Have you directly compared the live version against that in the AMP Cache? The idea that PSI preloads the page before starting the test doesn't make much sense

Have you directly compared the live version against that in the AMP Cache? The idea that PSI preloads the page before starting the test doesn't make much sense at all, and doesn't mesh with my own experiences (in the same space).