A single Lighthouse run is one sample, on one machine, against a simulated connection. It is useful for finding problems, not declaring victory. Here is how I read the numbers.
The audit panel is lab data. Your users generate field data from real devices and networks. When the two disagree, trust the field. Users do not run on your Wi-Fi.
For most products, start with LCP, INP, and CLS: when the main content appears, how the interface responds, and whether the layout stays put. A high Performance score with poor INP still feels slow.
Ignore the score when it becomes an aesthetic gate instead of a user signal. A 99 that costs weeks of churn-risk refactors on an internal admin tool may not be worth it. Optimize for the experience users feel. Keep the score as a smoke alarm, not a target.