Introducing RUM: A Methodological Contribution forEngineering Trustworthy AI Components in Industrial Systems

Martin Gonzalez, Loïc Cantat, Kévin PasiniAAAI 2025 · DOI

Abstract

We introduce RUM, a unified, lifecycle-aware framework for facilitating the engineering and assessing Trustworthy AI Components—software units embedding both pure and statistical functions within real-world systems. Unlike model-centric evaluations, RUM treats AI Components as indivisible units whose behavior must be understood across specification, development, operation, and updating phases. This announcement paper presents a series of research articles that establish the foundation of RUM: (1) a formal argument for the atomic nature of Trustworthy AI Components; (2) a structured set of novel trust metrics, many of them being of non-aggregative nature, spanning the component's lifecycle; and (3) an operational framework introducing AI Blueprints to support runtime monitoring, human-in-the-loop usage, and temporal maintainability while facilitating the evolution of AI Components at different stages. RUM offers a coherent alternative to fragmented evaluation tools, aligning with the needs of AI deployment in industrial contexts.

Introduction: The Challenge of Industrial AI

In laboratory settings, AI models are typically evaluated using static benchmarks like accuracy. However, in industrial environments, these models are embedded within complex AI Components (AIC) that include decision logic, state management, and calibration parameters. Traditional software engineering's "Modularity Hypothesis"—the assumption that systems can be decomposed into independent, pure functions—often fails for AI due to the "strong entanglement" between models and data. RUM (Robustness, Uncertainty & Monitoring) addresses this gap by shifting the focus from model-centric evaluation to a component-centric, lifecycle-aware methodology.

The RUM Methodology

RUM is built upon three core principles that move beyond simple model performance:

  1. Atomicity of AI Components: The AIC is treated as an indivisible unit comprising logic, statistical functions, datasets, and calibration. You cannot simply swap one model for another without re-evaluating the entire component.
  2. Lifecycle-Aware Trust Trade-Offs: Trustworthiness is evaluated through "non-aggregative" metrics across specification, development, and operation phases. These metrics represent trade-offs that stakeholders must weigh separately rather than averaging into a single score.
  3. Operation-Centric Orientation: The framework prioritizes the "Operate" stage, using AI Blueprints—reusable design patterns—to align AI behavior with operational constraints and human oversight.
The RUM lifecycle involves continuous loops between scoping, developing, and operating, ensuring the AI component evolves with its environment.

The framework follows a cyclic evolution: Operate (real-time monitoring and adjustment), Develop (targeted performance updates), and Scope (adapting the component to new contexts). This approach acknowledges that trust is not a one-time achievement but a moving target maintained through operational feedback.

Case Study: Visual Inspection in Automotive Manufacturing

The RUM framework is demonstrated through a visual inspection use-case involving the automated analysis of weld seams on vehicle chassis. This task requires high-frequency processing of image data characterized by varying luminosity, blur levels, and potential out-of-distribution (OOD) samples.

In the visual inspection use-case, the AI component provides anomaly status and confidence estimates to support human operator decisions and post-operation annotation.

By utilizing conformal prediction, the AIC provides a mathematical safety net, guaranteeing a high probability that the true class (OK/KO) is within the prediction set. This allows human operators to remain effectively "in-the-loop," focusing on high-uncertainty cases identified by the system while maintaining traceable records for quality control.

Key Insights and Results

RUM introduces the Indecomposability Principle, formalizing the need to evaluate the interaction between a model and its specific calibration context. Key results of this framework include:

  • Traceable Comparisons: RUM generates artifacts that allow operators to compare different versions of an AIC based on historical performance (e.g., "Version A would have reduced errors by X% over the last week").
  • Drift Visualization: Using techniques like HDBSCAN on UMAP-reduced latent spaces, RUM can graphically represent "data temporal drift," allowing teams to see exactly when the operational environment moves away from the training distribution.
  • Non-Aggregative Metrics: The framework moves away from "global scores," acknowledging that a system might be accurate but have poor temporal maintainability, requiring separate evaluation by stakeholders.

Conclusion and Future Work

RUM provides a structured alternative to fragmented AI evaluation tools, specifically tailored for mission-critical industrial applications. While the framework requires sophisticated infrastructure and cross-stakeholder coordination, it establishes a foundation for long-term reliability. Future work in this series will detail specific technical trust metrics and expand the framework's application to other use-case classes such as predictive maintenance and autonomous navigation.

Summary updated

Abstract

We introduce RUM, a unified, lifecycle-aware framework for facilitating the engineering and assessing Trustworthy AI Components—software units embedding both pure and statistical functions within real-world systems. Unlike model-centric evaluations, RUM treats AI Components as indivisible units whose behavior must be understood across specification, development, operation, and updating phases. This announcement paper presents a series of research articles that establish the foundation of RUM: (1) a formal argument for the atomic nature of Trustworthy AI Components; (2) a structured set of novel trust metrics, many of them being of non-aggregative nature, spanning the component's lifecycle; and (3) an operational framework introducing AI Blueprints to support runtime monitoring, human-in-the-loop usage, and temporal maintainability while facilitating the evolution of AI Components at different stages. RUM offers a coherent alternative to fragmented evaluation tools, aligning with the needs of AI deployment in industrial contexts.
BibTeX
@inproceedings{gonzalez2025introducing,
  author = {Martin Gonzalez and Loïc Cantat and Kévin Pasini},
  title = {{Introducing RUM: A Methodological Contribution forEngineering Trustworthy AI Components in Industrial Systems}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaaiss.v7i1.36881},
}
Figure