i would rely on the AD (allelic depth) of each sample in reads. By the time the reads are aggregated into the AF i wouldn't trust variant callers not to corrupt those numbers as they are expecting a biallelic genome. (GATK has become better at this)