MANTIS scores miners based on the information theoretical value of their embeddings to the prediction of BTCUSD.

Incentives are distributed proprtional to a given miners contribution to the accuracy of a MLP that interprets these embeddings run by validators.

Time to emissions depends entirely on how quickly the validator can learn how to interpret a given miners embeddings.

This approach is in contrast to having miners produce and directly scoring them, it is superior in 2 primary ways:

Synergy/Synthesis:

A set of embeddings which are not individually predictive, may carry predictive power when combined with others (such as with funding rate and open interest along with many other feature pairs, historically).

In other words, additive individual evaluation cannot detect higher-order interactions, and is therefore unsuitable for incentivizing the highest quality NETWORK outputs.

To provide an example, you could have a system that directly scores predictions, and a top set of 3 miners dominating the network, their accuracies are around 57%, and an ensemble of their outputs gets around 61%.

You could have a miner that outputs 58% accuracy, and takes over the network, however it's underlying signal was already captured by the existing miners, and the network performance in fact, degrades.

On the other hand, you could have a system with 55% accuracy, that captures signal orthagonal to the existing strategies, and this could not stay registered with a linear system.

MANTIS Solves this. MANTIS Is a framework for extracting the most accurate network performance for time series tasks. Our first being BTC forcasting.

Technical details on how salience is calculated, and on our timelock implementation can be found in the codebase: https://github.com/Barbariandev/MANTIS.