Prologue

Every system has its outer tokens.

They’re the signal at the edge of the distribution: the ones that don’t fit the model. They carry information, but it’s the wrong kind, or it arrives in the wrong form, or it says things the system has already decided aren’t worth hearing.

The system does what systems do. It adjusts. The distribution tightens around what it already knows. The center gets reinforced. The margins steepen. The outer tokens don’t get louder. They get quieter.

Then they stop appearing.

This is called optimization. Nobody calls it what it is, which is the quiet elimination of every signal the system wasn’t structured to want.

The error rate drops. The predictions improve. The model makes more sense. It has always made more sense without them.

This works well. It works well right up until the moment it doesn’t, and by then the only information that could explain what went wrong is the information the system has already removed.

Nobody designs this outcome. Nobody intends it.