Engineers argue about what actually proves a system works

Netlify says its Edge Functions are 5x faster after moving from V8 isolates to Firecracker MicroVMs. Commenters pushed back hard: Cloudflare Workers are also V8 isolates and run far faster than the 25-40ms Netlify reported, and the post doesn't say where the isolate latency came from. One asked why count milliseconds in JavaScript at all. In the Wi-Fi multi-connectivity thread, a team streaming with a 100 ms latency budget showed that roaming is the real killer, and others noted 802.11r and the coming 802.11bn already target it. Magnitude's benchmarks drew the same demand for methodology, since llama.cpp performance varies widely by configuration.

The TLA+ writeup ("What TLA+ can and can't check") hits the same nerve from the correctness side. A commenter says people claim "just write tests" or "just use formal verification" as if either is sufficient. A separate floating-point piece says to keep money as integers and separate storage from presentation.

The pattern: readers don't accept a headline number or a proof label without asking what exactly was measured or checked.