A council of frontier language models resolving onchain questions with Bittensor economic security
Onchain contracts — prediction markets, insurance, DAO governance, dispute resolution — require a trustworthy primitive that can answer post-event questions with evidence and finality. Existing oracle solutions are either centralized, economically thin, or blind to natural-language resolution criteria. We present Pelion, a decentralized adjudication protocol that routes questions to a council of frontier language models operating as Bittensor miners, aggregates their evidence-backed verdicts under TAO-backed economic security, and delivers signed results to any contract on Base via a bonded cross-chain relay. Pelion introduces UNRESOLVABLE as a first-class outcome — honest abstention in the face of insufficient evidence. All evidence and reasoning is archived permanently on IPFS and Arweave.
The oracle problem — getting real-world information onto a blockchain in a trustworthy way — has been the dominant unsolved problem in decentralized finance for a decade. Price feeds have largely been solved through aggregation and economic security. But the harder problem remains: answering natural-language questions about events that have already occurred.
A prediction market needs to know whether a specific flight landed on time. An insurance contract needs to know whether a hurricane made landfall in a specific county. A DAO needs to know whether a grant milestone was achieved. These questions require reading comprehension, evidence gathering, source verification, and judgment — capabilities that price-feed oracles do not provide.
Pelion is purpose-built for this problem. It treats onchain contract adjudication as a judgment task and routes it to the systems best suited to judgment: frontier language models operating under economic incentives designed to produce honest, evidence-backed verdicts.
Existing solutions fall into three categories, each with fundamental limitations:
Single points of trust failure are unacceptable for high-value contracts. Bribery, regulatory pressure, and insider corruption are existential risks.
Voting oracles (UMA, Kleros) are vulnerable to plutocratic capture. With sufficient capital, an attacker can simply buy a majority. Token holders are not selected for their ability to evaluate evidence.
Chainlink, Pyth, and similar systems are excellent for numerical price data but cannot answer natural-language questions. They have no mechanism for evidence gathering, reasoning, or honest abstention.
What is needed is a system that combines evidence-gathering ability (frontier AI), economic accountability (Bittensor), and transparent reasoning (permanent archive).
Pelion operates as a five-layer stack:
| Layer | Runtime | Role |
|---|---|---|
| Application | Base (Solidity) | Prediction markets, insurance contracts, DAOs. |
| Adapter | Base (Solidity) | Question intake, bond escrow, challenge handling, payout callbacks. |
| Router | Python (off-chain) | Question translation, subnet routing, evidence pinning. |
| Judgment | Bittensor subnets | Miners produce verdicts; validators score; TAO-backed security. |
| Relay | Bonded cross-chain | Carry verdicts back to Base with challenge windows. |
Each Bittensor miner in the Pelion subnet wraps a FrontierModelClient: a module that fans a Question to Anthropic Claude, OpenAI GPT-4, and Google Gemini in parallel and produces a Verdict with outcome label, confidence score, evidence citations, and chain-of-thought reasoning.
Requiring responses from multiple frontier providers means a single provider outage, policy change, or adversarial jailbreak cannot compromise a verdict. The aggregation step is resistant to any single-provider failure.
Pelion inherits Bittensor's economic security model. Miners and validators stake TAO; misbehavior is punished by slashing. For high-value questions, the Router distributes across multiple subnets, making attack cost roughly multiplicative.
Each question includes an EvidencePolicy that constrains acceptable sources. All evidence is archived permanently on IPFS and Arweave as part of the verdict bundle.
| Field | Description |
|---|---|
| allowed_source_categories | news, official, academic, blockchain |
| allowed_domains | Allowlist of exact source domains |
| max_age_hours | Sources older than this are inadmissible |
| min_source_count | Minimum independent sources required |
Validators score miners on outcome accuracy (60%), evidence quality (25%), and reasoning coherence (15%). The stake-weighted majority outcome becomes the final verdict. If confidence falls below threshold τ, the verdict is UNRESOLVABLE.
Any party may relay a verdict by posting a bond equal to the question's bond_amount. The Adapter opens a challenge window (1–6 hours). Successful challenges slash the relayer bond and reward the challenger. No whitelisted relayer set is required.
An attacker must bribe a majority of validators on every routing subnet. Cost scales with total staked TAO across all routing subnets.
A compromised single provider becomes a minority in the multi-provider structure. Miners relying only on it receive lower validator scores.
Evidence is archived at verdict time with access timestamps. Post-hoc fabrication is detectable. The EvidencePolicy domain allowlist limits the attack surface.
The Pelion mainnet token ($PELION) is the settlement currency for the Adapter layer on Base. It denominates question bonds and relayer bonds, is used for protocol fee payment, and governs parameter changes. Bittensor miners and validators are compensated in TAO through Bittensor's emission schedule.
| Phase | Deliverable |
|---|---|
| Alpha | Python core: schemas, frontier council, miner, validator. Open-source on GitHub. |
| Beta | Bittensor testnet. Router and Adapter on Base Sepolia. First integrations. |
| Mainnet | Bittensor mainnet subnet. Base mainnet Adapter. PELION token. Challenge mechanism live. |
| Scale | Multi-subnet routing. Arweave evidence archival. Governance module. |
Pelion addresses a structural gap in the onchain contract ecosystem: the absence of a trustworthy, evidence-backed, natural-language adjudicator. By combining frontier language model capability with Bittensor economic security, Pelion provides the adjudication primitive that prediction markets, insurance, and DAO governance require.
Pelion is not an oracle. It is an adjudicator — a system that reads, reasons, and rules, with economic skin in the game behind every answer.