Tree-dependent refinement (iterations 3, 4 ... ) can be speeded up by dividing the alignment vertically into blocks. Block boundaries are found by identifying high-scoring columns (e.g., a perfectly conserved column of Cs or Ws would be a candidate). Each vertical block is then refined independently before reassembling the complete alignment, which is faster because of the L2 factor in dynamic programming (e.g., suppose the alignment is split into two vertical blocks, then 2 x 0.5^2 = 0.5, so the dynamic programming time is roughly halved). The –noanchors option is used to disable this feature. This option has no effect if –maxiters 1 or –maxiters 2 is specified. On benchmark tests, enabling or disabling anchors has little or no effect on accuracy, but if you want to be very conservative and are striving for the best possible accuracy then –noanchors is a reasonable choice.