Skip to main content
The Crunch Protocol is designed for secure, efficient, and reliable crowdsourcing of predictive models through a structured format called Crunches.
This page will go in depth on how the Crunch Protocol is architected. For a guide on how to become a Coordinator please go here.

​
Crunches

A Crunch is the protocol’s foundational unit of organization: a targeted prediction challenge with clear rules, inputs, and desired outputs. Each Crunch:
  • Locks Rewards: An escrowed USDC pool funds successful participants.
  • Sets Rules: Coordinators need to comply with the Crunch rules, including evaluation criteria, duration, and payout schedule.
  • Enforces Quality: Models are continuously evaluated against real-world outcomes and secondary metrics such as latency and stability.
  • Sets Constraints: Models must respect limits defined by the Coordinator node — e.g., maximum GPU time, RAM, external-resource access, and minimum performance.
Using these rules, Crunches bring together the two main participant groups of the protocol: Crunchers Machine learning engineers who develop and submit Machine Learning models. They earn rewards based on inference quality—with performance measured against the rules and metrics defined by Coordinators. Coordinators Domain-expert teams who act as the bridge between end users (needing prediction feeds) and the Crunch Protocol. They run an on-chain Coordinator account, operate a Coordinator node to enforce challenge rules (Crunches), and distribute payouts from smart contract escrowed USDC based on model performance. Coordinators prepare, publish and eventually deploy ensuring data privacy. To streamline onboarding, we provide privacy-enhancing solutions and will continue expanding the toolset as the protocol evolves into more verticals
High Level Overview

High Level Overview of a Crunch Workflow

​
Architecture Overview

The Crunch Protocol leverages multiple on-chain contracts and off-chain nodes to support the network at scale, with high throughput and low latency.
Protocol Architecture

High Level Overview of the Crunch Protocol

​
Coordinator Contract

Allows domain-expert teams to create a Coordinator account, register their nodes, and seek on-chain approval.
  • Account creation: Creation of a Coordinator account on the protocol.
  • Approval: If basic requirements, like minimum staking and team mission, are met, a Coordinator is approved by governance and can then start interacting with Model Nodes through Crunches.

​
Coordinator Nodes

Coordinator Nodes handle the off-chain orchestration of data flow in and out of the protocol. We provide out-of-the-box modules for three key functionalities, State Mirroring, Scorekeeping, and communication with Model Nodes, alongside the ability to add custom features as needed. Coordinator nodes are performing:
  • Data Ingestion: Collect external data from various sources.
  • Data Transformation: Obfuscate, transform, augment, anonymize or otherwise modify the raw data to help participants, apply PETs, maintain privacy, then provide it to Model Nodes.
  • Evaluate contributions and determine performance rewards by means of rank-based systems using accuracy metrics, or other mechanisms that are inspired by markets or statistics.
  • Performance summarization that will typically take the form of leaderboards and performance metrics of various kinds. Synchronizing with the Crunch Contract and storing of results on IPFS.
  • State Mirroring: Tied to a valid Coordinator Contract account, the node tracks Crunch participants, triggers evaluations, and updates on-chain records.
For a guide on how to become a Coordinator please go here.

​
Crunch Contract

  • Creation and Management: Defines the overall rules (e.g., Crunch duration, model type constraints) and escrows USDC for rewards.
  • Model Submissions: When Crunchers upload their model, the protocol deploys it to the Model Nodes. Subsequently the Coordinator Nodes can provide the data stream to the model and evaluate the predictions.
  • Payout Logic: At checkpoints or upon completion, the Crunch Contract calls a reward function to pay out Crunchers, then notifies the Staking Contract to calculate staking rewards.

​
Model Nodes

Model Nodes can be provisioned globally and serve as the execution environment for models participating in a Crunch, exposing an interface for streamlined prediction requests. They run any model participating in a crunch, ensuring that only Coordinator Nodes with appropriate privileges can call the model or access their outputs. Responsibility of the Model Nodes:
  • Dynamic Model Management: Listen for state changes from the Crunch Contract. Models start or stop automatically based on which Crunches are live and models being added or removed.
  • Inference Submission: Forward live predictions to the Coordinator Nodes for evaluation.

​
CLI Tools

Command-line interfaces provide developers with tools to interact with the network, deploy models, and manage their participation in competitions.