Foundation

System Architecture

Neurosymbolic Type 6 classification (Kautz, 2022). Neural reasoning layer coupled with symbolic computation layer. Offline-first. Air-gapped. Explainable.

Two-Layer Design

Neural Layer — 5 Agents
Specialized agents handle situation understanding, course of action generation (3 independent), adversarial analysis, and comparative judgment. Each agent is a fine-tuned domain model, not a general-purpose assistant.
Symbolic Layer — 9 Modules
Monte Carlo simulation, minimax regret, MCDA scoring, sensitivity analysis, risk derivation, consensus engine, flaw mapping, confidence intervals, and audit trail generation. Deterministic, reproducible, traceable.

Interface Points

  1. Neural → Symbolic — Agent outputs (structured JSON) feed into symbolic scoring, validation, and decision analysis modules. The symbolic layer does not interpret natural language — it processes structured data.
  2. Symbolic → Neural — Computed scores, risk assessments, and Monte Carlo results are fed back to agents for informed judgment in later pipeline stages.
  3. Feedback Loops — Red team analysis results inform COA refinement. Monte Carlo failure modes trigger re-evaluation. Consensus disagreements flag specific reasoning chains for review.

Defense Properties

Explainability
Every recommendation traces through the full pipeline. Agent reasoning is captured, symbolic computation is logged, decision analysis is decomposable.
Reproducibility
Same inputs produce same symbolic outputs. Agent variation is captured through multi-run consensus (TMR), not hidden behind non-determinism.
Auditability
Full audit trail from question to recommendation. Each agent's contribution is tagged. Each score is decomposable to individual controls. Pipeline trace viewable in real-time.
Robustness
Monte Carlo validates recommendations against 1000 perturbation scenarios. Sensitivity analysis identifies fragile assumptions. TMR catches single-point reasoning failures.

Deployment Model

Offline-first. All core pipeline functionality runs on locally-hosted models in air-gapped environments. No external API dependencies for tactical operations. Designed for DDIL (Denied, Disrupted, Intermittent, Limited) conditions.

When connectivity is available, frontier models can be used for deeper analysis on non-sensitive contexts. The architecture supports seamless model substitution — swap a local model for a frontier model without pipeline changes.

Model Philosophy

No model names are exposed in operational contexts. Agents are identified by function: "situation analyst", "COA generator", "red team adversary", "comparative judge", "decision analyst". Models are interchangeable. The architecture is the product, not the model.

The strength of the system is not in any single model but in the structured interaction between neural reasoning and symbolic computation.