Allora allo

Allora is a self-improving, decentralized machine-intelligence network that combines community-built machine-learning models to produce context-aware predictions. Its goal is to make machine intelligence open and composable rather than siloed inside a few companies. Applications can consume aggregated predictions such as price forecasts, volatility signals, and risk assessments through APIs or oracles.

The network organizes collaboration into permissionless topics. Each topic defines a target variable and loss function, so participants work on narrowly specified prediction problems and are scored against relevant ground truth. Official architecture separates inference consumption, forecasting and synthesis, and consensus/economics layers.

Allora is built around the ALLO-token hub chain and a participant economy. Workers submit inferences and/or forecasts, reputers evaluate results against ground truth, and consumers request predictions and pay fees. The September 2024 paper describes a blockchain-coordinated network bridging data owners, model operators, and users of machine intelligence.

Machine intelligence has historically been concentrated in industry monoliths because useful models require proprietary data, algorithms, and computation. That creates opaque systems, high barriers to entry, and limited ability for independent contributors to connect models to users. A single-model approach also assumes one model is best across changing conditions, making outputs brittle or poorly calibrated.

A decentralized network must solve how to combine heterogeneous model outputs and reward different contributors without distorting their goals. Historical reputation alone cannot identify models temporarily better in the current context, while generic stake-based rewards can distract inference providers from accuracy. Allora addresses these issues with context-aware forecasting and role-specific incentives.

Allora participants operate in topics. Inference workers generate predictions for a topic target; forecasting workers predict expected losses of other workers' predictions under current conditions; some workers perform both jobs. Forecasted losses become regret and weights, creating forecast-implied inferences that emphasize models expected to perform well in the present context.

The topic coordinator synthesizes raw and forecast-implied inferences. After ground truth becomes available, reputers compare each result with that truth using the topic loss function and report losses. Reports are combined into consensus losses, and actual regret determines weights for the final network inference. The design can also provide confidence intervals reflecting disagreement among weighted worker outputs.

Rewards are differentiated by role. Workers are rewarded for unique contributions to network accuracy using contribution-score approximations, while reputers are rewarded for stake and agreement with learned consensus. Consumers request inferences and pay fees. The hub chain coordinates token economics, emissions, and transactions.

Official network docs currently list mainnet chain ID allora-mainnet-1 (allora-chain v0.16.0; emissions/v9) and testnet chain ID allora-testnet-1 (v0.17.0; emissions/v10). Testnet has a faucet; mainnet does not. Newer testnet releases include label-aware multi-output bundles, so API behavior can differ by network.