AVSSLP: A Six-Dimensional Evaluation Framework for Artificial Intelligence Systems
Year: 2026
Type: preprint
Abstract: This paper introduces AVSSLP, a six-dimensional framework for evaluating the performance, efficiency, and resilience of artificial intelligence systems. The framework consists of six dimensions: Actions per Second (APS) Value of Actions Performed Sustainability Stress Test Latency Power Usage AVSSLP expands conventional performance evaluation beyond throughput and response time by incorporating action value, energy efficiency, and system behavior under demanding operating conditions. The framework proposes operational definitions, measurement approaches, experimental considerations, and reproducibility guidelines for each dimension. It also examines how these measurements can reveal trade-offs that conventional performance benchmarks may overlook. A central principle of AVSSLP is that increased computational activity does not necessarily represent increased useful performance. Systems must be evaluated according to the value of their completed actions, the resources required to produce them, and their ability to maintain performance under stress. AVSSLP is proposed as a general-purpose evaluation methodology applicable to AI models, agentic systems, inference architectures, and adaptive resource-management systems. This work establishes the theoretical and methodological foundation for future AVSSLP-based experiments.
Author: Bigelow, Devin "River"
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7232302138
