* To pages which are fast. This is based on measurement of a specified benchmark, regardless of the implementation used to achieve those results.
* To pages which meet an AMP specification (do this, don't do this, etc) and are validated to conform to the spec. Publishers can bring their own hardware (CDN/Servers) and their own code (serverside language, javascript framework) to do it. Google's AMP servers as the defacto/reference implementation, but publishers/competitors are free to deviate from it. (Look at prebid.js as an example).
* To pages which use the AMP Framework code as-is (serverside and clientside) provided by the Google Team, but you can bring your own hardware or have to use a set of accredited hosting providers.
* To pages which use AMP and are hosted by Google.
Obviously the last one is the "simplest" in regards to getting it off the ground, but AMP has been around long enough and Google has enough resources that they could definitely do one of the other options which go much more to the open web.
Note: if you first thought says "verification of any of these things is way too hard and it's so subjective, so we shouldn't even try", then I'd recommend looking at Google's policies for when you can use their "skippable ads" (TrueView) on your video player. It's very manual and up to a lot of human interpretation when you get into the weeds: https://support.google.com/admanager/answer/3522024?hl=en#tr...
Or if you think "maintaining a list of approved vendors is too hard of a problem", look at their list of approved vendors for Advertising: https://developers.google.com/third-party-ads/googleads-vend...