Every feature, correlated against who eventually won, across every position ever recorded. Rank on the early column — late-game numbers are contaminated, because a squeezed king is how you win, so a strong late correlation can be the result showing through rather than a cause. A feature near zero early is not carrying information.
A personality is a search depth and a set of weights. Features are raw
magnitudes and the weight carries the sign: give
opp_king_mobility a negative weight to squeeze, give
own_king_mobility a positive one to stay free.
Plays the candidate against a fixed panel in mirrored pairs — the same game twice with the colours swapped — at fixed depth, so neither the first-move advantage nor how busy the machine is can colour the answer. The interval is what matters: if two personalities' intervals overlap, the difference has not been measured.
For each own/opp pair, the sum of the two weights. A pair summing to zero behaves exactly like a symmetric heuristic: both sides score it identically and it drops out of the comparison. That is the trap that made hand-tuning so frustrating in the nineties — a rule that feels important, and changes nothing.