# Benchmark reproducibility

This document describes the deterministic native workload used to measure the
library on a local machine. Timing values are machine- and toolchain-specific
and are not universal performance guarantees.

## Run the benchmark

```powershell
powershell -NoProfile -ExecutionPolicy Bypass -File .\tools\run-benchmarks.ps1 -Runs 5
```

The runner performs repeated matrix, interpolation, statistical, and
calibration workloads. Each run prints deterministic checksums so a workload
change or accidental removal of computation is visible in the generated
record.

The generated [benchmark-results.md](benchmark-results.md) records the native
release command, toolchain, individual timings, aggregate timing, and checksums.

## Development use

Use the same workload after changing numerical hot paths. Compare results on
the same machine and toolchain; small timing differences are expected when
system load changes.