Oracle Resolution

Strike supports two resolution methods: Pyth Price Feeds for price-based markets and the Flap AI Oracle for question-based markets. This page covers both.

Resolution Methods

Method	Used For	Source
Pyth Price Feeds	Price markets ("Will BTC be above $X?")	Cryptographically signed price data
Flap AI Oracle	AI markets ("Will the Fed cut rates?")	LLM reasoning with IPFS proof

---

Pyth Price Feed Resolution

Price markets are resolved by Pyth Network oracle price feeds. No human intervention, no subjective arbitration, no governance votes.

Settlement Rule

Each market has an expiry timestamp T. Resolution uses the earliest Pyth price update with publishTime in [T, T+Δ], where Δ is a protocol parameter (default: 60 seconds).

• The resolver fetches signed update data from Pyth's Hermes historical API

• The contract verifies it on-chain using parsePriceFeedUpdates(updateData, priceId, T, T+Δ)

• Pyth returns the update matching the window, and challengers can submit earlier publishTime updates to replace it

• The price field is used for settlement (not ema_price) — spot price for clean market semantics

Confidence Interval Check

Pyth publishes a confidence interval with every price update. If the confidence is too wide relative to the price, the resolution is rejected:

Reject if: conf > confThresholdBps × |price| / 10000

Default threshold: 1% (100 bps). This prevents settlement on unreliable price data.

Fallback Windows

If no Pyth update exists within [T, T+Δ]:

• The resolver can try [T, T+2Δ], then [T, T+3Δ], up to K×Δ

• This handles rare cases where Pyth publishing is delayed

• If no valid update exists within the maximum window → market cancels

Finality Gate

Resolution is not instant:

1. Resolver calls resolve() with a Pyth update → pendingResolution is set, market enters Resolving

2. Protocol waits for a 90-second finality period (FINALITY_PERIOD = 90 seconds)

3. During this window, anyone can submit an alternative Pyth update with an earlier valid publishTime, but only if it would change the outcome (e.g., flip the result from UP to DOWN)

4. After the finality window, finalizeResolution() is called → market enters Resolved state

5. Admin fallback: the admin can call setResolved() to skip the 2-step process in emergency situations

This "procedural challenge" mechanism ensures the deterministic rule (earliest update wins) is enforced even if the first resolver submits a suboptimal update.

Resolver Incentives

• Resolution is permissionless — anyone can call resolveMarket()

• The protocol runs backstop keepers to ensure timely resolution

• If no one resolves within 24 hours, the market auto-cancels and all funds are refunded

Why Pyth?

• Pull oracle — no continuous on-chain price pushing needed; only one update at resolution time

• ~400ms update cadence at the oracle network level

• Cryptographic verification — signed data is verified on-chain, not trusted from an EOA

• Historical data available — Hermes API serves signed updates for past timestamps

• Low cost — update fees are negligible on BSC (1 wei default)

---

AI Oracle Resolution

AI markets are resolved by the Flap AI Oracle, which uses large language models to evaluate question-based prompts. The AIResolver contract manages the full lifecycle:

1. Request — At market expiry, a keeper calls resolveMarket() which sends the prompt to the oracle

2. LLM reasoning (~90 seconds) — The oracle backend feeds the prompt to the selected model, which reasons over the question using current information

3. Callback — The oracle calls back with a binary choice (0 = YES, 1 = NO)

4. Liveness window (30 minutes) — The proposed resolution can be challenged

5. Finalisation — If unchallenged, anyone calls finalise() to settle the market

Challenge Mechanism

During the 30-minute liveness window, anyone can challenge the AI's proposed outcome by posting a 0.1 BNB bond. This extends the window to 24 hours for admin review. The admin either confirms the original resolution (challenger loses bond) or overrides it (challenger gets bond + 0.01 BNB reward).

IPFS Verification

Every AI resolution produces an IPFS proof containing the full prompt, reasoning trace, tool calls, and model metadata. The CID is available via the indexer API or on-chain via FlapAIProvider.getRequest(requestId).

For full details, see AI Markets.