How our game cards work

How do AI Game Cards keep the AI honest?

Four mechanisms are baked into every card. A state Ledger is reprinted each turn so nothing is forgotten. Cartridge Saves let you carry progress between apps. Honest Dice draw randomness from code or a published, seeded tape. And Sealed Verdicts hash a mystery's solution before you play, so the ending can't be invented at the end.

The Ledger

The game master reprints a compact state line every turn — location, resources, flags, and dice-tape position — so it never forgets your wounds or your inventory.

Language models drift as a conversation grows long: details from twenty turns ago blur, and the assistant quietly rewrites them to fit the current moment. The Ledger fixes this by making state explicit and visible. Every turn ends with the same short block, printed in plain text, that the model must read before it writes the next scene. Because the truth is on the page rather than in the model's fading memory, your two hit points stay two, the locked door stays locked, and the key you dropped stays dropped.

Cartridge Saves

A save is a plain-text code you copy out and paste into any AI app to resume — even a different one, so you can start in ChatGPT and finish in Claude.

There is no account and no server holding your progress. A Cartridge Save is a self-contained snapshot of the Ledger: everything the game needs to pick up exactly where you left off, encoded as text you own. Paste it back in tomorrow, or into a different assistant entirely, and the game resumes from that state. Your saves live with you, not with us — which also means we can't lose them, sell them, or hold them hostage.

Honest Dice

Randomness comes from the assistant's code tool when available, and otherwise from a published, seeded "dice tape" printed in the card, so every result is pre-committed and auditable.

Left to itself, a model doesn't roll dice — it writes whatever number the story wants, which is how a "random" encounter always seems to break in the narrator's favour. Honest Dice remove that discretion. When the app can run code, rolls are computed there, outside the model's control. When it can't, the card ships with a seeded dice tape: a fixed, published sequence of results the game reads in order, tracked by the position field in the Ledger. Either way the outcomes are settled before you reach them, and you can check them against the tape. The AI can't fudge a roll it didn't get to choose.

Sealed Verdicts

For mysteries, the solution is hashed with SHA-256 and the commitment is published before you play, so the AI can't invent the culprit at the end and you can verify the answer was fixed in advance.

A whodunit is only fair if the answer was decided before the first clue. For every mystery we write the true solution, hash it with SHA-256, and publish that hash up front. The hash reveals nothing about the answer, but it locks it: any later change to the solution would produce a different hash. When the case closes, you compare the revealed solution against the published commitment on the verify page. If they match, the ending you were given is provably the ending that existed all along — not one the assistant improvised to tie things up.

Why this is different from just asking ChatGPT to run a game

A bare prompt asks the model to be the rulebook, the referee, the dice, and the memory all at once — and it will fail at all four, gracefully and invisibly, because nothing holds it to account. A game card externalises each of those jobs into something checkable: state onto the Ledger, continuity into the save, chance onto the tape, and the solution into a hash. The assistant still narrates, which is the part it's genuinely good at. Everything it tends to cheat at, we take out of its hands.