v1.0.60

v1.0.60

Three changes from the September 2026 search-efficiency audit (Notion: Search Efficiency Audit — September 2026; Linear NET-1154).

Strength

• NET-1192 — history learning rate raised to min(2048, 16·d²) (PR #85). All four history tables (butterfly, countermove, follow-up, capture) were updated with depth² on a 16,384 gravity scale, 20–40× smaller than peer engines; the history-modulated LMR thresholds were almost never reached. Divisors, thresholds and decay are unchanged.
  • SPRT vs v1.0.59, 1+0.01, 1 thread, Hash 128, paired 8moves_v3 openings, bounds [0, +10], α=β=0.05: H1 accepted at 2,792 games, 1123-1004-665, +14.8 ± 11.2 Elo, LOS 99.5%.
  • Measured mechanism (instrumented bench): the gain comes mainly from the malus side. Nearly half of LMR decisions now get the extra reduction ply for bad history (was 1%); the positive threshold is still rarely reached. Bench (depth 12) 2,725,178 → 1,619,564 nodes.
  • Limitation: measured at 1+0.01 only; no longer time control test yet.

Speed (node-identical, shipped in v1.0.59, listed for the record)

• NET-1188 pre-make futility/LMP rejection (+8.7% NPS) and NET-1189 move-list copies removed (+1.5% NPS) were released in v1.0.59.

Experiments run and not merged (recorded on their Linear tickets)

• NET-1190 root fail-high exit: +5.0 ± 11.4 Elo at the 3,000-game cap, inconclusive, parked.
• NET-1191a null move only when static eval ≥ beta: +6.4 ± 9.3 Elo at the 4,000-game cap, inconclusive, parked for a later revisit against this release.

Bench signature

bench (16 positions, depth 12): 1,619,564 nodes, 1 thread.