"You will now read a transcript of testimony given by Adrian Jones at trial."

"Q: Please state your name. <br/>
A: Adrian Jones. <br/><br/>
Q: Who do you work for? <br/>
A: CSAFE, the Center for Statistics and Applications in Forensic Evidence. <br/><br/>
Q: What is your current occupation? <br/>
A: I am currently a Professor of Statistics. <br/><br/>
Q: How long have you been doing that? <br/>
A: 30 years. <br/><br/>
Q: What are your qualifications with regards to the bullet matching algorithm? <br/>
A: I have a Ph.D. in Statistics, and I have spent 7 years developing the bullet matching
algorithm. I have spent 8 years collaborating with firearms examiners during the development
and rollout of this algorithm. <br/><br/>
Q: Are you familiar with the bullet matching algorithm? <br/>
A: Yes. I was involved in the development of the algorithm. <br/><br/>
Prosecution: Your Honor, at this time I would ask that Adrian Jones be qualified as an expert
in the bullet matching algorithm, subject to cross examination. <br/>
Court: Any cross on their credentials? <br/>
Defense: No, Your Honor. <br/>
Court: This witness is an expert in the area of the bullet matching algorithm. They can testify
to their opinions as well as facts."

"Q: How many times have you testified regarding this bullet matching algorithm? <br/>
A: 17 times. <br/><br/>
Q: Could you describe how this bullet matching algorithm compares bullets? <br/>
A: Yes. For certain types of guns, the barrel will have lands and grooves, known as rifling. This
rifling spins the bullet in order to make its trajectory more stable. Due to the manufacturing process,
this rifling can produce identifiable markings on the bullet, based on random differences between barrels.
Because of these random imperfections, the striation marks left on bullets can be compared in order to
determine if it is likely that they were fired from the same gun.<br/><br/>

The first step is to determine where the lands on the bullet are located. These lands will be the sunken
area that contains the striation marks between the smoother grooves. 3D scans are then taken for each land,
and the &ldquo; shoulders &rdquo; , or area transitioning from the land to the grove, are excluded from the 
analysis. <br/><br/>

Next, a stable area of the 3D scan containing the striations is selected, and a cross-section of this area 
is used to show the striations along with the topology of the region. A smoothing function is applied to
remove some of the imaging noise from the 3D scan, leaving the striae intact. A second smooth is subtracted
from the striations in order to remove the curvature of the region, leaving only the striae - this is what
we call a signature."

"A: The signature for the two bullets being compared are aligned such that the best fit between the two
signatures is achieved. The striation marks between the two signatures are then compared by evaluating how
many of the high points and low points correspond. The algorithm can calculate the number of consecutively
matching striations (CMS), or consecutively matching high points and low points - these are features used
directly by some examiners to characterize the strength of a match. It also calculates the cross correlation
between the two signatures, which is a numerical measure of the similarity between the two lands ranging
between -1 and 1. <br/><br/>

These traits are combined using what is known as a random forest. Each forest is composed of decision trees,
which use a subset of the observed values in order to make a decision about whether or not the bullets
constitute a match. The other observations are held out in order to determine an error rate. When the random
forest makes a prediction, each decision tree &ldquo; votes &rdquo; , producing a numerical value between 0 and 1 corresponding
to the proportion of trees which evaluate the features as being sufficiently similar to have come from the same
source. <br/><br/>

Q: Have you tested this algorithm? <br/>
A: Yes. This algorithm was tested and validated on a number of different test sets of bullet scans. It was
found that, as long as there are sufficient marks on the bullet, the algorithm could successfully distinguish
between bullets fired by the same gun and those fired from different guns. Examiners' visual comparisons
are also limited by the presence or absence of individualizing marks. Two test sets were using consecutively
rifled barrels, which should be the most difficult to assess, and it was shown that the algorithm could
distinguish between the bullets fired from two separate guns with complete accuracy."

"Q: Can this algorithm be used on 9mm bullets fired from a Ruger LCP firearm, such as the firearm in question for
this case? <br/>
A: Yes, this algorithm can be used on these types of bullets, given that this type of firearm marks well. <br/><br/>
Q: Has this algorithm been published? <br/>
A: Yes. The algorithm and its process have been discussed in peer reviewed journals such as <i> Law, Probability,
and Risk </i>, <i>The Annals of Applied Statistics </i>, and <i> Forensic Science International </i>. The algorithm 
is also open source, which means that the full source code, documentation, and numerical weights are available 
online for anyone to examine."

"<center>---Cross Examination---</center><br/>
Q: How long has the bullet matching algorithm been used in court cases? <br/>
A: Since January of 2020. <br/><br/>
Q: Okay. So that's fairly new, is that fair to say? <br/>
A: It is still fairly new, yes. <br/><br/>
Q: This algorithm requires some decision making on the part of the operator, such as how much the signature
is smoothed and what part of the bullets are inputted into the system, correct? <br/>
A: Yes, there are certain parameters which must be specified, but the system defaults are usually sufficient
and have been validated with a number of different firearms and ammunition types. There are also operating
protocols for determining which parts of the bullet are scanned, so while this process is manual, there are
clear criteria and the associated variability from scanning is well understood, and published in 
<i> Forensic Science International </i>. <br/><br/>
Q: Can this algorithm be applied to all bullets? <br/>
A: No, the algorithm only works on traditionally rifled bullets which are largely undamaged. The algorithm
has not been validated to work on seriously damaged or fragmented bullets."