Ship with a provenance trail

A verified result that cannot show its work is not meaningfully different, to an outside reader, from an unverified one. The provenance trail is what turns "we checked this" from a claim you make into a claim a reader can independently confirm.

At minimum, that trail records which model or models were involved in producing a piece of work, what was asked of them, and a same-day, append-only note of what happened, not a reconstruction written later from memory. Append-only matters specifically: if a line turns out to be wrong, the correction is a new line, not an edit to the old one, so the record shows what was actually believed at the time rather than a cleaned-up version of it. A dashboard that just says "current status" with no history is not a provenance trail, it is a snapshot that will eventually contradict itself.

The discipline behind this is the same one that governs any regulated or high-stakes engineering practice: nothing ships because it sounds right, it ships because there is a record of what was checked, when, and by what method, and that record survived an attempt to poke holes in it. A same-day ledger line costs a few minutes. What it buys is a log that is queryable evidence later, instead of a vague sense that the work was probably careful.

Start today: for the next AI-assisted deliverable you produce, write one line the same day, before you move on to the next task. What model, what you asked it to do, what you personally checked before trusting the output. Keep that line even after the deliverable ships. Especially after it ships.