I don't understand why the source code itself is the distinction between something being verifiable or non-verifiable. It seems the results of the machine under standard testing would.

If we stood your biocube up to the similar tests that any real device would go though (being presented with samples with a known alcohol percentage) your system would get some right, some wrong and you could then say, of a sample size X, the device accurately determined the BAC of Y samples with a accuracy window of Z.

This data (as well as the device itself - generally) would be available for use during any case which this was used in. For example, if (for the sake of argument) the initial tests came out with 100% accuracy - the defense would have the device available to conduct another set of tests (And another, and another) until they could show evidence that the measuring device itself was faulty - or at least inconsistent. In an ideal system, this would lead to reasonable doubt that the test worked, and the defense would win the case.

I don't see why the real device wouldn't be put up to the same test regardless of what drives the device (hardware, software, vanilla pudding, whatever) to determine if it is providing data in an accurate enough way to be used as evidence. If the device is known to have inconsistencies like the summary states, it would seem like testing would be able to show this (caused by software or not) - hence reasonable doubt of any readings. It's up to us (as people who would be on a jury) to determine whether Y correct out of X samples is accurate enough to convict someone in our society.

I'm all for the company to release the software - but I'm not 100% convinced that a software bug being shown in court would cause a device like this to fail on the above criteria.

Lets say (for the sake of argument) that the device consists of:
a physical sensor of some sort
a way to adjust the input of the sensor (since there's always variance on those sorts of things),
software which reads this adjusted input and displays the result (perhaps averaging samples over time, etc, etc).

A normal production model would go through the testing described above, have the input adjusted until the results displayed by the software met standard criteria, then the adjustment would be locked and the unit would ship.

Even in the case where the software could have a significant bug, like always adding .15 to every read, the system as a whole would be adjusted to compensate for this.

And while the bug in this case would definitely be used in court - I have the feeling it would be more of the "get an guilty person free" rather than the "showing true reasonable doubt of the system" type of use - since the device as a whole would be accurate and precise, whether that bug existed or not.