[kplpwfzv] Self play bad evaluations

Play a chess program against itself in a number of games (with some nondeterminism to avoid repeating the same game.)

Given the final outcome, how "wrong" was the evaluation earlier in the game?  For example, for the winning side, what was the most negative evaluation of any position earlier in the game?

In principle, a computer should never be wrong when playing itself: it knows how to model the opponent perfectly.

These instances possibly serve as indicators of flaws in the program.  They may also serve as a way to measure and compare different programs, though I don't know what the measurement means.

See the diagrams at Glauring chess games , particularly the few that first go up then down, or vice versa.