SYSTEM ARCHITECTURE // V2.1

CONFLICT INTELLIGENCE ENGINE

The core Tacitus engine that fuses OSINT scraping, deep-document RAG, and a full conflict ontology into a living model of political, organizational, and interpersonal disputes. It turns unstructured communication into a structured conflict graph that downstream agents can query for leverage points, veto paths, and realistic corridors of agreement.

01. SYSTEM TOPOLOGY

>> The Intelligence Pipeline

The Concordia Engine operates across a structured analytical chain: Ingestion → Agentic Compute → Strategic Output. Each layer is fault-tolerant, independently scalable, and designed to model conflicts ranging from family disputes to geopolitical crises.

INPUT LAYER

DATA INFRASTRUCTURE

  • > Web search APIs (Tavily / Serper / GDELT)
  • > Firecrawl local media + PDF scraping
  • > Curated Conflict Library (timelines, actor dossiers, agreements)
  • > RAG over authoritative reports & research
  • > Supabase / Vertex AI vector search
  • > Object storage for private case files
COMPUTE LAYER

AGENTIC SWARM

  • > Planner Agent + Agentic RAG Orchestrator
  • > Fact-Discovery Agents (LangGraph)
  • > Contradiction & propaganda filters
  • > Ontology Engine (actors, drivers, narratives)
  • > Strategy / policy agents
OUTPUT LAYER

STRATEGIC ARTIFACTS

  • > Situation reports (JSON / PDF)
  • > Stakeholder & actor maps
  • > Conflict graphs & timelines
  • > Recommendations & options memos
02. INGESTION VECTORS

>> Data Infrastructure

Tacitus uses three major pipelines, each tailored to a different temporal and epistemic profile:

Pipeline A: Velocity

TARGET: High-Frequency OSINT

Tavily / Serper / GDELT / global wires feed a real-time event stream. A deduplication engine filters noise every 3–12 hours, using text-similarity hashing and temporal clustering.

Example sources: Reuters, AP, Al Jazeera, Anadolu, BBC Monitoring.

Pipeline B: Specificity

TARGET: Local Media / Niche Actors

Firecrawl extracts content from community radio, local newspapers, armed-group statements, municipality bulletins, and NGO field reports.

Use cases: local outlets, research groups, neighborhood committees, syndicates, and grassroots platforms.

Pipeline C: Authority

TARGET: Slow-Moving High-Authority Corpus

A curated library and vector store of official communiqués, policy papers, peer-reviewed research, legal frameworks, and vetted long-form reporting. Indexed and chunked with Supabase or Vertex AI for long-horizon recall.

>> Pipeline Orchestration

All three pipelines flow through a LangGraph orchestrator and a Cloud Run microservice layer that schedules asynchronous scraping jobs, handles retries, and writes normalized events into the conflict library and vector stores.

03. AGENTIC PROCESSING

>> The "Concordia" Swarm

Tacitus does not summarize — it reconstructs reality through a structured, multi-agent council. Agents cooperate, challenge each other, and escalate contradictions.

LAYER 1 — FACT DISCOVERY

Developments Agent

Synthesizes the last 24–72h of political, military, and humanitarian developments from the velocity and specificity pipelines.

Contradiction Checker

Runs cross-source validation via a ReAct-style chain. Flags propaganda, inconsistencies, and manipulated media, and can request fresh retrieval when signals conflict.

LAYER 2 — DEEP ONTOLOGY

Conflict Ontology Engine

Converts events into an actor graph: state actors, militias, ministries, NGOs, business actors, and individual influencers. Tracks interests, constraints, leverage points, and red lines on a per-actor basis.

Narrative Miner

Extracts legitimacy claims, grievances, power deltas, and red lines. Distinguishes drivers of conflict from surface rhetoric and updates the curated conflict library with evolving narratives.

LAYER 3 — POLICY GENERATION

> INPUT: User defines stakeholder (multilateral agency, foreign ministry, NGO network, corporate leadership).

> PROCESS: Strategy agents adjust analysis to mandate, leverage, and risk tolerance, using the latest graph plus curated historical baselines.

> OUTPUT: Prioritized to-dos, mediation pathways, pressure levers, and anticipated reactions across multiple time horizons.

AGENTIC RAG — RETRIEVAL AS A TEAM

In Tacitus, Retrieval-Augmented Generation is not a single vector search. A supervisor agent decomposes the question, plans a retrieval strategy, dispatches specialized retriever agents towards different corpora (news stream, local media, curated library, private uploads), and then reconciles the results before any analysis is written.

  • > Planner agent — chooses which pipelines and tools to call, in which order.
  • > Retriever agents — run corpus-specific queries (velocity, locality, authority, private files).
  • > Critic / judge agent — scores relevance, discards hallucinated or off-topic chunks.
  • > Merger agent — assembles the working context window used by analysis and strategy agents.

This Agentic RAG pattern makes the system less brittle than naive RAG: if sources conflict or a query is too broad, agents can iterate, narrow the scope, or surface uncertainty instead of fabricating confident answers.

04. CONFLICT VISUALIZER

>> Visualizing Conflicts Across Scales

The Concordia Engine converts any conflict — geopolitical, organizational, familial — into a dynamic Actor → Interest → Constraint → Relationship graph. The examples below show three different scales: a regional détente, a boardroom governance crisis, and a multi-party inheritance dispute.

>> Node Details

Hover a node to see its Interests, Constraints, Role, and (where defined) Leverage and Red Lines.

>> Deploy the Engine

The Concordia Engine is available for institutional pilots. Supports JSON export, graph visualization, and multi-agent workflows.

Request Access