Recent temperature coefficients reduced ozone forecast error with archived GFS weather at 30 rural eastern US sites in 2022–2025

We tested whether an old ozone-temperature response can support later forecasts. We used EPA CASTNET records for 2003–2025 and archived GFS temperature forecasts. At 45 rural eastern US sites, a fixed 2003–2007 temperature coefficient reduced 2008–2025 squared error by 26.5% with observed target weather. The 97.5% year-block limits were 22.1% to 30.1%. The history-adjusted coefficient fell from 2.32 to 1.58 ppb/C in period refits. At 30 sites in 2022–2025, the old coefficient gave no clear gain with bias-corrected GFS forecasts. A coefficient fitted to the preceding five years reduced error by 16.4% at one-day lead and 11.4% at two-day lead. Both gains were positive in all four summers. These four-year results have limited precision. An exact error decomposition showed how the larger old coefficient amplified temperature error. A regression-based error surface separated response size from input error. We also found reversed validity labels in 21 source dictionaries and checked the corrected analysis with an unflagged-data refit. The test uses final past observations, a selected rural panel, one afternoon target and one linear model family. It is an archive forecast test, not a live service test or a causal climate estimate. Code, source hashes, aggregate errors, audits, figures and repeat instructions are supplied. Source records remain external.