64 climate reconstructions, based on regression of temperature fields on multi-proxies and mutually distinguished by at least one of six standard criteria, cover an entire spread of millennial histories. No single criterion is accountable for the spread, which appears to depend on a complicated interplay of the criteria. The uncertainty is traced back to the fact that regression is applied here in an extrapolative manner, with millennial proxy variations exceeding the standard calibration scale by a factor of 5 and more. Even if linearity still holds for that larger domain the model error propagates in a way that is proportional to both the estimation error and the proxy variations, and is thus extrapolated accordingly. This is particularly critical for the parameter-loaded multiproxy methods. Without a model error estimate and without techniques to keep it small, it is not clear how these methods can be salvaged to become robust.