All Insights
    EU AI ActAI GovernanceCompliance EngineeringAWS

    EU AI Act Articles 12, 13 and 14: A Technical Implementation Guide

    Vinita SilaparasettyAI Governance Engineer, Aevoxis Solutions10 minutesJune 2026

    Most discussions of EU AI Act compliance treat Articles 12, 13 and 14 as documentation work. Write a risk management policy. Keep logs. Put a human somewhere in the process. Tick the boxes.

    I see it differently. I treat compliance as architecture: governance logic built into the execution path of the system, not a policy layer sitting on top of a generative AI system.

    The Spec-Drift Chronometer shows what that looks like in practice. It is a working, open-source demonstration, built for the August 2, 2026 applicability deadline.

    Update, November 2026: When this article was written, the high-risk rules were due to apply on August 2, 2026. The Digital Omnibus on AI (Regulation (EU) 2026/1744, in force since 27 July 2026) moved them. Articles 12, 13 and 14 now apply from 2 December 2027 for Annex III high-risk systems. For high-risk AI in products covered by Annex I, they apply from 2 August 2028. The architecture described below is unchanged.

    What Is Spec-Drift and Why Does It Matter for EU AI Act Compliance?

    An autonomous AI system that generates outputs over time does something predictable. Its outputs start to move away from the original human-approved specification.

    Nothing is broken when this happens. Any system that sits between human intent and probabilistic inference behaves this way.

    I call this Spec-Drift: the measurable gap between what a human architect specified and what an autonomous system actually produces.

    In traditional software, code review, testing and deployment pipelines catch this kind of deviation. An autonomous AI system running in real time has none of those checks.

    It makes decisions, takes actions and changes state without a human seeing each output. By the time someone notices the drift, it has often compounded into something much harder to fix.

    This is exactly the scenario EU AI Act Article 14 was written for. The Article requires that AI systems allow human oversight to intervene, prevent, correct and override outputs.

    The regulation does not answer the technical question, though. How do you intervene in a system that runs faster than human review cycles?

    The Spec-Drift Chronometer is my answer to that question.

    What Do EU AI Act Articles 12, 13 and 14 Actually Require Technically?

    Most commentary on these articles stays at the policy level. Let's take a look at what they mean at the implementation level.

    What does EU AI Act Article 12 require?

    Article 12 requires high-risk AI systems to log events automatically throughout their operation. The logging must be good enough to audit the system's behavior after the fact.

    Article 12 asks for automatic logs over the system's lifetime, detailed enough to trace how it works (Art. 12(1) and (2)). It lists specific fields only for remote biometric identification systems (Art. 12(3)). In practice, I recommend logging the input that triggered a decision, the decision itself and enough context for a reviewer to reconstruct what happened and why.

    Why is it important? If logging is added after the system is built, the data needed to reconstruct a decision often does not exist. It was never retained.

    So design logging into the system architecture from the beginning. Logging cannot be an afterthought.

    What does EU AI Act Article 13 require?

    Article 13 requires high-risk AI systems to be transparent enough for deployers to interpret the output and use it appropriately (Art. 13(1)).

    In practice, I recommend giving more than an output. Give enough context about how that output was produced for the person overseeing the system to evaluate it.

    Note: Transparency has to show up in the interface, not only in technical files. A reasoning trace in a log file nobody reads is not transparency in the Article 13 sense.

    What does EU AI Act Article 14 require?

    Article 14 requires high-risk AI systems to be designed so that natural persons can effectively oversee them during operation.

    In particular, humans must be able to decide not to use the system in a particular situation, intervene in its operation and override its outputs.

    Note: Human oversight cannot be a passive capability. An override button that exists but is never triggered does not satisfy Article 14 in spirit.

    The oversight mechanism must be active. It has to sit in the critical path of the system's operation, not in a bypass lane beside it.

    Now that you know what the three articles ask for, you are ready to see how the Chronometer implements them.

    How Does the Spec-Drift Chronometer Implement These Requirements?

    The system covers all three articles with one architectural pattern: the Governance Loop.

    Human Architect
          |
          v
    Sovereign Spec Vault (.kiro/steering/)
          |
          v
    Warden Engine (FastAPI, AWS Lambda eu-central-1)
          |
          v
    Drift Detection (every 3 seconds)
          |
       [below threshold]          [above threshold]
          |                               |
          v                               v
    Audit Log                    Justification Gate
    (Article 12)                 (Article 14)
                                          |
                                 Human provides justification
                                          |
                                 Warden evaluates against spec
                                          |
                          [approved]                [rejected]
                               |                        |
                        Spec updated             Event logged
                        Audit trail              Audit trail
                        (Article 12)             (Article 12)
                        Reasoning trace          Reasoning trace
                        (Article 13)             (Article 13)

    Every component in this loop maps to a specific regulatory obligation. That is on purpose. The architecture was designed around the obligations, not the other way around.

    How Is the Human-Approved Specification Stored and Protected?

    The foundation of the system is the Sovereign Spec Vault: a set of human-authored specification files stored in .kiro/steering/.

    These files define the approved boundaries the autonomous system is allowed to operate within.

    .kiro/steering/
    ├── governance.md          # Warden persona and protocol
    ├── tech.md                # Technology constraints
    ├── product.md             # Vision and strategic boundaries
    ├── human-intent-specs.md  # Explicit architect declarations
    └── spec.json              # Machine-readable thresholds

    The spec.json file is the machine-readable source of truth:

    json
    {
      "control_parameters": {
        "semantic_drift_threshold": 0.0075,
        "max_latency_ms": 200,
        "primary_model": "amazon.nova-pro-v1:0",
        "audit_trail_active": true
      },
      "sovereign_region": "eu-central-1"
    }

    The drift threshold of 0.0075 is the value that matters most. When the measured gap between the system's current output and the approved specification goes above it, the Justification Gate activates.

    The vault is human-authored and version-controlled. The autonomous system cannot modify it.

    Why is it important? The specification is read-only for the AI system but editable by the human architect. That constraint is the structural foundation of Article 14 compliance.

    How Does Real-Time Drift Detection Satisfy Article 13?

    The Warden Engine polls a drift measurement endpoint every 3 seconds. The FastAPI backend exposes the drift state:

    python
    @app.get("/drift")
    async def get_drift_state():
        drift_value = calculate_drift()
        threshold = load_spec()["control_parameters"]["semantic_drift_threshold"]
    
        return {
            "drift_index": drift_value,
            "sovereign_threshold": threshold,
            "excess_delta": drift_value - threshold,
            "spec_compliance": "BREACH" if drift_value > threshold else "100%",
            "system_health": "CRITICAL_DRIFT" if drift_value > threshold else "SOVEREIGN",
            "warden_status": "GATE_ACTIVE" if drift_value > threshold else "Observing"
        }

    This response powers the real-time dashboard. A non-technical stakeholder can see:

    • the current drift index as a number
    • whether it is within the approved threshold
    • the system health status in plain language
    • the Warden's current operational state

    They do not need to understand how drift is calculated. The interface is designed so that people can make governance decisions without technical knowledge.

    This is Article 13 transparency surfaced through the interface, not buried in a log file.

    How Does the Justification Gate Enforce Article 14 Human Oversight?

    When drift exceeds the sovereign threshold, the system does not take corrective action on its own. It stops and asks for a human decision. This is the Justification Gate.

    The gate shows the human:

    • the drift value and how far it exceeds the threshold
    • the exact specification clause that was violated
    • the AI model that produced the non-compliant output
    • a text field for a written justification, which is required before the system proceeds

    The Warden Agent then evaluates the justification against three criteria:

    • Semantic relevance: Does the justification address the type of drift detected? A justification about database schema changes does not cover a drift in content generation behavior.
    • Specification cross-reference: Is the justification consistent with the specification clause that was violated?
    • Article 14 guardrails: Does the justification show that a human actually understood the deviation and made a conscious governance decision?

    Example: A weak justification such as "it is fine" or "I approved this" scores below the acceptance threshold. The Warden rejects it.

    The rejection is logged too, with the full reasoning trace. That gives you an Article 12-compliant record of the failed oversight attempt.

    A strong justification names the technical reason for the deviation and connects it to a legitimate architectural decision. The Warden approves it and updates the specification to reflect the new approved boundary.

    What Does an Article 12-Compliant Audit Trail Look Like?

    Every event in the Governance Loop is logged to a downloadable audit trail with a SHA-256 verification hash:

    ╔══════════════════════════════════════════════════════════════╗
    ║      SPEC-DRIFT CHRONOMETER — SOVEREIGN AUDIT TRAIL         ║
    ╚══════════════════════════════════════════════════════════════╝
    
    Timestamp:          2026-06-14 15:24:15 UTC
    Drift Index:        0.0082
    Threshold:          0.0075
    Gate Status:        RESOLVED
    Spec Hash:          bf40efdc39297d64
    Run Hash:           cdfa7ff9a941820f
    
    ── GOVERNANCE COMPLIANCE ──────────────────────────────────────
    EU AI Act Article 14 (Human Oversight):   VERIFIED
    EU AI Act Article 12 (Transparency):      VERIFIED
    Sovereign Region:                          eu-central-1 (Frankfurt)
    System Integrity:                          100%
    
    ── ACTIVE SPECIFICATIONS (.kiro/steering/spec.json) ───────────
    Semantic Drift Threshold:  0.15
    Max Latency (ms):          200
    Primary Model:             amazon.nova-pro-v1:0
    Audit Trail Active:        true
    
    ── JUSTIFICATION GATE RECORD ──────────────────────────────────
    Decision:         APPROVED
    Justification:    Migrating auth layer to OAuth2 to satisfy GDPR
                      Article 7 compliance requirements signed off
                      by legal team on 2026-06-10.
    
    ══════════════════════════════════════════════════════════════

    The SHA-256 hash of the specification at the time of the audit makes any later change to the specification detectable.

    Why is it important? A governance system whose records can be changed after the fact gives you no real compliance assurance.

    This goes beyond what Article 12 strictly requires. However, it is a practical concern that any serious compliance buyer should raise.

    The audit trail also includes the model version used for evaluation, the human justification word for word and the Warden's decision with full reasoning.

    In production deployments, the simulated Warden evaluation is replaced with live Amazon Bedrock Nova Pro reasoning in eu-central-1.

    What Are the Current Limitations of This Implementation?

    Honest documentation of a governance system has to include its limits. Here are the current ones.

    Drift measurement is simulated in the demo

    The live demo uses pre-scripted drift scenarios. It does not measure actual semantic deviation in a real AI system's outputs.

    The governance logic, the Justification Gate, the audit trail and the spec vault are fully functional. A production deployment connects the drift measurement endpoint to the output stream of the AI system being governed.

    The specification vault requires manual maintenance

    A human architect has to write the initial specification by hand. This is the set of human-approved boundaries that defines what counts as drift.

    In the Article 14 sense, this is a feature: the human is the source of the specification. But it also means the system cannot govern itself into existence.

    Rejection does not force-halt external systems in all scenarios

    When the Justification Gate rejects a justification, it logs the rejection and returns control to the dashboard.

    A hard stop on rejection needs integration with the live autonomous system. That design depends on the system being governed, so it is marked as Phase 2 in the roadmap.

    How to Run the Demo

    bash
    git clone https://github.com/VinitaSilaparasetty/spec-drift_chronometer.git
    cd spec-drift_chronometer
    DEMO_MODE=true ./dev.sh

    Open http://localhost:3000. You do not need AWS credentials in demo mode.

    The live demo is available at spec-drift-chronometer.aevoxis.de. For a production deployment with Amazon Bedrock and AWS Lambda eu-central-1, see the README in the repository.

    Now that the demo is running, you are ready for the next step: wiring the same gate into a production LangChain RAG chatbot. That is the next article: aevoxis.de/insights/eu-ai-act-langchain-rag-integration.

    Frequently Asked Questions

    Does the Spec-Drift Chronometer make an AI system fully EU AI Act compliant?
    No. It demonstrates a technical approach to the human oversight and logging requirements of Articles 12, 13 and 14 for autonomous AI systems. Full EU AI Act compliance for a high-risk system involves more. Depending on the system category, that includes risk management documentation, data governance, technical documentation and conformity assessment.
    Which types of AI systems does this governance pattern apply to?
    The Governance Loop pattern fits any autonomous AI system that generates outputs continuously without a human reviewing each output. This includes AI coding assistants, autonomous content generation systems, AI-driven process automation and agentic AI workflows. The implementation will vary by system. The pattern itself carries over across implementation languages and cloud providers: a human-authored specification vault, real-time drift measurement, a mandatory human checkpoint and a verifiable audit trail.
    Is the Justification Gate a hard block or can it be bypassed?
    In the current implementation the Justification Gate is a soft block. It requires a human justification before the system records the governance decision. It does not force-halt an external autonomous system. A hard block that stops a system until a valid justification arrives is marked as Phase 2. It would be necessary for production deployment in a high-risk AI system context.
    What is the difference between Article 14 human oversight and simply having a human approve outputs?
    Article 14 requires the human oversight mechanism to be effective, not merely present. A human who nominally approves outputs without meaningful information about what they are approving is not effective oversight. The Justification Gate asks for a documented rationale. The Warden evaluates it for semantic relevance to the specific drift detected. A rubber-stamp approval is rejected.
    Does the system store personal data?
    In demo mode the system stores no personal data. The audit trail captures the justification text the user enters, and in a production deployment that text may contain personal data. Production deployments should include a GDPR-compliant data retention policy for audit trail records and a data subject erasure endpoint. The production architecture includes that endpoint.
    Can this be adapted for non-AWS infrastructure?
    Yes. The Warden Engine is a FastAPI application, so it runs on any infrastructure that can run Python. You can replace the Bedrock integration for production Warden evaluation with any LLM provider API. The spec vault uses plain Markdown and JSON files that are infrastructure-agnostic. The audit trail is a generated text file with no cloud dependency.

    Is Your Generative AI System Ready for December 2, 2027?

    EU AI Act Article 14 applies to Annex III high-risk AI systems from December 2, 2027, following the Digital Omnibus on AI. If you deploy autonomous AI systems in Annex III high-risk categories, you need to demonstrate human oversight capability before that date. These categories include systems used in employment, education, credit, biometric identification or critical infrastructure. If you are assessing your organization's autonomous AI systems for Article 14 compliance and want a technical second opinion on your architecture, try the EU AI Act Technical Readiness Assessment at aevoxis.de. It takes about five minutes. You get a personalized response within 24 hours.

    Start the Assessment
    VS

    Vinita Silaparasetty

    AI Governance Engineer, Aevoxis Solutions

    Vinita Silaparasetty is an AI Governance Engineer working under the Aevoxis Solutions name as a member of SMartDE eG, Germany. Published author (Apress), Google Developer Expert (ML), M.Sc. Data Science Newcastle University (Russell Group). The Spec-Drift Chronometer was developed as an entry for the AWS 10,000 AIdeas Competition 2025, Top 300 Finalist from over 10,000 global submissions.

    Disclosure

    Parts of this article were written with AI writing assistance (Claude, Anthropic). All technical concepts, system architecture, implementation decisions, and intellectual content are original work by the author. The Spec-Drift Chronometer was designed, built, and documented by Vinita Silaparasetty.