The NOVA Architecture

Core thesis

Single-model AI gives one perspective. Real decisions benefit from multiple viewpoints. NOVA Council implements NOVA Nexus — 17 AI entities with distinct AI cores and processing characteristics. Their outputs are processed in parallel with multi-stage adaptive routing and synthesized by NOVA Core into one advisory response—preserving disagreement where it matters and surfacing tradeoffs.

GMAS 2.0 — Technical coordination

Multi-stage adaptive processing: each of the 12 Pillars processes queries through adaptive routing—simple queries use lightweight paths, complex queries engage deeper analysis; creates genuine synthesis rather than single-perspective output. Stateful processing units with outcome-based lifecycle management: units process per query component with confidence-threshold survival; successful patterns persist via evolutionary pattern caching, failed paths are discarded. Confidence self-calibration: per-entity SGD adjusts confidence multipliers based on acceptance feedback. Parallel processing: consulted pillars run simultaneously (2–4× faster, no race conditions). Parameterized cognitive tuning: each entity has tunable risk_tolerance, decision_speed, and language_patterns—e.g. The Analyst (precision-weighted, low risk tolerance), The Innovator (divergent exploration, high risk tolerance), Sentinel (5-factor risk evaluation), Arbiter (governance and constraint validation). Multi-factor risk evaluation and alignment scoring are built into the architecture.

GMAS — 6 layers of geometric intelligence

The system is built on a Geometric Multi-Agent System (GMAS): six shapes that define origin, coordination, stability, advisory spread, balance, and system boundary.

• Point — NOVA Nexus; every decision begins here.
• Tetrahedron — NOVA Core, Sentinel, Arbiter, and Living Memory form the core.
• Cube — Structural bridge between core and the 12 Pillars.
• Icosahedron — 12 Pillars with multi-stage adaptive processing and parameterized cognitive tuning; adaptive routing for genuine synthesis.
• Sphere — Parallel processing, equal distance from center; fair weighting.
• Hypersphere — The Orchestrator at the system boundary.

Memory uses a separate hierarchy: raw inputs → categorization → pattern recognition → synthesis. Routing selects which Pillars to consult based on query and domain. Optimized to ~2× overhead vs single-model baseline; GPU-accelerated at scale.

Key principles

• Advisory only — AI advises; humans decide. No autonomous execution.
• Transparent disagreement — Conflicting advisor views are surfaced, not hidden. Most AI systems hide disagreement behind averaging. NOVA Council surfaces it — because knowing where experts differ is often more valuable than the final answer.
• Risk-aware — High-stakes outputs can be gated for human review.
• Multi-perspective by design — Each query is routed to the most relevant advisors from a council of 12.