Trying to make use of JointSNVMix2's output

[lethalfang]

I am testing JointSNVMix2 and used the joint_snv_mix_two workflow:

python jsm.py classify joint_snv_mix_two \
/PATH/references/human_g1k_v37_decoy.fasta \
/PATH/Normal_Sample.bam \
/PATH/Tumor_Sample.bam \
/PATH/trained.exome.parameter.cfg \
/PATH/exome.jointsnvmix2.tsv

It has an output for every single chromosome coordinate, so I tried to screen for somatic mutation calls:

cat exome.jointsnvmix2.tsv | awk -F "\t" '$10+$11>.99'

And I get a bunch of things like this:

chrom	position	ref_base	var_base	normal_counts_a	normal_counts_b	tumour_counts_a	tumour_counts_b	p_AA_AA	p_AA_AB	p_AA_BB	p_AB_AA	p_AB_AB	p_AB_BB	p_BB_AA	p_BB_AB	p_BB_BB

1	12907379	T	C	156	12	125	24	0.0000	1.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000
1	12907380	C	A	156	10	128	24	0.0000	1.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000
1	16890771	G	A	237	12	267	25	0.0000	1.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000
1	16890800	A	T	53	2	44	8	0.0000	1.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000
1	16893070	C	T	466	34	457	39	0.0000	1.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000

All those have p_AA_AB (i.e., probability of AA genotype in normal --> AB genotype in tumor) of 1, but when you look at the counts, they hardly look like something with a somatic probability of 1.

Is there a model in JointSNVMix (there are a few) that works better than this one?
When you assign probability=1 for a bunch of sites with minimal evidence for somatic mutation, then this model for somatic mutation calling is useless for our sample.