Reppo is specifically built for the next phase of agentic AI, where models require real-time feedback loops rather than static datasets. Autonomous agents operating in physical or financial environments cannot afford to train on stale data; they need reward signals that update as reality shifts. Reppo's prediction market mechanism provides a live 'opinion contract' that agents can subscribe to for continuous alignment and verification.
In the agent stack, Reppo occupies the data and evaluation layer. By providing a decentralized infrastructure for reinforcement learning reward signals, they enable developers to build agents that are less susceptible to reward hacking. Their use of the $REPPO token as a universal denomination for independent data markets makes them a coordination layer for the diverse data needs of a multi-agent ecosystem.
The central thesis of Reppo is that AI training data is currently a static deliverable being treated as a living commodity. In the current market, companies like Scale AI or Labelbox manage quality through administrative processes: audits, guidelines, and inter-annotator agreement scores. Reppo argues this approach is insufficient for frontier models because annotators have no economic incentive to be right. They get paid the same whether their work is accurate or merely looks accurate to an auditor. This lack of skin in the game leads to reward hacking, where models learn to exploit the blind spots of their verifiers rather than solving tasks correctly.
Founded in 2024 by Caroline Cai and Raghav Ramadya, Reppo is a decentralized network that replaces the annotation factory with programmable prediction markets. The company, which spun out of the Protocol Labs Venture Studio, recently raised $2.2 million in seed funding to build out what they call 'Datanets.' These are on-chain environments where participants use the $REPPO token to bet on the quality and accuracy of specific data points, or 'Pods.'
The architecture of a Datanet involves four distinct roles. Owners register a Datanet as an NFT and set the economic parameters. Publishers submit raw data, such as robotics trajectories, code snippets, or medical images. Voters lock tokens to gain voting power and stake that power on 'opinion contracts' regarding the data’s quality. Finally, AI subscribers—ranging from robotics startups to frontier labs—pay in $REPPO to access the resulting datasets.
This setup addresses the verification crisis in reinforcement learning from human feedback (RLHF). Because voters lose capital if they approve low-quality data, the system creates a self-correcting signal. Reppo calls this 'skin in the game.' Unlike traditional labeling pipelines where data is collected, labeled, and then shipped as a static file, Reppo's data flows are continuous. Every 48 hours, an epoch closes, votes are tallied, and the dataset updates. For a robotics company training an autonomous system, this means they can subscribe to a live pipeline of verified data rather than a snapshot from several months ago.
Competitively, Reppo is an infrastructure layer. By turning a curated dataset into a tradeable NFT with a verifiable on-chain P&L, Reppo allows builders to treat data curation as a business. If a robotics engineer builds a high-performing Datanet that attracts consistent subscription revenue, that Datanet can be sold or traded.
The team brings experience from organizations that defined decentralized infrastructure, including Protocol Labs, AWS, and Filecoin. Headquartered in Canada with a global remote team, the company is positioning itself for a future where autonomous agents require reward signals that update in real-time. The bet is that markets will always outperform committees at separating signal from noise. As AI models move into complex physical environments, the bottleneck is no longer compute power, but the human-verified truth required to align those models. Reppo provides the economic rails to produce that truth at scale.
Programmable prediction markets for AI training data and evaluation.