Performance Gates

Performance claims require measured evidence.

Required Baselines

Every benchmark must declare:

• fastest dense baseline available;
• safe Rust backend;
• Python wrapper overhead;
• ripser or GUDHI when installed;
• same-budget random selector for sparse methods;
• same-budget locality selector for sparse methods.

Sparse Attention Gate

For topology-derived sparse attention:

\[ \text{relative L2} = \frac{\|y_{dense} - y_{sparse}\|_2}{\|y_{dense}\|_2} \]

Promotion requires:

• relative L2 at or below declared tolerance;
• cosine similarity at or above declared tolerance;
• schedule build time reported separately;
• kernel execution time reported separately;
• comparison against dense SDPA or FlashAttention, not only slow CSR.

Claim Policy

A README speed claim must point to raw benchmark output. Do not publish a speed claim without raw benchmark output and the script that produced it.