Confidence-Gated Arbitration of Automated Control Actions for Stability under Compounded Failures in High-Concurrency Platforms: A Comparative Empirical Study

Abstract

High-concurrency platforms rely on automated control actions such as auto-scaling, request retries, circuit breaking, and failover to absorb load and recover from disturbances. Under compounded failures these same actions frequently amplify rather than dampen instability, producing retry surges, oscillating scaling decisions, and recovery trajectories that stall in degraded states. This study presents a comparative empirical evaluation of three operating configurations on a microservice benchmark driven by production-derived workload characteristics: naive automation, queue-based smoothing only, and a full governance arrangement that adds observability-driven confidence gating and rate constraints over the automated actions. Three fault classes are injected under controlled load: dependency timeout, node failure, and traffic spike. Across ten repetitions per configuration-fault pair, the governance arrangement lowers tail latency and recovery time while suppressing retry amplification and scaling oscillation, with the magnitude of the gain depending strongly on the fault class. Queue smoothing alone recovers most of the latency benefit under traffic spikes, while confidence gating contributes the larger share under timeout and node-failure conditions where self-reinforcing retry behavior dominates. The findings delineate where arbitration of automated actions pays off and where buffering suffices, offering an engineering-grade basis for stabilizing large-scale digital infrastructure.