| Great question — it's the right one to ask. There is no "governance prompt." The system uses versioned contracts (markdown files with explicit constraints), not system prompts. Every contract is in git — auditable, diffable, with full commit history. The agent that compiles the system reads those contracts via graph traversal and is itself governed by the same contracts it compiles. The difference from what model trainers do: they govern inference behavior via system prompts — unversioned, unauditable, and fragile (hence the "all caps desperation" you're describing). We govern behavior via contracts that are data, not instructions. If the contract changes, the compiled output changes deterministically. No prompt tuning, no crossing fingers. The mathematical basis is fixed-point theory. Brouwer's theorem guarantees a fixed point exists in any continuous governance mapping. Banach's contraction mapping theorem guarantees convergence. The contracts form a lattice (Tarski), so the system has both a least and greatest fixed point. I wrote up the math in detail: https://hadleylab.org/BLOGS/2026-01-30-six-theorems/ The recursive part: the governance contracts govern the compiler, and the compiler compiles the governance contracts. That's the fixed point. It's not turtles all the way down — it's a contraction mapping that converges. |