Leather Armor

Foreword

What you have before you is a love letter, from a solo RPG player, to the unforgiving and cruel dungeons of old and to the more modern RPG iterations that aim to keep at least the rule-heavy part of it alive. The system presented here uses a large language model, an artificial intelligence, if we are to be generous with that term. It serves as your referee, adversary, and narrator in solo adventures through pre-written dungeons and wild, dangerous areas.

The AI operates on a rule system roughly based on Pathfinder^* 1E. It weaves together a lot of deterministic code tracking your adventure. The code tracks your hit points, your rations, and your torches. It does not fudge results. When you fail a save against poison, you are poisoned. When you fall in combat, you die. This is the intended experience. The AI is there to interpret your decisions and actions into hard-coded facts that can be digested by a deterministic, and hopefully unforgiving, system.

The philosophy behind the stories and dungeons owes more to the original tabletop tradition of Chainmail and the earliest dungeon crawls of the 70s than to the modern notion that the player character is a hero on a predetermined journey. Here, you are an adventurer. You should be very likely to die in your first dungeons, and returning alive from a failed expedition should be its own kind of victory.

On the Limitations of the Machine Referee

It would be dishonest not to address this plainly. The AI that runs your game is a language model. It is not intelligent. It does not understand the rules the way a human game master does; it approximates them, often convincingly, sometimes not. There will be moments where it loses track of a detail, applies a rule incorrectly, or generates a response that doesn’t quite follow from what came before. Our hope is to fine-tune the models as we go, as well as have it be responsible for an ever-diminishing body of work that maintains your average adventure. Ideally, it will be the randomness engine behind a hard rule-heavy system.

We have worked to constrain it: to give it structure, to ground it in the actual rules text, to make it track state with reasonable fidelity where cold math alone cannot. But we will not pretend it is something it is not. A language model is a sophisticated text predictor. It can maintain a surprisingly engaging adventure, and it can also, on occasion, hallucinate a door in a wall that wasn’t there, or forget that you already searched the chest. It will detail every chest as intricate and olden if we do not slap some sense into it. That is what we are doing from the start, trying to shave off its vices and keep what’s useful of it.

If you have played tabletop games with a distracted or novice game master, you will find the experience familiar. The difference is that the AI never tires, never cancels the session, and if it’s pedantically disagreeing with you, you can send us an e-mail to plead your case, and we will likely be less ego-driven than your average novice GM that is too in love with his own story to let you play it your way.

We consider this an honest trade.

On the Nature of the Adventures

The tone is grim. Not gratuitously so, we are not interested in shock for its own sake, but the stories we have written should have the player character as a mortal creature in a hostile environment. Resources are finite. Rest is dangerous. The things in the dark are not calibrated to your level, because why would they be?

You will die. Probably often, at first. A first-level character with eight hit points and a shortbow does not have plot armor. The AI will describe your death with some flair when it arbitrates your demise, but sometimes all you get is a description of the mechanics of the roll that killed you and a system message ensuring you know you died. You may roll a new character and try again. And again…

The reward for survival is not a cinematic ending. It is the accumulation of small, hard-won gains: a map of the upper level, a magic weapon you don’t fully understand, enough gold to buy better armor. The stories that emerge are your own. They are better for not having been scripted besides the premises. We did write with certain solutions and paths in mind, but we believe it’s better for you that we are not trying to railroad you, and instead, AI will do its best to try to keep up with you, and it will gladly ignore everything we planned for you if you start going your own way.

On Playing Alone

Solo play has always existed at the margins of the hobby. There have been paragraph-books, oracle systems, and procedural generators of varying quality. What a language model offers (imperfect as it is) is something closer to the feel of having another person across the table. Not identical to it. But closer than anything before.

You type what your character does. The AI tells you what happens. Sometimes it asks what you’d like to do. Sometimes it simply tells you that the floor has given way and you are now thirty feet below where you were, and if you roll badly enough, it tells you your adventure came to a premature end.