What broke when our LLM Game Master could touch game state, and what still breaks
Matt, solo developer. October 2026. The free demo is playable now on Steam: Play the demo. It is part of Steam Next Fest, October 19 to 26, 2026 (the Fest ends at 10:00 am Pacific on October 26), and it stays up after the Fest. The full game launches November 10, 2026.
I'm a solo dev shipping a solo RPG built on Ironsworn's tabletop rules. The Game Master is a language model. The dice, the moves and every consequence live in a deterministic engine, and the model's output is parsed before anything reaches the character sheet. This is what that split cost to learn.
What broke when the model held state
- It named a place without binding to it. One tester sat on the same backdrop for 61 turns while the prose visited five locations, because the phrase "head outside" didn't match the parser's move list. The model was right. The seam wasn't.
- It narrated past a consequence the engine had locked. A reroll after a refused turn described dice whose effects had already been rolled back.
- It moved a progress track because its own text said so, with no move behind it.
- It believed a typed roll. A player wrote "(rolled a 9)" in the text box with no roll made, and the GM narrated the hit.
- It reintroduced an NPC the player had met two scenes earlier, because the context window keeps a pinned head and a sliding tail and the meeting had fallen between them.
What catches it now
The parser turns the model's tags into typed events and the engine refuses any tag with no move behind it. A cue that comes back without its effects trips a gate. The client tells the server when a roll cue was skipped, so the model is told no roll happened instead of guessing. The content screen refuses without touching the dice.
One receipt, from my own session
Undertake a Journey on Wits: action die 1, plus 3, against challenge dice 8 and 5. A miss. Supply went 5 to 3 and momentum 3 to 2 before a word of narration arrived. The number moved first. The prose followed. That order is the whole design.
What's still hard
Memory across a long session. One tester reminded the GM of 16 facts over a hundred turns. When I pulled the prompts, 13 of those facts had never been in the context at all. The chronicle recorded what started and never what finished. The current build keeps each NPC's own exchanges and the open vow's subject in view. Whether that's enough is the next thing I measure.
Try to break it
The demo is free on Steam during Next Fest and after: store.steampowered.com/app/5156570. The full game is at store.steampowered.com/app/4927450. Reports go in the Discord's #feedback channel, linked from the store page, or to [email protected]. The press kit is at isentroth.com/presskit.