We have studied the only the supermassive black hole binary black hole
datasets in challenge 3.2.  Our general approach has been to conduct the
data analysis within the framework of Xspec 
(http://heasarc.gsfc.nasa.gov/docs/xanadu/xspec/index.html) a
publicly available, and widely used in spectral analysis of high-energy
astronomical data.  The gravitational wave component is in the early 
experimental stage, and is not part of the standard publicly available 
software.  

For these gravitational-wave searches, we applied Xspec's standard fitting 
tools, primarily applying the tempered Metropolis-Hastings Markov-Chain 
Monte-Carlo (MCMC) algorithm.  For this round we had modest goals, which
we have only partially fullfilled. We made extensive use of the MLDC 
dataset generation pipeline codes with the hope to avoid excessive 
difficulties with conventional consistency for source parameters and 
LISA specifc, including the FastSpinBBH code and Synthetic LISA.  The latter
is code is somewhat ill-suited to likeihood estimation, intented as a general-
purpose modeling tool for LISA rather than a rapid gravitational-wave response 
calculator. Though we intended this as a "zeroth" step to provide a baseline
for verifying the consistency of a more efficiency with the MLDC pipelines, we
have not yet succeeded in producing an efficient and compatible code. Therefore
we utilized synthLISA directly for our black hole searches. Even so, we did not
leave time to test the code thoroughly on the training data, an oversight which
is likely to be evident in the challenge evaluation.

Our slow code limited us to analyzing rather short samples of the
data, each containing 1400-1500 frequency bins.  For BH1 we studied
a 500000 s segment of data, restricting initially to frequencies less 
than 1.5mHz, and then more narrowly to the frequencies which the signal 
was significant. A number of chains were run, and analyzed by hand in an 
effort to improve the fit. Our pedestrian approach, and the very limited
time we allowed, left a significant residual with an estimated reduced 
$\chi^2$ of nearly 5, even over our limited data sample.  

We applied a similar approach to BH2, using 1.2 million seconds of data 
below 1.25mHz . In this case our fit is considerably worse, 
with reduced $\chi^2$ near 20.  It is likely that the low frequency
portions of this signal are non-negligibly impacted by BH3 which is merges
only slightly later, with comparable signal frequencies.  Though Xspec 
allows simultaneous searches for multiple sources, we made no attempt for
a combined fit this time.  We did eliminate a larger portion of the 
low-frequency part of the signal for most of the parameter search.
 
The last merger we found was at a somewhat weaker signal, making it easier to
fit.  This time we estimated a reduced $\chi^2$ of about 1.1 for our sample
taken from 1.5 million seconds with frequencies limited to below 0.8 mHz.

In total our search utilized a number of short chains (<1000steps) and a few 
longer chains (up to 20000 steps) with a estimated total of less that 200000 
likelihood evaluations conducted over 1.5 days on a MacBook Pro.

We plan to continue development of out search tools with hope of a more 
credible effort next time.