This category gathers NeoMundi’s foundational reference articles on AI Governance Metrology. It forms the public conceptual and methodological foundation used to design, document and interpret measurements of AI system behavior in production.
Governing AI systems in production cannot rely solely on average performance, declared compliance or classical technical observability. It requires repeated, documented and traceable measurements of actual behavior during execution.
This corpus establishes the core distinctions and concepts of behavioral metrology:
• Classical observability vs behavioral metrology
• Signal vs verdict
• Stability vs factual consistency
• Deceptive stability
• Runtime risk signals
• Reproducible measurement campaigns
• Weekly barometers and monthly cartographies
• Baselines in behavioral metrology
• Evidence trails and traceability
• Multi-signal analysis
• Longitudinal monitoring and drift detection
• Integrating runtime metrology into governance processes
• Limitations of behavioral measurement in production
• Privacy-first design and data minimization
• Human oversight and decision authority
• Connecting controlled campaigns with continuous production monitoring
• Organizational maturity in AI governance metrology
The articles in this series provide the definitions, principles, methodological frameworks and limitations required to rigorously measure the behavior of deployed AI systems, while maintaining a clear separation between observation and decision-making.
Introduction AI governance metrology is not only a technical capability. It is also an organizational practice that develops over time. Organizations differ in how systematically they measure AI behavior, interpret signals, and integrate observations into decision processes. This article outlines a maturity perspective on AI governance metrology. It describes progressive levels of capability and highlights […]
Introduction AI behavioral metrology generates structured observations of generative systems in production. Because these observations are produced from real interactions, the design of the measurement system itself raises important questions related to data protection, privacy, and proportionality. This article examines how a privacy-first approach and the principle of data minimization can be applied to AI
Introduction Runtime metrology produces structured behavioral signals from AI systems in production. These signals can inform governance processes, but they do not replace human or organizational judgment. Maintaining clear decision authority is a foundational requirement of responsible AI governance. This article examines the role of human oversight in systems that incorporate runtime metrology. It clarifies
Introduction AI behavioral metrology can be applied in two complementary modes: controlled measurement campaigns and continuous production monitoring. These two modes are complementary rather than sequential. Each mode serves distinct purposes, operates under different constraints, and generates different types of insight. This article clarifies the relationship between these two approaches. It explains how controlled campaigns
Introduction In AI behavioral metrology, comparison with a reference helps interpret measurements and their variations. Without a clear reference point, observed variations remain difficult to interpret. This reference point is commonly called a baseline. This article defines the concept of a baseline in the context of AI governance metrology, explains its role in longitudinal monitoring
Introduction In AI governance, the value of a measurement depends not only on what is observed, but also on the ability to reconstruct how that observation was produced. Without traceability, signals remain difficult to audit, compare, or defend. This article defines the concept of evidence trails in the context of AI behavioral metrology, explains why
Introduction Runtime metrology provides structured behavioral signals from AI systems in production. However, the value of these signals depends on how effectively they are integrated into existing organizational governance processes. This article examines the practical considerations involved in connecting a runtime metrology layer to governance workflows. It focuses on integration principles, common challenges, and approaches
Introduction AI behavioral metrology provides structured methods for observing the behavior of generative systems in production. Like any measurement discipline, it is subject to important limitations. Recognizing these limitations is essential for responsible interpretation and for avoiding overconfidence in the resulting signals. This article outlines the principal limitations of current approaches to AI behavioral measurement
Introduction As organizations move generative AI systems into production, the need for continuous, reproducible behavioral measurement becomes critical. Traditional observability tools focus primarily on infrastructure metrics, while static evaluations provide only periodic snapshots. A dedicated runtime metrology layer is required to bridge this gap. ControlTower is NeoMundi’s runtime metrology layer designed to measure and surface
Introduction Relying on a single behavioral metric is rarely sufficient for effective AI governance in production. A model may appear stable while showing factual-risk signals, or it may produce coherent outputs that nevertheless drift over time. Multi-signal analysis addresses this limitation by combining several complementary behavioral indicators to form a richer and potentially more robust
