Two rules: if we can’t measure it, we don’t claim to have measured it — and we publish method and raw numbers with every review so you can disagree with the maths, not just the verdict.
What “measured” means here
Software & AI: objective benchmarks on public, reproducible data — e.g. Word Error Rate for speech-to-text against printed-book ground truth, tokens/s and memory for local models, import fidelity for apps. The scripts and audio/Text samples stay in our repo and we describe them in each article.
Hardware: instruments — plug-in power meters for standby and cycling, USB power meters for chargers, fio for sustained storage writes. Where an instrument is still on order, the article says so and is labelled a guide until the numbers exist.
Services: repeated runs on our own connection (speed, leaks, bitrate actually delivered via ffprobe or browser devtools), with the variance published, not a cherry-picked best number.
Prices are checked at retailer pages and dated at publication; we re-check before re-featuring a test.
Verdict vocabulary: hardware gets Buy it / Wait / Don’t buy, software and services get Use it / Situational / Skip. No 8.7/10 theatre.