ELO    -5.90 +- 2.77 (95%)
SPRT   10.0+0.10s Threads=1 Hash=16MB
LLR    -2.95 (-2.94, 2.94) [0.00, 2.00]
GAMES  N: 13024 W: 3025 L: 3246 D: 6753
PENTA  [1, 1510, 3714, 1283, 4]

Search: Fitted 50-move eval damping Replace the linear (234 - hmc) / 234 eval damping with a curve fitted on game results: 1 - 0.221 x - 0.713 x^2 with x = hmc / 100, and 1 - x for pawnless positions. The multipliers are a table built at startup (indexed by pawnless and half-move counter), so the search does one lookup, a multiply and a shift. Fit: for each half-move-counter bucket, the slope k of the logistic E(k * eval) that best predicts the game result, relative to HMC 0-7, using the raw static eval of the master net. 69M positions from OpenBench STC games, 2.1M from LTC games and 13.6M from selfplay data give nearly the same curve (multiplier at HMC 48: 0.70 / 0.61 / 0.71, at HMC 80: 0.36 / 0.32 / 0.40; previously 0.79 and 0.66). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>