# Benchmark Data Notes

This repository includes a small reproducible data file at
`data/benchmark-runs.csv`. The rows are intentionally modest: they are not a
claim of laboratory performance, but evidence that the library is exercised with
real command output and imported sample scenarios.

## Collection Method

The local rows were collected with:

```bash
moon run cmd/main
```

The command prints summary CSV, JSON distribution data, cleaned imported sample
reports, and a baseline regression result. The CSV file records the numeric
fields from those runs so reviewers can inspect concrete data without rerunning
the command on the same machine.

## Observations

- Local Windows runs produced means around `35 ms` to `37 ms` for the demo loop.
- The demo budget was widened to `80 ms` because the default `env.now()` clock is
  millisecond-grained and host load can move the example by several milliseconds.
- The imported data row demonstrates the non-clock path: callers can pass their
  own elapsed samples, remove a `0 ns` coarse-clock artifact and an `80 ms`
  outlier, then generate percentiles and regression reports.
- The `2026-07-17-local-2` row is kept even though it used the older stricter
  demo budget. It documents why the example budget was adjusted instead of
  hiding an inconvenient run.

## How The Data Maps To Features

- `summary_csv` writes the `count,total_ns,min_ns,max_ns,mean_ns,spread_ns`
  fields.
- `distribution` supplies `p50_ns`, `p95_ns`, and `p99_ns`.
- `filter_samples`, `drop_warmup`, `clamp_samples`, and `trim_samples` prepare
  raw imported measurements before reporting.
- `quality_report` turns the same data into a compact stability line for CI
  logs and release notes.