3. Technical Architecture

Decentralized Computing Network Platform Design

3. Technical Architecture

3.1 Decentralized Computing Network Platform Design

The cAIToken decentralized computing platform is built on globally distributed device nodes. Its core goal is to split AI training tasks from traditional centralized data centers into on-chain verifiable task units, executed by participating nodes to earn cAI rewards from the unique mining pool. The platform adheres to the principle "single pool, zero permissions, output only from real contribution," centering around the release logic of 80,000,000 pre-minted tokens.

The underlying architecture adopts a modular "Task Slicer + Distributed Execution Layer" design, decoupling task segmentation, gradient computation, verification, and reward calculation to enable high concurrency, multi-node collaboration, and heterogeneous computing power execution. All task execution produces on-chain verifiable "computing proofs" as the sole basis for mining pool rewards.

cAI platform supports GPUs, CPUs, edge devices, and other node types without requiring specialized mining farms. Device detection, weight allocation, task distribution, and reward settlement are fully transparent, ensuring fair recording and incentive for all computing contributions. The platform uses an EVM-compatible architecture, enabling high scalability and portability of pool, task, and verification contracts.

3.2 Compute-PoW + AI-PoW

cAIToken's consensus model centers on the unique mining pool output mechanism --- 100% of token release comes from two verifiable actions: PoW computing power mining and node verification rewards from staking.

(1) Compute-PoW (Basic Computing Output)

Nodes perform basic computation via GPU/CPU and submit verifiable computing proofs (PoC) to receive basic pool output. Output depends only on:

Real device computing power

Online duration

Task execution quality

Computing reputation (historical performance)

Meaningless computation is not rewarded; all verifiable results must be reproducible, auditable, and tamper-proof.

(2) AI-PoW (Training Task Output)

Nodes execute training fragments, gradient calculations, inference tasks, etc., validated through the Proof-of-Training model. Greater training contribution earns higher rewards from the pool.

(3) Stake-Boost (Staked Task Enhancement)

Nodes staking cAI gain task acceptance and verification weight, allowing them to:

Earn additional verification rewards

Increase task allocation priority

Enhance long-term contribution value

Staking does not directly produce tokens but increases task and verification weight, fully conforming to the second release method emphasized. All rewards come 100% from the pool with no other issuance source.

3.3 Mining Pool Contract and Token Release Protocol

The mining pool contract is the ultimate core of cAIToken, adhering to three principles: unique, zero permissions, full-process transparency.

(1) Single Pool Architecture

80,000,000 cAI are minted once before mainnet launch and written entirely into the pool contract;

No team reserve

No private sale allocation

No investor share

No token control entry

This forms the basis of cAI's trustworthy distribution mechanism.

(2) No Admin Keys

The pool has renounced all administrative rights:

Cannot mint

Cannot pause issuance

Cannot modify release rules

Cannot adjust reward ratios manually

On-chain rules are permanent once written and can only be upgraded by DAO governance in future versions.

(3) Dual Release Logic (100% from Pool)

PoW: Output from real computing contributions

Staked Nodes: Earn ecological rewards as task executors/validators

Smart contracts calculate final rewards based on task execution, verification quality, computing reputation, and stake weight.

3.4 Security and Privacy Protection Mechanism

As the foundational infrastructure hosting global training tasks, cAIToken's security framework focuses on four dimensions: verification, anti-fraud, isolation, and transparency.

(1) Pool Security

Tokens have renounced permissions

Reward release is fully automated

No one can pause, modify, or intercept

Reward calculation and distribution are fully auditable

(2) Computing Power Authenticity Protection

Behavior analysis, replayable task verification, and GPU fingerprinting detect:

Fake computing submissions

Simulated GPUs

Replayed tasks

Forged gradients or inference results

Detected violations are automatically penalized, downgraded, or removed.

(3) Node Isolation and Reputation System

Each node has an on-chain reputation profile recording:

Successful tasks

Failure rate

Reward receipt

Stake status

Anomaly judgments

Higher reputation yields higher task priority.

(4) Privacy Protection

Training data is executed locally on nodes only, not uploaded or broadcast. Only the following is stored on-chain:

Parameter hashes

Contribution proofs

Verification results

User device information, data content, and model weights remain private.