Skip to content

JACE v1.5.2

Latest

Choose a tag to compare

@github-actions github-actions released this 16 Nov 20:47
51fb9c5

JACE 1.5.2

JACE 1.5.2 gives the accepted baseline a distinct version for external testing,
including CCRL submission. Author: Jeffrey S. Groby, United States.
Open source under MIT; based on Bluefever Software's VICE tutorial engine,
with AI-assisted development using OpenAI ChatGPT/Codex. See CREDITS.md.

Changes since 1.5.1

  • Faster SEE attacker enumeration from the exchange square, preserving the
    original equal-value ordering and king-safety checks (PR #31).
  • macOS CI moved to the supported macOS 15 ARM runner.
  • Updated UCI version identification and release documentation to 1.5.2.

The nonblocking hash-store experiment (PR #32) and direct attack-detection
experiment (PR #33) were closed without merging and are not included.
This release introduces no further search or evaluation changes beyond the
already accepted baseline cb32a87.

Validation and strength

The accepted SEE change matched 20,853 reference capture scores and 647,040
attacker sets. Seven alternating fixed-depth benchmark passes on Apple M1 Max
preserved search scores, principal variations and node counts while reducing
median search time by 11.43%. A 256-game paired match against 1.5.1 at 60+1,
four threads and 64 MB hash scored 76 wins, 56 losses and 124 draws (53.91%;
+27.20 +/-31.63 Elo). All games ended normally and were independently verified.
The interval includes zero: this is not a conclusive strength gain or an
absolute CCRL rating. Evidence is in validation/see-enumeration and PR #31.

JACE remains independently unrated. Earlier local opponent testing suggests
only a very provisional 2000-2200 Elo range; it does not calibrate this release
across thread counts, hardware or time controls.

Downloads and configuration

Packages target Windows x64/ARM64, Linux x64/ARM64 and macOS ARM64. Choose the
archive matching the test machine; Windows x64 uses jace-windows-x64.exe.
Each archive includes instructions, copyright/license notices, build provenance
and SHA-256 checksums. The staged release executable is smoke-tested before
packaging. macOS binaries are unsigned; Linux builds target Ubuntu 24.04.

Lazy SMP supports Threads 1-4 (default 1), with Hash shared across workers.
For matches use MultiPV=1, Ponder=false, Book=false and TaperedEval=true, and
explicitly set Threads and Hash to the tester's chosen configuration.
MultiPV greater than 1 and console mode search serially. Optional Syzygy WDL/DTZ
probing uses Fathom; set SyzygyPath to separately downloaded table files.
Books and tablebase data are not bundled or required to run the engine.

Source and downloads: https://github.com/jsgroby/jace/releases/tag/v1.5.2