In short, they're accurate 90-97% of the time depending on the population. This means that 3-10% of the time they are classified as 'unknown'.
The COMT haplotypes are gamete phase-dependent. This is a fancy way of saying the series of 4 possible bases, ACGT, used for speed assignment is determined by a single parent contributing those specific 4 alleles. Current genotyping technology cannot distinguish this level of detail and requires more focused sequencing to sort this out.
Genomic research requires this level of detail but does not always have access to this more expensive sequencing method. Thus, software programs compute the most statistically likely pattern an individual has inherited and assigns them to a haplotype. This inferred haplotype analysis is typically population dependent. It can vary dramatically for a specific area of the genome among different ethnicities, or it can be stable across populations.
Remember, it's important to realize this speed labeling does not necessarily reflect real-life situations. It does not take into account environmental influences. It is purely based on speed averages measured in a research setting.