Sport Intelligence Benchmark (v1.0.0)

Ensemble V2 LightGBM meta-learner methodology and a synthetic reproduction dataset for football match outcome prediction, reported as two distinct, deliberately never-reconciled results (D-07). Reproducible: `python run_benchmark.py --seed 42` against the shipped synthetic sample (`data/derived_sample.csv`, itself generated via `python export_dataset.py --rows 4000 --seed 42 --output data/derived_sample.csv`) prints a Brier score (3-class summed, range 0.0-2.0) that any third party can verify end to end. Production: Brier 0.5783 on the real 97,000-match production corpus — this figure is measured on non-redistributable data and cannot be verified by third parties; it is reported for transparency, not for independent reproduction. See DATA_PROVENANCE.md for the full reconciliation. Code: MIT. Contents of `data/`: CC-BY-4.0.