KG-next engines — Python Board capture hash fix

KataGo engines for KG-next, built from fork master commit 0feaf24d35f75d6b424b7db3f24063d5a78fa1be after merging PR #10.

Changes since kg-next-f18995833951

• Backport the Python Board capture-hash fix from upstream PR #1256, authored by zliu1022. Removing a captured stone now XORs the Zobrist value of that stone's actual color, restoring the incremental position/situation hash and capture-undo contract.
• Add six focused regression cases, including seeded play/undo games, captures by both colors, equivalent final positions and a pass control, with a lightweight Python board CI workflow.
• This is a Python Board correction. C++ search, CUDA dependencies, GUI board logic, models and engine performance are unchanged by the PR. The Windows executables are rebuilt to identify the exact current master commit; they do not contain the Python Board module.

Downloads

• windows-nvidia: CUDA 12.8 / cuDNN 9.8, Windows x64.
• windows-nvidia-cuda13: CUDA 13.2 / cuDNN 9.24, Windows x64.

The two archives contain katago.exe, GTP/analysis configurations, licenses and the original source-release.json receipt. NVIDIA runtime DLLs and models are separate. ENGINE-SHA256SUMS.txt contains the checksums. The Python fix is available in this tag's source tree and GitHub source archives.

Verification

• Python board regressions: focused capture, incremental-hash and undo regressions on the actual merged master.
• Platform builds: platform builds/tests, including the stripped Linux release-copy checks.
• Windows CUDA builds: both profiles built from this exact release commit, including dependency/PE audits, isolated-path startup and native runtests/runoutputtests.
• CPU root-reuse regression: native invariants, probe lifecycle tests, Fox SGF checks and real-engine probes across six color/thread/visit-cap configurations.
• Downloaded archive/member hashes, executable/source identity, runtime profiles and uploaded release asset SHA-256 digests were verified.

Original hosted-build receipts retain hardwareAcceptanceStatus: PENDING_HARDWARE. Hosted CI does not have an NVIDIA GPU. Local hardware checks, when recorded, apply only to their stated GPU, model and configurations and do not change these original receipts.
Local hardware check: both exact release executables passed 39 root-reuse cases each on an NVIDIA RTX 4090, driver 610.88, with the g170-b6c96 test model, White, four threads, ownership and root visit cap 32. This is scoped smoke/regression evidence, not full GPU acceptance across colors, threads, models or long sessions.

KG-next engines — hardened probes and verified Linux release copies

KataGo engines for KG-next, built from fork master commit f18995833951a994dcc7e7d16e201dba05f98466 after merging PR #9.

Changes since kg-next-a1bc99c7209f

• Harden the root-reuse validation probe: stream raw logs to disk, bound capture queues and diagnostic tails, and propagate reader, encoding, write and flush errors before reporting PASS.
• Reject Python optimized modes that would disable assertions. Parse and retain validation of every analysis report without duplicating report text in command responses.
• Strip only the Linux release copy after the normal build and tests, then check symbol removal, version identity and native tests. The global CMake configuration and other platforms are unchanged.
• Search C++, training conversion, CUDA/cuDNN versions, and Windows bundle acceptance logic are unchanged by PR #9.

Downloads

• windows-nvidia: CUDA 12.8 / cuDNN 9.8, Windows x64.
• windows-nvidia-cuda13: CUDA 13.2 / cuDNN 9.24, Windows x64.

Both archives contain katago.exe, GTP/analysis configuration files, licenses, and the original source-release.json receipt. NVIDIA runtime DLLs and neural-network models are separate. See runtime profiles and packaging. ENGINE-SHA256SUMS.txt contains the download checksums. The Linux stripping change is exercised by platform CI; this engine release continues to publish the two Windows CUDA profiles.

Verification

• Windows CUDA builds: both profiles built from the exact release commit, with dependency/PE audits, isolated-path startup checks, and native runtests / runoutputtests.
• CPU root-reuse regression: probe capture/lifecycle regressions, native invariants, Fox SGF checks, and 234 real-engine GTP cases across six configurations covering both colors, one/four threads, ownership, and root visit caps.
• Platform builds, including the stripped Linux release-copy checks, on the same master commit.
• Downloaded archives verified against build summaries, including archive/member SHA-256 hashes, source commit, executable identity and runtime profile. Uploaded asset digests were checked against those verified files.

The original hosted-build receipts retain hardwareAcceptanceStatus: PENDING_HARDWARE. Hosted CI has no NVIDIA GPU; its build/native test success does not establish GPU inference acceptance. Any local hardware smoke checks are separately scoped evidence and do not alter the original receipts.

Local hardware check: both exact release executables passed 39 root-reuse cases each on an NVIDIA RTX 4090 with driver 610.88, using the small g170-b6c96 test model, White, four search threads, ownership output and root visit cap 32. This covers those two profile/configuration combinations; it does not establish all-color/thread/model or long-session GPU acceptance. Original hosted-build receipts are unchanged.

KG-next engines — root reuse, SGF fixes and verified CI caches

KataGo engines for KG-next, built from fork master commit a1bc99c7209fd83d6aa82476ef204c588c89f672 after merging PR #8.

Changes since kg-next-2039e2d71b1a

• PR #6: Stop processing Fox SGFs immediately after handicap-komi exclusion, and provide an empty Q-target entry for each indexed turn when converting human SGFs. Add real-command regression coverage for exclusion, output, and accounting.
• PR #7: Preserve the root visit-cap snapshot and interval when the effective current-player root mask is unchanged, including equivalent empty/all-zero masks and changes that only affect the other player. Strengthen native invariants and probe lifecycle/final-report validation.
• PR #8: Add conservative documentation-only CI planning, compiler and dependency-download caches with validation, and fork-specific README/CI documentation. CUDA/cuDNN versions remain unchanged; cache hits still run configuration, builds, and tests.

Downloads

• windows-nvidia: CUDA 12.8 / cuDNN 9.8, Windows x64.
• windows-nvidia-cuda13: CUDA 13.2 / cuDNN 9.24, Windows x64.

Both archives contain katago.exe, GTP/analysis configuration files, licenses, and the original source-release.json receipt. NVIDIA runtime DLLs and neural-network models are not included; KG-next provides the separately matched runtime. See runtime profiles and packaging. ENGINE-SHA256SUMS.txt contains the download checksums.

Verification

• Windows CUDA builds: both profiles built from the exact release commit, with dependency/PE audits, isolated-path startup checks, and native runtests / runoutputtests.
• CPU root-reuse regression: native invariants, probe lifecycle/final-report regressions, real-command Fox SGF checks, and 234 real-engine GTP cases across six configurations covering both colors, one/four threads, ownership, and root visit caps.
• Platform builds and ONNX backend checks on the same master commit.
• Downloaded archives verified against their build summaries, including archive and member SHA-256 hashes, source commit, executable identity, and runtime profile.

This release revision has not been tested on an NVIDIA GPU. Both original build receipts retain hardwareAcceptanceStatus: PENDING_HARDWARE; build success and native CPU tests are not GPU inference acceptance. GPU acceptance of older releases is not carried forward to these binaries. Publishing this release does not update the KG-next GUI engine catalog.

KG-next engines — scored-SGF near-end detection fix

KataGo engines for KG-next, built from fork master commit 2039e2d71b1af2decf40b4d6dd1515b42fb3b443 after merging PR #4.

Changes since kg-next-664c4d78c80c

• Fix ownership-based near-end detection for scored SGFs. The three unsettled/ownership counters now use an exact, widened integer comparison, avoiding integer truncation and floating-point boundary errors while preserving the strict threshold.
• Add 6,524 native assertions covering ten square and rectangular board shapes through the same predicate used by the production code.
• Keep the terminal-leaf visit-cap regression silent by removing its redundant diagnostic output; all assertions remain active.
• Retain the existing root-tree reuse and symmetry-preserving restriction behavior.

Downloads

• windows-nvidia: CUDA 12.8 / cuDNN 9.8.
• windows-nvidia-cuda13: CUDA 13.2 / cuDNN 9.24.

Both Windows x64 archives contain katago.exe, GTP/analysis configuration files, licenses, and the original source-release.json receipt. NVIDIA runtime DLLs and neural-network models are not included; KG-next provides the separately matched runtime. See runtime profiles and packaging. ENGINE-SHA256SUMS.txt contains the download checksums.

Verification

• Windows CUDA builds: both profiles built from the exact release commit, with dependency/PE audits, isolated-path startup checks, and native runtests / runoutputtests.
• CPU root-reuse regression: native tests and 198 real-engine GTP cases across six configurations covering both colors, one/four threads, ownership, and root visit caps.
• Downloaded archives verified against their build summaries, including archive and member SHA-256 hashes, source commit, executable identity, and runtime profile.

This release revision has not been tested on an NVIDIA GPU. Both original build receipts retain hardwareAcceptanceStatus: PENDING_HARDWARE; build success and native CPU tests are not GPU inference acceptance. GPU acceptance of older releases is not carried forward to these binaries.

KG-next engines — root symmetry reuse improvements

KataGo engines for KG-next, built from commit 664c4d7 (merged by PR #3).

Preserves root symmetry pruning when move restrictions leave the symmetry representatives unchanged. Representative-changing restrictions retain the existing safe fallback.

• CUDA 12.8 / cuDNN 9.8: verified on an RTX 4090 with a real model, including restriction transitions.
• CUDA 13.2 / cuDNN 9.24: build and dependency audit passed; this source revision has not been hardware-tested on CUDA 13.
• Both archives are statically linked with zlib and contain engine/config/license files, not the NVIDIA runtime or model. KG-next bundles those separately.
• Build provenance: https://github.com/22nsuk/KataGo/actions/runs/37179612695
• Original source-release.json receipts are preserved, including their CI hardware acceptance status.

KG-next custom CUDA engines — CUDA 12.8 / CUDA 13.2

KG-next의 CUDA12·CUDA13 선택 번들이 사용하는 최종 사용자 지정 KataGo 엔진입니다. 두 파일은 같은 소스 커밋 54268eb6330a4c2b7d63c9abe828c8e23d25333e에서 빌드했습니다.

선택 번들	기본 런타임	엔진 자산
CUDA12	CUDA 12.8 / cuDNN 9.8.0.87	katago-source-54268eb6330a-windows-nvidia.zip
CUDA13	CUDA 13.2.2 / cuDNN 9.24.0.43	katago-source-54268eb6330a-windows-nvidia-cuda13.zip

두 엔진 모두 kg-reuse-root-tree / root-only-v1 탐색 트리 재사용 확장을 포함합니다. ZIP은 엔진·설정·라이선스·소스 및 파일 해시 기록으로 구성되며, NVIDIA 런타임 DLL은 KG-next의 해당 선택 번들에서 별도로 준비합니다. zlib·libzip 및 엔진의 MSVC 런타임은 정적으로 연결했습니다.

검증 결과:

• 최종 Windows CUDA CI에서 두 프로필 빌드, 전체 런타임 DLL 의존성 감사, runtests, runoutputtests, ZIP 파일 목록·해시 검증을 통과했습니다.
• CPU 트리 재사용 CI에서 양쪽 색상·1/4스레드·소유권·방문 수 제한 조합을 통과했습니다.
• Windows / RTX 4090 / 드라이버 610.88 / B11 모델에서 최종 두 실행 파일의 실제 GPU 검사 총 72개 사례가 통과했습니다. 초기화를 생략한 네 가지 반례도 해당 assertion으로 정확히 거부했습니다.
• 두 선택 번들의 실제 engines/katago/windows-x64 폴더 구조에서 KG-next Java helper의 출처·정적 zlib 검사가 통과했습니다. CUDA13은 실제 추론 사전 검사와 성공 캐시 재사용도 통과했습니다.
• 실제 사용한 NVIDIA DLL 16개씩의 해시가 빌드 의존성 기록과 일치합니다. gpu-acceptance.json에 실행 파일·모델·런타임 해시, 소스 및 GUI 커밋, 개별 결과를 제공합니다.

이 버전에는 내장 검사의 종료 노드 처리도 포함됩니다. 두 번 연속 패스하여 종료된 노드는 자식 선택 분포가 없으므로 스냅샷을 요구하지 않고, 기보 경로를 재생해 실제 종료 상태와 NN/자식/스냅샷 부재를 확인합니다. 일반 노드의 기존 스냅샷·분포 assertion은 유지합니다. 탐색·평가 알고리즘의 변경은 없습니다.

빌드 manifest의 PENDING_HARDWARE는 GPU 없는 CI가 생성한 원본 기록으로 보존합니다. 이후 수행한 GPU 결과는 별도 gpu-acceptance.json으로 연결됩니다. 실제 GPU 확인 범위는 위 RTX 4090 환경입니다.

이 릴리스가 이전 kg-next-216741b44b9f를 대체합니다. 연결 PR: KG-next #2, KataGo #2.

KG-next custom engines: CUDA 12.8 / CUDA 13.2

이 릴리스는 kg-next-54268eb6330a로 대체되었습니다. 기존 파일과 해시는 재현을 위해 보존합니다.

KG-next의 CUDA12·CUDA13 선택 번들을 위한 사용자 지정 KataGo 엔진입니다. 두 파일 모두 이 저장소의 같은 소스 커밋 216741b44b9fd97b40456c6d334dd7bed7994fbf에서 빌드했으며, kg-reuse-root-tree / root-only-v1 탐색 트리 재사용 기능을 포함합니다.

자산	기본 런타임	ZIP SHA-256
katago-source-216741b44b9f-windows-nvidia.zip	CUDA 12.8 / cuDNN 9.8.0.87	2713c7a0d0183292ff459bb03d18a6a44c4453790ce3af78ff5bad43c5157189
katago-source-216741b44b9f-windows-nvidia-cuda13.zip	CUDA 13.2.2 / cuDNN 9.24.0.43	1c000f494e444703002dceb5f3c0e5c5b6e4d6658c6b62244ee8657e1a74b5b8

ZIP은 KG-next가 사용하는 엔진·설정·라이선스·출처 기록입니다. NVIDIA 런타임 DLL은 KG-next의 해당 CUDA 번들에서 별도로 준비합니다. 실행 파일은 zlib·libzip 및 MSVC 런타임을 정적으로 연결합니다.

검증 결과:

• Windows CI에서 두 프로필 빌드, 전체 런타임 DLL 의존성 검사, runtests, runoutputtests, ZIP 파일 목록·해시 검증 통과.
• Windows / RTX 4090 / 드라이버 610.88에서 B11 모델 실제 GPU 검증 72개 사례 통과. 두 프로필의 트리 재사용·깊은 제한 변경 시 초기화·방문 수 제한을 확인했습니다.
• 초기화를 생략한 반례 네 가지를 검사에서 정확히 거부했습니다. 검증 스크립트는 1ea5e04b703e6c2f4a46447e9e29b96021693678이며 엔진 C++와 빌드 코드는 빌드 커밋 이후 변경되지 않았습니다.
• KG-next Java helper가 두 실행 파일의 출처와 정적 zlib를 확인했습니다. CUDA13의 실제 추론 사전 검사와 성공 캐시 재사용도 통과했습니다.
• 사용한 런타임 DLL의 실제 해시가 빌드 의존성 기록과 일치합니다. 자세한 결과는 gpu-acceptance.json을 참고하세요.

빌드 manifest의 PENDING_HARDWARE는 GPU 없는 CI가 생성한 원본 기록으로 보존했습니다. 이후 수행한 실제 GPU 결과는 별도 gpu-acceptance.json에 실행 파일 해시와 함께 기록했습니다. 검증한 하드웨어는 위 RTX 4090 환경입니다.

연결 PR: KataGo #2, KG-next #2.