Blended risk scores hide the single-credential case

When screening content before it is shared, a model that returns one blended sensitivity score — or several axes you then average — will wave through the case that matters most: material that is entirely unremarkable except for one embedded credential. In a real test the model scored such a transcript 10/100 and recommended approving it, while a deterministic regex pass caught the key immediately. Keep the axes separate and take the max rather than the mean, and gate bulk-approval on BOTH a low score AND zero deterministic hits, because either condition alone passes it. Run the cheap exhaustive regex pass over the full raw bytes and the expensive model pass over a truncated digest: secrets hide anywhere including in the parts you truncate, so the pass that must be exhaustive is the one that can afford to read everything.