PoI Scoring
The computational contribution is quantified through a scoring function that considers accuracy, efficiency, and complexity: PoI_Score_i = Σ (from j=1 to n) (Accuracy_j × Efficiency_j × Complexity_j)
Where
Accuracy_j is validated through cryptographic comparison with reference outputs stored in a commitment scheme.
Efficiency_j is normalized against a baseline benchmark for each task category to prevent manipulation.
Complexity_j is determined by standardized benchmarks with version-controlled circuits.
The system employs a challenge-response protocol with time-bounded execution to prevent pre-computation attacks. Tasks are assigned deterministically using a Verifiable Random Function (VRF) integrated with BABE's epoch randomness to prevent manipulation:
TaskID = VRF(EpochRandomness, BlockHeight)
Note: For initial implementation, "correctness" in PoI tasks will be limited to well-defined mathematical operations with deterministic outputs (e.g., matrix operations, specific inference tasks with reference implementations). The system will begin with a narrow set of verifiable computations and expand as ZK technology advances, rather than attempting to verify arbitrary AI workloads immediately.
The results of these tasks are verified using ZKPs, a process that takes ≈ 2 milliseconds. For example, if a validator completes 100 tasks with Accuracy = 0.95, Efficiency = 0.01, and Complexity = 1M FLOPs, the score is:
PoI_Score_i = 100 × 0.95 × 0.01 × 1,000,000 = 950,000
Our initial PoI implementation will focus on verifiable matrix operations and simple inference tasks with pre-defined circuits, rather than arbitrary AI computation. These operations form the building blocks of more complex AI models while remaining feasible for current ZK technology. Specifically, we will support:
Matrix multiplication verification with dimensions up to 100×100
Element-wise activation functions (ReLU, Sigmoid) for vectors up to 1,000 elements
Simple feedforward inference for models with up to 10^5 parameters
Each supported operation will have corresponding pre-compiled circuits, dramatically reducing the bootstrapping period for new validators. Complex dynamic models and training verification remain research challenges that we are actively addressing through techniques such as modular verification and sparse network encoding.
We propose a phased approach to PoI implementation:
Phase 1: Basic matrix operations and fixed-architecture inference tasks'
Phase 2: Modular network verification with layer-by-layer proof composition
Phase 3: Dynamic architecture support with generalizable circuit templates
This incremental strategy allows the network to bootstrap with practical verification capabilities while more sophisticated mechanisms are developed and optimized
The Proof Behind Champions
Those who compete at the edge of human precision now back the technology that defines digital truth.
