JumpStar Chinese Checkers Engine

This is the public engine and benchmark repository for JumpStar, a self-play-trained Chinese Checkers AI created by Joe Philleo. As of the public launch on 2026-05-23, JumpStar_60 is, to the best of this project's public evidence, the strongest openly documented Chinese Checkers AI engine under the CCERL-2P10-v2 benchmark. That claim is benchmark-scoped: stronger private engines may exist, and CCERL is designed so public challengers can test the result.

This repository intentionally contains only the public research surface:

• the dependency-light C++20 CCZero engine core
• tests and JSON schemas
• the CCERL benchmark rules, protocol, positions, and baseline definitions
• public leaderboard and direct-match evidence
• a small benchmark runner/tool subset
• a lightweight static replay viewer

It does not contain the hosted website, Vercel deployment code, private training automation, or internal experiment workspaces.

Build

cmake -S . -B build -DCCZERO_USE_ACCELERATE=OFF
cmake --build build --parallel
ctest --test-dir build --output-on-failure

On macOS, the Makefile also supports:

make release
make test

Start Here

• docs/benchmark/README.md: benchmark documentation.
• docs/benchmark/CCERL_LEADERBOARD.md: current public leaderboard snapshot.
• docs/benchmark/CCERL-2P10-v2.md: current ruleset.
• docs/benchmark/CCP_PROTOCOL.md: stdin/stdout protocol for external engines.
• benchmarks/ccerl-v1/positions/official_elo_v2.jsonl: frozen v2 position schedule.
• experiments/public_benchmark/ladder_v2_latest_champions_s64/elo.json: public Elo summary.
• experiments/public_benchmark/champion_match_iter063_vs_iter060_s768_c24_v2/champion_match.md: direct champion-defense match summary.

Quick Commands

Print the benchmark manifest:

python3 tools/ccbench.py manifest

Run a tiny native ladder sample after building build/cczero:

python3 tools/ccbench.py baseline-ladder-native \
  --engine build/cczero \
  --positions benchmarks/ccerl-v1/positions/official_elo_v2.jsonl \
  --out-dir build/public_ladder_smoke \
  --label random \
  --label greedy \
  --limit 2 \
  --simulations 8 \
  --workers 1 \
  --force

External engines can implement CCP and be run through:

python3 tools/ccbench.py referee-ccp --help

Replay Viewer

A small static replay viewer is included at viewer/replay.html. It can load the included sample game or any JSONL produced by:

./build/cczero match --rules strict --p0 greedy --p1 random --seed 7 --max-plies 80 --log build/game.jsonl

Model Artifacts

The public repository records exact model labels, paths, sizes, and hashes used for the benchmark. The .ccpv model files themselves are not stored in Git history. If model downloads are published, they should be attached as GitHub Release assets and restored to the paths documented in benchmarks/ccerl-v1/baseline_pack/MODEL_NOTES.md.

License

This repository is licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). See LICENSE.

Model checkpoints are not stored in Git history. If model files are published as release assets later, their download page should state whether the same license or separate model-artifact terms apply.