Categories specify a definition about what a model should be targeting and what its output should be. Category output schemas address the latter by defining what models targeting a given category should provide as output.
Category output schemas will be defined following the JSON standard. As an example, let’s look at what a potential output schema for Uniswap v3 will look like:
Figure 3: Sample Output Schema
Output schemas contain all of the data required. It defines the web3 application, category, and the expected output. In this example, the output is a list of three elements. Each element represents a specified range for the API3/ETH liquidity pool for Uniswap v3 for a given risk assessment, tied together with an expected APY.
Every model for this category is expected to produce output based on this schema and every category is expected to have a single output schema defined. This standardizes the output, homogenizing the results between different models.
With output schemas, Credmark offers a simple way for web3 applications to look at the attestation registry and instantly know exactly how to interact with the output data they’ll be receiving. This provides clarity and transparency, which in turn speeds up adoption and integration.
The model selection strategy defines which model is the current best for a given category. A strategy specifies the weights that are used for a category’s performance metrics. The analysis then selects the model that best matches reality over time for the most number of outputs as defined by the schema.
For example, a model selection strategy for Uniswap v3 can choose to weight live performance validation highly due to the fact that there is limited historical data at the time this whitepaper is written. Other possibilities for model selection strategies include whether one prefers to weigh consistency of APY over time more heavily compared to maximized APY over time but more varied, or vice versa.
If a category is competitive, but the old model selection strategy is too specialized, the Credmark community can come together to decide on an improved strategy. Because of the implications of how this will end up affecting the system as a whole, the decisions for choosing model selection strategies for categories will be determined 9 by governance. CMK token holders will be able to vote to change model selection strategies for any category, bringing agility to the system as whole. This way, Credmark selects models that most accurately reflect reality, so users get the most accurate available results.
With a decentralized model selection strategy mechanism, the system preempts selection bias, a common issue in centralized providers. CMK token holders are key for the success of the system, with this being one of their essential roles.