I'm a graduate student in a research lab that does research on solid state sensors for volatile organic compounds, which is essentially what breathalyzers are. I also do the programming of the instruments acquiring data (in Labview). I would find it pretty unlikely that the source code itself is especially hard to write, unless there is something unusual going on. Either the company is just stonewalling because they can, or the code either: A. Reveals some proprietary method of detection through the calculations performed or B. The corrections that must be gone through are so involved that they are actually worried that their sensor may be invalidated.
I wouldn't be surprised at all if it was B, as there is a reason that breathalyzers have to be recalibrated relatively frequently. It is pretty difficult to make a sensor that doesn't degrade over time, and correcting for this degradation can be quite complicated.

Although this repetition you speak of is important, it is far less beneficial than exponential. (0.993)^3 is actually less accurate (0.979), I would guess you meant to do (1-(1-0.993)^3) = 0.9999997. Also, re-measuring data does not give an exponential error decrease. Most signal to noise levels go up as the square root of the number of samples taken. This is assuming the error isn't systematic.