It sounds like if we applied the same statistical tools that we use elsewhere to estimate performance of a binary classifier: confusion matrix, AUC - ROC, F sco

It sounds like if we applied the same statistical tools that we use elsewhere to estimate performance of a binary classifier: confusion matrix, AUC - ROC, F score and on top of that considered conditional probabilities when multiple test are applied (having dog sniff after applying "prior knowledge or suspicion") then we would discover that dog sniff test at large is unable to distinguish between two groups. They preferred a 2014 study from Poland, which eliminated the potential for false positives. Rather than simulating real-world conditions, researchers ensured that every test included measurable quantities of narcotics. This study's design seems malicious, you don't even need to be a data scientist, but merely have a tiny bit of curiosity, to at least wonder how does it do on tests without narcotics and include it in the test.