobotics MCP/ROS2 converterscoped database operationinternal approval workflowlocal tool execution