Summary

A 750-run study demonstrating that runtime governance can operate as domain-invariant infrastructure, producing consistent enforcement, tamper-evident records, and fully reconstructable decision trails across policy domains and model providers.

Key findings

• 100% of governed runs produce complete, gap-free evidence chains with fully reconstructable decision trails.
• Decision trails are recoverable from stored evidence alone with 100% fidelity, without access to execution traces.
• Governance infrastructure operates identically across domains (operational, security, organizational) and models.
• Three distinct enforcement profiles (DENY-dominant, REQUIRE-dominant, balanced) emerge from the same underlying system.
• Governance separates cleanly into invariant infrastructure and domain-specific policy, enabling reuse across environments.

Abstract

As AI systems begin executing real-world actions, governance becomes an infrastructure requirement rather than a model property. This paper evaluates a runtime governance system across 750 executions spanning three policy domains, two model providers, and five system architectures. The focus is not task performance, but whether governance infrastructure itself is correct, invariant, and complete. The results show that every governed action produces a tamper-evident Policy Restriction Evidence (PRE) record, that evidence chains are gap-free, and that full decision trails can be reconstructed from stored evidence alone with 100% fidelity. Critically, the same governance engine, schema, verification path, and reconstruction algorithm operate identically across operational, security, and organizational domains, as well as across different model providers. The findings establish a separation between policy (domain-specific) and enforcement infrastructure (domain-invariant), demonstrating that runtime governance can function as a single architectural layer independent of application domain.