I've submitted multiple titles to the ESRB, so I've only seen this process from the outside, but I do want to talk briefly about my experiences.
Over the last decade, the ESRB rating system has become much more predictable. Predictability in ratings is a GOOD THING. The quality of your content doesn't matter; only the experience itself. Crudely modeled mass slaughter will be rated the same as highly detailed mass slaughter if both are meant to convey the same experience.
That said, it would be helpful if the ESRB would play at least a small portion of the game, even in a pre-release state, prior to assigning ratings. The ESRB is the only rating system worldwide that does not play the game at least in part before assigning the ratings.
That isn't to say they don't play the games. They pick games at random after release and verify the questionnaire and video properly represent the final product. Games that don't line up can get fined, re-stickered, etc.
Regarding ratings "tweaks," they do happen, but there are reasons why the raters wouldn't be involved. Publishers and developers work with representatives at the ESRB as part of an appeal process and many times we'll get a higher rating or a descriptor because of a single item on the questionnaire or in the video. We will negotiate via phone to decide what needs to change in order to get a lower rating or lose a descriptor, we'll make the change, the publisher will sign a sworn affidavit saying the content or code was removed, and then we get the changed rating.
While I agree that a T-rating split could be beneficial, I don't see changing M as being beneficial at all. Right now, M and AO are perfect corollaries for R and NC-17, and I wouldn't mess with that dividing line.