Anyway, the way he explained  it to me was that the list of categories that can lead to false alerts is simply too extensive. Similar to polygraphs, where the best "interpreters" will tell you that the machines have critical limitations, so the reliability of the results rests more with the operator, then with the equipment. And, since it's their interpretation of what the squiggles mean, it becomes more an art than a science. So, you don't have to be genius to imagine all the ways it can go wrong.  It's not science. Same with dogs.

First of all, the propensity for "sniffing" certain contraband is not inherited. There's no way to know which dog will be effective. Mutts are every bit as dependable (or, undependable) as pure breds. Secondly, certain dogs sniff only certain items, and the "strength" of the alerts fluctuate from day to day. Some days they'll alert positively 9 out of ten times, other days 6 out of ten. Thirdly, each dog has a different emotional attachment with their handler. When certain dogs don't alert on an actual contraband item, they sense the disappointment in the handler, and they start alerting on anything. Don't discount the fact that we're dealing with senses, and not cognitive abilities.  And we know so little of how that works. Basically, the dog doesn't specifically want to find the contraband, it wants to please its handler. So put the dog desire, together with senses, and one can imagine there's a lot of room for mischief. In other words, the handler is sending signals he/she doesn't know are being sent.

Finally, and very important, the dogs do successful alerts in streaks. So, if a dog, let's say has a 90% success rate over a large number of trials, he/she may be successful 10 out of ten times, three times in a row, and then 6 out of ten, then 7, and so on. So, obviously, sniffing dogs are erratic, unreliable, and have a much greater propensity to go on alert than not. That means that if you have actual contraband in a building and test the dog ten times, they may score quite highly. Then if you give them ten tests, with no contraband, they will never "stand down" to the degree they alert to in the first test. They will always alert to a greater degree when there is actual contraband, then not alert when there is no contraband.