simex
How often chess engines choose the same move in positions where several moves are about equally good. Each engine finds such positions in its own games, all of them go into one shared pool, and every engine is tested on every position in the pool. The method is Ed Schröder's SIMEX, run continuously and on a shared pool: the idea and the test are his.
Move agreement
Similarity map
The closer two engines are, the more often they pick the same move. Positions are worked out from the agreement grid by a standard statistical method (multidimensional scaling, explained here). Only distance matters: up, down, left and right mean nothing. Hover over an engine to see the engines it agrees with most.
Agreement by search limit
One line per pair of engines, with the search limits in config order. Lines that rise from left to right show the engines converging as they search for longer.
Pool size
Positions in the shared pool over time.
Phase bands
New positions are kept so that the pool moves toward the target share of each band (see Settings). Agreement is the mean over all pairs of engines of how often both chose the same move.
Positions by game phase
Fruit's phase: knight or bishop 1, rook 2, queen 4, counted for both sides. 24 is the full set of pieces, 0 is kings and pawns only (pawns don't count).
Engines
Eval key: HCE = hand-crafted evaluation. For networks: 768 = 12 piece types × 64 squares; x16 = 16 king buckets; hm = horizontally mirrored; threats and pawn pairs are extra input features (which pieces attack or defend which, pawn pairings); HalfKA = king-relative inputs; each → is a layer, with its size; ×8 = 8 output buckets (usually by piece count).
The equish rate is the share of analysed game positions that passed the filter. Positions tested is how many pool positions the engine has been tested on so far (new engines catch up over time). Engines with no games can’t produce usable MultiPV results (see Notes), so they are tested on the pool but don’t add positions to it.