Frozen temperature models did not improve British butterfly timing-gap forecasts in two tests through 2024

We tested whether published frozen climate models predict differences between annual timing anomalies, rather than a shared early or late season. We used signed annual statistics for 43 British butterfly species in 2005-2024 and 35 in 1995-2024, with Swiss spring plant phases and butterfly growth as checks. An exact squared-error identity separates shared-mean and pair-gap skill. Butterfly timing-gap skill was -0.154 and -0.189 against no gap change. Joint two-year block ranges were -0.276 to -0.062 and -0.302 to -0.081. Shared-mean timing skill was 0.417 and 0.631. Even hindsight global amplitude correction gave maximal gap skill below 0.007. Plant gap skill was 0.261 and 0.279, much less than pooled date skill above 0.80. Fixed-window plant gap point scores were higher. Species deletion, matched growth sets and bias removal retained the butterfly gap failure. Climate forecasts beat the tested linear-trend gap baseline, so baseline choice matters. This secondary analysis uses overlapping test periods and different site networks. The gaps are differences of aggregate anomalies, not local ecological mismatch or fitness loss. Code, exact input hashes, aggregate results and reproduction instructions are included.