Expand description
Cross-engine benchmark harness (port of the Python benchmarks/ package).
Compares the minimax, MCTS, beam-search, and uniform-random engines on a
shared, versioned, checksummed position dataset under methodologically
consistent conditions. See docs/BENCHMARKS.md.
Modulesยง
- adapters
- Uniform engine adapters with effective-work observations.
- agreement
- Agreement and cost aggregation for benchmark move selections.
- book_
export - Bulk export of solved benchmark references into the SQLite opening book, and lookup of stored references to short-circuit repeated solves.
- bundle
- Reproducible benchmark result bundles (port of
benchmarks/bundle.py). - canonical
- Python-compatible canonical JSON encoding.
- checkpoint
- Directory-based resumable checkpoint storage for long benchmark runs.
- contracts
- Contract-oriented projections for benchmark artifacts.
- correctness
- Correctness preflight for benchmark inputs and adapters.
- dataset
- Shared, versioned position dataset for cross-engine benchmarks.
- head_
to_ head - Paired, side-balanced head-to-head games from shared positions.
- metrics
- Statistics helpers for benchmark reports (port of
benchmarks/metrics.py). - reference
- Exact game-theoretic references for benchmark positions.
- report
- Markdown report generation from benchmark result bundles.
- stability
- Across-seed stability of benchmark move selections.