Active learning for interactive prompt clarification in safety-critical programmable logic controller code generation

Large language models (LLMs) enable automated program synthesis, but their application in safety-critical Industrial Control Systems (ICSs) is constrained by specification imprecision and non-deterministic logic drift. This study addresses these challenges by integrating Active Learning (AL) as a decision-governance layer for Interactive Prompt Clarification (IPC) in Programmable Logic Controller (PLC) code generation. We treat requirements-to-code synthesis as a soft-computing problem involving reasoning under uncertainty, in which missing specifications are quantified as first-class uncertainty signals. The proposed methodology, IPC + AL, implements a policy-level controller that measures logical ambiguity using entropy derived from multi-validator disagreement. This approach transforms code synthesis into a budgeted value-of-information (VOI) policy that determines when to request clarification and when to terminate the synthesis loop deterministically. By embedding uncertainty quantification with fail-closed gating, IPC + AL blocks candidate-code emission unless the predefined safety and dialect gates pass. Controlled feasibility studies on Batch Mixing and Robot Pick-and-Place show improved gate pass rates and committee-assessed artifact scores relative to IPC and standard prompting baselines. Cohen’s κ and Krippendorff’s α are used as evaluator-agreement measures and are not interpreted as evidence of executability, physical-process safety, or certification readiness. The results position IPC + AL as an auditable pre-commissioning governance layer that requires downstream vendor compilation, simulation, hardware-in-the-loop validation, formal verification, and expert approval.