Parent: theory · Reference: good-regulator-theorem
Three ways "control = modeling" lands on the harness:
- The positive reason for the observability ceiling (convergence-needs-an-observable-target): control requires a model; the unobservable part of the system (hidden variable / mood) can't be modeled → can't be regulated → you converge only to the observable marginal. "Must model" ⟺ "can't converge on what you can't model."
- The LLM is the model: an LLM-agent regulating a task is, by this theorem, necessarily a model of the task domain — its common sense / world-knowledge is the model it regulates with (tacit-spec-as-spec-compression). Regulation quality = model fidelity.
- Upstream of "control = destroy variety / converge" (control-methods, recursion-convergence-contraction): variety-matching (Ashby) → matching requires a model (Good Regulator).