Probability scores kept Swiss plant timing gains but gave no British butterfly growth gain in two tests from 2005 to 2024

We tested whether climate models that fail to improve mean forecasts can improve probability forecasts. We used published annual statistics for Swiss plant dates and British butterfly dates and growth. Ten years set forecast bias and spread. The next ten years supplied scores. Student t forecasts gave mean CRPS skill of 0.559 and 0.458 for spring plants in 2005-2014 and 2015-2024. Later plant skill was 0.261 and 0.279. All four plant intervals were above zero after correction for eight tests. Butterfly date skill was 0.014 and 0.007, with intervals across zero. Growth skill was -0.086 and -0.060. Only the first growth interval stayed below zero after correction. Gaussian and empirical residual forecasts kept the same plant-growth order. A location-scale score decomposition assigned most plant gains and growth losses to location. Later spring intervals missed 24.2% of dates on the early side and 1.1% on the late side, despite positive relative skill. Thus proper probability scores did not recover growth gain in this scope, and relative skill did not ensure calibrated tails. This is new analysis of overlapping published records, not new field evidence. Short calibration and test periods, observed target weather, fixed climate increments, changing site coverage and conditional uncertainty limit inference. Code, performance statistics, exact input hashes, and reproduction checks are included.