EU AI Act Compliance — Azure Maritime Intelligence

Product: Azure AI Weather (Azure Maritime Intelligence V2) Document version: 1.4 Last reviewed: 2026-07-16 Contact: contact@azureai.pt

Operator: see Operator below.

Governance index: docs/README.md · AI System Card · Model Registry · Versioning Policy · Model Changelog

Plain Language Summary

• Azure AI Weather is an AI-assisted decision-support system.
• It does not control vessels, engines or autopilots.
• It never replaces the skipper.
• GO / CAUTION / NO-GO are advisory recommendations.
• Weather forecasts are probabilistic and conditions may change rapidly.
• The final navigation decision always belongs to the skipper.
• Human oversight is always present.
• Azure is currently documented as a proportionate limited-risk AI system and is reviewed whenever major features are added.

For everyday users

This document is written primarily for transparency and regulatory documentation (auditors, partners, and internal governance).

For day-to-day use you only need to understand that:

• Azure AI Weather provides advisory support.
• It does not replace official weather services.
• It does not replace skipper judgement.
• Always cross-check critical decisions with official forecasts and safe seamanship.

The full sections below keep the technical and compliance detail required for AI Act alignment.

Executive Summary

Azure Maritime Intelligence is a maritime decision-support application for recreational and professional skippers. It combines weather forecasts, fused marine data, deterministic risk scoring, and optional large-language-model (LLM) narration. The application does not control vessels, autonomously navigate, or issue mandatory operational commands. The skipper remains solely responsible for every navigation and safety decision.

This document records our proportionate alignment with the EU Artificial Intelligence Act (Regulation (EU) 2024/1689). Based on current functionality, Azure is not classified as a High-Risk AI System under Annex III for the use cases implemented today. Obligations are applied where relevant — especially transparency (Article 50) and technical documentation (Article 11) — without unnecessary bureaucracy.

Estimated AI Act readiness: Moderate–Good for a limited-risk / general-purpose decision-support product. Core deterministic logic is documented; transparency measures have been strengthened in UI and prompts; formal conformity assessment is not required at this classification level.

Table of contents

1. Plain Language Summary
2. For everyday users
3. Executive Summary
4. Operator
5. Human Override Principle
6. AI Limitations
7. AI system description
8. Risk classification (EU AI Act)
9. Architecture
10. Confidence Score
11. Component inventory
12. Black Box
13. Data sources
14. Model inventory
15. AI Lifecycle
16. VISUS historical navigation pattern analysis
17. Human oversight
18. Risk assessment (proportionate)
19. Transparency measures (implemented)
20. Limitations (user-facing summary)
21. Monitoring process
22. Governance Review
23. Version management
24. Future compliance roadmap
25. References
26. Revision history

Revision history

Operator

Diogo Botelho Moniz

Trading as

Azure AI Weather

A FrontSea Intelligence product

Portugal

Contact: contact@azureai.pt

Additional legal identity details (address, NIF) appear in the Privacy Policy and Terms of Service.

Human Override Principle

This principle underpins Skipper, Black Box, chat, Skipper Insight, Old Salt, VISUS historical navigation pattern analysis, and Historical Navigation Insight. No output constitutes a navigation order, steering command, or mandatory route.

AI Limitations

1. AI system description

1.1 Intended purpose

• Provide coastal and river sailing weather intelligence (wind, sea, tide, pressure, patterns).
• Offer an advisory GO / CAUTION / NO-GO planning layer derived from an Operational Risk Index.
• Record operational context during voyages (Black Box) for replay, debrief, and manual model development.
• Collect structured Field Check feedback from Founders cohort skippers to improve forecast alignment.
• Answer maritime weather questions via an AI assistant (home chat, Skipper Insight, Old Salt replay commentary).
• Provide Historical Navigation Insight — situational awareness from historical AIS and similar-weather patterns (never navigation instructions).

1.2 What the system does not do

1.3 Primary users

• Recreational and semi-professional skippers (Free, Skipper, Pro tiers).
• Founders validation cohort (Field Check).
• Internal development team (manual VISUS navigation model training).

2. Risk classification (EU AI Act)

2.1 High-risk assessment

Conclusion: Not high-risk under current deployment, subject to periodic review.

Annex III high-risk categories commonly cited for maritime systems include autonomous transport and safety-critical control. Azure:

• Does not perform autonomous navigation or traffic management as a safety component of transport.
• Does not issue binding operational decisions to machinery.
• Presents outputs as decision support with human oversight and override.

If future features add autonomous routing commands, COLREGS automation, or safety-critical control without human confirmation, re-classification would be required.

2.2 Applicable obligation areas

2.3 GPAI / foundation model note

Generative text uses HuggingFace Inference API with Qwen/Qwen2.5-7B-Instruct. Azure does not fine-tune or deploy this model; it is constrained by system prompts, context injection, and output guards (core/security/prompt_guard.py). Provider terms and EU GPAI obligations are primarily HuggingFace's; Azure documents usage and limitations here.

3. Architecture

flowchart TB
    subgraph inputs [Data inputs — non-autonomous]
        IPMA[IPMA / national forecasts]
        OWM[OpenWeather marine]
        ERA5[ERA5 K-means pattern models]
        IOT[Ecowitt station Azure_Station_V1]
        TIDE[Tide models]
    end

    subgraph deterministic [Deterministic core — no LLM]
        DG[Decision graph]
        OR[Operational Risk 0–1]
        CONF[Confidence 0–1]
        GNG[GO / CAUTION / NO-GO advisory]
    end

    subgraph generative [Generative layer — LLM narration]
        CHAT[Home chat assistant]
        INSIGHT[Skipper Insight briefing]
        OLDSALT[Old Salt replay commentary]
    end

    subgraph human [Human oversight]
        SKIPPER[Skipper decision authority]
        BB[Black Box confirm / override]
        FC[Field Check ground truth]
    end

    inputs --> DG --> OR --> GNG
    OR --> CONF
    GNG --> INSIGHT
    OR --> CHAT
    GNG --> BB
    SKIPPER --> BB
    FC --> GT[(founders_ground_truth.jsonl)]
    BB --> LOG[(blackbox JSONL)]

3.1 Confidence Score

Confidence represents the quality and consistency of available information (coverage, forecast horizon, source agreement). It is shown alongside Operational Risk in Skipper.

Technical detail: OPERATIONAL_RISK_METHODOLOGY.md · CONF-001 in MODEL_REGISTRY.md · confidence/confidence_engine.py.

When confidence is below ~0.45, the GO/NO-GO layer may record a low_confidence factor so users treat the advisory with extra caution — still without changing the meaning of Operational Risk itself.

4. Component inventory

4.1 AI / ML components (summary)

4.2 Non-AI components (deterministic / data)

4.3 Black Box — operational recording & replay

The Black Box is an operational recording and replay system.

Authoritative policy: blackbox/datasets.py · AI_SYSTEM_CARD.md — Black Box recording policy.

5. Data sources

Personal data (accounts, chat memory for Pro) is handled under standard privacy practices; voyage logs may contain location tracks — see deployment data layout (deploy/PERSISTENT_DATA.md).

6. Model inventory

See MODEL_REGISTRY.md for the complete registry (LLM-001, ERA5-PT/GL, STORM-001, RISK-001, CONF-001, VISUS shared models) including version, status, validation, and oversight per model.

6.1 AI Lifecycle

Every AI model follows the governance lifecycle below. No AI model is deployed directly into production. Every production model is validated before release.

Research
    ↓
Validation
    ↓
Internal Testing
    ↓
Founder Beta
    ↓
Production
    ↓
Continuous Monitoring
    ↓
Periodic Review
    ↓
Retirement

Related: VERSIONING_POLICY.md · MODEL_CHANGELOG.md · §12 Governance Review.

6.2 VISUS historical navigation pattern analysis

VISUS AI models used in Black Box are documented as:

Earlier internal wording such as “Speed/heading hints” is replaced by historical navigation pattern analysis.

What VISUS historical models do

• Analyse historical AIS observations
• Compare similar weather situations
• Identify statistical navigation patterns
• Provide situational awareness only

What they do not do

• Calculate mandatory routes
• Generate steering commands
• Replace skipper judgement

Historical Navigation Insight (Black Box card) surfaces confidence, observation counts, similar-weather classification, and historical traffic density. Optional corridor highlighting represents a statistical density zone, not a route to follow. Wording is descriptive (e.g. “Historical AIS data indicates that many similar vessels navigated through the highlighted corridor”) and never prescriptive.

This feature does not modify GO / CAUTION / NO-GO logic, Operational Risk calculations, or weather fusion. It does not control navigation, optimise routes, or generate steering commands.

7. Human oversight

8. Risk assessment (proportionate)

9. Transparency measures (implemented)

9.1 User interface

• Chat: AI disclosure + forecast limitation disclaimer (frontend/shared/locale.js).
• Skipper: Advisory notice on GO/CAUTION/NO-GO pill; operational risk + confidence visible; Skipper Insight labeled as AI-generated briefing.
• Black Box: Operational recording / replay system — not an AI decision engine or autonomous navigator.
• Historical Navigation Insight (Black Box): Permanent Historical AI Insight label; info icon disclosure that patterns are derived from historical AIS and similar environmental conditions for situational awareness only; never framed as a navigation instruction.
• Field Check: Documented as ground-truth validation for quality improvement.

9.2 LLM system prompts

Shared rules in core/ai_act_advisory.py injected into chat and Skipper Insight prompts:

• Decision-support only; no orders or guarantees.
• Cautious, probabilistic language.
• Encourage official forecast cross-check.
• Skipper retains full responsibility.

9.3 Output guards

• guard_user_prompt() — injection resistance.
• guard_llm_output() — blocks instruction leakage; redirects to authority consultation.
• Insight coherence guards — align narrative with deterministic GO status and wind/sea context.

9.4 Documentation

• docs/README.md — governance index
• AI_SYSTEM_CARD.md — system card (limitations, oversight, lifecycle)
• MODEL_REGISTRY.md — model inventory
• VERSIONING_POLICY.md — semver, rollback, deprecation
• MODEL_CHANGELOG.md — change history
• This file (AI_ACT_COMPLIANCE.md)
• OPERATIONAL_RISK_METHODOLOGY.md
• meteorologia_base/ YAML doctrine (forecast limits, risk communication)
• blackbox/datasets.py data governance policy

10. Limitations (user-facing summary)

1. Forecasts are probabilistic — not guarantees; conditions can change rapidly.
2. Spatial resolution — grid models may miss local harbour effects.
3. Advisory status — GO/CAUTION/NO-GO is a planning envelope, not permission to sail.
4. LLM narration — may omit nuance; numbers in chat/insight must match backend context when guards succeed; always verify critical decisions against official sources.
5. Coverage — optimized for Iberian Atlantic / Tagus operations; other regions may have thinner ground truth.
6. IoT station — point measurement; not a substitute for route-wide assessment.
7. Confidence — low confidence means greater uncertainty in available information, not automatically “dangerous conditions.”
8. Historical patterns — statistical AIS/weather context only; never navigation instructions.

See also the AI Limitations callout above.

11. Monitoring process

12. Governance Review

AI documentation is reviewed:

• Annually
• Whenever major AI functionality changes
• Whenever regulations change
• Before commercial public releases

Review outputs typically include updates to this file, the AI System Card, Model Registry, and Model Changelog as applicable.

13. Version management

See VERSIONING_POLICY.md for semantic versioning, dataset versioning, model artifact rules, rollback, and deprecation.

Quick reference:

• Application: 2.0.0 (core/app_factory.py)
• Risk thresholds: skipper/go_nogo.py (git-versioned)
• LLM: HUGGINGFACE_CHAT_MODEL env (default Qwen/Qwen2.5-7B-Instruct)
• Change log: MODEL_CHANGELOG.md

14. Future compliance roadmap

15. References

• Regulation (EU) 2024/1689 — Artificial Intelligence Act
• docs/README.md — governance documentation index
• AI_SYSTEM_CARD.md
• MODEL_REGISTRY.md
• VERSIONING_POLICY.md
• MODEL_CHANGELOG.md
• OPERATIONAL_RISK_METHODOLOGY.md
• deploy/README.md — deployment and access tiers
• blackbox/datasets.py — voyage data governance
• meteorologia_base/previsao/limites_previsao.yaml — forecast limit doctrine
• meteorologia_base/risco_e_decisao/avaliacao_risco.yaml — risk communication doctrine

*This document is internal governance and transparency material. It does not constitute legal advice. Consult qualified counsel for formal regulatory opinions.*