LATENCY BUDGETING FOR STREAMING MACHINE LEARNING DECISION SUPPORT SYSTEMS

Real-time decision support systems operate under strict latency and resource budgets. Adding reliability estimation and adaptive control layers must not violate these constraints. This paper presents a microbenchmarking study of a reliability-driven multimodal pipeline. We formalize the concept of «operational budgets» (latency, update frequency, false alarms) and report empirical latency measurements. Results confirm that the proposed reliability estimation adds only ≈150 μs overhead per step, and the full adaptive pipeline remains well within millisecondrange requirements for real-time monitoring. The analysis is positioned in the context of multimodal fusion [1, 2], concept drift adaptation [3], and tail-latency constraints in large-scale systems