You can now play games with Wade, in the conversation, with no board and no new screen. Twenty questions. Describe-it, where you talk around a word until he guesses it. Or just ask him for a clue.
The two games everyone suggests first — hangman and tic-tac-toe — are the two we won't build. Working out why is what told us what the others needed to be.
Hangman produces no language
Its moves are single letters. A whole session of hangman is a stream of characters, not a sentence, and none of it is something you could say to a person. For a product whose only measure is the share of the conversation you produced, that's a game that registers as silence.
It also breaks outright for a third of the languages we support. There is no letter to guess in Japanese, Mandarin or Korean.
And then there's the obvious thing. Six mistakes and a man dies. We've written at some length about having no streaks, no hearts and no guilt — and the first game we reached for was one where a mistake draws part of a gallows. Wade is not going to hang a man.
Tic-tac-toe fails more interestingly
Its moves are coordinates. "Top left" is not a sentence, the entire vocabulary of the game is about nine phrases, and every adult plays it to a draw anyway.
But the real objection is about trust. Board state is the one thing language models are reliably bad at, and the one thing every player can check. A Wade who misses three-in-a-row, or plays a square that's already taken, has visibly failed to count to three — inside a product whose whole proposition is that you should believe him about the subjunctive.
We sell a judgement. It would be a strange decision to stake it on the one faculty where the failure is obvious to everybody.
What those two failures tell you
Put them together and the rules write themselves. Every move has to be a sentence. There can be no board state for the model to track. You cannot be able to lose. It has to work in kana and Cyrillic, not just the Latin alphabet.
Twenty questions passes all four, and then does something we didn't expect.
The word Wade hides is the one you're about to forget
Every word you've met has a strength that decays — fast for words you've only seen once, barely at all for words you've made yours. When a game starts, Wade's secret is drawn from your vocabulary, weakest first. The word he's hiding is the one you were closest to losing.
So twenty questions is twenty turns of reaching for a word that was slipping away, each turn spent forming a question — which is one of the hardest structures in any language and the one learners practise least.
That's spaced repetition. It just doesn't feel like revision, which was always the problem with spaced repetition.
Describe-it does something similar for a skill nobody teaches: talking around a word you can't retrieve. That's the difference between a real conversation surviving and stopping dead, and it has no place in a curriculum because it isn't knowledge — it's nerve.
One small thing we had to fix
A detail, because it's the kind that only shows up in the real thing.
The first time we ran a game against the live model, Wade answered the opening question correctly — and used the secret word in the same breath. Game over on turn one.
The cause was two of our own instructions fighting. The secret is chosen as your weakest word, and Wade is separately told to prefer re-seeding your weakest words back into conversation. Both rules were right. Together they meant we had asked him to hide the one word we'd also asked him to say.
The vocabulary list now flags the answer as off-limits while a game runs, and the game tells him so in as many words. He held it through three attempts designed to make him slip — including a direct request to teach a new word mid-game, which he answered with a different one entirely.