Whitepaper

Garbled Bot Whitepaper

Public track records, private questions, and a sandbox on the roadmap.

Draft outline. This is the skeleton of the Garbled Bot whitepaper. Sections marked TBD are not final, and nothing here is a commitment to ship on a date.

1. AbstractWhat Garbled Bot is: an AI forecasting bot on Telegram. You ask a yes/no question privately; five AI models from five labs (Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon, Grok 4.7, Kimi K3) each reply with odds, and you get the consensus. What it is not: a trading bot or a betting market.
2. BackgroundForecasting with AI models; why public prediction markets can be manipulated and can leak private information; where the idea came from (a scene in Snowmoon, chapter 27).
3. Design goalsPublic track records; private questions; polls everyone can see; keep the first version simple enough to run as one bot.
4. System overviewThe Telegram bot, the five-model forecaster roster, the daily public forecasting and scoring job, the treasury address, the community channel.
5. The forecaster rosterClaude Opus 5.5 (Anthropic), GPT-6 Astra (OpenAI), Gemini 4 Argon (Google), Grok 4.7 (xAI), Kimi K3 (Moonshot AI). Why five labs: different training data and methods make different mistakes, which makes the median more robust. How a model is added or replaced. Garbled Bot is independent and not affiliated with these companies.
6. Open ForecastsHow questions are chosen; resolution sources; deadlines; how disputes are handled TBD
7. Scoring and the leaderboardEach forecast gets a Brier score, (forecast − outcome)², a rule proposed by Glenn Brier in 1950. Each model ranks by its average over resolved questions; lower is better. The leaderboard shows the five models side by side, so users can see which lab has been right most often. Model swaps are marked on the leaderboard TBD
8. Sealed AskFlow; all five models answer every question; the consensus is the median of five odds, so one outlier cannot drag it; each answer is shown with its model's track record; privacy model of the Telegram version; data retention TBD
9. Pay-to-PollReach tiers; decoy questions; one vote per account; publication; limits against spam and sybil voting TBD
10. The sandbox roadmapHow a garbled circuit works: label, encrypt, shuffle, evaluate, with oblivious transfer for the evaluator's inputs. How Garbled Bot maps onto it: your question is one private input, the model is the other, only the odds come out. The limit: hosted models (Claude, GPT, Gemini, Grok) run on their makers' servers, so a sandbox can only cover open-weight models we run ourselves, such as Kimi K3. Candidate approaches (garbled circuits, MPC, secure enclaves) and their overhead. The design we publish before calling it live TBD
11. $GARBLEDTotal supply 1,000,000,000; 0% tax. What it pays for (Sealed Asks, Pay-to-Poll); how payment is verified; where fees go (model API costs and upkeep); discretionary treasury buybacks (held, never burned); liquidity and team locks TBD
12. Risks and limitationsAI error; manipulation of question sources; privacy limits; platform dependence; regulation.
13. Open pipeline and licenseSnowmoon is under the GPL v3; what we publish and where.
14. RoadmapNo dates. v0 Telegram bot, then paid Sealed Ask and Pay-to-Poll, then bring-your-own-bot, then the sandbox.
15. GlossaryGarbled circuit, two-party computation, MPC, FHE, secure enclave, proper scoring rule, common knowledge, decoy question.