A treatise on agent safety.
Safety is not a policy you write. It is a space you cannot leave.
The prevailing approach asks an agent to behave, then watches to see whether it did. This is governance by hindsight — a ledger of harms already done. Axiom proposes the opposite discipline. We do not instruct the agent in what it ought not do; we shape the world so that the forbidden act has no path to occur. The rule is compiled into the boundary of action itself, evaluated before any irreversible step, and accompanied always by the proof of its own enforcement. What cannot happen needs no monitoring.
The Axioms
- I.
Axiom I.
A constraint is evaluated before the act, never after.
Permission is decided at the threshold. An action that has already occurred can be logged, regretted, or audited — but it cannot be prevented. Axiom moves the decision earlier in time.
- II.
Axiom II.
What is forbidden cannot be reached, not merely discouraged.
A rule that depends on the agent's cooperation is a suggestion. A constraint that bounds the space of possible actions is a law. We deal only in the latter.
- III.
Axiom III.
Every prohibition carries its own proof.
When an action is refused, the refusal is accompanied by the predicate that refused it and a record sufficient to reconstruct the judgment. Compliance is not asserted; it is demonstrable.
Figure 1 — a constraint as predicate
constraint("pii.never.leaves.eu") {
when agent.action
∈ {export, send}
then reject · log · prove
}The predicate wraps every session. Violation is structurally impossible.
Conclusion
What cannot happen, needs no policing.
Q.E.D.
Correspondence
Begin with one constraint.
Write the rule once; the boundary holds it thereafter.