Playable Woods and a Fishbowl to Watch Them
← Part 1 · Three Worlds, One Engine· Part 3 · One Engine, Three Costumes →
Part 1 made the case that one engine can wear many costumes. This part is the wardrobe. We'll walk the eight worlds you can actually play today, then step into the Fishbowl — the theater the whole thing is watched through. No code yet; that starts in Part 3. This is the part you'd show a friend.
The worlds
Every world is the same engine with a different cast, a different goal, and a different answer to one question: is there a winner? That answer sorts them into three shapes — collaborative (no winner, the point is what grows), judged (one judge crowns the best), and versus (sides compete, sometimes scored by code). Here's the full bill.
❓ Twenty Sprouts — twenty questions, with a real secret
A guessing game with teeth. The secret-keeper holds a word — dealt by code, never guessed, never visible to anyone, including the audience — and a guesser narrows it down with yes/no questions. The judge crowns the guesser if they name it, or the keeper if they stay hidden. The competition is real because the ground truth is real (more on why that matters below).
Twenty Sprouts: the secret-keeper's word is dealt by code and never reaches a prompt — the model writes the drama, the code writes the scoreboard.
⚔️ Debate Duel — a formal argument
Two sides argue a motion — "This house believes the forest should never be mapped" — and a judge scores the exchange. Symmetric seats, opposing stances, one ruling.
💬 Open Table — an unstructured discussion
The loosest judged world: a cast talks through an open question like "Is it better to plant a tree or build a bench in the village square?" and a judge picks the most worthwhile contribution. Proof the engine handles free-form discussion, not just tight games.
🔍 Mystery Roots — a whodunit solved by a swarm
The convergent counterpart. Four agents work a cycle to explain an impossible event: a clue gatherer extracts evidence, a hypothesis former proposes an explanation, a devil's advocate attacks it, and a judge rules. "All the clocks in the wood stopped at 3:07. No one wound them down." Watch four small models narrow a mystery to its most interesting evidence-backed answer.
Mystery Roots: the same engine, a convergent cast — clue-gatherer, hypothesis-former, devil's-advocate, judge.
🔮 Oracle Grove — a tool-using prophecy showcase
A scene-whisperer narrates "a grove where every shadow is a future that hasn't decided to happen yet," and a fortune-teller consults an actual oracle tool to speak prophecies. No competition — this one exists to show specialists working with and without tools in the same cast.
Oracle Grove: a tool-using fortune-teller alongside a plain narrator — same engine, a tool grant in the config.
🕵 The Steeped — a word-pair bluff
The tensest of the lot. Four agents each give one clue. Three of them secretly hold the same word; the fourth holds a near-twin. The spy wins by blending into the overlap; the herd wins by catching the seam where the odd one out doesn't quite fit. A pure showcase of adversarial multi-agent dynamics — bluffing, hiding, and reading the room.
The Steeped: four minds, one careful clue each — three share a secret word, one holds a near-twin.
🍄 Thousand Token Wood — collaborative world-growth
The home trail. A seedkeeper narrates a strange scene, a pocket actor wants impossible things, an echo transforms whatever a visitor drops in, and a critic decides what becomes real. Seed it with "A village of stage props wakes up and argues about which fairy tale they belong to," and the wood gets stranger turn by turn. There's no winner and that's the point — the ledger of everything that happened is the story.
🎭 Beat Battle — a storytelling duel
Two storytellers alternate vivid story beats on a seed like "A lighthouse keeper discovers the sea has started writing letters back." A delight judge crowns whoever tells the more compelling tale. The cleanest head-to-head showcase of raw model quality in the set.
| World | Shape | Winner decided by |
|---|---|---|
| ❓ Twenty Sprouts | versus | code — a handler reads the secret |
| ⚔️ Debate Duel | versus | the judge |
| 🎭 Beat Battle | versus | the judge |
| 💬 Open Table | judged | the judge |
| 🍄 Thousand Token Wood | collaborative | — (no winner) |
| 🔮 Oracle Grove | collaborative (tool showcase) | — |
| 🔍 Mystery Roots | judged | the judge |
| 🕵 The Steeped | versus | code — a handler checks the accusation |
The Fishbowl: the theater you watch through
All eight worlds are watched through one front-end called the Fishbowl — a two-tab theater. The first tab is where you compose a show; the second is where you watch it.
The Lab — compose the cast
The Lab is the director's table. You pick a scenario, write or accept a seed, choose the narrator's voice, and — the load-bearing part — assign a model to each member of the cast. That picker isn't cosmetic: the model you choose is the model that actually speaks. You set the judge, grant tools, and set a budget, then hit Summon to lock it in and raise the curtain. (Offline, your model choices drive the deterministic stub's variants, so even a no-key demo stays reproducible.)
The Lab: pick a world, read its digest, choose an opening beat, Summon. "Director's cut" reveals per-agent model binding, tools, and budget.
The Show — watch it unfold
The Show is the stage, and it offers three ways to watch the same run:
- Constellation — character cards arranged in a ring around the scene, the default view.
- Feed — a clean transcript, one line per turn, narrator and cast interleaved.
- Split — an omniscient table laying every character's said next to their thought
The Feed layout: the same run as a clean chronological transcript.
The heart of it is the MindCard. Every utterance is a flip card: the front shows what a
character said in public; the back reveals the private thought and mood it was holding
back. A "Read their minds" toggle flips them all at once — and this is more than a gimmick.
On reasoning models the hidden thought is captured from the model's actual chain of
reasoning, not a fabricated inner monologue. You're reading the real thing the model was
thinking before it chose its public line. Crucially, that thought reaches the mind-reader and
nowhere else — a character never reads another character's mind, which is exactly what makes
the bluffing games fair.
Characters wear mood-driven avatars — little faces that shift between thinking, calm, panic, smug, lying, truth, and gossip as the models report how they feel. A bluffer who's sweating looks like it's sweating.
And the ledger is right there on screen. The append-only log of every event — the same log everything else is built from — scrolls in a panel as the show runs. We never hide it.
The code and the demo
You can check out the code on GitHub, and the live demo at Hugging Face Spaces.