Data Architecture

How EVE AI Core stores, verifies, and governs data across the governance pipeline.

Overview

EVE's data architecture is purpose-built for deterministic governance. Five specialized data layers provide the storage, retrieval, and verification infrastructure required to enforce charter rules, maintain immutable audit trails, and anchor policy configuration to a stable identity.

Working Context

Active request context and governance pipeline state (Redis)

Episodic Audit

Timestamped governance events and decision records (MongoDB)

Semantic Index

Vector-embedded policy knowledge and fact retrieval (Pinecone)

Identity Anchor

Governance policy configuration with bounded drift (PostgreSQL)

Ledger Memory

Immutable audit trail — 7+ year regulatory retention (PostgreSQL)

Working Context Layer

Active Pipeline State — Redis

Working context holds the current request state as it flows through the governance pipeline. It includes:

Request Context: The active governance query being evaluated
Pipeline State: Current stage in the TVE (Truth-Verified Execution) pipeline
Active Entities: Charter rules, claims, and trust evaluations in scope
CRD Scores: Confidence, Risk, and Domain calibration values for the current request

Working context is short-lived (seconds to minutes) but provides fast, immediate access to governance-critical data during request processing.

Episodic Audit Layer

Decision Records — MongoDB

Episodic audit stores timestamped governance events and decision records. Each episode captures:

When: Precise timestamp of the governance decision
What: The charter rule evaluated, action proposed, or claim assessed
Verdict: COMPLIANT, HARD_BLOCK, SOFT_BLOCK, or ESCALATION outcome
Evidence Chain: Supporting data that led to the decision

Episodic records enable full governance traceability — auditors can reconstruct exactly why any decision was made at any point in time.

Record Consolidation

Governance records undergo consolidation to balance storage efficiency with audit completeness:

Recent records retain full detail including raw request payloads
Older records are compressed to summary form while preserving verdict and evidence hash
HARD_BLOCK and ESCALATION records are never compressed — full detail is retained permanently
Routine COMPLIANT records may be aggregated into daily summaries after 90 days

Semantic Index Layer

Policy Knowledge & Fact Retrieval — Pinecone

The semantic index stores vector-embedded governance knowledge for fast retrieval:

Charter Rules: All 15 charter rules indexed for natural-language policy queries
Verified Facts: Ground truth facts used by the CRD scoring pipeline
Historical Precedents: Past governance decisions indexed for similarity search
Domain Calibration Data: Domain-specific confidence thresholds and trust profiles

Semantic search enables the governance engine to find relevant precedents and policy context in sub-millisecond time.

Identity Anchor Layer

Governance Policy Configuration — PostgreSQL

The Identity Anchor is EVE's governance policy configuration store. Rather than static config files, policy parameters are stored as bounded traits with drift protection:

Policy Traits: Configurable governance parameters (e.g., risk tolerance, confidence thresholds) with min/max bounds
Protected Invariants: 13 invariants that can never be modified (core values, ethical boundaries, identity name)
Drift Budget: Daily (10%), weekly (30%), and monthly (50%) limits on cumulative policy changes
Bounded Drift: Any policy change must pass through the Self-Modification Governance approval chain

The Identity Anchor ensures governance policy cannot silently shift over time. Every change is budgeted, audited, and reversible.

Ledger Memory Layer

Immutable Audit Trail — PostgreSQL (7+ Year Retention)

The Ledger is EVE's ground truth. It is an append-only, tamper-evident audit trail designed for regulatory compliance:

Append-Only: Records can never be modified or deleted — only new entries are written
SHA-256 Hash Chain: Each entry includes the hash of the previous entry, forming a cryptographic chain
7+ Year Retention: Meets SOX, HIPAA, and financial services regulatory retention requirements
Chain Verification: Any auditor can verify the integrity of the entire ledger by recomputing the hash chain

Tamper Evidence: If any record in the ledger is altered, the hash chain breaks at that point. Chain verification via GET /api/claims/chain/verify detects any tampering instantly.

Truth Store

SHA-256 Hashed Facts for CRD Verification

The Truth Store is a curated set of verified facts used by the CRD (Confidence-Risk-Domain) scoring pipeline to ground governance decisions in reality:

Curated Facts: Verified facts across multiple categories (world leaders, current events, country facts, science, technology)
Content Hashing: Each fact is SHA-256 hashed for deduplication and integrity verification
Confidence Decay: Facts that are not re-confirmed lose confidence over time, preventing stale data from influencing governance
Post-Response Validation: Responses are validated against the Truth Store before delivery — incorrect claims are corrected at the sentence level

The Truth Store ensures that governance decisions are anchored to verified reality, not to LLM training data that may be outdated or incorrect.

Claims Ledger

Cryptographic Claim Provenance & Calibration

The Claims Ledger provides cryptographic provenance for every factual claim made through the governance pipeline:

Claim Registration: Every claim is registered with its evidence chain, source, and confidence level
"Prove It" Queries: Any claim can be challenged — the ledger returns the full evidence chain that supports it
Hash-Chain Integrity: Claims are linked via SHA-256 hash chain, making tampering detectable
Brier Score Calibration: Resolved claims feed back into calibration scoring, tracking how well-calibrated EVE's confidence levels are over time
Source Trust Weighting: Sources are tracked from VERIFIED (1.0) through UNTRUSTED (0.2), with trust scores influencing claim confidence

Calibration Feedback Loop: When claims are later confirmed or refuted, the Brier score is updated. This calibration data is used by the Trust Dial to adjust confidence thresholds across all governance domains.

Data Retention by Plan

Data retention varies by subscription tier:

Design Partner Pilot

90 days governance logs

Enforcement License

7-year signed audit retention

Enterprise Governance

7-year retention + custom compliance policies

Sovereign Infrastructure

Custom retention + on-premise control

Data Governance & Privacy

Your Data, Your Control: Organizations can view, export, or delete their governance data at any time through the Sovereign SDK API or account settings. Ledger records are retained for compliance but can be exported and purged per GDPR requirements.

Data Controls

View Governance Logs: Full transparency into all governance decisions and audit trails
Export Data: Download governance records, claims, and audit trails in JSON or CSV format
Selective Deletion: Remove specific episodic records (ledger entries are retained for compliance)
GDPR Compliance: Full data export and deletion workflows for regulatory compliance

How Data Architecture Supports Governance

The five data layers work together to provide a complete governance infrastructure:

Working context enables sub-millisecond governance decisions during request processing
Episodic audit provides full traceability for every governance decision
Semantic indexing enables fast policy lookup and precedent matching
Identity anchoring prevents silent policy drift through bounded change budgets
Ledger memory provides tamper-evident, regulatory-grade audit trails
Truth Store grounds decisions in verified facts, not stale training data
Claims Ledger ensures every factual claim has cryptographic provenance