Despite its fame, The Edifice is a fairly rough-edged affair, and I seem to have bumped into just about every one of those rough edges.
As winner of IF Comp 1997 (the third ever), this game looms large in the early post-commercial history of the form. It has what may be the second-most famous puzzle in the history of parser games, the first being the climax puzzle of Spider and Web. I had tried The Edifice a few times before and never got very far, but this time I was determined to stick it out.
After an hour of mucking around in release 2, I took a look at the walkthrough. This is where I discovered that something I'd tried to do before, specifically (Spoiler - click to show)>SHARPEN STICK, will work, but you need to be holding something else in your inventory. The default failure message gives absolutely no clue about this, nor any clue that the verb in question can take an indirect object.
In the second scenario, I promptly managed to crash the game in a manner similar to what the release 2 notes say had been fixed. The same bug did not recur after a >RESTORE, and I was able to get to the part that makes this game notable. Author Lucian P. Smith took advantage of some rarely-used affordances in Inform 6 to create an NPC with his own language, which the PC (and by extension you, the player) must learn in order to communicate.
This puzzle is brilliant, and nearly unique. The only other game I know with a comparable task is Absence of Law. The constructed language used for the NPC consists of a relative handful of words -- only 25 are recognized by the game -- but the amount of interaction that is supported is surprisingly large. The "foreign language" parser takes a few shortcuts to simplify its implementation, which can result in confusing replies on occasion. It's not necessary to completely understand every word in order to solve the puzzle, and, in fact, even a complete understanding is not sufficient to do so.
Smith has been extremely open about his design and implementation of the puzzle, having provided an extensive overview of the former in XYZZY News issue #16 and an analogous implementation of the latter's code (using a different invented language) after the game's release.
For all of the effort that has been put into NPC conversation over the years, what's achieved here feels the most "real" of any that I've seen. The effect is almost totally an illusion, one created and maintained by the player's own slowly dissolving ignorance, but the breakthrough moments are extremely satisfying in a way that few puzzles based on comparably complex systems manage to achieve.
Unfortunately, the rest of the game isn't on a par with the centerpiece, though the apparent gap is substantially widened by relatively minor issues. As always, bringing a modern perspective to an older game results in certain small flaws being much magnified -- what was once an easily overlooked oversight can seem like a gaping hole in the interaction today. It's easy to forget that modern standards are informed by many years of lessons learned, and that a series of small improvements can result in a subjective improvement that is more than the sum of its parts. That said, my instincts run toward both two stars (for the numerous small issues) and four stars (for the brilliant centerpiece), so I'm splitting the difference.
By current standards, some of the frustration that a modern player might feel is very much not the player's fault, so for those trying the game my advice would be to seek hints as needed without worrying about spoiling the experience. Those interested in the game purely as an item of study might prefer to go straight to the materials linked above; these are undoubtedly worth the time to examine in detail.