We introduce Reppo, a novel ML modelling protocol designed as an open infrastructure to incubate, build, and deploy modular and composable risk & reputation systems i.e. systems that allow end users to progressively delegate trust. Users of networks and assets that have a significant off-chain component, such as Decentralized Physical infrastructure Networks(DePINs), Real World Assets (RWAs), and DeFi networks, leverage models built on Reppo to make decisions about investing, service usage, and staking.
The Reppo protocol incentivizes a community of AI and ML researchers, referred to as Reppo Modellers, to combine on-chain data with predictive modeling to determine which stakeholders for a given network are more likely to engage in “bad” behavior such abusing block rewards through self-dealing, being non-responsive to serving client requests, or responding with lower quality of service than they committed to. The modellers stake Reppo tokens to commit models to the Reppo pool of Models and are rewarded a royalties paid by products and apps that end users and stakeholders can interact with depending on their specific use case.
Traditional physical infrastructure systems are designed and controlled by centralized entities, such as Uber or AWS. These vertically integrated entities allow for fast scaling and quality control simultaneously. However, participants of these networks lack control over their benefits and value received because power lies with these conglomerates that take a significant cut in the value chain. Decentralized Physical infrastructure Networks (DePINs) are a new paradigm in physical infrastructure that leverages blockchain technology and decentralized networks to create more efficient, transparent, and resilient systems. DePINs give value back to the operators of physical infrastructure. Networks like Filecoin and DIMO provide full transparency and control to its supply side actors, allowing them to directly benefit from the value they create. At a market cap of 2.2 trillion USD today, the DePIN sector is projected to grow to 3.5 trillion USD in 2028.
As such projects work towards liberating our built-environment, they face the hard challenge of verification i.e. the oracle problem of verifying physical sensor data from hardware (or objects in case of RWAs), referred to as one of the toughest problems in crypto. Without meaningfully addressing the verification challenge, networks face the Verification Trilemma i.e. the trade off between decentralization, scalability, and quality deal-making on the network, the latter often being rated limited by the behavior of supply side actors, often referred to as validators and/or node operators. The aforementioned supply side actors, as profit-maximizing stakeholders, have strong incentives to engage in abusing block rewards through self-dealing, being non-responsive to serving client requests, or responding with lower quality of service than they committed to.
This risk is especially high for protocol and decentralized networks that have significant off-chain components, such as facilitating services off-chain, like data transfer and storage for decentralized storage networks, claims on usage of clean energy to operate the nodes etc.
While on-chain activities are easy to track, off-chain quality of service is less transparent and verifiable. Networks currently succumb to centralized KYC / KYB checks to weed out suspicious actors and de-risk the network but the trade off on decentralization is significant.
Risk and Reputation are inherently subjective.
To discern between honest and malicious parties, and progressively delegate trust, a reputation system is needed. Traditional reputation systems aggregate a number of metrics and follow a scoring mechanism to produce scores to establish trust between buyers and sellers using transaction feedback. Eg. Yelp, e-commerce review systems etc. However, these types of reputation systems rely on a centralized infrastructure, and are therefore unsuitable for Web3 networks where the calculation, governance, and verifiability of such systems is paramount.
This is where Decentralized Risk & Reputation Modles come in picture.
Decentralized Risk & Reputation Models provide a way for networks to focus on scalability and decentralization without sacrificing the quality of their network. With verifiable models, built by domain knowledge experts, networks can focus on hyperscaling the demand and supply sides of their network and offload the job of building and maintaining reputation systems that incentivize accretive good behavior on their network to Reppo. Bad actors will naturally incur bad reputation/high risk making it easier for a specific network to disincentivize and eliminate such actors from the ecosystem due to either “starvation” (i.e. no clients want to make deals with bad actors) or governance (e.g. get kicked out by the community).
Few risk and reputation models exist today for DePINs, DeFi, and RWAs. Some are via traditional methods such as user-driven rating systems, while others are via random-sampling bots that mimic clients to gather empirical data on node operators and enforce consequences on malicious actors. While these methods are effective to a certain extent, they are centralized in nature and are prone to single point of failure and collusion. A malicious actor would be willing to pay the rating system or the bot operators to get better ratings, as long as the cost is smaller than not cheating.
Reppo’s approach to build a decentralized modelling network makes it economically unwise for a malicious actor to cheat. A large community of Reppo Modellers, each of whom use their own heuristics and techniques, are constantly and actively competing to build the most accurat models based on the data being generated by the validators/node operators. While validators can cheat by intervening in what data is shared in the short term, they have a strong incentive to engage in good behavior i.e. not engaging in self-dealing in the medium to long-term as the reppo flywheel spins. This is because the iterative nature of ML models, as well as the diversity of such models in the pool of models, will result in identifying bad actors. Reppo protocol weaves various layers and tools necessary to build models and consume them downstream including but not limited to access to:
Metric pools - On-chain and off-chain data that are measured against stated goals.
Provenance layer - The layer allowing modellers to prove ownership of their models. We plan to leverage Witness.co to achieve this.
zkML Enabled Inference Layer - The layer that allows the risk reputation consumers to verify that the results of the models aren’t doctored. We are partnering with Modulus.xyz and Vannalabs.ai to enable a seamless experience.
The aforementioned components aren’t exhaustive and we see future opportunities to partner with ML training protocols such as Gensysn.ai making Reppo the single-stop decentralized reputation modeling network that anybody in the world can rely on.
The Reppo network further achieves robustness by measuring the quality of work committed by the Reppo modelers and taking a marketplace approach giving the modelers full ownership and control of their models and associated monetization pathways. The developer community also act as a quality & assurance layer in the network selecting Reppo models based on parameters relevant to addressing threat vectors.
If an actor i.e. the validator/node operator is well-behaved and motivated to improve its business, different risk & reputation models albeit having different heuristics should converge to the finding that this is a good actor useful to the network. This eliminates biases that current centralized risk & reputation systems have such as brand value, size of the business etc.
Reppo Modeller: participates in the Reppo network as nodes. They are data scientists, AI modellers, ML researchers, and anyone who wants to be on the supply side of the model economy.
Raw Data Pool: Reppo provides the holistic experience for building ML models, starting with data from on chain and off chain sources for various networks. This includes chain snapshots, indexed on-chain data and off chain data from sources such as https://novaenergy.ai/