I re-ran the data for Elo 2600--2800 after modifying one line of the Stockfish 7 code to forbid move-count-based pruning at PV nodes. Doing so greatly reduced the numerical issue shown by my tables at depth 5, but seems not to have completely removed it. I have updated my post
https://rjlipton.wordpress.com/2016/01/ ... l-at-zero/ on the Gödel's Lost Letter blog accordingly. Here are the old and new tables.
Old, using the official Stockfish 7 release:
New:
Still registers 6--10 times more large mistakes at depth 5 on the same moves as Komodo 9.3 does. Perhaps I may combine the single-line edit stated in the post with one or more suggestions given here.
[/img]