Frozen weather models ranked poor British butterfly growth years above chance in 2005-2024

We tested whether weather models can rank poor growth years when growth magnitude forecasts fail. We used annual forecast outputs from a prior SCI study, with no new field records. For 36 British butterfly species in 2005-2024, the mean AUC for the lowest fifth of growth anomalies was 0.631. Its shared-year two-year block interval was 0.565 to 0.688. Mean squared-error skill was -0.155. AUC for growth below the demographic baseline was 0.608 among 35 eligible species. Tail AUC stayed positive within both test decades and after separate linear trend removal. Fixed weather terms and an earlier training split also retained full-period discrimination. Joint forecast-year permutations supported alignment, but a circular-shift test had only 20 possible shifts and gave p = 0.05. The result qualifies a claim of no growth forecast value: poor magnitude prediction can coexist with modest rank information. Growth below a fitted baseline is not necessarily population decline. Tail labels use test-period ranks. The study does not establish a causal climate effect or an operational warning system. Fits, data selection and source observations overlap with prior work. Code, source hashes, aggregate results, checks and figures are included.