How we measure installer size and startup time

The reproducible method used to measure Tablyne's installer size, startup time and idle memory.

This page describes the measurement method. It does not publish a number until the corresponding release gate has a checked value and the raw data is available.

Test environment

The current reference environment is Windows 11 x64 on an Intel Core i7 desktop with 64 GB of RAM. Windows Defender remains enabled during the run, and CPU idle is recorded before and after each batch. macOS and Linux are not measured yet, so this label stays explicit until those platforms have their own data.

Installer size

Each release installer is measured as a byte count and reported in MiB. The raw file, platform, build, and command are kept with the release record. The published claim uses the measured installer artifact, not an estimate of the application bundle.

Startup time

Startup ends only after persistence bootstrap resolves, the React tree mounts, the real application shell is present in the DOM, and the following animation frames have run. The measurement reports wall time as the primary user-facing value and process time as a secondary diagnostic.

Two startup classes are kept separate:

  • warm: the process is closed, while the disk and WebView2 runtime are already resident.
  • fresh-webview: a new WebView2 profile is used without killing unrelated WebView2 processes.

Each class records N, warmup runs, failures and the raw samples. Results report the median, p90 and p95. Timeouts count as failures. Outliers are not silently discarded, and retries are not used to hide a failed run.

Idle memory

The complete process tree is measured, including WebView2 descendants. Working Set and Private Bytes are reported separately. Idle memory uses a sampling window and reports the median rather than the first observation. A startup timeout is a failed run and does not contribute a memory sample.

What is not measured

This page does not claim a cold start after reboot or a clean disk cache. It does not claim macOS or Linux performance, a full parameter-by-parameter build matrix, or the size of every final installer format. Those require separate runs on the relevant platform and are listed as limitations until measured.