Measuring Bounded Autonomy for an LLM Decomposition Planner on a Live 5G-Core Testbed

Abstract

Large language models (LLMs) are entering network-management control loops, yet their plans can fall outside the executable action space. Whether hard deterministic bounds around such a planner have real effect, or merely restate a definition, is an empirical question. We measure it on a live 5G-core testbed, where an LLM planner decides when to decompose control-plane procedures into stateless handlers inside a bounded loop: a JSON schema constrains the output, a projection-repair validator re-checks it, a deterministic fallback guarantees liveness, and every decision is persisted. Three measurements anchor the paper. First, the two tested models, run unbounded with the allowed action sets stated in the prompt, still violated intent-handler rules in 13 of 44 standard calls. Every violation broke an intent-handler combination rule that the deployed flat schema cannot encode (schema enforcement as deployed caught 0/13); a union schema encoding the rule removes the class by construction. Second, with the bounds in place, all 33 decoded live plans were feasible without repair, and a matched-condition probe put the in-distribution unbounded rate at one violation in 20 windows. Third, exploratory: on the design set the planner matched an author-specified selection reference that a threshold policy and three a priori codings each missed. Transfer to held-out scenarios remained partial (5-6 of 10), and the planner's two-clause distillate could replace the runtime planner itself. Together, the results suggest elicitation and distillation, not runtime control, as the LLM's role.