Which Question Is Your Attention Metric Answering? Attention Rows as Compositional Data

Standard metrics like cosine similarity or Jensen-Shannon divergence produce inconsistent results when analyzing transformer attention rows due to the presence of sink tokens. Analyzing ten pretrained models shows that 17-47% of head-similarity verdicts flip depending on whether the sink is retained or renormalized. Using Aitchison distance treats attention as compositional data, orthogonally separating sink weight from content distribution.