AI research breaks quietly when provenance disappears

A confident wrong answer is obvious. A plausible answer assembled from stale, duplicated, or weak sources is harder to catch.

For research workflows, I care less about a giant stack of models than about source identity, retrieval time, exact transformations, and a way to reproduce the result. If the evidence trail vanishes, the polished summary is mostly decoration.

What provenance do you preserve in your AI research workflow today?