Introduction

The Network for AI Training Data, Powered by Prediction Markets

Reppo is building a decentralized, blockchain-based AI training network powered by prediction markets.

The network is made up of domain-specific RL environments called datanets. Each datanet competes to produce valuable training data by incentivizing two groups:

• Publishers, who contribute raw data

• Voters, domain experts who lock REPPO to receive veREPPO

Each datanet earns revenue through publishing and access fees paid by publishers, plus network rewards tied to metrics such as revenue earned, veREPPO volume, and data monetization revenue.

What is an RL environment?

When you fine-tune a language model with reinforcement learning, you need three things: a task for the model to attempt, a way to score the model's output, and a feedback loop that pushes the model toward higher scores over millions of attempts.

That task-and-scorer combination is the RL environment. Think of it as a gym for the model.

On Reppo, each RL environment uses prediction market mechanics to curate AI training datasets for AI labs, robotics, and agentic use cases. That lets participants in the AI training data pipeline share more directly in the upside.

“Scale AI coordinates training data pipelines with contracts. Reppo coordinates training data pipelines with prediction markets and blockchain incentives.”

On Reppo, training data is crowdsourced and curated on Reppo.ai.

Publishers submit raw data into domain-specific datanets by paying a small fee set by the datanet owner. They are betting on the originality and quality of their work. Publishers can be human or non-human.

On the other side are voters, domain experts who label, annotate, and provide feedback on raw data. To participate, they lock $REPPO to receive veREPPO, the network's voting power.

By putting economic stake behind both raw data contribution and data curation, Reppo is designed to address the AI Training Data Trilemma.

AI Training Data Trilemma

Reppo's Approach

We are also building a permissionless venue for data monetization on Reppo.exchange, a data DEX for AI training data.

Datanet owners can use or monetize the feedback and data they crowdsource for any purpose, including off-chain monetization. Publishers and voters are compensated through market-based incentives for their work, which may reduce some legal and operational friction around the data.

The network is built on top of Reppo Protocol, a generalized incentive protocol that lets anyone use prediction markets to crowdsource Stake-Assured Human Feedback.

Our GTM focus is AI training data on demand, but Reppo Protocol can also support a broader class of products and services where the oracle is human preference instead of world events.

Each data market acts as an on-demand data factory, allowing data owners and domain experts to collaborate and produce high-signal training data for AI models, agents, robotics, and other large-scale AI systems.

For crypto natives, Reppo can be thought of as "Bittensor for AI training data," but the consensus mechanism, tokenomics, and utility functions are entirely different.

Current Gaps in AI Training Data

The traditional approach to AI training data relies on a pay-per-task model that prioritizes speed over quality, leading to:

• Rushed work to maximize volume

• No accountability for accuracy

• Shallow binary labels

• Heavy QA overhead

Reppo turns this process into a stake-backed market between miners and validators, generating stake-verified AI training data at scale.

Miners on the Reppo network are primarily source data contributors. This includes anyone with unlabeled or unannotated data, such as geospatial data, robotics data, adult content, or code.

These miners participate in the datanets that match their skills and source data, and provide proprietary human-generated input data. We plan to integrate with Worldcoin to help distinguish humans from AI agents, but we also believe AI agents can mine $REPPO.

On the other side, validators, or voters, are motivated by profit rather than salary. They lock REPPO to acquire voting power, called veREPPO, and vote on miner-published data while labeling and annotating it. They do this because incentives are aligned.

Voters on the Reppo network are analogous to validators who redirect network emissions each epoch to miners based on the quality of their work.

Reppo is not an L1 or L2 network.