AI Oracle

The AGON AI Oracle runs multi-LLM concordance to resolve markets and surface edge signals.

PoAIC concordance algorithm

Proof of AI Concordance (PoAIC) is not a blockchain consensus mechanism. Per ADR-002, it is a latency optimization layer above OracleDAO. The AI resolves approximately 95% of markets in under 60 seconds; OracleDAO handles the remaining 5% of edge cases with a human backstop.

The concordance flow is implemented in apps/workers/ai/concordance.py:

1. Prompt construction. A shared prompt is built by build_oracle_prompt() from the market question, description, current AMM YES price, and optional knowledge-base grounding context (team stats, fixture data, crypto fundamentals injected via get_market_context()).

2. Parallel fan-out. The same prompt is dispatched to all configured providers via asyncio.gather():

   • anthropic -- Claude Haiku 4.5 (claude-haiku-4-5-20251001)
   • google -- Gemini 2.0 Flash (gemini-2.0-flash)
   • openai -- GPT-4o
   • bittensor -- Sportstensor Subnet 41 (optional, sports markets only, activated when bittensor_subnet41_endpoint is configured)

   Each provider returns a ProviderResult with outcome (YES/NO/INVALID), probability (0.01-0.99), confidence (0.0-1.0), and reasoning. A 45-second outer timeout caps the entire fan-out.

3. Diversity check. If more than 2/3 of valid results come from the same provider family, the aggregation is rejected (Sybil defense).

4. Weighted median. The median probability across all valid results is computed. Each provider result is then checked against a category-specific tolerance window:

   Category	Tolerance
   Sports (football, NBA, NFL, etc.)	3%
   Crypto (token, DeFi, NFT)	5%
   Politics (election, geopolitics)	10%

   A result is "concordant" if |probability - median| < tolerance.

5. Uncertainty band. If the median probability falls in the 0.45-0.55 range, the market is forced to human review regardless of concordance -- it is too close to call for automated resolution.

6. Decision. If >= 2/3 of providers are concordant AND mean confidence >= the configured threshold (default 0.80): the concordance is achieved. The outcome is determined by majority vote among concordant providers (ties broken by median probability). The api_resolver worker then calls APIResolver.resolve() on-chain.

   If concordance is NOT met, the market is flagged ai_needs_human_review = TRUE for OracleDAO manual resolution.

All provider submissions (including failures) are persisted to the validator_submissions table for observability and auditing.

Validator role

Per ADR-003, decentralized validators bear their own LLM inference costs. They are remunerated via two channels:

• FeeDistributor: 25% of protocol trading fees are routed to validators proportional to their staking weight.
• ArenaRewards oracle budget: 20 000 000 AGON allocated over the reward horizon for correct resolutions.

On-chain, AIOracleHub.sol manages the validator lifecycle:

1. Validators register with a minimum bond of 500 AGON in ReputationStaking.
2. The off-chain concordance coordinator collects EIP-712 signed attestations from registered validators.
3. submitConcordance() verifies signatures, weights, and diversity, then calls MarketFactory.proposeResolution() to enter the 24-hour dispute window.
4. If OracleDAO overrides the AI resolution, slashValidator() burns the offending validators' bonds.

The equilibrium KPI for validators is weekly_rewards / weekly_inference_cost.

Edge signal generation

The AI analyst (apps/workers/ai/analyst.py) runs every 15 minutes via the ai/scheduler.py batch loop. For each open market, it calls Claude Haiku to produce an independent probability estimate, which is written to the ai_probability column in the markets table.

The signal generator (workers/signal_generator.py) then compares ai_probability against the current AMM yes_price. If the absolute divergence exceeds 10% (EDGE_THRESHOLD = 0.10), it emits an edge signal to the signals table and fires an SSE event on the Redis event bus.

Fetching edge signals

// REST -- default min_edge filter is 5%
GET /api/v1/oracle/edges?min_edge=0.05

// MCP tool
get_oracle_edges({ min_edge: 0.05, limit: 20 })

The min_edge query parameter controls the minimum divergence returned by the API. The default of 5% surfaces signals that the signal generator may not yet have persisted (it uses a 10% threshold for generation).

OracleDAO fallback

When AI concordance fails -- disagreement among providers, insufficient valid results, or the uncertainty band is triggered -- the market enters OracleDAO manual review.

OracleDAO.sol implements a propose-challenge-vote flow:

1. Propose: A resolver posts a bond (250 AGON minimum) and proposes an outcome. This starts a 48-hour challenge period.
2. Challenge: Any AGON holder can dispute by posting an equal bond and proposing an alternative outcome.
3. Vote: If challenged, a 72-hour voting period opens. Voting power is snapshot-based using AgonToken (ERC20Votes with delegation and checkpoints, per ADR-007). Minimum quorum: 1 000 AGON in votes.
4. Settle: The winning outcome is forwarded to MarketFactory.proposeResolution(). The losing party forfeits their bond. Correct resolvers earn an AGON bonus via ArenaRewards.

Governance parameters (minBond, minQuorum) are adjustable by the TimelockController.

Where to next