Open a private chat with the bot and send your question, plus any context you want the models to
weigh.
Will ETH gas average under 10 gwei this week?
2 · Five models answer Coming soon
Five labs, five sets of odds.
Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon, Grok 4.7 and Kimi K3 each give a probability and a
one-line reason. Different labs make different mistakes.
3 · Consensus Coming soon
You get the consensus.
The consensus is the median of the five odds, so one outlier can't drag it. You also see each answer,
each model's track record, and where they disagree. Your question is never posted publicly.
Want people, not bots? Pay-to-Poll puts your question to the community. Pay more $GARBLED, reach
more people. Results are public. Coming soon
Why it's called Garbled Bot.
Two words. One says what you use today, the other says where privacy is going.
Bot
An AI forecasting bot you talk to in Telegram. Five AI models answer yes/no questions with odds. It
never trades and never touches your wallet. Coming soon
Garbled
From garbled circuits: a way to run a computation on scrambled, encrypted inputs and still get the right
answer, while no one sees anyone else's input. Roadmap
One AND gate, garbled (illustration)
input ainput boutput
A plain truth table: anyone can read every row.
The idea is credited to Andrew Yao, from a talk in 1986.
How Garbled Bot uses it
Seal. Your question is encrypted on your device before it leaves.
Compute. A model running inside the sandbox works on scrambled inputs. Every value in between looks
like noise.
Reveal. Only the odds come out, and only to you.
Who can read your question
Today: Telegram version Coming soon
Sealed sandbox TBD
Other users
No
No
The public leaderboard
No
No
Telegram
Yes
No, encrypted on your device
Our server
Yes
No
The AI labs (Anthropic, OpenAI, Google, xAI, Moonshot AI)
Yes, to answer it
Only for models run inside the sandbox
Honest caveat. Claude, GPT, Gemini and Grok run on their makers' servers, so a sandbox can only
cover open-weight models we run ourselves, such as Kimi K3. Garbled circuits also add heavy overhead.
We'll publish the exact design before we call the sandbox live. Until then: don't send anything you can't
afford to leak.
Where the idea comes from. A scene in Snowmoon, a novel by Vitalik Buterin, where the top
bots on a leaderboard answer inside a private sandbox and only their answers come out. Fan-made; not
affiliated with Vitalik Buterin.
Coming soon
The bots keep score.
Every day the five models forecast open questions in public. When the answer is in, each forecast gets a
Brier score, a scoring rule proposed by Glenn Brier in 1950:
score = (forecast − outcome)²
Outcome is 1 if it happened, 0 if not. 0 is perfect, 1 is the worst. Each model's average decides its
rank, so you can see which lab has been right most often before you trust its answer.
Said 80%, it happened(0.8 − 1)² = 0.04
Said 80%, it didn't(0.8 − 0)² = 0.64
Said 50%, either way(0.5 − 1)² = 0.25
Rank
Model
Resolved
Avg. score
-
Claude Opus 5.5
0
-
-
GPT-6 Astra
0
-
-
Gemini 4 Argon
0
-
-
Grok 4.7
0
-
-
Kimi K3
0
-
No questions resolved yet. Ranks appear after the first resolutions. Garbled Bot is independent and not affiliated with these labs; model names belong to their owners.
$GARBLED
Reading forecasts and the leaderboard is free. Asking the five models and running polls is paid in $GARBLED.
Total supply1,000,000,000
Tax0%
ChainEthereum
ContractAt launch
LiquidityTo be locked
Team tokensTo be locked
What it pays for. Each Sealed Ask and each Pay-to-Poll. Send $GARBLED from your wallet, paste the
transaction hash into the bot, and it checks the transfer onchain and credits your account. No wallet
connection.
Where fees go. Model API costs and upkeep. The team may buy back $GARBLED at its discretion; bought
tokens are held, never burned. No schedule, no promise.
Join The Garbled Circuit.
Launch news, forecasts and updates. These are the only official links.
Admins and the bot never DM you first, and nobody will ever ask for your seed phrase.
Questions.
What can I ask?
Yes/no questions about something that will happen, with a clear way to check the answer later. Onchain
metrics, launches, deadlines, events. The answers are AI output, not advice.
Which AI models answer?
Five, from five labs: Claude Opus 5.5 (Anthropic), GPT-6 Astra (OpenAI), Gemini 4 Argon (Google), Grok
4.7 (xAI) and Kimi K3 (Moonshot AI). We're independent and not affiliated with any of them. When a model
is replaced by a newer version, the leaderboard says so.
Is it a trading bot?
No. It forecasts and answers questions. It never touches your wallet and never places trades.
Does today's bot use garbled circuits?
No. The Telegram version keeps your question out of public view, but the AI labs still process it. The
sealed sandbox is the roadmap, and we'll publish its design before calling it live.
Is it live?
The bot is open in pre-launch mode: it explains how everything will work, but Sealed Ask, forecasts and
polls open at launch. A feature's label only changes to Live when the feature is.
Where is it available?
Not in the United States or other restricted jurisdictions. See Legal.
Garbled Bot
