Native Benchmark Evidence¶

The repository now keeps native benchmark evidence as one public trust surface, not as one hidden maintainer-only utility.

For the practical command workflow, start with native benchmark review. This page explains how to interpret the benchmark family once those reports exist.

What The Benchmark Family Covers¶

The native maximum-likelihood benchmark family covers:

runtime scaling
peak memory behavior
topology and likelihood recovery on governed truth cases
wrapper comparison against cached external reference lanes
one unified suite that combines those dimensions into one review surface

The public CLI family is:

benchmark native-maximum-likelihood-speed
benchmark native-maximum-likelihood-memory
benchmark native-maximum-likelihood-accuracy
benchmark native-maximum-likelihood-suite
benchmark maximum-likelihood-wrapper-comparison

Why This Exists¶

Parity alone is not enough for a serious native runtime story.

The repository needs to show:

where native code is stronger
where native code is slower but still behaving as expected
where model assumptions explain a difference honestly
where one difference is a real native bug
where one case remains unsupported instead of being silently rounded into a pass

How To Read The Unified Suite¶

benchmark native-maximum-likelihood-suite is the main public summary lane.

Its governed status classes are:

native-advantage
native-bug
expected-assumption-difference
unsupported-case

Those categories are intentional. The suite is a review surface, not one generic performance score.

What This Does Not Claim¶

it does not claim that one green suite closes every scientific question
it does not claim that wrapper tools are irrelevant once native ownership exists
it does not claim that a slower native result is automatically incorrect
it does not claim that every model family and every benchmark lane are at the same maturity level