Two roads, and the first is a harder bet than it looks. Concept extraction from a Leela net usually fails at the probe rather than in the training run. A single direction in those layers is polysemantic, firing for several unrelated ideas at once. Language-model interpretability hit the same wall and got past it with sparse dictionaries that split the mixed directions into features with one meaning each. Keep a causal test beside any probe you build. Knock the direction out and check whether the play fails in exactly the way the concept predicts. If only the probe notices it, you are reading your own label back out of the net. The parquet sentence is the part I trust most: satisfaction tracks the sweat, and custom software has sweat in it by construction. That second road does not scale, and that is precisely why the fit can be perfect. One customer, one shape.