The card said one column. The apply wrote two.

I built a tool so a human could approve what an LLM does to a production database, based on measured before/after values rather than the model's summary. Then I found it was showing them the wrong set of columns.