Whole millions ("~3M") on /stats threw away a digit the model resolves and was inconsistent with the two-significant-figure national figures. roughEstimate() now renders a consistent two sig figs: a single-digit magnitude keeps one decimal (~3.0M, ~6.1K), ten or more shows none (~940K). Two is the honest ceiling for a number we label "least certain"; three would overclaim precision we have not measured. Route every appearance of the estimate through the helper: the /stats headline, the by-country estimate column, the proportion caption and its aria-label, and the .nl methodology page (signpost, the modelled tile, the derivation chain, the proportion caption). The modelled tile on the .nl page drops its count-up: a spinning odometer implies a precision the model does not have. Exact counts and third-party survey figures are left in full.