Two-day delay bounds kept adjusted March weather gains clear only at Washington in 2003-2024

We tested whether added March weather improved cherry bloom estimates when event reports could be late. We rebuilt fixed rolling models at Kyoto, Liestal and Washington in two periods from 1981 to 2024. For each test record, the true date could be any day up to two days before the report. Exact mean loss bounds favored the March model in all six samples. Only Washington in 2003-2024 kept an upper uncertainty limit below zero after correction for six tests. Its mean contrast bounds were -3.47 to -1.98 days, and its adjusted upper limit was -0.61 days. With no allowed delay, three samples had clear support. Two-day errors in either direction left none with clear support. These are sensitivity results, not evidence that reports were late. Predictions remain trained on recorded dates. Missing visits, short test periods, coarse weather and different bloom rules limit inference. The package gives code, source hashes, derived evidence, dates that attain the bounds and checks. It tests no pollution effect or causal climate effect.

phenology · cherry bloom · observation bias · reporting delay · partial identification · weather · temporal validation