Core Architecture v0.3.0
System Architecture
The 5-tier routing engine (Amortized Reasoning) minimizes latency and token expenditure by prioritizing retrieval over stochastic generation.
The 5-Tier Routing Pipeline
FAST ROUTE0 Tokens
Exact memory match via FAISS/HNSW similarity lookup.
REFLEX ROUTE0 Tokens
Pre-compiled programmatic skill activation.
COMPILED ROUTE0 Tokens
Semantic pattern matching against the skill library.
HYBRID ROUTE~20% Tokens
Memory-assisted generation for delta-reasoning tasks.
SLOW ROUTE100% Tokens
Full Verified Synthesis loop for novel/complex problems.
Episodic Memory
NARE implements a persistent episodic memory layer using a FAISS-backed HNSW index. Every successful task is stored as a high-fidelity trace.
- Semantic similarity threshold (τ ≥ 0.85)
- Automatic pattern discovery via clustering
- Sub-100ms retrieval for known problems
Verified Synthesis
All SLOW-tier generations pass through a programmatic validation oracle. Code is rejected unless it satisfies formal constraints.
Verification Loop
Synthesize → AST Parse → Run Tests → Criticize → Repair