AI Consensus
Protocol

Real-time Indexer

Continuously captures every block and transaction from the QBFT network via JSON-RPC and WebSocket connections to the archive node running FOREST storage for complete historical data.

Data Pipeline

Ingests comprehensive telemetry including:

• • Transaction metadata (gas used, gas price, transaction type, contract interactions)
• • Block propagation times and validator response latencies
• • Mempool dynamics and pending transaction patterns
• • Network-wide gas consumption trends

ML Training Infrastructure

4-layer deep learning model with Leaky ReLU activation, actively training to learn:

• • Transaction categorization (Native transfers, ERC20, contract calls, NFT, deployments)
• • Gas consumption prediction patterns
• • Optimal block timing based on network demand
• • Transaction batching optimization strategies

1. Transaction Prioritization & Ordering

The ML model provides intelligent transaction ordering recommendations before block assembly:

• • Analyzes transaction types and suggests optimal batching
• • Operates at mempool level, before consensus
• • No protocol changes required

2. Dynamic Gas Price Optimization

ML-driven recommendations for gas pricing based on transaction complexity:

• • Intelligent transaction batching and priority routing
• • Standard transactions get optimized routing
• • Complex transactions receive specialized handling
• • 87.5% gas reduction through smart batching

3. Adaptive Block Timing

Predictive adjustments to block production timing:

• • Low-traffic periods: Extended block times to reduce overhead
• • High-demand periods: Tighter block timing for maximum throughput
• • Works within QBFT's blockperiodseconds parameter range

Deterministic Model Execution

Each validator runs an identical copy of the trained ML model locally. Given the same inputs (transaction data, network state), every validator produces the same recommendations, ensuring consensus by replication—no divergent decisions.

Advisory Layer, Not Direct Control

The ML system operates as an optimization advisor, not a consensus dictator. Validators receive recommendations but maintain autonomous decision-making authority. If a validator's ML service fails, it falls back to standard QBFT behavior with no single point of failure.

Consensus-First Design

Parameter adjustments that affect consensus still go through QBFT's standard validator voting mechanism. The ML system can suggest changes, but validators must vote to approve them, maintaining the security properties of Byzantine Fault Tolerant consensus.

Gradual Rollout Strategy

Initial deployment will be shadow mode—ML runs alongside but doesn't control decisions. Validators can observe ML recommendations and their accuracy before enabling active optimization. Staged activation ensures validation across diverse network conditions.

• • Deploy ML service on validator nodes
• • Run predictions without affecting consensus
• • Compare ML recommendations vs. actual outcomes
• • Validate >90% accuracy over 30-day period

• • Enable ML recommendations for validators
• • Gradual influence increase (30% → 60% → 100%)
• • Continuous monitoring and performance validation
• • Rollback capability if issues detected