TIKL-GRACP: A Context-Aware Transport Augmentation for AI-Driven and Virtualized Networks

The rapid growth of AI-driven systems has introduced new challenges for transport-layer performance and stability. Modern AI workloads involve large-scale data movement, continuous telemetry exchange, and distributed execution across heterogeneous, virtualized, and resource-shared environments. In such settings, traditional congestion control mechanisms often misinterpret host-side artifacts-such as CPU scheduling delays, virtualization noise, and wireless variability-as network congestion, leading to unstable throughput, unfair resource sharing, and excessive queuing delay. This paper proposes an AI-native transport augmentation that explicitly separates path diagnosis from rate actuation. The framework consists of two cooperative components: Telemetry-Informed Kinetic Learning (TIKL), which performs bounded active and passive measurements to infer network and hostside conditions, and Governed Rate Adaptation and Chunk Pacing (GRACP), which regulates data transmission using explicit stability, fairness, and delay constraints. Unlike existing congestion control algorithms that rely solely on passive endto-end feedback, the proposed approach incorporates context awareness to distinguish true network bottlenecks from hostinduced artifacts. The proposed framework is designed as an extension layer compatible with existing transport protocols such as TCP and QUIC. Experimental evaluation demonstrates improved throughput stability, reduced queuing delay, and enhanced fairness under dynamic bandwidth, virtualization, and mixed-traffic scenarios, making the approach well-suited for modern AI-centric networking environments.