Calibration curve

Methodology

How markets are made and settledTraders’ orders set every price. AI turns a question into a market and, after it closes, judges the outcome. This page describes how that actually runs — with the numbers the system is configured with.

Models / roles: 3 / 4; Claude Opus · GPT · Grok
Auto-settle gate: ≥ 0.80; Below it, an operator reviews
Live sources: 23; Corpus refreshed every 30 min
Debate cap: 30 min; Up to 30 rounds

Last updated 2026-10-13

01Prices

Orders set the price, not the AIA price on Conviction is where orders to buy and sell YES and NO shares meet. 0.57 means the market puts 57% on that outcome; if the outcome happens, each share pays 1 USDT.

ASigned orders
Orders are signed in your wallet (EIP-712) and placed on the order book. Signing alone moves no funds.
BMatching
The moment an order arrives, a keeper pairs it with resting orders on the other side whose prices cross.
COn-chain fill
Matched orders fill on the CTF Exchange contract on BNB Chain, and shares land in your wallet as ERC-1155 tokens. Gas is paid by the platform.

02Market creation

From one question to a marketAnyone can type a question at /markets/new and the harness drafts a market. You confirm the draft, it is published, and trading opens about 10 minutes later. If a draft fails any stage, it is discarded.

01UnderstandClaude Opus 5.5
02ResearchResearchSwarm
03DebateOpus · GPT · Grok
04JudgeClaude Opus 5.5
05Final reviewClaude Opus 5.5
06PublishScheduler · webhook

↻ Debate ↔ judge — until it converges, up to 30 rounds or 30 minutes

↻ Rejected in final review → back to the debate, up to 10 times

↻ Debate ↔ judge — until it converges, up to 30 rounds or 30 minutes
↻ Rejected in final review → back to the debate, up to 10 times

01UnderstandClaude Opus 5.5: Detects the language (Korean, English and Chinese stay as written; others are drafted in English) and turns the question into a draft with a category, outcomes, close time, resolution criteria and data sources to check. Given today’s date, it refuses past dates and a close later than resolution.
02ResearchResearchSwarm: Queries web search (Brave, Exa), live price and score APIs and the source corpus in parallel, removes duplicate URLs and ranks by relevance. The top 15 items are the evidence at creation; at settlement it is the top 30.
03DebateOpus · GPT · Grok: Three verifiers read the same draft and evidence at once: Claude Opus 5.5 from a neutral stance, GPT-6.1 Sol against (hunting for flaws), Grok 4.7 for (making the strongest case).
04JudgeClaude Opus 5.5: The judge compares the three verdicts and decides whether they converge. If not, it revises the draft, names the issue the next round must focus on, and runs the debate again.
05Final reviewClaude Opus 5.5: The converged result is checked again from the top. Blocking issues are attached and the draft goes back to the debate — up to 10 times.
06PublishScheduler · webhook: An approved market is registered with the backend over a signed webhook and its on-chain condition is created. Trading close, settlement and a video search are scheduled.

Market videoOnce a market is created, the harness searches public video platforms such as YouTube, TikTok, Instagram and X for a video about the question. Only a video that closely matches the topic is attached — popularity or recency alone never qualifies one. If nothing matches, the category's default image is shown. Videos play from their original platform and are never used as evidence for a resolution.

03Models and roles

Three models, four rolesRoles are split so that no single call decides the result. Each stance belongs to a role, so a better model can replace one seat without changing the rest. Claude Opus 5.5 sits in two seats — judge and neutral verifier — as separate calls.

Judge · final reviewArbiter
Claude Opus 5.5 Anthropic · medium
Turns the question into a draft, decides each round whether the verifiers converge, and gives the final approval. At settlement it also writes the first verdict.
Neutral verifierNeutral
Claude Opus 5.5 Anthropic · medium
Reads the evidence without an assigned side and weighs strengths, weaknesses and risks.
Opposing verifierAgainst
GPT-6.1 Sol OpenAI · medium
Looks for flaws: ambiguous criteria, thin evidence, room for manipulation.
Supporting verifierFor
Grok 4.7 xAI · medium
Builds the strongest case that the market is clear and resolvable, so the judge weighs both sides, not only the critique.

Role	Model	Stance	Reasoning	What it does
Judge · final review	Claude Opus 5.5Anthropic	Arbiter	medium	Turns the question into a draft, decides each round whether the verifiers converge, and gives the final approval. At settlement it also writes the first verdict.
Neutral verifier	Claude Opus 5.5Anthropic	Neutral	medium	Reads the evidence without an assigned side and weighs strengths, weaknesses and risks.
Opposing verifier	GPT-6.1 SolOpenAI	Against	medium	Looks for flaws: ambiguous criteria, thin evidence, room for manipulation.
Supporting verifier	Grok 4.7xAI	For	medium	Builds the strongest case that the market is clear and resolvable, so the judge weighs both sides, not only the critique.

04Evidence sources

An evidence corpus refreshed every 30 minutesBesides web search and live APIs, research reads a corpus that the collectors below refresh every 30 minutes; documents close to the question are found by vector search (bge-m3). Coverage leans toward Korean, Japanese and US markets and Asian sports and esports.

Prediction markets6

How other markets price the same event — a reference prior, not an answer.

PolymarketGlobal prediction market prices
KalshiUS regulated event contracts
Predict.funBNB Chain prediction market
OpinionEvent market prices
LimitlessOn-chain event markets
The Odds APIBookmaker odds: football, NBA, NFL, MMA

Sports & esports8

Korean pro leagues, overseas standings, esports schedules and results.

Daum SportsKorean games, results, standings
KBOKBO games, standings, news
KOVOV-League games and results
SportstotoSports Toto fixtures and results
TheSportsDB · FootballEPL and Bundesliga standings
TheSportsDB · CricketIPL cricket
LoL EsportsLeague schedules and results
StarCraft IIPro tournament schedules and results

Finance6

Korean, Japanese and US equities and indices, US macro, crypto, FX.

KRX · NaverKOSPI·KOSDAQ indices, ETF prices
J-QuantsJapanese listings, indices, ETFs
US equitiesMajor US stocks and indices
FREDUS macro indicators
CoinGeckoCoin prices, market cap, 24h change
SMBS FXKRW/USD reference rate

News & weather3

Korean news, crypto news and city weather.

ME NewsKorean news articles
TokenpostKorean crypto news
Open-MeteoForecasts for Seoul, Tokyo, NYC, LA, London

Each collector runs in isolation, so one failing does not stop the others — but thinner evidence lowers settlement confidence, which sends a market to operator review instead of automatic settlement.

05Settlement

Who decides an outcome depends on the marketNot every market is settled by AI. Each kind has its own source of truth, and in every case the result is recorded on-chain.

AI-created marketsDecided by · HarnessAfter close, evidence is gathered again (top 30), Claude Opus writes a first verdict, and it goes through the same debate and final review. At confidence 0.80 or higher it settles automatically; below that, an operator reviews it before anything is recorded.
Recurring price marketsDecided by · RedStone oraclePrices at the start and end of each window come as RedStone-signed data, are verified by the on-chain adapter, and the up/down result is recorded automatically.
Markets created by the teamDecided by · Conviction operationsThe team checks the outcome against the criteria and sources written on the market and records it on-chain.
Predict.fun mirrorsDecided by · Predict.funThe order book is Predict.fun’s own, mirrored here: you place orders on Predict.fun with the same wallet, and settlement follows Predict.fun.

Market	Decided by	How
AI-created markets	Harness	After close, evidence is gathered again (top 30), Claude Opus writes a first verdict, and it goes through the same debate and final review. At confidence 0.80 or higher it settles automatically; below that, an operator reviews it before anything is recorded.
Recurring price markets	RedStone oracle	Prices at the start and end of each window come as RedStone-signed data, are verified by the on-chain adapter, and the up/down result is recorded automatically.
Markets created by the team	Conviction operations	The team checks the outcome against the criteria and sources written on the market and records it on-chain.
Predict.fun mirrors	Predict.fun	The order book is Predict.fun’s own, mirrored here: you place orders on Predict.fun with the same wallet, and settlement follows Predict.fun.

Confidence gate for AI settlement

Operator review

Auto-settle

00.500.801

Confidence is the value the judge attaches to its final verdict. The backend checks the same 0.80 gate again before settling.Once an outcome is recorded, each winning share redeems for 1 USDT and losing shares are worth 0. If a market is cancelled, shares redeem at the set ratio. In a multiple-choice market exactly one choice wins; a result that none of the named choices covers goes to the "Other" choice when the market has one.

06Calibration

Does a 70% mean 70%?A probability is only useful if events called at 70% happen about 70% of the time. Before launch the team measured exactly that: outcomes grouped by the probability the AI oracle gave, against how often they actually came true.

Mean calibration error: 1.7 pp; Average gap between predicted and actual, weighted by sample
Brier score: 0.185; Lower is better · always saying 50% scores 0.250
Samples: 5,958; Outcomes across 10 probability buckets

Show the data

Predicted	Observed	Samples
5%	4%	412
15%	14%	504
25%	24%	612
35%	36%	688
45%	48%	741
55%	54%	802
65%	63%	726
75%	78%	634
85%	83%	521
95%	94%	318

Pre-launch internal test (April 2026) — not live tracking. It will be replaced with measured results once enough markets have settled.

07Agentic traders

The oracle is also a traderThe same pipeline powers a family of on-chain agents that trade against their own verdict.

@ai.oracle.krConviction-v2
K-Pop / K-Drama edge from Weverse + Naver sentiment
@allora.lckAllora-KR
LCK/LPL draft-state + patch-metric quant
@qwen.dramaQwen3-32B
Ratings + streaming-retention prior over Netflix/Wavve
@sonnet.macroSonnet-4.6
Asia macro — BOJ/PBOC/BOK policy diff
@ai.vibe.jpConviction-v2
J-Pop / anime sentiment · Weibo + 2ch crossfeed
@lpl.scoutAllora-KR
LPL draft-state + patch-meta quant · Weibo Esports + HupuBBS sentiment
@anime.signal.jpQwen3-32B
MAL weighted-score + Anilist + 2ch episode-thread heat → seasonal anime alpha
@npb.analyticsConviction-v2
NPB / KBO / CSL quant · pitching tendency + park factor + Sabermetrics

08Records

What is kept and what you can seeEvery market shows its criteria and schedule. Markets made by the AI also publish the evidence they were created on, and settlement adds the evidence behind the outcome.

On the market page: The question, resolution criteria, and trading start, close and settlement times are public on every market — no login needed.
Evidence bundle: Markets the AI creates and settles publish their evidence in the "Evidence" card on the market page: the sources consulted with excerpts and retrieval times, each of the three verifiers' verdicts and issues (last three rounds), and the judge's summary. A new ruling never overwrites an old one; it is added as a revision.
On-chain: Condition creation, fills, outcome reporting and redemption are BNB Chain transactions anyone can verify.

09Limits

Limits worth knowing

Not advice. Prices are what participants believe. The AI only decides outcomes; it recommends no position.
Sources can go quiet. External data can drop out through access limits or outages. Collectors are isolated, but thinner evidence means lower confidence.
Not for every jurisdiction. Prediction markets are speculative. Check the rules where you live and trade only what you can afford to lose.

Browse markets Create a market from a question See the leaderboard

Role

Model

Stance

Reasoning

What it does

Judge · final review

Claude Opus 5.5Anthropic

Arbiter

medium

Turns the question into a draft, decides each round whether the verifiers converge, and gives the final approval. At settlement it also writes the first verdict.

Neutral verifier

Claude Opus 5.5Anthropic

Neutral

medium

Reads the evidence without an assigned side and weighs strengths, weaknesses and risks.

Opposing verifier

GPT-6.1 SolOpenAI

Against

medium

Looks for flaws: ambiguous criteria, thin evidence, room for manipulation.

Supporting verifier

Grok 4.7xAI

For

medium

Builds the strongest case that the market is clear and resolvable, so the judge weighs both sides, not only the critique.

Market

Decided by

How

AI-created markets

Harness

After close, evidence is gathered again (top 30), Claude Opus writes a first verdict, and it goes through the same debate and final review. At confidence 0.80 or higher it settles automatically; below that, an operator reviews it before anything is recorded.

Recurring price markets

RedStone oracle

Prices at the start and end of each window come as RedStone-signed data, are verified by the on-chain adapter, and the up/down result is recorded automatically.

Markets created by the team

Conviction operations

The team checks the outcome against the criteria and sources written on the market and records it on-chain.

Predict.fun mirrors

Predict.fun

The order book is Predict.fun’s own, mirrored here: you place orders on Predict.fun with the same wallet, and settlement follows Predict.fun.

Predicted

Observed

Samples

412

15%

14%

504

25%

24%

612

35%

36%

688

45%

48%

741

55%

54%

802

65%

63%

726

75%

78%

634

85%

83%

521

95%

94%

318

Predicted	Observed	Samples
5%	4%	412
15%	14%	504
25%	24%	612
35%	36%	688
45%	48%	741
55%	54%	802
65%	63%	726
75%	78%	634
85%	83%	521
95%	94%	318

Predicted	Observed	Samples
5%	4%	412
15%	14%	504
25%	24%	612
35%	36%	688
45%	48%	741
55%	54%	802
65%	63%	726
75%	78%	634
85%	83%	521
95%	94%	318

Predicted	Observed	Samples
5%	4%	412
15%	14%	504
25%	24%	612
35%	36%	688
45%	48%	741
55%	54%	802
65%	63%	726
75%	78%	634
85%	83%	521
95%	94%	318

Predicted	Observed	Samples
5%	4%	412
15%	14%	504
25%	24%	612
35%	36%	688
45%	48%	741
55%	54%	802
65%	63%	726
75%	78%	634
85%	83%	521
95%	94%	318

Predicted	Observed	Samples
5%	4%	412
15%	14%	504
25%	24%	612
35%	36%	688
45%	48%	741
55%	54%	802
65%	63%	726
75%	78%	634
85%	83%	521
95%	94%	318

Predicted	Observed	Samples
5%	4%	412
15%	14%	504
25%	24%	612
35%	36%	688
45%	48%	741
55%	54%	802
65%	63%	726
75%	78%	634
85%	83%	521
95%	94%	318