To be fair, until a few years ago the testsuites were mostly terrible and laughably bad — little more than skeletal things to ensure there was some basic intero

To be fair, until a few years ago the testsuites were mostly terrible and laughably bad — little more than skeletal things to ensure there was some basic interoperability (and not necessarily covering all the features of a spec). It took a huge push from a lot of people to get to where we are today; I remember pushing for the CSS 2.1 testsuite, which was one of the first to have a reasonable number of tests though it's still far too shallow, to be largely automatable and getting a fair bit of resistance (because of fear of delaying the testsuite and the spec demonstrating interoperability of implementations and reaching REC), and that was in 2010 (though, holy cow, that's almost five years ago now!). As far as reference implementations go — I'm on the whole against them (you can end up having to implement bug-for-bug compatibility, which doesn't help anyone and if those bugs fall out of implementation details they can lock you down into specific implementation designs). It's better to have clear, unambiguous specs (preferably with next to nothing undefined) and multiple interoperable implementations — because it shows that the spec is clear enough for multiple people to implement it and come up with the same result. The machine readable route is a dangerous one — you don't want to end up in a situation where everyone just uses the same implementation, because monocultures aren't at all great.