How Verifiable AI Oracle Works

Discover how Walmarket combines Nautilus TEE, GPT-5, and cryptographic proofs to automatically verify prediction market outcomes with unprecedented accuracy, security, and transparency.

The Challenge

Traditional prediction markets rely on centralized resolution mechanisms, creating single points of failure and potential manipulation. Walmarket solves this with a verifiable AI Oracle system powered by Nautilus TEE that:

✓Automatically verifies outcomes using multiple trusted data sources
✓Runs AI inference in Trusted Execution Environments with cryptographic attestation
✓Stores all reasoning and proofs permanently on Walrus
✓Eliminates trust requirements through TEE signatures and on-chain verification

System Architecture

Data Collection

AI agents monitor multiple trusted data sources in real-time

📊

→

AI Analysis

GPT-5 processes normalized data using fixed prompts and JSON schema

🤖

↓

Cryptographic Proof

Evidence bundle is hashed and stored permanently on Walrus

🔐

→

On-Chain Verification

SUI smart contracts verify TEE attestation and aggregate results

⛓️

🔐 Nautilus TEE: Verifiable AI Execution

Walmarket leverages Nautilus Trusted Execution Environment to ensure AI oracle operations are tamper-proof and cryptographically verifiable:

🛡️ Execution Integrity

AI inference runs inside secure enclaves with measurement (mrenclave) verification, preventing operator tampering

🔑 Cryptographic Attestation

Each report is signed with enclave-bound keys and includes remote attestation proofs

📋 Enclave Registry

On-chain whitelist of authorized enclaves with epoch-based key rotation for security

⚡ Replay Protection

Timestamps and nonces prevent report replay attacks, ensuring fresh data only

TEE attestation eliminates the need to trust oracle operators - cryptographic proofs guarantee execution happened as claimed.

🎯 Single Model Policy

Walmarket uses a unified approach with GPT-5 as the core AI model to ensure:

✓ Consistency

All predictions use the same model version, reducing variance and improving reliability

✓ Reproducibility

Fixed prompts, schemas, and model versions enable exact result verification

✓ Cost Efficiency

Token caps and caching optimize API costs while maintaining quality

✓ Easy Auditing

Single model simplifies verification, debugging, and community review

Off-Chain: Nautilus TEE Reporter

1.Data Normalization: Collect and standardize sources from trusted APIs
2.Prompt Template: Apply fixed system/user prompts with JSON schema enforcement
3.TEE Execution: Run GPT-5 inference inside Nautilus enclave with execution integrity proof
4.Hash Computation: Calculate SHA256 of input, output, and evidence bundle
5.Walrus Upload: Store bundle permanently with blob_id and blob_hash
6.TEE Attestation: Sign report digest with enclave key and generate remote attestation

On-Chain: SUI Aggregator

1.Enclave Registry: Verify enclave_pubkey or mrenclave against whitelist and epoch validity
2.TEE Signature: Validate cryptographic signature using registered enclave public key
3.Hash Verification: Confirm blob_hash matches Walrus anchor and report digest
4.Policy Check: Ensure model_id, prompt_hash, parser_hash, schema_hash match whitelist
5.Aggregation: Calculate median value from multiple verified reporter submissions
6.Challenge & Slash: Optimistic disputes with evidence bonds, immediate slashing for violations

📋 Resolution Schema

Every AI Oracle output follows a standardized JSON format for transparency and verifiability:

{
  "round": 12345,
  "task": "binary",
  "resolution": {
    "value": 1,           // 1=YES, 0=NO (or numeric value)
    "confidence": 0.87    // 0.0 to 1.0
  },
  "sources": [
    {
      "id": "reuters:article123",
      "url": "https://...",
      "quote_hash": "0xabc..."
    }
  ],
  "rationale": "Candidate A won with 52.3% of votes...",
  "controls": {
    "model_id": "gpt-5-thinking@2025-11-POC",
    "prompt_hash": "0x...",
    "parser_hash": "0x...",
    "schema_hash": "0x..."
  },
  "tee_proof": {
    "enclave_id": "0x123...",
    "enclave_pubkey": "0xabc...",
    "mrenclave": "0xdef...",
    "sig": "0x789...",
    "attestation": "0x456...",
    "timestamp": 1672531200,
    "nonce": "0x999...",
    "h_in": "0xaaa...",     // SHA256(input)
    "h_out": "0xbbb...",    // SHA256(output)
    "blob_id": "0xccc...",
    "blob_hash": "0xddd..."
  }
}

🛡️ Security & Trust

Trusted Execution Environment (TEE)

AI inference runs in secure enclaves with remote attestation, ensuring operators cannot tamper with results

Immutable Evidence Storage

All input data, AI reasoning, and source citations stored permanently on Walrus for post-hoc auditing

Optimistic Challenge System

Community can dispute resolutions with evidence bonds, creating a secondary safety layer

Economic Penalties

Automatic slashing for format violations, hash mismatches, or invalid source citations

Why This Matters

⚡

Real-Time Resolution

Markets resolve instantly when conditions are met, no manual intervention required

🔍

Full Transparency

Every decision is auditable with complete source data and AI reasoning stored on-chain

🎯

87.3% Accuracy

AI Oracle achieves industry-leading accuracy through multi-source verification and confidence scoring

Ready to Experience AI-Powered Markets?

See our AI Oracle in action on live prediction markets

Explore Markets