Why Ashby is the trunk of this whole thread — the mapping from his conservation laws onto the harness:
| Ashby | our thread |
|---|---|
| Requisite Variety (suppression ≤ regulator capacity) | convergence rate ≤ controller capacity = verification bandwidth (recursion-convergence-contraction / verification-is-the-new-latency) |
| regulation = destroy variety | the eight controls = ways to destroy the executor's variety (control-methods) |
| Good Regulator (must model) | the observability ceiling (convergence-needs-an-observable-target) |
| ultrastability (search viable configs) | the CSP/collapse/MCTS search (csp-formalization, stages-gates-bound-the-search) |
| essential variables within limits | constraint satisfaction / P×C |
The whole story — convergence capped by controller capacity, the eight controls as variety-weapons, search as finding a viable configuration — sits on Requisite Variety. Remove Ashby and half of it collapses.
Honesty about our use: we lean on Requisite Variety as a ceiling concept (a real necessary condition — "convergence can't outrun controller capacity"). Legitimate as intuition, but we don't actually measure an LLM's "verification-bandwidth variety" — so it's a rigorous concept used as a necessary condition, not a quantitative law for the harness. Claiming otherwise would make us the fetishists.