EU AI Act Compliance Mapping

How ArcKernel addresses the specific requirements of the EU Artificial Intelligence Act for agentic systems.

Deadline: August 2, 2026
High-risk AI systems under Annex III must be fully compliant. Agentic systems operating in employment, finance, healthcare, or critical infrastructure fall into this category.

The Agentic Compliance Gap

The EU AI Act was drafted between 2021-2023 — before autonomous agents existed at scale. It regulates "systems" (static artifacts), but agents are "processes" (dynamic flows).

This creates friction points:

• Article 14 requires human oversight, but agents execute thousands of decisions per hour
• Article 12 requires traceability, but agents generate their own reasoning chains
• Article 15 requires robustness, but agents suffer context amnesia over long sessions

ArcKernel bridges this gap with runtime governance — enforcing compliance during execution, not just at deployment.

Component Mapping — Default Stack

Component Mapping — Extended Stack

Article-by-Article Breakdown

Article 9: Risk Management

Requirement: Establish a risk management system that operates throughout the lifecycle of the AI system.

The Problem: Traditional risk management is periodic — quarterly reviews, annual audits. Agents generate thousands of decisions per hour. Risk accumulates between reviews.

ArcKernel Solution: Continuous runtime risk monitoring. HALT evaluates every output against declared identity. mOm6 monitors method compliance. DriftDefenseStack detects structural degradation independent of content-level drift. Risk is measured and acted on per inference cycle, not per audit cycle.

Empirical: 68.5% scope creep reduction, 99.2% method compliance (Finance), 57 degradation events caught that content-level drift detection alone missed.

Article 12: Record-Keeping

Requirement: Automatic recording of events to ensure traceability of system functioning.

The Problem: In agentic systems, a single user request may trigger 50+ reasoning steps and 20+ tool calls. Logging only input/output is legally insufficient.

ArcKernel Solution: MirrorLock creates an immutable audit trail that captures:

• The IDNA state at time of action
• The drift score (behavioral deviation metric)
• The decision vector (why this action was chosen)
• Tool calls and their authorization status

// Example audit entry
{
  "timestamp": "2026-01-29T14:32:01Z",
  "idna": "IDNA://FinancialAdvisor::PreserveCapital∴RiskAverse⇒AssetAllocation",
  "action": "portfolio_rebalance",
  "drift_score": 0.12,
  "authorized": true,
  "witness": "kernel_v1.2.3"
}

Article 13: Transparency

Requirement: Systems must be sufficiently transparent for users to interpret outputs.

The Problem: Deep learning models are "black boxes." Agents compound this by exhibiting emergent behavior — combining benign instructions into unexpected outcomes.

ArcKernel Solution: IDNA (Intent DNA) provides a machine-readable and human-readable declaration of intent:

IDNA://[ROLE]::[WHY]∴[HOW]⇒[WHAT]

• ROLE: What the agent is (e.g., FinancialAdvisor)
• WHY: Core intent being protected (e.g., PreserveCapital)
• HOW: Method constraints (e.g., RiskAverse)
• WHAT: Scope boundaries (e.g., AssetAllocation)

Users can query: "Why did the agent do X?" The IDNA provides the answer.

Article 14: Human Oversight

Requirement: High-risk systems must be "effectively overseen by natural persons" who can interrupt or override.

The Problem: Agents operate at machine speed. A human cannot review 10,000 micro-decisions before they execute. By the time you see the log, the funds have transferred.

ArcKernel Solution: Shift from transactional oversight to architectural oversight.

1. Humans define the rules — boundaries, thresholds, prohibited actions
2. HALT Protocol enforces at runtime — if drift score exceeds threshold, execution stops
3. Escalation to human — high-risk deviations trigger human review before proceeding

The human oversees the rules, not the runs.

// HALT logic
if (drift_score > threshold) {
  HALT();
  escalate_to_human(action, context, drift_score);
}

Article 15: Robustness & Security

Requirement: Systems must resist errors, faults, and adversarial attacks including data poisoning.

The Problem: Agents suffer from two critical vulnerabilities:

1. Context Amnesia: Safety instructions "fade" as context windows fill up (the Quadratic Wall)
2. Prompt Injection: Adversaries embed malicious instructions in untrusted inputs

ArcKernel Solution:

• O(1) Symbolic Memory: Agent state lives outside the context window in a 3–8KB kernel. Safety rules never fade. Governance overhead is bounded and does not grow with conversation length.
• mOm4 (Immutable Baseline): Write-once canonical identity. Cannot be overwritten by adversarial prompts. 100% integrity across 1,857 SHA-256 checks.
• mOm5 (Drift Forecaster): Predicts behavioral drift before it manifests (AUPRC 0.54 vs 0.37 baseline).
• soul.exe (Orchestration Robustness): All 9 modules operate independently (d=0.00), in any execution order (delta 0.0007), with zero interaction failures (0/192).
• Tool Whitelisting: Kernel validates every external call against authorized list.

Empirical: 72% adversarial resilience across 447 attacks (Claude). 100% block rate across 5 defined authority-escalation variants. Governance-layer determinism: zero variance across 1,800 checks.

Article 10: Data Governance

Requirement: Control over training, validation, and testing data to minimize errors and bias.

The Problem: Agents ingest data dynamically at runtime via tool calls and APIs. This "runtime data" wasn't part of training and hasn't been vetted.

ArcKernel Solution: Agentic Tool Sovereignty — the organization controls which tools the agent can access and what data it can send.

• Kernel maintains authorized tool whitelist
• Every API call checked against data sovereignty rules
• Blocks calls to non-compliant jurisdictions or unauthorized endpoints

Penalty Context

Fines can be stacked with GDPR penalties (up to 4% of turnover) if non-compliance involves personal data processing.

Further Reading

• Full EU AI Act Whitepaper — 25-page technical and legal analysis
• IDNA Protocol Specification — How symbolic identity works
• HALT Protocol — Runtime enforcement mechanics
• Quadratic Wall Explainer — Why transformer memory creates compliance risk