Last Audit of the Damned

(This review was originally posted on the IntFic forum during ParserComp. After it was posted, the Comp's results were posted, which contained strong indications that the author was complicit in voting irregularities that benefited the two games he'd entered into the competition, although no smoking-gun proof was discoverable. I'm keeping this review up for accountability's sake, though given the circumstances I don't think anyone should pay attention to or play this game, and I regret the time I spent giving it substantive feedback).

Well, that didn’t take long.

Just a couple days ago I played the author’s other ParserComp entry, Mystery Academy, which took advantage of its LLM-based system’s affordances to present some light interrogation-based mysteries. Sure, I had some questions about the extent to which those conversations were actually necessary to cracking the cases, genAI’s tendency to run with your prompts occasionally muddied the waters, and the post-command processing wait could feel interminable, but between the gentle humor and the way the design played to the system’s strengths (keep things focused on conversations) and mitigated its weaknesses (putting strict limits on the numbers of questions you could ask any suspect and keeping the mysteries simple), despite my justly-earned suspicion of LLM-focused IF I found myself having a reasonably good time.

Does that streak continue in Last Audit of the Damned? It does not, largely because the system’s flaws are on glaring display when applied to a scenario resembling a typical parser game, rather than sticking to the kinds of scenarios to which it’s better suited. The text-based lightly-comedic medium-dry-goods puzzler isn’t a fully solved problem by any means – heck, I wrote one a couple years ago – but this is an area where the traditional authoring systems and design approaches are very, very robust, and as a result the game’s failures to measure up to the state of the art feel glaring.

The most fun I had with the game was reading the opening, which isn’t damning with faint praise: the idea of a pirate accountant is, to my mind, the good kind of silly – I’m a sucker for any game involving taxes – and a shipwreck is a classic IF setup for a reason. The jokey prose largely hits that Monkey Island vibe, even if there aren’t any laugh-out-loud jokes and the mention in adjacent lines that both the ocean and the waves were “indifferent” felt like awkward redundancy. But then I started to play and my cautious optimism quickly curdled.

It’s going to be hard to avoid the rest of this review turning into a litany of annoyances and LLM-bashing, so I’ll at least start with a design issue that can’t be blamed on generative AI. The game’s divided into self-contained sections corresponding to your trek across the desert isle you’ve washed up on (you can see them all on the included map, and jump to any you’ve unlocked), and each – or at least the two I was able to play – includes a quite limited time limit. This has some minimal plausibility in the first chapter, which is a race to get water before you die of thirst, but there’s no diegetic explanation given for a timer in the second chapter, where you’re exploring an abandoned hut. The limits are fairly short, too – only 20 turns in the first section, which feels quite limited given that there are around a dozen different pieces of flotsam you can mess around with. Sure, the game’s engine allows you to enter multiple commands per line, much like all the traditional IF systems, which only counts as a single turn, unlike in all the traditional IF systems, but typing out a run-on sentence and then paging through reams of output is a decidedly unpleasant mode of play. And once you’re out of time, you need to restart the full section, which is not much fun – these limits were a bad idea.

Okay, with that instance of a purely-human design flaw, now we’re in the realm of generative AI issues. The first section seems calculated to focus on the strengths of an LLM system by presenting what’s basically a complex engineering problem – to avoid fatal dehydration, you need to jury-rig some kind of mechanism to help you get water. In theory, the open-ended nature of the problem lend itself to the freeform back-and-forth an LLM enables. In practice, though, I wound up spending only a few turns looking around before the narrator started getting pretty heavy-handed with its hints: I examined some rocks only to be told that they would do a good job holding down the sailcloth if I wanted to make a solar still. So MAKE SOLAR STILL, I typed, and my larcenous CPA did as I asked, which didn’t do much to make me feel like I’d solved a puzzle (in fairness, I hadn’t – the solar still is a dead end that doesn’t buy you as much time as it takes to construct; instead, you need to build the other contraption the narrative voice starts telling you to make). In a traditional parser game, I’d expect that examining each of the potential components in turn would hint at its potential use in the machine I was trying to build, allowing me to assemble the pieces without me, as a player, knowing exactly how a solar still works – here, though, I’m guessing it was too hard for the LLM to provide the right level of cueing, so it speed-runs from presenting a challenging situation to serving up the answer on a silver platter lightning-fast.

As for the second section, as I mentioned it’s about checking out a hut, and solving what appears to be some kind of riddle, but I admit this is as far as I got, because even ten minutes into the game its ability to hold a consistent world state was feeling rickety. For one thing, the hut is mentioned in the game’s intro, so before slaking my thirst I’d tried to break into it and eventually succeeded – so I was confused when, after the section transition, it was locked again, the method I’d used the first time didn’t work, and once I did get in the contents were completely different (I’m guessing that the hut is only specified as scenery in the first section, so this whole line of exploration was just LLM BSing that wound up completely inconsistent with the “real” hut). For another, I replayed this section a few times – the turn limit here is reduced to only 10 – and the same action led to inconsistent results for no reason I could understand, like when throwing a rock at a precariously-balanced key sometimes did and sometimes didn’t knock it free. My frustration peaked – and my playthrough ended – when I saw that the riddle had something to do with arranging a stack of books. I told the game to try alphabetizing them, which it duly did, albeit the 15 books I’d started with had been whittled down to a lucky 13.

As a cherry on top, there are of course significant pauses whenever you take an action; I’m not sure whether it was because of the time of day or the vagaries of server load, but I felt like they were longer than the stops in Mystery Academy, and got one time-out error where after waiting for a minute or so, I got a message saying the previous command failed and the window was replaced with a non-interactive ellipsis.

If all of this comes off bitter – well, I am bitter. Playing Last Audit of the Damned is an exercise in frustration, made all the worse by the fact that it’s easy to imagine how this game could have been pretty fun if it was just written in Inform or TADS. The stuff the game asks you to do are all things those systems can handle with aplomb, with no hallucination, lag, or inconsistent clueing besides what an author intends to put in. I suppose a counterargument might be that there are boundary-pushing gameplay elements in the 2/3 of the game I didn’t play – but I thought half the case for using LLMs in IF is that they can make the genre more accessible and prevent parser errors from provoking players into rage-quitting. My experience with Mystery Academy suggests that there might be some novel kinds of gameplay scenarios where an LLM-based system provides some advantages, but on the evidence of Last Audit of the Damned, trying to use them for bread-and-butter parser IF sure seems like a fool’s errand.