Architecture Overview¶

Archivus is built as a seven-layer intelligence stack where each layer addresses a specific aspect of verifiable enterprise AI.

The Full Stack¶

graph TB
    subgraph Input["Input Layer"]
        DOC[Documents]
        VOICE[Voice Sessions]
        CONN[Data Connectors]
        API[External APIs]
    end

    subgraph KG["Knowledge Substrate"]
        ENT[Entities]
        REL[Relationships]
        CLAIMS[Claims]
        PROV[Provenance]
    end

    subgraph REASON["Symbolic Reasoning"]
        GRAPH[Graph Traversal]
        CONTRA[Contradiction Detection]
        INFER[Inference Chains]
    end

    subgraph VERIFY["Verification Layer"]
        GOLAG[Evolutionary Agents]
        QV[Calibrated Confidence]
    end

    subgraph ORCH["Orchestration"]
        DAG[Multi-Step Workflows]
        HITL[Human-in-the-Loop]
    end

    subgraph COMPLY["Compliance Backbone"]
        AUDIT[Audit Logs]
        EVIDENCE[Evidence Bundles]
        RETENTION[7+ Year Retention]
    end

    subgraph TRUST["Trust Layer"]
        HASH[Hash Chains]
        HEDERA[Hedera Anchoring]
    end

    subgraph FED["Federation"]
        SHARE[Verified Facts]
        BOUNDARY[Data Sovereignty]
    end

    Input --> KG
    KG --> REASON
    REASON --> VERIFY
    VERIFY --> ORCH
    ORCH --> COMPLY
    COMPLY --> TRUST
    TRUST --> FED

Layer Responsibilities¶

1. Input Layer¶

What it does: Ingests intelligence from any source

• Documents: PDFs, Office files, images with AI extraction
• Voice: Real-time transcription and intelligence capture
• Data Connectors: Structured data from external systems (POS, reviews, APIs)
• Research: Web intelligence with automated fact-checking

All inputs flow into the Knowledge Graph, not into isolated silos.

2. Knowledge Substrate¶

What it does: Transforms data into queryable knowledge

Core model: Quadruples (not triples)

(Entity, Relationship, Entity, CONTEXT)

Context includes:
- Temporal validity (when was this true?)
- Geographic scope (where does this apply?)
- Provenance (who said it? what's the source?)
- Supporting evidence (what sentences prove this?)

This is what makes "Who is the president?" answerable across time.

Learn more →

3. Symbolic Reasoning¶

What it does: Queries the graph BEFORE asking the LLM

Traditional AI: User question → LLM → Response (may hallucinate)

Archivus: User question → Graph traversal → Verified facts → LLM grounded in truth → Response

Capabilities: - Entity recognition and linking - Relationship discovery ("If A employs B and B authored C...") - Contradiction detection (symbolic, not probabilistic) - Temporal reasoning ("What was true in Q3 2024?")

This is the "symbolic before neural" principle.

4. Verification Layer¶

What it does: Ensures AI decisions are calibrated and improve over time

Powered by GOLAG (Game-Oriented Lagrangian Agent Governance): - Agents have finite confidence budgets - Quadratic voting forces honest calibration - Overconfident agents exhaust budgets - Well-calibrated agents accumulate influence - System gets smarter by knowing what it doesn't know

The system learns epistemic humility.

Learn more →

5. Orchestration Layer¶

What it does: Executes complex multi-step intelligence pipelines

Not "document workflows"—intelligence workflows: - Multi-step verification processes - Human-in-the-loop approval gates - External tool integration via MCP - Parallel execution with dependency resolution

24 node types from AI processing to approval requests.

Learn more →

6. Compliance Backbone¶

What it does: Makes AI decisions auditable and compliant

• Audit Logs: Every agent decision recorded
• Evidence Bundles: Self-verifying compliance exports
• 7+ Year Retention: S3 Parquet for long-term storage
• Content Addressing: Tamper-evident by design

This isn't "document analytics"—it's the infrastructure that makes AI trustworthy in regulated industries.

7. Trust Layer¶

What it does: Enables independent verification without trusting Archivus

Three-layer architecture: - Local: SHA256 hash chains (tamper detection) - Tenant: MotherDuck compliance backbone (analytics-ready audit trail) - Global: Hedera Consensus Service (public ledger anchoring)

The insight: When Enterprise B receives claims from Enterprise A, B doesn't need to trust A's database. Cryptographic proof via Hedera replaces institutional trust.

Learn more →

8. Federation Layer¶

What it does: Intelligence flows across organizations, data stays home

What flows: - Verified claims with provenance - Entity references (canonical IDs) - Relationship signals (anonymized) - Trust scores from calibrated verification

What stays home: - Documents (always) - Raw data (never leaves) - PII (protected)

This is the endgame: cross-organizational intelligence without compromising data sovereignty.

Learn more →

Architectural Principles¶

1. Verifiability Over Fluency¶

A confident wrong answer is worse than an uncertain right one. Every claim must be traceable to its source.

2. Symbolic Before Neural¶

Query the knowledge graph first. Detect contradictions symbolically. Build inference chains with logic. Then let the LLM make it fluent.

3. Data Sovereignty is Sacred¶

Raw data never leaves an enterprise's boundary. Only verified facts—with provenance—flow through federation.

4. Trust Must Be Explicit¶

Every fact has a source type. Every relationship has a confidence score. Every cross-enterprise interaction has a trust chain. No implicit access.

5. Calibrated Confidence¶

Confidence is a finite resource. Systems that treat every claim as maximally certain undermine integrity. Quadratic voting enforces honest calibration.

Technology Stack¶

Multi-Tenant Security¶

Every layer enforces tenant isolation: - Application: JWT validation, middleware enforcement - Service: Tenant ID validation on all operations - Database: Row-Level Security (RLS) on 149 tables

Security is defense-in-depth, not a single checkpoint.

Learn more →

What Makes This Different¶

Next Steps¶

Explore specific architectural layers:

• Knowledge Graph - The substrate that makes AI verifiable
• Evolutionary Verification - How agents get smarter over time
• DAG Orchestration - Multi-step intelligence pipelines
• Trust Layer - Cryptographic verification architecture
• Federation Protocol - Cross-org intelligence sharing
• Voice Intelligence - Real-time capture from conversations
• Multi-Tenant Security - Defense-in-depth isolation

Architecture designed for verifiable intelligence at enterprise scale.