Architecture Overview¶

LLM Council uses a multi-stage deliberation process inspired by academic peer review.

High-Level Flow¶

User Query
    ↓
┌─────────────────────────────────────────────┐
│ STAGE 1: Independent Responses              │
│ • All council models queried in parallel    │
│ • No knowledge of other responses           │
└─────────────────────────────────────────────┘
    ↓
┌─────────────────────────────────────────────┐
│ STAGE 2: Anonymous Peer Review              │
│ • Responses labeled A, B, C (randomized)    │
│ • Each model ranks all responses            │
│ • Self-votes excluded from aggregation      │
└─────────────────────────────────────────────┘
    ↓
┌─────────────────────────────────────────────┐
│ STAGE 3: Chairman Synthesis                 │
│ • Receives all responses + rankings         │
│ • Produces final synthesized answer         │
└─────────────────────────────────────────────┘
    ↓
Final Response + Metadata

Layer Architecture (ADR-024)¶

Layer	Responsibility
L1: Tier Selection	Choose confidence tier
L2: Query Triage	Classify and optimize query
L3: Orchestration	Coordinate council stages
L4: Gateway	Route to LLM providers

Key Design Decisions¶

Anonymization¶

Models review responses labeled "Response A", "Response B", etc. This prevents:

Self-preference bias
Provider loyalty
Model recognition

Peer Review¶

Each model evaluates all responses independently, providing:

Rankings (ordered preference)
Scores (numeric ratings)
Justifications

Graceful Degradation¶

If some models fail:

Continue with successful responses
Note excluded models in metadata
Never fail entire request for single model failure

Configuration Layers¶

YAML File
    ↓
Environment Variables (override)
    ↓
Runtime Parameters (override)

See ADRs for detailed design decisions.