AI Markets

AI markets are prediction markets resolved by a large language model (LLM) instead of a price feed. They extend Strike beyond quantitative price markets to cover qualitative events — geopolitics, culture, politics, sports, and more.

How They Differ from Price Markets

	Price Markets	AI Markets
Question	"Will BTC be above $90k at expiry?"	"Will the Fed cut rates in May?"
Resolution source	Pyth price feed	Flap AI Oracle (LLM)
On-chain data	priceId (bytes32) + strike price	prompt (string) + model ID
Creation fee	None (beyond gas)	BNB fee (varies by model)
Trading	Identical — same orderbook, same FBA mechanism	Identical

Supported Models

The Flap AI Oracle supports multiple LLM models. Market creators choose the model at creation time and pay the corresponding fee in BNB.

Model	Model ID	Fee (BNB)	Best For
Gemini 3 Flash	0	0.01	High-frequency, lower-stakes markets
Claude Sonnet 4.6	1	0.05	Complex reasoning, nuanced judgment
DeepSeek R1	2	0.03	Balanced cost and quality

Note: Fees are queried from the oracle contract at runtime via getModel(modelId).price. The values above are current as of launch but may change.

Market Lifecycle

1. Creation

A market creator calls createAIMarket(prompt, modelId, expiryTime, minLots) on MarketFactory, sending the model fee in BNB. The prompt is stored on-chain in the AIResolver contract.

Example prompts:

• "Count tweets from @cz_binance between March 20-27, 2026 using X/Twitter. Answer YES if ≥50, NO otherwise. Include the raw count."

• "Search federalreserve.gov for the May 2026 FOMC statement. Answer YES if it announces a rate cut, NO otherwise. Quote the relevant text."

• "Check Steam and PlayStation Store for GTA VI availability. Answer YES if available for purchase, NO otherwise. Include links."

2. Trading

Trading works identically to price markets — the same 4-sided orderbook (Bid, Ask, SellYes, SellNo), the same Frequent Batch Auction clearing, the same USDT collateral. Traders buy YES or NO positions based on whether they think the answer to the prompt is affirmative or negative.

3. Expiry

When the market expires, a keeper calls AIResolver.resolveMarket(marketId). This sends the prompt to the Flap AI Oracle with the pre-deposited BNB fee.

4. AI Resolution (~90 seconds)

The oracle's off-chain backend:

1. Receives the prompt via an on-chain event

2. Feeds it to the selected LLM with current information

3. The LLM reasons over the question and returns a binary choice (0 = YES, 1 = NO)

4. Pins the full reasoning trace to IPFS

5. Calls back AIResolver.fulfillReasoning(requestId, choice) on-chain

5. Liveness Window (30 minutes)

After the oracle callback, the proposed resolution enters a 30-minute liveness window. If no one challenges, anyone can call finalise(marketId) to settle the market.

6. Finalised

The market resolves with the AI's proposed outcome. Winning positions pay out normally via the Redemption contract.

Challenge Process

During the 30-minute liveness window, anyone who disagrees with the AI's proposed outcome can challenge it.

How to Challenge

1. Call AIResolver.challenge(marketId) with a 0.1 BNB bond

2. The liveness window extends to a 24-hour review period

3. A protocol admin reviews the challenge

Challenge Outcomes

Outcome	What Happens
Admin confirms (AI was correct)	Original resolution stands. Challenger loses their 0.1 BNB bond (sent to treasury).
Admin overrides (AI was wrong)	Resolution is corrected. Challenger receives their 0.1 BNB bond back plus a 0.01 BNB reward.

Timeline

Expiry → resolveMarket() → [~90s] Oracle callback
  → Proposed resolution (30-min liveness)
    → No challenge → finalise() → Resolved
    → Challenge (0.1 BNB) → 24h admin review
      → Confirmed → Resolved (original outcome)
      → Overridden → Resolved (corrected outcome)

IPFS Verification

Every AI resolution produces a verifiable proof pinned to IPFS. The proof contains:

• Full prompt sent to the LLM

• Reasoning steps — the model's chain-of-thought

• Tool calls — any real-time data the model fetched (e.g., price lookups)

• Model metadata — model version, temperature, and other parameters

• Final choice — the numeric outcome returned

How to Verify

1. Get the IPFS CID from the indexer: GET /v1/markets/{id}/ai-resolution → ipfs_cid field

2. Fetch the proof: https://ipfs.io/ipfs/{cid}

3. Review the full reasoning trace

The CID is also available on-chain via FlapAIProvider.getRequest(requestId) (decode the struct to extract the CID field).

Writing Good Prompts

AI prompts must be actionable instructions, not vague questions. The LLM needs clear guidance on:

1. What data to find — specific, measurable criteria

2. Where to look — which tools or sources to use

3. How to report — include raw data in reasoning

4. What to answer — explicit yes/no criteria

Bad vs Good Prompts

❌ Bad (vague)	✅ Good (actionable)
"Will CZ tweet a lot this week?"	"Use X/Twitter search to count how many times @cz_binance tweeted between March 20-27, 2026. Include the exact count in your reasoning. Answer YES if the count is 50 or more, NO otherwise."
"Will the Fed cut rates?"	"Search for the official FOMC statement from the May 2026 meeting on federalreserve.gov. Answer YES if the statement announces a rate cut, NO otherwise. Quote the relevant text."
"Did the proposal pass?"	"Check the Uniswap governance portal for proposal #42 final results. Include the vote counts (for/against/abstain). Answer YES if the proposal passed quorum and majority, NO otherwise."

Example Prompts by Category

Category	Market Question	AI Prompt
Social	Did CZ tweet 50+ times this week?	"Use X/Twitter to count tweets from @cz_binance between March 20-27, 2026. Include the raw count. Answer YES if ≥50, NO otherwise."
Events	Did GTA VI release?	"Search for official Rockstar Games announcements and Steam/PlayStation store listings for GTA VI. Answer YES if the game is available for purchase/download, NO otherwise. Include sources."
Sports	Did Argentina win the match?	"Search for the final score of the Argentina vs Brazil match on March 26, 2026. Include the score and source. Answer YES if Argentina won, NO otherwise."
Governance	Did the proposal pass?	"Check the Uniswap governance portal for proposal #42 final results. Include the vote counts. Answer YES if the proposal passed, NO otherwise."
Macro	Did the Fed cut rates?	"Search federalreserve.gov for the May 2026 FOMC statement. Answer YES if it announces a rate cut, NO otherwise. Quote the relevant sentence."

Note: For price-based markets (e.g., "Will BTC be above $100k?"), use Price Markets with Pyth oracle instead of AI markets.

Prompt Guidelines

1. Be specific about thresholds — "above $750" not "high price"

2. Specify date ranges — "between March 20-27" not "this week"

3. Name your sources — "from CoinGecko" not "from somewhere"

4. Request evidence — "include the exact count" forces verifiable reasoning

5. Define YES/NO clearly — "YES if ≥50, NO otherwise" removes ambiguity

Contract Reference

AI markets are managed by the AIResolver contract, which extends FlapAIConsumerBase. See AIResolver.sol for the full contract reference.

Key Addresses

Contract	Network	Address
Flap AI Oracle	BSC Testnet	0xFfddcE44e8cFf7703Fd85118524bfC8B2f70b744
Flap AI Oracle	BSC Mainnet	0xaEe3a7Ca6fe6b53f6c32a3e8407eC5A9dF8B7E39