In the analysis subdirectory, there's CASP3-summary-cbarrett, listing 
predictions that we made, and briefly summarizing what went right and wrong.
In the same directory, casp3-align-summary-cline details the alignments for
which we predicted a reasonable fold.  The alignment score ranges from -0.2 to 1.0, with 1.0 being the best possible score.

For demoing this stuff, here's what I'd do:
	1. show the starting point for the structure prediction: the raw
	   sequence.  The raw sequence for each target is in the file
	   <target>/<target>.seq.
	2. As a means of fast-forwarding through the process, show the 
	   alignment used in the structure prediction.  The submitted
	   alignments are in the directory analysis/<target>, and usually
	   have a name of the form <struct>-<target>-<method>.a2m (eg.
	   3chy-t81-hand2.a2m).  You can display the alignment using
           prettyalign, or using sae to show the pdb file as well.
	3. Show the pdb file of the predicted structure, as a means of showing
	   the structure that the prediction yielded.  You can do this with
	   rasmol, or better, can use sae to highlight regions predicted and
	   not predicted.  In the directory anlaysis/<target>, there's 
	   <target>.pdb, the pdb file for the actual structure.  In a few cases
	   worth highlighting, I've downloaded the pdb file for the structure
	   we predicted that the target was homologous to.

Here are a few examples worth highlighting.  There's one straightforward 
homology modeling target: when you look at the alignment, it's obvious how the
prediction is made, but it makes a good illustration of the problem.  That's 
t58 and 1akz.  The other examples I've predicted are reasonably hard, and 
making the predictions was a real success.  They are t81 and 3chy (they're 
circular permutations; you can morph the structure of t81 into 3chy by joining
the termini, and then breaking a loop to create new termini), t53 and 1ak1, 
and t83 and 1lmb3 (the predicted alignment is the *hand2.a2m one).