"AI Dungeon" I believe attempted this. I recall hearing it did indeed make a good parser, but the major issue was around context depth.

I wonder what it'd take to use an LLM as a parser, with ground truth available from a separate game instance, and the LLM's role limited to providing descriptions and handling player input. That might be a good way to get structure and consistency while not losing the improvisational, storytelling sort of vibe.

Yep. AI Dungeon proved this a while ago (you can do anything you say you can), but the engine suffers from a lack of constraints. It's like a D&D game played by drunk edgelords DMed by someone on acid.

You as King Arthur giving up on Excalibur and slaying a dragon with the BFG 9000 before dying on your chamberpot from dysentery is allowed to happen.

That's where Inform[ish] comes in. You can implement a level of sanity to the story that the verbs have to work around. Doing something ridiculous like `ride the Rat King` is still going to kill you if he's flagged as aggressive (as opposed to working through your differences and going on a journey together, bonding with the trusty steed you just subjugated).

You cannot engage in rat diplomacy. They are never to be trusted!

“would be... could...” seems unnecessarily cautious language given that AI text adventures/RP are very much a thing.

(Another strength is that you can have a usable and entertaining one just with little more than text for the starting point, and description of the tone/premise, and a handful of sample interactions.)