The model answers questions. Code keeps score.
Zebra splits every call in two. A small language model on the device answers narrow questions: does this count, how likely was it, how close are these two things. Ordinary game code turns those answers into points, applies the house rules, and keeps the ledger.
Each question is small and specific, so a phone-class chip can answer it in about the time it takes to say the next word.
Answers are generated the same way every time and cached for the game. The same word gets the same call for every player.
Because code does the scoring, Zebra can always say exactly why a call is worth what it's worth.
The device.
Specs are design targets for a first dev kit, not a finished bill of materials.
- Size
- Palm-sized puck, about the footprint of a coaster. [FINAL DIMENSIONS]
- Microphones
- Four-mic array with beamforming. The direction each voice arrives from maps it to a seat.
- Display
- Round screen about the size of a watch face, under the top glass.
- Light ring
- Lit whenever it's listening. Changes color for calls, challenges and timers.
- Speaker
- Full-range driver behind the striped band, loud enough for a noisy room.
- Compute
- Phone-class chip with a neural processor. Runs a small language model, a speech recognizer and an embedding model locally.
- Storage
- The frozen judge model, plus rules modules loaded from game cards.
- Inputs
- NFC reader for seat tokens and game cards. The top surface is the challenge button.
- Camera
- Optional clip-on module with a physical shutter.
- Connection
- None needed to play. Wi-Fi only for updates and the Daily.
- Power
- Battery sized for a long game night. Charges over USB-C.
- Parts cost
- [BOM TARGET]
From voice to call.
Eight steps, all on the device. The target is a call within about a second and a half of the end of an answer, and faster for simple yes-or-no rulings.
- 1Listen
Voice activity detection wakes the pipeline only during a turn. The ring lights up.
- 2Transcribe
Streaming speech recognition turns speech into text as it's said.
- 3Attribute
Mic direction plus the seat map from token tap-in tells Zebra who spoke.
- 4Normalize
Spelling, plurals, sound-alikes. When unsure it asks on the display: “Did you say tarsier?”
- 5Judge
The game's rules module asks one or more of the eleven calls a narrow question.
- 6Score
Plain code turns the answer into points and applies the house rules.
- 7Explain
A reason in under 12 words, shown on the display and spoken aloud.
- 8Lock
A short challenge window opens. If nobody taps, the call is final.
Eleven calls, and how each is worked out.
Every game is assembled from these. A game designer decides which questions to ask and what the answers are worth; the judge answers them the same way on every Zebra.
| Call | The question | How it’s worked out | Used in |
|---|---|---|---|
| Counts or not | Does this answer follow the rule? | The model gives a probability that the answer is valid. It counts above a set threshold. Inside a narrow band around the threshold, the player gets 10 seconds to defend it, and the defense is judged too. For factual categories like animals or countries, a curated list is checked first and the model second. | Categories+, Taboo, Just One, Five Seconds, Dumb It Down |
| How rare | How unlikely was it, given that it counts? | Points are bits of surprise: round(−log₂ P), kept between 1 and 5, where P is the chance a table gives that answer. P starts as the model's estimate and is replaced by real answer frequencies as more Zebra tables play. | Categories+, Outlier, Survey Says |
| How close | How related are these two things? | Cosine similarity between embeddings, rescaled against a reference set so 50 means a typical pair and 90 means near-synonyms. | Missing Link, Close Call, Mind Meld, Reverse Engineer, Codenames |
| What a crowd would say | What would 100 of these people answer? | The model answers 100 times as the named crowd. Duplicates merge into slots, ranked by count. The board is fixed for the whole game so it can't shift mid-round. | Survey Says, Who's the Crowd, Wavelength |
| What you meant | What is this person describing? | The model guesses from the description alone and reports its top guesses with a confidence for each. Used wherever the judge plays the listener. | Dumb It Down, Technically Correct, Reverse Engineer |
| Whose style | How much does this sound like the voice on the card? | A classifier score shown as a percentage, so the call reads “91% detective” rather than a vibe. | Ghostwriter, Balderdash, Apples to Apples |
| Against the rubric | How does this do on each named criterion? | Fixed rubrics of three or four criteria, each scored 0 to 3 against written examples. The judge must quote the line that earned or lost each point. | Flag on the Play, Bars, Yes And, Hijack, Scavenge |
| Secret rule | Does this fit the rule nobody's been told? | The rule is drawn at the start and its fingerprint (a hash) is shown on the display. At the end the rule is revealed, and anyone can check it matches the fingerprint. | Secret Rule |
| Contradictions | Does this clash with anything said before? | Every statement goes into a ledger. Each new one is checked against it, and a conflict is reported with both lines quoted side by side. | Cross-Examination, Alibi, Hijack, Binding |
| Keeping the books | Whose turn, what's owed, who's out? | Plain game state in code: roles, votes, timers, promises, trades. No model involved, which is the point. | Werewolf, Binding, Trap Words, Taboo |
| What's in view | What's in front of the camera? | Object and gesture recognition from the optional camera, processed on the device and never stored. | Scavenge, Draw & Act |