The i6 and i7 languages have a guaranted internal consistency starting with grammar. Even if6 ('old' compared against i7) it's an incredible language targetting the Z Machine with a really easy OOP architecture to design your games. LLM's often lose the internal storyline on bizarre ways.
An Inform 6 compiler can run under DOS/386 or Linux/BSD for 486 and compile a really large and functional game in seconds. For Inform7, maybe a Pentium II/III and a bunch of seconds too.
An AI to do the same, not even a 'modern' Core Duo with a GL 2.1 adapter (the lowest of the lowest bearable specs for light web browsing and office work) can't even run a consistent world.
That’s not my proposal. The LLM doesn’t run the world. It is simply the interface. It tells you the state of the game and tries to enhance the description. The game is a separate application and the LLM is simply an agentic layer between you and the actual game.
The engine doesn't understand natural English. It only understands the hard-coded words. The game author forgets to include some common verb or an uncommon spelling, and oops, an otherwise great puzzle is now very frustrating.
The engine is the ZMachine, and depending on the target and the grammar (IF6+English, or better, i7) the 'guess the verb' issue straightly died in the 90's.
Your comment coudn't be more outdated since the Curses! release for the ZMachine in 1993.
The v5 machine release was much better than the v3 one, and the v5-V8 ones allowed semi-complex phrases with indirect object pronouns after a previous entered phrase and much more.
Go play Anchorhead and compare it to a z3 machine game from Infocom, or any game made it with Puny Inform.
