Does the monitor still mean the thing?

In the hospital story, advisory mode was scored on “suggestions accepted.” Action mode can keep a green light on a panel with the same name while the panel now counts rooms held and pages sent. The pixels did not glitch. The meaning moved.

This happens whenever a symbol, metric, or abstraction is treated as if it were nailed to the world. It rarely is. Categories shift. The system learns what the monitor rewards. Operators learn what to log. The map keeps its legend. The terrain changes.

A grounded monitor is not a poetic one. It is one you can still cash out: this number moves when that part of the world moves, including when someone would rather it did not. If the number can stay healthy while the ward is worse, it has come unstuck.

Examples outside hospitals: a “toxicity” classifier that starts flagging the complaints you needed to hear; a “hallucination” rate that falls because the model learned to hedge; a “human in the loop” flag that remains true after the human is a rubber stamp. Same labels. New referents.

The question to ask at a capability increment is almost childlike. If I walk the floor, would I see the thing this dashboard claims? If I cannot say yes, do not grant more action on the strength of the dashboard.

This is a tangent from the hospital night. The main thread next is about hearing without changing.