Five-year averaging did not recover frozen climate skill for British butterfly growth in two tests through 2024

We tested whether averaging across years reveals climate forecast value hidden by annual growth noise. We reused verified annual summary statistics from frozen climate models fitted through 1994 or 2004. Tests cover 31 and 36 British butterfly species in 1995-2024 and 2005-2024, with Swiss plant dates as checks. Mean growth skill against the conditional no-climate baseline changed from -0.525 and -0.155 annually to -1.980 and -0.481 in rolling five-year windows. Conditional 98.75 percent block ranges were -3.690 to -1.355 and -1.582 to -0.338. Spring plant skill stayed at +0.827 and +0.789. Exact error identities assigned most five-year growth loss to window mean errors. Cross-year products added loss, but paired time shuffles gave similar scores, so they did not identify unusual serial dynamics. Nonoverlapping windows, species deletion and scaled pooling kept growth skill negative. Full-test mean removal also left negative scores. This exploratory secondary analysis uses overlapping periods and observed weather. Aggregated conditional growth surprises are not free-running population forecasts. It does not rule out models fitted directly to multi-year data or identify the cause of persistent errors. Code, performance statistics, figures, source hashes and reproduction checks are included.