Policy Enforcement — References

The architectural vocabulary for separating a rule from the code that applies it, and the frameworks that increasingly require an organisation to say which rules it is applying.

Separating rule from mechanism

• Policy/Mechanism Separation in Hydra — Roy Levin, Ellis Cohen, William Corwin, Fred Pollack and William Wulf, ACM Symposium on Operating Systems Principles, 1975. The original argument that the component enforcing a decision and the component making it should be distinct, which is the whole justification for a policy engine rather than conditionals in application code. Cited by name.
• eXtensible Access Control Markup Language (XACML) Version 3.0 — OASIS Standard. Cited here for obligations and advice: a decision that carries a required action for the enforcement point to perform, which is the formal model for "append this disclosure" as a policy outcome rather than as product code. Cited by name.
• Open Policy Agent — included for its bundle distribution and decision-logging mechanics, which are the parts that make rule versioning and evaluation records practical rather than aspirational.

Being required to have rules at all

• NIST AI Risk Management Framework — cited here for the expectation that policies governing a deployed system are documented, assigned to an owner, and periodically reviewed, which is the organisational precondition for anything on this page being maintainable.
• Regulation (EU) 2024/1689, the Artificial Intelligence Act — cited here for the transparency obligations in Article 50, which are a concrete example of a rule that must fire on specific interactions, be provable after the fact, and vary by jurisdiction. Cited by name; consult the official consolidated text for applicability and operative wording.

No verified reference is recorded for policy versioning with validity windows in language-model products, for deriving enforcement test suites from policy prose, or for measuring rule coverage across request paths. Access-control standards supply the vocabulary and stop short of the authoring and change-management problem, which is where this capability actually fails.