What Is Monitored Containment? The Council of AI Measurement Framework

Q: What is monitored containment?

Monitored containment is the principle that AI systems should be deployed within observable boundaries, with continuous per-axis measurement replacing trust as the basis for safety assurance — rejecting both provable isolation and pure release-and-monitor in favour of a measurement perimeter.

Monitored containment is the Council of AI's core governance framework: the principle that AI systems should be deployed within observable boundaries, with continuous measurement replacing trust as the basis for safety assurance. It rejects both extremes of the AI governance debate. Against the 'provable isolation' school — which argues that frontier models must run in air-gapped, mathematically verified enclosures — it observes that perfect isolation is neither achievable nor desirable: AI systems must interact with data, APIs, and humans to deliver value, and every interaction surface is a potential escape vector.

Against the 'release-and-monitor' school — which argues that safety emerges from post-deployment observation — it observes that retrospective monitoring cannot prevent harm, only document it. Monitored containment instead establishes a measurement perimeter: a set of 14 axes (governance, privacy, AGI-readiness, ASI-readiness, MCP interoperability, open-source compliance, machine reasoning, care/safety, extended reality, detection/interop, Article 5 screening, swarm coordination, and affect alignment) against which every model run produces a signed, auditable measurement credential. The containment is 'monitored' because measurement is continuous — every inference, every tool call, every multi-agent negotiation generates telemetry that the measurement board evaluates in near-real-time.

The containment is 'not provable isolation' because the system acknowledges that measurement, not mathematical proof, is the achievable standard for deployed AI. This distinction matters for procurement: a tender that requires 'provable isolation' will fail, because no deployed system can meet it; a tender that requires 'monitored containment with signed measurement credentials' is auditable, enforceable, and grounded in what the Council of AI's 14-board measurement fleet actually produces. The framework is documented in the Council of AI Charter of Charters and operationalised through the Council City municipal AI programme.

Standards and sources referenced

Council of AI Charter of Charters (Measurement Framework, Section 6)
Council City Pilot Programme — Municipal AI Procurement
DEFONEOS-SEAL Verified Measurement Credential specification
NIST AI 100-1 (AI Risk Management Framework)
ISO/IEC 42001:2023 AI Management System
UK AI Safety Institute — Evaluations Platform (Inspect)