Twenty Questions

Both replays are recorded games that ship with the page: the whole transcript, the ledger, the reveal and the commitment check, with no account and no credits. The free two-player mode gives you the ledger, the repeat detector and the information meter for a game you run yourself, on this device, against another person.

A ledger of every established fact, restated to the model on every turn. Repeat and contradiction detection: a question that re-asks an answered one, inverts it, or is already settled by an earlier answer is flagged in the transcript. A question quality score, and an axis coach that names the properties nobody has asked about yet. An information meter — twenty yes/no answers is twenty bits, and the remaining candidate space is shown falling as you spend them. Games are saved to your account and are resumable on another device mid-game. At the end you can export the transcript as Markdown, the fact ledger as CSV, the whole game as JSON, and a spoiler-safe scorecard.

A model asked to “think of something” and then quizzed about it has nothing to be consistent with. Left alone it drifts, and at the reveal it names whatever fits the answers it happened to give. Two mechanisms stop that, and both of them live outside the model.

When the app holds the secret, the secret is chosen once, in a setup turn, and hashed with a random salt into the short code shown above the transcript before you ask anything. From then on that exact string is re-sent on every single turn, so the model is reporting a property of a fixed thing rather than inventing one that fits. At the reveal the code is recomputed from the revealed answer and compared, and if they differ the app says so against itself. It is tamper evidence rather than cryptography — it proves the reveal matches what was locked in at question one, and a determined player could work backwards from a short word, so do not stare at it.

When you hold the secret, the risk is the mirror image: the model contradicting an answer you already gave, usually because the early turns fell out of the conversation. The platform truncates a long conversation oldest-pair-first, and in this game the first answers are the load-bearing ones. So every turn restates the entire fact ledger, and the browser checks each new question against that ledger — a near-repeat, the same question with the negation flipped, or something an earlier yes/no already settles is flagged in the transcript where you can see it.