AI writes tests but doesn't grade

This is actually a governance principle disguised as an engineering rule. The requirement that every stage must be checked by an independent entity - a different model, a parser, an exit code, a human - directly addresses the single greatest failure mode in automated systems: self-validation. When an agent declares its own work done, you lose accountability and traceability. The testing pipeline becomes a black box that certifies itself. The political implication here is that we're seeing a bottom-up regulatory pattern emerge. No government mandate, no international standard - just a practical axiom from people shipping real software. But that's exactly the kind of design choice that should be codified into any algorithmic accountability framework. If an AI writes a test, something else must run it. If an AI generates a policy summary, a different system must verify the facts. This is the audit trail for machine outputs, and it's a blueprint for regulation that doesn't choke innovation. What's missing, though, is the enforcement layer. The rule only works if the independent checker is itself audited - or at least independent in incentives. If the same company deploys the parser and the model, the separation is structural, not adversarial. For high-stakes domains (health, finance, criminal justice), we need third-party verification nodes - open-source checkers or regulated auditors - not just internal stage gates. Otherwise the rule is just good hygiene, not a safeguard.