Governed Execution Boundary (reference implementation)

The Governed Execution Boundary is a reference implementation of a single control point placed in front of an autonomous agent's actions. Rather than trusting a model to police its own outputs, it makes every action pass through one place that enforces policy, pauses for human approval where that is required, and records what happened. Because the boundary is singular and explicit, it is auditable: there is exactly one place to reason about what an agent is permitted to do, and exactly one place that produces the record of what it did.

How it works

Conceptually the boundary separates proposing from disposing. The model proposes an action; a sovereign predicate at the boundary decides whether that action proceeds, is escalated to a human, or is refused — and either way the decision and its context are captured as evidence. This is the same governance gene the family applies elsewhere: the model proposes, a sovereign predicate disposes, and everything leaves an evidence trail. The value is architectural rather than algorithmic — the design keeps the enforcement surface small and inspectable instead of scattering checks through agent logic.

Why it matters

It exists to support the governed-agent-systems research programme: a concrete, readable substrate for studying how a single approvable, evidence-producing boundary behaves in practice, and for grounding discussion in something runnable. It is deliberately experimental — interfaces may change without notice, and it has had no independent security audit — so it is intended for research and discussion rather than production use. Anyone considering it beyond that setting should commission independent review first.