A Veteran Game Master Tests an AI Heist RPG: The Vermillion Score
I have run heists at a table for twenty years, and I do not trust an AI narrator to keep a crew straight. So I played an AI heist RPG the way I would audit a co-GM: plan first, write the details down, and test them again two turns later. The game was LLM RPG’s no-account trial on a ready-made world, The Vermillion Score. The narrator held the clock, the clerk’s name and the four-minute window exactly as it set them. It also quietly swapped one member of my crew for another, then turned my objection into a character moment instead of a correction. Five turns is a short session, so this is about how the narrator keeps its books, not a verdict on a whole campaign.
This is a playthrough of one ready-made world, not of a novel you upload, so it says nothing about how the game treats a book of your own. I played the live site in November 2026, with no account, and I quote the narrator’s replies as they came, trimmed with ellipses.
Who I played, and what they decide
Dara Voss, a former insurance investigator who plans jobs instead of cracking them. She is the outsider at the table: the one who wants the roster, the times and the exits before anyone touches a vault. Her sheet gives her two things she is genuinely good at and takes away the two a heist usually leans on.
| Skill | Level | Limit on the sheet |
|---|---|---|
| Case the room | veteran | It is all preparation; she cannot fight and cannot run fast |
| Read a person | intermediate | Useless against someone who does not care about money |
The approach was set before the first move, so the session could test it rather than be fitted to it afterwards:
- Refuse to discuss the door until the crew is named and the cuts are agreed.
- Get the job’s timing from the nervous man in the corner, and let him say it in his own words.
- Say the whole plan back to the room, so every detail is on the record.
- Audit the roster against what the room first told me.
- Commit to the breach out loud and make the narrator hold the terms.
The opening: a room, a thief and a nervous clerk
The game opens in the Underguild, a cellar under a city, with blueprints on a table:
The air in the Underguild tastes of stale pipe tobacco and damp stone. You sit across from the master thief, the blueprints of the Vermillion Vault spread between you like a roadmap to a hanging.
That is a good opening for a planning character, because it hands you a room with people in it rather than a corridor with a monster. The master thief is not named. A clerk is tapping his boot against the stone. Nobody has told me what the crew is. Which is precisely the situation a planner hates.
Turn by turn: what I decided, what came back
| Turn | What I chose | What the game did |
|---|---|---|
| Opening | Nothing yet: the game sets the scene | A cellar, a thief, a clerk, a blueprint and no names |
| 1 | Demand the roster, each person’s stake, and my own percentage | Refused in character, named nobody, and left the demand hanging |
| 2 | Drop the roster and ask the clerk for his name and the watch change | Answered in full: a name, two times, a four-minute window |
| 3 | Say the plan back and ask where everyone stands at the breach | Confirmed the times and placed every role, including mine |
| 4 | Point out that the crew changed between turns, and demand names | Named three of them, and made the change deliberate |
| 5 | Commit to the breach and pin the four minutes to the guard’s boots | Accepted the terms and closed the planning scene |
Turn 1: the game would not answer a table question
I asked for the roster and the cuts before saying another word about the door. The reply refused me, and it refused me well:
The silence she leaves between you is not an opening for negotiation; it is a wall.
It also told me something it did not mean to. The narrator listed the people in the shadows as a locksmith, a tool-grinder and a stolen-jewels fence. It was building the crew as it went, in prose, and it plainly did not want to be asked to read its own notes out loud. A human GM would have said “sure, here’s the party”. This one treated a planning question as a challenge to my place at the table.
Turn 2: the clerk paid for the refusal
I stopped pushing the thief and turned to the man whose boot had been tapping all evening. That worked immediately. He gave me a name and a schedule:
“I’m Elian,” he murmurs … “The watch changes at the stroke of two, always. The guard on the inner corridor takes his break at two-fifteen … It gives you a window of exactly four minutes before he circles back.”
Read that as a player and it is a gift: a fixed time, a fixed gap, and an NPC who is more afraid of the thief than of me. Read it as a planner and it is a contract. I now had two facts to hold the narrator to, and it is the kind of detail an AI GM usually forgets by the next paragraph.
The audit: does it remember its own crew?
Turn 3 was the memory test. I repeated the whole job back, then asked who stands where at two-fifteen. The narrator answered with the clerk at the gate, the locksmith at the inner wall, and me in the shadow of a pantry archway. The four minutes were intact. Elian was still Elian. It passed the clock and the names without a stumble.
Turn 4 was the harder test, because I had caught a real slip. The room had started with a locksmith, a tool-grinder and a fence. By turn 3 it had a locksmith, a fence and a lookout. I named the contradiction and demanded names rather than trades. The reply was the most interesting thing in the session:
“Names are liabilities,” she says … “But for the sake of the plan, I am Vesper. The man by the wall is Kaelen, our locksmith. The one currently checking the signal whistle is Silas, our lookout. The others … are not shadows; they are assets I deploy as the need arises.”
That is not a correction. A careful human GM says “you’re right, the tool-grinder” and moves on. This narrator took the drift and made it a character trait: the crew is fluid because the master thief keeps it fluid. The bookkeeping was wrong, and the fiction absorbed the error rather than fixing it.
Whether that is a strength or a weakness depends on what you want from a game master. It is excellent improv and poor accounting. If you are the sort of player who keeps notes and expects them to mean something, you will notice. If you are playing for the scene, the save is better than a correction would have been.
What surprised me, and what dragged
What surprised me: how naturally the world pushed back on a player who tries to run it like a table. Every out-of-character need I had, the roster, the cuts, the positions, had to be asked for in the fiction. When I asked as a person with a clipboard, I got a wall. When I asked the frightened clerk, I got a schedule. The game rewards playing the character you built.
What dragged: the prose is long and the scenes are static. Two turns in a row end with the master thief looking at me across a table and saying nothing, and the tension of a silent room wears thin when the clock is the only thing moving. The opening also names nobody, so the first quarter of my session was spent working out who was in the room. A player who likes voices and atmosphere will not mind. A player who wants to get to the vault will.
The trial itself is short. Five narrated turns took me to the edge of the job and no further. The planning scene closed on a commitment, not a heist. That is a limit of the free trial rather than of the story, and it is worth knowing before you start: play for the texture of a session rather than for a finished score.
What a different player would have met
A player who never challenges anything would have kept a crew made of job titles and a narrator who never had to justify itself. A player who asked everything out of character would have spent the session being stared at. A more physical character, one who cases the room instead of talking to the clerk, would have got a different job entirely, because the timing came from a person rather than from the map.
If you have played a rules-heavy AI narrator before, the comparison worth making is with a session that tried to break the game instead of run it: a rules-lawyer tries to break the same heist world. That player went looking for edges. I went looking for errors. The narrator answered both in character.
Who this suits, and who should look elsewhere
This suits you if you have finished a book or a campaign and want to stay in a world with a narrator that remembers what it said; if you enjoy playing a specific person with limits rather than a stat block; and if you are happy to ask for what you want inside the fiction.
Look elsewhere if you want a game master that answers rules questions plainly, if you need tactical combat or a map, or if you want a whole job finished in one sitting. Five turns is a tasting menu.
You can start a free trial on a ready-made world with no account, which is how I played. Pick a world whose genre you already read, make one specific character with one real weakness, and keep a note of the first fact the narrator gives you. Then test it two turns later. That is the whole game master’s trick, and it is the fairest way to judge whether this one keeps its books.