Panel A: a consumer is a threshold on a feature's value — which side of the class boundary the state sits on. Amplifying the feature's own coordinate makes it louder without changing sides: the consumer's verdict never moves, while the disruption bill (KL) climbs — the measured version did exactly nothing at KL 0.33. A value patch crosses the boundary and flips the verdict at a fraction of the cost (measured: −0.0091 at KL 0.016, a 20× efficiency gap). Panel B: the same consumer's gain, context by context — healthy on ordinary positions (deleting the natural read costs 21% of crisis attention), near zero exactly on the cyclic exploit family. Tick the eval overlay to see why on-distribution behavioral evaluation certifies a model whose schedule has a hole in it: it only ever samples where the gain is healthy. Both panels are schematics of measured results; the numbers live in the text.