> show that SIEVE outperforms all state-of-the-art eviction algorithms on more than 45% of the traces.

Doesn't that mean that over half the time, 55%, it's sometimes equal and, more likely so, worse?