A ten-year correction reduced later butterfly growth prediction error, but its added climate gain was uncertain

We tested whether calibration can explain failure of old climate models. We used published annual predictions from models fitted through 1994. We fitted a slope and mean correction in 1995-2004 and tested it in 2005-2024. For 32 UK butterfly growth units, the skill against the original baseline changed from -53.8% to +14.1%. The gain against a baseline with its own mean correction was 7.1%. A test-year block interval was 4.6-9.6%. When calibration years were also resampled, the interval crossed zero in 31 complete units. The fixed five-term weather model gave no added gain. Swiss spring and later plant dates retained large gains under calibration resampling. Butterfly flight dates gave a small uncertain gain against the corrected baseline. Thus, original model failure can contain a correctable signal, but a ten-year record did not establish a reliable growth correction. This is a retrospective analysis of prior data, not independent field validation. Short records, fixed source fits, changing surveys and observed-weather inputs limit inference. Code, aggregate results, figures, hashes and reproduction instructions are included.