Model Attestation Registry

The model attestation registry is a smart contract running on Ethereum that keeps a public record of all of the models currently available on the Credmark platform.
Each model attestation entry in the registry is comprised of:
  • The compact hash of the model (aka model ID).
  • The Nix hash of the model.
  • The category of the model.
  • The output schema of the model.
  • The model NFT ID.
This data provides context to investigate model metadata. An auditor can understand which category the model belongs to, what the output results of the model are (which is helpful for integration with web3 applications), and the model NFT ID. Each model attestation entry acts as a secure pointer to a given model on the Credmark platform.
Whenever a model is submitted to the platform it goes through the model attestation pipeline.
  1. 1.
    The model contributor packages the model.
  2. 2.
    The model is run through the deterministic hashing function.
  3. 3.
    An NFT is generated to enable transferrable model ownership.
  4. 4.
    Attestation metadata is attached (hash, category, NFT, output schema).
  5. 5.
    The contributor submits the attestation to the Ethereum blockchain with staked CMK tokens.
Figure 4: Model Attestation Registry
CMK tokens are required to be staked when submitting a model attestation to the registry. This prevents resource abuse and spam, and ensures the platform can scale gracefully. Furthermore, it provides a clear and straightforward way to keep a public record of the current set of available models being run.
If a contributor wishes to remove their model from the attestation registry then they must burn the NFT. This allows deletion of the model from the registry, unlocking the staked CMK.

Model NFTs

Holding a Credmark Non-Fungible Token, known as an NFT, allows the holder to claim the rewards of the underlying model’s usage. With this mechanism, Credmark has created a new asset class of non-fungible tokens in which value is calculated as the future earning potential of an executable data model.
All models added to the attestation registry require an NFT that represents it. This NFT is officially recognized as soon as the attestation is added to the registry. Consequently, models running on the Credmark platform are fully tokenized. This enables transferable ownership of income generating tokens.
When a contributor submits a new model to the attestation registry, they have officially attested that their NFT represents their new model. If the contributor is well known and has previous models that have done well, this NFT already has intrinsic value upon minting.
Given that new models are subject to a live performance validation period before they’re eligible to produce returns, the model NFT acts as the future until that period expires. Model contributors can sell their NFT shortly after attestation where the risk is higher for the buyer, meaning that the reputation of the contributor and performance of their previous models is most likely to account for the NFT pricing. They could also hold the NFT for passive income, or sell it after the income has been proven. As the live performance validation period comes to an end, the real performance of the model and its likelihood to earn returns starts to solidify, which gets priced into the NFT.
This income-generating model NFT also ends up being an attractive financial asset for potential buyers. The NFT has guaranteed returns for the foreseeable future, in addition to acting as one of the less crypto-market correlated assets with a completely novel risk and reward profile. Thus crypto portfolios may look to purchase battle-tested revenue generating model NFTs for diversification purposes.
The potential applications of model NFTs in the grander DeFi ecosystem appear to be quite far reaching, with likely many opportunities for collaboration and interoperability with existing web3 applications opening higher order financial primitives for everyone.

Model Leaderboard

The model leaderboard is a smart contract that exists on-chain and keeps track of the top 3 models for every category. The purpose of the leaderboard is threefold:
  • To provide end users with an easy to find, on-chain reference for the current best models so they can easily query for them.
  • As a transparent and timestamped history of the performance of the given models in each category (via looking back into the Ethereum blockchain history to cut down on storage fees rather than storing past rankings within the contract itself).
  • For other smart contracts and web3 applications to use within their protocols as an on-chain source of truth, which enables dynamic model switching to improve risk assessment for their users.
Whenever a new model outperforms one of the current top 3 models in a given category based on the governance chosen model selection strategy, it enters into the top 3, replacing whichever models were currently there. This is automated by the Credmark platform, which posts new transactions updating the leaderboard any time that the top 3 of a category changes.
For model creators, the leaderboard acts as a public proof of their competency and dictates which models earn money for the category. When the model is executed on Credmark, if the model is in the top 3 of that application’s model category, then the model’s NFT owner will receive the reward according to governance.
  • First Place: 80%
  • Second Place: 15%
  • Third Place: 5%
By rewarding even the unused models, Credmark prevents a winner-takes-all scenario, motivating contributors to create models by offering potential income streams. Even a 3rd place model in an extremely popular category can offer decent returns, which also means a decently priced model NFT could be sold on the open market.
Given the financial incentive, eligibility for the leaderboard must take into account game theory and be resistant to attacks. That’s why a model must have passed governance, determined in the live performance period, to be eligible for the leaderboard. Until it has done so, no matter how good the other performance metrics may be, it will not be added to the leaderboard.
This prevents model contributors from gaming the system and producing models that can get them on the leaderboard quickly, yet fail to show any predictive ability. This is a core requirement for the protocol to work, increasing the trust in the system by providing stronger guarantees for end users.