Verify the bench — in your own browser.

The formula does the work — but if you suspect we fudged the bench, this is where you check. A 2-layer LoRA-shaped MLP (rank-32) trains on TR-MRV-Bench via WebGPU on your GPU in ~3 seconds. No backend, no telemetry, no account. The model ties the formula at ±0.3% in 5-outer LODO; if the bench were rigged, this page would show it.