Winnex Definitive Benchmark v1.0

Complete empirical evaluation of the Winnex AI Stack vs FAISS baselines across 3 datasets, 16 methods, and 12 metrics.

Summary

This benchmark establishes exactly what the Winnex AI Stack delivers by measuring all methods against FAISS baselines on validated datasets (SIFT-1M, News Category, Synthetic uniform sphere).

Methods Tested (16 variants)

• Winnex: MadhavaCore [64,128], [32,64], MadHybrid np=5/10/15, HMC Hierarchical, H4+M10 Gate
• FAISS: HNSW ef=32/64/128/256, IVF nprobe=1/10/20/50, PQ m=16, FlatIP (exact)

Key Results

SIFT-50K

MadhavaCore [64,128]: NDCG=1.000, Latency=1.42ms, Build=0.09s, Zero bound violations

Synthetic 100K

MadhavaCore [64,128]: NDCG=0.998, Build 23x faster than HNSW (0.23s vs 15s)

Near-tie

H4+M10 Gate: flp_bad=0 across all gap levels, Spearman rho=0.9996

Verified Advantages

1. Mathematical bound guarantee per excluded document (unique to Madhava)
2. Build speed: 5-65x faster than HNSW
3. Zero bound violations across 254M+ query-vector pairs
4. Deterministic and CPU-only inference
5. H4+M10 eliminates false positives in near-tie scenarios

Limitations

1. Latency 3-10x higher than HNSW (Python vs C++/SIMD)
2. NDCG degrades on uniform data at N > 100K
3. MadHybrid adds complexity without consistent gains over simple MadhavaCore

Kaggle: https://www.kaggle.com/code/kleniopadilha/winnex-definitive-benchmark-v1-0

License: BSL 1.1 | pay@winnex.ai