Although I can't find the exact quote any longer in the CCC archives I remember these prophetic words quite well. The realization changes could be proven by playing 10,000-20,000+ bullet games was part of his rapid progress. And for that one needed of course hardware that was capable of doing that which in those days waan't richly available as it is now.
Secondly that indeed some old ideas rejected earlier as a regression seems to work after all now that testing became more reliable. And in reverse, that in the past programmers made changes which in retrospect weren't improvements after all, but regressions and even worse never were looked back at. All in all it make sense to re-test every change once made now that accurate testing is possible.
In this blog (whiich I hope to update frequently) I will propose a number of ideas, old and new, that worked for me which necessarily doesn't have to work in any other engine, ideas that eventually were small regressions that might work for other engines and ideas that gave such a small improvement (1-2 elo) the LOS remained under 95% which is the line drawn in the sand for me to reject or accept changes.
So here we go.
With LMR we have a number of restrictions not doing LMR because it's too dangerous to reduce. Most known cases, king in check, giving a check, captures etc. I found 2 new ones, one about King Safety, one about Passed Pawns.
King Safety - if the evaluation of the king safety for white is greater than 0.50 and the king safety for white has gone up with 0.25 in comparison with the previous ply then don't LMR. Likewise for black.
Same for passed pawn.
Passed Pawn - if the passed pawn evaluation for white is greater than 0.50 and the passed pawn evaluation for white has gone up with 0.25 in comparison with the previous ply then don't LMR. Likewise for black.
Progress about 5-7 elo, LOS okay.
Note that the 0.50 and 0.25 values were first best guesses and need tuning.