Nullstar 032

Nullstar 032

Third public binary release of Nullstar, an experimental Windows x64 UCI
chess engine by Jasper Sinclair.

Changes

• adds bounded soft and hard time limits with a low-clock reserve;
• preserves the last fully completed iterative-deepening result;
• adapts the soft limit using best-move stability, score trend, and root-move
  effort;
• performs periodic deadline checks in the main search and quiescence search;
• hardens initialization and resolves the remaining static-analysis
  diagnostics;
• retains the embedded stm-base epoch 29 768x256 side-to-move NNUE network;
• includes MSVC and MinGW-w64 PGO and non-PGO builds.

Testing

10,240-game gauntlet against an expanded 17-engine local strong pool:

• 5108 wins, 2347 draws, 2785 losses
• 61.34%
• Ordo rating: 3112 (error: 6.4)
• zero Nullstar time forfeits and zero illegal moves

All six binaries passed UCI identity, readiness, termination, perft, and
deterministic benchmark verification. Cross-compiler search results matched,
and the hash-table and adaptive time-manager probes passed.

For most modern Windows systems, begin with
nullstar_mingw_avx2_pgo.exe or nullstar_msvc_pgo.exe. The native
MinGW builds are intended for the build computer or a compatible processor.

See the repository documentation for checksums, build instructions, test
conditions, network provenance, and training details.

Licensed under GPL-3.0-only.

Nullstar 027

Second public binary release of Nullstar, an experimental Windows x64 UCI chess engine by Jasper Sinclair.

Changes

• introduces a new 768x256 side-to-move NNUE network;
• embeds the network directly in each executable;
• aligns quantized NNUE inference with the trainer;
• includes search-correctness improvements and a hardened transposition table;
• includes MSVC and MinGW-w64 PGO and non-PGO builds;
• expands the reproducible NNUE training, export, and parity-verification workflow.

Testing

10,240-game gauntlet against a fixed 16-engine pool averaging 3020:

• 4682 wins, 2326 draws, 3232 losses
• 57.08%
• Ordo rating: 3081 (error: 5.9)
• 59 Ordo Elo above Nullstar 026 under the same conditions

All six binaries passed UCI, readiness, termination, perft, and deterministic benchmark verification. Cross-compiler search results also matched.

For most modern Windows systems, begin with nullstar_mingw_avx2_pgo.exe or nullstar_msvc_pgo.exe. The native builds are intended for the build computer or a compatible processor.

See the repository documentation for checksums, build instructions, test conditions, network provenance, and training details.

Licensed under GPL-3.0-only.

Nullstar 003

Initial public binary release of Nullstar, an experimental Windows x64 UCI chess engine by Jasper Sinclair.

Changes

• introduces the Set 006 768x256 NNUE network;
• embeds the network directly in each executable;
• includes MSVC and MinGW-w64 PGO and non-PGO builds;
• includes the complete NNUE training and embedding workflow.

Testing

10,000-game gauntlet against an approximately 2800-rated pool:

• 5099 wins, 1456 draws, 3445 losses
• 58.27%
• estimated rating: 2873

All six binaries passed UCI, readiness, termination, and deterministic benchmark verification.

For most modern Windows systems, begin with nullstar_mingw_avx2_pgo.exe or nullstar_msvc_pgo.exe. The native builds are intended for the build computer or a compatible processor.

See the repository documentation for checksums, build instructions, test conditions, network provenance, and training details.

Licensed under GPL-3.0-only.