Early winter temperature cut Great Lakes ice prediction error in a frozen 2011-2026 test

We tested whether a winter temperature response retained prediction skill in five Great Lakes. Separate linear models fitted to 1973-2010 were held fixed for 2011-2026. November-December temperature reduced equal-lake mean squared error by 38.10% against the prior ten-year mean. Conditional 95% three-year block limits were 14.11% to 52.62%. A correction for three early information windows retained a positive interval. November alone gave no clear gain. November-January temperature gave 60.82% gain. Full November-February temperature gave 74.41%, but most ice peaks were already past; this is a diagnostic, not a forecast. A recent-only 2019-2026 check did not resolve the November-December gain. A conditional error budget states the weather precision needed under independent unbiased errors. Revised inputs, unknown release delays, coarse temperature boxes, 16 shared test winters and absent current-ice controls prevent a claim of verified operational skill. Code, aggregate evidence, source hashes, uncertainty draws and retrieval steps are supplied. External records remain outside the package.