I'm not sure how the algorithm works but perhaps the positives need to stack to a certain amount before the CRAM delivers the yellow or red light. If it takes a certain number of readings per second and it needs 1,000 positive readings in a certain time frame to validate a "decision" about the driver's vigilance, the chances of having a correct judgment by the device are elevated.
I'd have to look at the data again to know how many times the device read alert when the driver is sleepy... That's the worst case. If it errs I would hope that it does so on the side of being too cautious.