On 2026-07-05, re-verifying the numbers before publishing the 17-plugin benchmark, eslint-plugin-unicorn came back 0 of 40 — a suspiciously clean zero that would have read as proof a well-regarded plugin was worthless. It wasn't: a stale Node version in my shell had silently swallowed the run, and the real number was 22 of 40, F1 51.8% (the full forensic walkthrough lives in Bias in Measurement; this article is about the process that made me go looking for it).
That's the kind of mistake a vendor benchmark doesn't usually surface — vendors don't go looking for ways their own test caught nothing. Every ESLint benchmark on this site compares my own plugins to competitors, and I'm the one who decides what counts as a bug — which means every benchmark from a plugin author, including this one, has that exact problem. A benchmark that's only ever been run once, by the person who built the tool it favors, hasn't been tested. It's been asserted. This article is the concrete process that keeps mine from being just that — not a disclaimer, a process you can go verify yourself, with the two times it caught something wrong.
The problem, stated plainly
I'm a solo creator. There's no lab, no independent research team, no third party benchmarking Interlace on my behalf. If "credible self-benchmark" requires a formal external audit, every benchmark I publish fails before I write the first fixture — not because the numbers are wrong, but because that category of evidence doesn't exist for someone working alone.
