> ## Documentation Index
> Fetch the complete documentation index at: https://docs.pelion.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Introduction

> Bittensor Decentralized Adjudication for onchain contracts.

<Warning>
  The canonical technical specification is complete. Core Python modules are built and under active test coverage. The adapter contract, token, and cross-chain relay are specified but not yet deployed.
</Warning>

Pelion is Bittensor Decentralized Adjudication for onchain contracts. Prediction markets, insurance contracts, DAO decisions, dispute escalations all need the same primitive: a trustworthy adjudicator that answers post-event questions with evidence and finality. Pelion provides that primitive using a council of frontier language models, with Bittensor economic security behind the verdicts and a mainnet token and adapter in front of them.

## The problem Pelion solves

UMA's Optimistic Oracle is today's incumbent onchain adjudicator. It has a structural flaw. Token-weighted voting lets large UMA holders vote against factual accuracy when they hold positions in the underlying market.

In July 2025, UMA's Zelensky-suit market resolved "No" despite mainstream media consensus that Zelensky wore a suit at the NATO summit. The market had $242M in stakes. UMA's total market cap was $95M. Honest defense was economically weaker than dishonest attack, and that is the predictable behavior of token-weighted voting once market stakes exceed the voter base's market cap.

Three other UMA failure modes compound the first. Voting incentives reward matching the perceived majority (reflexive consensus, not truth). Every market must eventually resolve YES or NO, so genuinely ambiguous questions get forced into one bucket. UMA's response to these problems (managed proposers, whitelisted resolvers) trades decentralization for quality without addressing the underlying structure.

See [UMA's structural failures](/problem/uma-failures) for the full failure-mode breakdown.

## Why a frontier-model protocol works now

Three independent lines of progress converged in 2025 and later, and Pelion sits on top of all of them.

Frontier language models are now accurate enough for post-event adjudication. Published work (Capponi et al. 2025 and subsequent studies) shows models match human resolution on roughly 95% of disputed Polymarket markets when given retrieval access. The remaining failure cases are small and characterizable, which is what makes `UNRESOLVABLE` a real answer rather than a hedge.

Bittensor is a working marketplace for AI work, with subnets already running forecasting (SN6 Numinous and Infinite Games), time-series prediction (SN8, SN50), and financial oracles (SN28). Post-event judgment is an open niche that needs a dedicated subnet.

Cross-chain infrastructure is production grade. Bonded relay patterns from Across, Hyperlane, and LayerZero settle real value at real scale. Bittensor's EVM layer bridges to EVM chains without exotic infrastructure.

See [Why now](/problem/why-now) for citations and details.

## How Pelion works

The protocol runs as a five-layer stack.

<CardGroup cols={2}>
  <Card title="Application" icon="layer-group">
    Prediction markets, insurance contracts, DAOs on mainnet that submit questions and consume verdicts.
  </Card>

  <Card title="Adapter" icon="file-contract">
    Solidity contracts that accept questions, escrow bonds, handle disputes, and relay verdicts to consumers.
  </Card>

  <Card title="Router" icon="route">
    Off-chain services that translate questions to Bittensor queries and aggregate responses across subnets.
  </Card>

  <Card title="Judgment" icon="scale-balanced">
    Bittensor subnets where miners produce verdicts and validators score them. Economic security via TAO staking.
  </Card>

  <Card title="Relay" icon="bridge">
    Bonded cross-chain messengers that carry verdicts back to mainnet with permissionless challenge windows.
  </Card>

  <Card title="Token and treasury" icon="coins">
    Mainnet ERC-20 funded by trading tax. Treasury earns Bittensor emissions and executes buyback/burn.
  </Card>
</CardGroup>

A question submitted on mainnet is routed off-chain to one or more Bittensor subnets, where AI miners produce verdicts with evidence and reasoning. Validators score them into a weighted consensus. A bonded relay carries the verdict back to mainnet, where anyone can challenge with a counter-bond during a challenge window. After finalization, payouts fire on the requesting contract.

See [Architecture](/how-it-works/architecture) for the data flow and trust boundaries.

## What's different

**Multi-subnet routing** for independent economic security domains. High-stakes questions can require consensus across multiple subnets. Attack cost scales roughly multiplicatively.

**`UNRESOLVABLE` as a first-class outcome.** Genuinely ambiguous questions are not forced into YES or NO. The protocol acknowledges uncertainty rather than manufacturing false resolution.

**Bonded cross-chain relay with permissionless challenge.** Anyone can submit a counter-verdict with a bond. On-chain signature verification decides.

**Public dataset.** Every question, verdict, evidence bundle, and reasoning trace is published permanently on IPFS and Arweave. Both a credibility anchor today and, over time, canonical training data for AI adjudication.

**Treasury-backed token flywheel.** Mainnet token holders gain indirect Bittensor emission exposure without taking on direct TAO complexity. Trading tax funds the treasury, treasury stakes alpha across subnets, emissions swap back to buy and burn.

## Headline claims

Accuracy at or above UMA on standard questions (at or above 99% on historical Polymarket sample).

Better accuracy than UMA on contested questions. The goal is to resolve Zelensky-type cases correctly.

Lower per-query cost than UMA (fractions of a cent for AI compute, versus UMA's \$750+ bond threshold for dispute).

Ability to return `UNRESOLVABLE` with escalation paths when evidence is genuinely inconclusive.

Transparent reasoning and evidence attached to every verdict, published permanently.

The first three claims become testable artifacts via the [Priority 1 benchmark](/whats-built/benchmark-methodology), which runs the built AI council against 18 months of historical Polymarket markets.

## Further reading

<CardGroup cols={2}>
  <Card title="See what's already built" icon="code" href="/whats-built/repository">
    Four Python modules, canonical schemas locked, multi-provider AI council operational.
  </Card>

  <Card title="Understand the architecture" icon="sitemap" href="/how-it-works/architecture">
    Five-layer stack, data flow, and trust boundaries in one page.
  </Card>

  <Card title="Read the token mechanics" icon="coins" href="/token/mechanics">
    Trading tax with auto-decay, utility, and the buyback/burn flywheel.
  </Card>

  <Card title="Jump to the full spec" icon="file-lines" href="/reference/full-spec">
    The canonical technical specification, for readers who want everything.
  </Card>
</CardGroup>