A benchmark you cannot reproduce is marketing. We split our numbers into two tiers and keep both honest.
CPU smoke tier
Small problem sizes that run on a laptop or CI in seconds. They verify correctness - closed-form derivatives match autodiff and finite differences, and the backends agree bit-for-bit. These run on every push.
JAX_PLATFORMS=cpu python -m bench.laplacian_scaling.dimension_sweep \
--dims 3 12 30 --hidden 32 --batch 64 --repeats 3GPU headline tier
Full-fidelity runs on data-center GPUs, produced off-band and transcribed verbatim into the docs. Every row records the hardware class (memory tier), the precision, and the agreement tolerance. Where a number is pending, we say so rather than estimate.
The rule we never break
Every speedup is reported against an identical numerical answer. If the methods disagree beyond float64 round-off, it is not a benchmark - it is a bug.
Cross-backend parity
Because every backend imports the same pure-Python polynomial coefficients, a given (activation, order) pair is float64-ULP-equal across PyTorch, JAX, and Keras 3. That parity is checked per release by a dedicated suite - it is what makes backend-portable certified numerics possible at all.
Why we bother
Scientific users do not adopt a kernel because a landing page says "fast". They adopt it because they re-ran the smoke tier on their own machine, saw the same answer, and watched the wall-clock drop. That is the only marketing that works on this audience.