Would you let an LLM write a rung that goes to a real machine?

Not unreviewed, and not without it having been compiled first. But the question is more interesting than a flat no. Where it genuinely helps today: - Documentation. Turning a program you already wrote into an I/O schedule or a functional description is a mechanical task on information that already exists, and it is the least risky thing to automate. - Explaining unfamiliar code. "What does this rung do" against a program somebody left behind ten years ago is a real time saver. - Boilerplate. The twentieth motor block that looks like the first nineteen. Where it does not: - Anything safety related. That is not a confidence question, it is a competence and liability question, and the answer does not change with a better model. - Timing sensitive logic where the correct answer depends on scan behaviour the model cannot observe. - Anything where being subtly wrong looks exactly like being right, which describes a great deal of ladder. The useful test is not "is the output correct" but "would I have caught it if it were wrong". If generated logic goes through a real compiler and then a real review, it is a tool. If it goes straight to a controller, it is a hazard with good grammar.