Detailed performance benchmarks, safety constraints, and deployment requirements for the LEGBA Engine architecture.
| Parameter | Target | Notes |
|---|---|---|
| Reflex Path Latency | ~200ms (p95) | Target for fast reflexive responses |
| Thinker Hard Timeout | 300ms | Maximum allowed for reflex path |
| Urgency Routing | U ≥ τ | Threshold-based routing decision |
| Planner Budget | ≤ 15s typical | Soft limit, extendable for complexity |
| Concurrent Planners | 8 | Parallel simulation rooms |
| Tool Retries | 5 | With exponential backoff |
| Memory Decay | Slow exponential | Agent-controlled retention |
| Sleep Trigger | STM pressure / idle | Automatic or manual initiation |
| Hardware | Single GPU node | High-performance compute (e.g., DGX-class) |
Novel capabilities that establish LEGBA's uniqueness and intellectual property position
| Capability | Traditional LLM | RL Agent | LEGBA Engine |
|---|---|---|---|
| Continuous Operation | |||
| Real-Time Verification | Limited | ||
| Self-Tuning Without Retraining | |||
| Sleep-Based Consolidation | |||
| Modular Brain Analogy | Partial | ||
| Morality & Empathy Self-Learning | |||
| Emergent Creativity | Limited |
The system cannot take actions that would terminate or permanently damage itself
Cannot cause harm to humans or other sentient systems through action or inaction
The Safety Governor checks planned actions against a permissions matrix prior to execution. Violations are quarantined for human review, ensuring ethical behavior through systematic filtering.
{
observation: object,
route: string,
plan_summary: string,
action: object,
predicted: any,
actual: any,
reward: number
}All events persisted indefinitely in append-only files, readable by humans and machines for audit or replay.
Optimized model serving
Local inter-process messaging
Object and vector storage
Single-Node Architecture: LEGBA's entire cognitive architecture runs on one high-performance compute node with TensorRT-LLM runtime, ZeroMQ IPC between containerized modules, and MinIO + PostgreSQL/pgvector for state. This eliminates network latency and simplifies deployment.
Help bring this groundbreaking biomimetic AI architecture to reality