For T0248, we liked one of the Robetta models better than we liked the
models that undertaker was generating, so reoptimized it using undertaker.
The result is not great, but scores well with an unconstrained or
lightly constrained undertaker cost function.  Only model 1 uses a
Robetta model as a base.

The best-scoring model (try16-opt2) had (we believe) a broken
N-terminal helix, so we hastily patched the helix and reoptimized.  We
did not have time to get the helix to pack well against the rest of
the protein, and the result does not score quite as well as the model
with the broken helix, but we believed strongly enough in the
predicted helix that we are submitting our second-best scoring model
as model 1.

This is try17-opt2.