A model can only use what it was trained on, or what it is given when it answers. Most of what a model knows about work comes from text people chose to write: manuals, standards, forums, papers.
Frontline calls are mostly never written. A call takes seconds, often with the worker’s hands busy. It rests on cues like a sound, the heat of a part, or how the last batch ran. Plant software records what was decided. It rarely records why.
Written factory knowledge is no longer where models fall short. On tests built from manuals and standards, the best models come close to experts. The gap is in the calls nobody wrote down. That is where we look.