Active Learning for Interactive Prompt Clarification in Safety-Critical PLC Code Generation

This document provides the companion reproducibility package for the manuscriptActive Learning for Interactive Prompt Clarification in Safety-Critical Programmable Logic Controller Code Generation. It delivers (i) a complete implementation-level specification of the IPC+AL governance architecture and (ii) a transparent, task-scaled replication workflow for benchmarking uncertainty-governed synthesis against IPC and a baseline prompting suite including RaR, APE, AP, ReAct, and MTPR. To ensure independent reproducibility, the package consolidates full prompt specifications, hard-gate definitions, acquisition policies, and artifact schemas used in the study. Executable prompt instances are provided for two industrial case studies—Batch Mixing Process and Robot Pick-and-Place—across IEC~61131-3 dialects. By fixing the uncertainty governance protocol and varying only task instances, external researchers can deploy IPC and IPC+AL on alternative PLC benchmarks while preserving methodological consistency. The documentation further formalizes the end-to-end evaluation stack, including LLM-in-the-Loop validation with interchangeable validator committees and Human-in-the-Loop expert adjudication. Inter-rater reliability is computed at the gate-decision level using Cohen’s kappa and Krippendorff’s alpha, strictly across tasks rather than across prompting methods, thereby preserving the rater–item structure required for statistically valid agreement estimation. Overall, this package functions as a standalone replication and audit manual. By externalizing uncertainty thresholds, acquisition rules, and fail-closed emission criteria, it enables regeneration of all intermediate artifacts, recomputation of statistical metrics, and verification that uncertainty-governed Active Learning consistently improves safety admissibility and reliability under controlled, auditable conditions.