Proposed sovereign AI assurance initiative · Andorra

AI decisions need a memory.

Andorra AI Safe Haven is a proposed independent evidence and replay infrastructure designed to help organisations prove what an AI system knew, what it was asked, what it did and why.

Currently in design and stakeholder consultation. Government endorsement, regulatory recognition and certification are objectives to explore and should not be represented as existing approvals.

The core business question

AI can make decisions. Can you reconstruct them?

Modern AI applications can involve models, agents, prompts, documents, databases, APIs, policies, retrieval systems and human intervention.

Can you reconstruct the state of information, rules, models and evidence that existed when an AI decision was made?

  • Q01

    What information was available to the AI?

  • Q02

    What did the organisation believe to be true at that moment?

  • Q03

    Which model and model version were used?

  • Q04

    Which system prompt and user prompt were active?

  • Q05

    Which documents or RAG chunks were retrieved?

  • Q06

    Which tools and APIs were invoked?

  • Q07

    Which business policies and controls applied?

  • Q08

    Was a human involved, and did they approve, reject or override the output?

  • Q09

    Has the evidence changed since the event?

Logs tell you that something happened. Safe Haven is designed to reconstruct the state in which it happened.

Bitemporal thinking

There is more than one kind of time.

Valid Time
When something was true in the real world.

Transaction Time
When the organisation's systems knew or recorded it.

  1. 10 March

    Customer's actual status changes to LOW RISK.

  2. 14 March

    AI makes a decision while the system still shows HIGH RISK.

  3. 18 March

    Correct information reaches the organisation.

At the decision point

Reality
LOW RISK
System knowledge
HIGH RISK

An AI decision should be understood against the information and rules actually available when the decision occurred.

What was true? — Valid Time

PastAI DecisionFuture

What was known? — Transaction Time

Safe Haven preserves both.

Architecture

An independent evidence plane for AI.

A proposed database-neutral temporal evidence service positioned alongside AI systems rather than inside them.

Layer 01

AI Applications

  • LLMs
  • Agents
  • ML
  • RAG
  • Automated Decisions

Layer 02

Safe Haven Connect

  • API
  • SDK
  • Gateway
  • Events

Layer 03

AI Evidence Envelope

  • Data
  • Model
  • Prompt
  • RAG
  • Tools
  • Policies
  • Identity
  • Human Intervention
  • Output
  • Action

Layer 04

Temporal Replay Engine

  • Valid Time
  • Transaction Time
  • Corrections
  • Version History
  • Temporal Relationships

Layer 05

Immutable Evidence Vault

  • Cryptographic Integrity
  • Digital Signatures
  • Tamper Evidence
  • Retention Policies

Layer 06

Audit & Replay

  • Historical Reconstruction
  • Investigation
  • Compliance
  • Counterfactual Analysis
  • Authorised Third-Party Access

Three forms of replay

Replay is not one capability.

  • Recorded Replay

    What actually happened?

    Display the evidence captured around the original AI event without rerunning it.

  • Historical Reconstruction

    What did the system know at that time?

    Reconstruct relevant data, policy, model, prompt, retrieval and knowledge state.

  • Counterfactual Replay

    What would happen using what we know now?

    Re-evaluate historical decisions using corrected data, changed policy or a different model.

Exact computational reproducibility may depend on the continued availability and deterministic behaviour of external models and services. Safe Haven therefore distinguishes recorded evidence from computational re-execution.

AI Evidence Envelope

Capture the context, not simply the output.

  • Model

    Provider, model family, exact version or checkpoint where available, and configuration.

  • Prompt

    System prompt, user prompt, template and template version.

  • Knowledge

    Documents, vector results, database queries and source records.

  • Data

    Relevant business data with its temporal state.

  • Agents

    Agent identity, task, hand-offs and workflow metadata.

  • Tools

    APIs, functions and external services invoked.

  • Policy

    Business rules, guardrails and regulatory controls active at the time.

  • Identity

    Person, service or AI process initiating the activity.

  • Human Oversight

    Approval, escalation, rejection or override by a person.

  • Decision

    Model output plus the actual downstream business action.

  • Integrity

    Hashes, signatures and evidence-chain verification.

Cryptographic trust

Evidence should be verifiable, not merely stored.

The proposed architecture is intended to support cryptographically protected evidence capable of showing whether historical records have been altered.
  • Hardware Security Modules (HSMs)
  • Protected signing keys
  • Digital signatures
  • Immutable / WORM storage
  • Evidence hashes
  • Hash chains or Merkle structures
  • Trusted timestamping
  • Strong administrator separation and role segregation

Architecture target: support for FIPS 140-3 Level 3 hardware-backed cryptographic key protection where required. Deployed FIPS compliance is not claimed until the final architecture and validated components are in service.

AI systems

Where evidence and replay matter most.

  • Automated Decision Systems

    Reconstruct why an automated decision occurred.

  • AI Agents

    Follow agent actions and tool calls across workflows.

  • RAG Systems

    Identify which source content was supplied to a model.

  • Regulated AI

    Maintain evidence for governance, compliance and investigation.

  • Human + AI Decisions

    Capture where responsibility transitioned between system and person.

  • Model Migration

    Compare outcomes before and after model changes.

  • Data Corrections

    Identify historical decisions potentially affected by incorrect data.

Financial services · second use case

Built from a problem banking has understood for decades.

Gartner's 2008 research identified temporal data management as a significant banking problem and reported client estimates that temporal handling could represent 20%-40% of application development and transformation costs.
  • Credit Decisions

    What information and policy applied when credit was granted or declined?

  • Fraud & Financial Crime

    What information was available when a transaction or customer was flagged?

  • Pricing

    Which rates, balances, customer state and pricing rules applied?

  • Payments

    Which rules and system state existed when an authorisation or risk decision occurred?

  • Customer Records

    Distinguish when something was true from when the bank learned it.

  • AI in Banking

    Create evidence around AI-assisted financial decisions.

Financial services detail

A roadmap vertical on the common temporal evidence platform, not a delivered product at launch.

Trust framework

From software to a trust framework.

Five proposed assurance levels.
  • 01

    Recorded

    A defined evidence record exists for AI events.

  • 02

    Reconstructable

    Historical information and system state can be reconstructed.

  • 03

    Verifiable

    Cryptographic controls provide evidence-integrity assurance.

  • 04

    Independently Auditable

    Authorised independent parties can interrogate evidence.

  • 05

    Sovereign Assured

    Deployment satisfies defined jurisdictional, governance and infrastructure requirements.

The final Safe Haven assurance framework would be developed with appropriate governmental, regulatory, legal, technical and industry stakeholders.

Jurisdiction

Why Andorra?

Andorra has the opportunity to establish itself as a trusted jurisdiction for responsible adoption of artificial intelligence and digital infrastructure. Its scale, proximity between government, financial institutions and business, established financial-services sector and European position create an environment in which a sovereign AI assurance initiative could potentially be developed collaboratively.

Strategic engagement

Government, banking or technology stakeholder?

We are currently developing the Safe Haven architecture and exploring potential pilot, policy, infrastructure and funding partnerships.