Type-Safe Decision Frameworks for Agentic 5G Control: A Theory-Driven Testbed Characterization of Where They Can Be Applied

2026-09-27

AI Summary

This study addresses the vulnerability of free-text decision-making in 5G agent control to formatting errors and misjudgments, alongside the unclear applicability boundaries of type safety. We evaluate three type-safe frameworks on an Open5GS testbed and propose a theoretical formulation that translates timeliness and type consistency into verifiable predicates. By integrating Jev hosting, Laya fine-tuned encoders, and AnyJev zero-label adaptation with conformal risk control, we design an optimal action/escalation/abstention gating mechanism. Our contributions include a mapping from generative frameworks to 5G decision classes, demonstrating that while type safety eliminates formatting errors, it cannot guarantee semantic correctness. Furthermore, we reveal that fine-tuned encoders exhibit error rates up to 80% on change-related queries, whereas question-reading frameworks maintain error rates below 0.143%.

Abstract

This paper presents a theory-driven characterization of type-safe decision frameworks for the agentic control of 5G networks, where every decision must be an element of a declared option set rather than free text. Three design points are evaluated on an Open5GS/UERANSIM testbed with a closed core-policy loop, namely a hosted typed model (Jev), an open fine-tunable typed encoder (Laya), and a zero-label retrofit of a general language model (AnyJev). The proposed theoretical framework turns timeliness, type conformance, certification cost and cardinality into checkable applicability predicates, supported by an optimal act/escalate/abstain gate, an escalation-feasibility floor, a co-location stability condition, per-type conformal risk control with a certification label floor, and a type-mismatch bound. Measuring every predicate yields an applicability map from framework to 5G decision class. Type safety removes format failures but not the question: the fine-tuned typed encoder returned its training answer for 98-99.5% of changed questions, and its calibrated gate then acted wrongly on up to 80% of them, whereas the question-reading frameworks acted wrongly on at most 0.143 (Jev) and 0.137 (AnyJev) of any changed question, but were either hosted and 11-29 times slower (Jev) or reliant on an 8B language model (AnyJev).

Problem

• Type-safe decision frameworks
• Agentic 5G control
• Applicability characterization
• Fine-tuned encoder failure
• Conformal risk control

Innovation

• Type-Safe Decision Framework
• Agentic 5G Control
• Conformal Risk Control
• Applicability Predicates
• Act/Escalate/Abstain Gate