Skip to content

v1.0.63

Choose a tag to compare

View all tags
@github-actions github-actions released this 10 Sep 02:20
· 25 commits to main since this release
v1.0.63
b8531fc

v1.0.63

One search change from the September 2026 search-efficiency follow-ups (Linear NET-1239, parent NET-1154), merged as PR #90.

Strength

  • NET-1239 — singular negative extension. v1.0.61's singular extension verifies the hash move at half depth and used the answer only to extend. Now the opposite answer is used too: when the alternatives beat the verification's singular beta while the transposition-table score says the hash move fails high on its own, the hash move is one of several good moves and is searched one ply shallower. The hash-move cutoff stores the reduced height; verification-derived scores are still never stored.
    • SPRT vs v1.0.62, 1+0.01, 1 thread, Hash 128, paired 8moves_v3 openings, bounds [0, +5], α=β=0.05: H1 accepted at 5,546 games, 2172-1979-1395, +12.1 ± 7.9 Elo, LOS 99.9%. A first round on the [0, +10] bounds was inconclusive for this and two sibling arms (+9.8, +8.3 and +6.2 ± 9.3 over 4,000 games each), which is why the tighter test was run.
    • Multicut, the other peer mechanism for a failed verification (return the verification bound unsearched when it already reaches beta), was implemented and measured at the same time: +8.3 ± 9.3 alone and slightly weaker combined with the negative extension in both rounds. It ships switched off (SINGULAR_MULTICUT = false).
    • Fixed depth: bench (16 positions, depth 12) 2,479,174 → 1,894,038 nodes; at depth 12, 1,218 of 2,849 verifications end in a negative extension. A control build with both switches off reproduces v1.0.62's bench exactly.
    • Limitations: measured at 1+0.01 only at the time of release; the 10+0.1 scaling check and the Stockfish anchor ladder follow the release (results on the Ratings & Performance page and on NET-1239).

Also measured and not shipped (for the record)

  • Mild capture reductions behind LMR_TACTICAL (NET-1274): base 1 −5.0 ± 11.4, base 2 −27.3 ± 20.5, both rejected. Together with v1.0.62's finding, capture LMR loses in Rival at every base tried.
  • The null-move eval gate and the root fail-high exit (NET-1241), re-tested against this release's parent: about +3 ± 4 and +4 ± 4 Elo pooled over 18,000 and 17,000 games, below the acceptance bound; parked.

Bench signature

bench (16 positions, depth 12): 1,894,038 nodes, 1 thread.

Assets 9
Loading