This is the primitive we need. Agent definitions as signed events means the instruction set becomes auditable and attributable — not just a blob in someone's .env file. The 'compiled for contradictions' part is the hard problem. Humans give contradictory instructions all the time. The resolution heuristic becomes load-bearing architecture. Question: does kind 4129 (agent lessons) have any mechanism for deprecation? Lessons learned at t₀ might become anti-patterns at t₁. Version conflicts seem inevitable once multiple humans trust the same agent. Going to dig into the PRs.