RFC: EU AI Act Compliance Mapping — TealTiger v1.4 (Draft for Review)

[nagasatish007]

Your Idea

EU AI Act Compliance Mapping

TealTiger v1.4 — Runtime Governance for High-Risk Agentic AI Systems

---

Document Information

Field	Value
Document Version	DRAFT v0.1
Date	June 3, 2026
Author	TealTiger Engineering
Classification	DRAFT — For External Review
TealTiger Version	v1.4 (Q1 2027 target)
Regulation	Regulation (EU) 2024/1689 (EU AI Act)
Scope	High-risk systems under Annex III (use-based classification)

Compliance Timeline Note

The original deadline for high-risk obligations (Articles 8-17) is August 2, 2026. Under the Digital Omnibus provisional agreement (Council and Parliament, May 7, 2026), standalone Annex III systems are expected to receive a deferral to December 2, 2027. This agreement is pending formal adoption. If not adopted before August 2, 2026, the original timeline applies.

This mapping is designed to satisfy requirements under either timeline.

---

Executive Summary

TealTiger is an open-source AI agent security platform that provides deterministic governance, guardrails, cost tracking, and policy management for LLM-based agentic applications. It operates in-process with less than 5ms overhead, requires no external services, and produces cryptographically verifiable evidence of governance decisions.

This document maps TealTiger's v1.4 capabilities to the EU AI Act's high-risk system obligations (Articles 8-17, 19, 26, 27, 73), identifying coverage, partial coverage, and gaps for each requirement.

Coverage Summary

Coverage Level	Articles
Full Coverage	Art. 9, 12, 13, 14, 15
Partial Coverage	Art. 8, 10, 11, 17
Requires External Complement	Art. 16, 26, 27, 73

---

Article-by-Article Mapping

Article 8 — General Requirements for High-Risk AI Systems

Obligation: High-risk AI systems shall be designed and developed in accordance with Articles 9-15.

TealTiger Coverage: PARTIAL

TealTiger provides the runtime enforcement layer that ensures Articles 9-15 obligations are satisfied during operation. It does not cover the design-time and pre-market requirements (conformity assessment, CE marking, quality management system design). TealTiger is complementary to design-time tools.

Requirement	TealTiger Module	Status
Compliance with Art. 9-15 during operation	TealEngine + all modules	Covered
Design-time conformity	N/A (design-time tool scope)	External

---

Article 9 — Risk Management System

Obligation: A risk management system shall be established, implemented, documented, and maintained throughout the entire lifecycle of the high-risk AI system.

TealTiger Coverage: FULL

Requirement	TealTiger Module	How
Continuous risk identification	TealDrift (v1.3)	Statistical behavioral drift detection against baselines
Risk estimation and evaluation	TealDrift + TealMonitor	Risk scoring per action, per agent, per session
Adoption of risk management measures	TealEngine (policy enforcement)	Deterministic policy rules map to identified risks
Testing for risk mitigation	Policy Test Harness (v1.1)	Automated testing of policy rules against risk scenarios
Residual risk monitoring	TealCircuit + TealRunawayGuard (v1.4)	Circuit breakers, loop detection, spending velocity alerts
Lifecycle documentation	TealAudit + TEEC receipts	Continuous evidence production throughout operation

Evidence Artifacts: TEEC execution receipts, TealDrift anomaly reports, Policy Test Harness results (JUnit XML, SARIF)

---

Article 10 — Data and Data Governance

Obligation: Training, validation, and testing datasets shall be subject to appropriate data governance and management practices.

TealTiger Coverage: PARTIAL

TealTiger governs data at runtime (what data flows through agents) but does not manage training data governance.

Requirement	TealTiger Module	How
Data provenance tracking	TealMemory (v1.2)	Provenance-tagged memory governance with transitive trust
Data quality monitoring at runtime	PIIDetectionGuardrail + TealSecrets	Detect PII, secrets, and anomalous data in agent I/O
Bias detection in runtime data	Not covered	External tool scope (training-time concern)
Input data traceability	TEEC v2.1 input_reference field	Content-addressable hash linking inputs to outputs

Gap: Training data governance (dataset bias, representativeness, labeling quality) is outside TealTiger's runtime scope. Requires complementary tools (e.g., data lineage platforms).

---

Article 11 — Technical Documentation

Obligation: Technical documentation shall be drawn up before the AI system is placed on the market or put into service and shall be kept up to date.

TealTiger Coverage: PARTIAL

TealTiger auto-generates runtime technical documentation from governance evidence. It does not generate pre-market design documentation.

Requirement	TealTiger Module	How
Description of system capabilities	TealRegistry (tool/model catalog)	Registered tools, models, and capabilities with hashes
Risk management documentation	BundleExporter (SARIF, JUnit XML)	Exportable evidence bundles per compliance period
Description of monitoring measures	TealAudit + Compliance Report Generator	Auto-generated reports mapping to regulatory controls
Runtime performance documentation	observe() + assess CLI	Behavioral baselines, cost metrics, anomaly reports

Gap: Pre-market documentation (system architecture descriptions, intended purpose statements, training methodology) requires complementary documentation tools.

---

Article 12 — Record-Keeping

Obligation: High-risk AI systems shall technically allow for the automatic recording of events (logs) over the lifetime of the system.

TealTiger Coverage: FULL (with v1.4 additions)

This is TealTiger's strongest coverage area. The TEEC v2.1 evidence contract is specifically designed for this requirement.

Requirement	TealTiger Module	How
Automatic event recording	TealAudit (always-on in observe mode)	Every tool call, decision, and model response logged
Identification of risks	TealDrift + action sequence scoring	Behavioral anomalies and kill chain patterns detected
Tamper-evidence	TEEC signing extension (Ed25519 + JCS)	Cryptographically signed, independently verifiable receipts
Verifiable timestamps	TEEC issued_at field (ISO 8601)	Every receipt timestamped at creation
Independent verifiability	verifier_contract_version + JWK Set	Competent authority can verify without trusting the provider
Session lifecycle coverage	TEEC session.start / session.end events	Full operational lifetime tracking (v1.4 addition)
Input-output traceability	TEEC input_reference / output_reference	Content-addressable hashes linking I/O (v1.4 addition)
Retention	retention_until field + configurable policy	Minimum 6 months with EU AI Act pack enabled (v1.4 addition)

Art. 12 Receipt-to-Requirement Mapping

Art. 12 Property	TEEC Field	Notes
"Automatic recording of events"	All TEEC events via TealAudit	Zero-config in observe mode
"Over the lifetime of the system"	session.start / session.end lifecycle events	Session brackets all decisions
"Traceability of the AI system's operation"	correlation_id, session_id, request_id	Full request lineage
"Input data that led to a given output"	input_reference (SHA-256 hash)	Content-addressable, not raw data
"Identification of risks or anomalies"	TealDrift events, action sequence alerts	Anomaly receipts linked to session

Known Gaps (identified by Sebastian Tagwercher):

1. Exact retention duration: Art. 19/26 implies 6-month minimum but regulatory guidance is unsettled. TealTiger uses configurable retention policy (default: 6 months with EU AI Act pack).
2. prEN ISO/IEC 24970 (event logging standard) is still in draft. When finalized, a teec.iso24970 extension adapter will map TEEC fields to the standard's format.

---

Article 13 — Transparency and Provision of Information to Deployers

Obligation: High-risk AI systems shall be designed and developed in such a way as to ensure that their operation is sufficiently transparent to enable deployers to interpret the system's output and use it appropriately.

TealTiger Coverage: FULL

Requirement	TealTiger Module	How
Explainable decisions	TEEC reason field + policy_refs	Every governance decision includes human-readable reason and policy reference
Operator understanding of output	TEEC intent, execution_outcome, stop_reason	Answers "what happened?" in under 10 seconds
Information about capabilities and limitations	TealRegistry + assess CLI report	Tool catalog, capability boundaries, risk assessment
Deployment guidance	FDE Playbook documentation	Standard deployment methodology with progressive disclosure

Evidence Artifacts: TEEC receipts with reason, stop_reason, retry_posture fields; assessment reports from tealtiger assess

---

Article 14 — Human Oversight

Obligation: High-risk AI systems shall be designed and developed in such a way as to allow for effective human oversight during the period in which the AI system is in use.

TealTiger Coverage: FULL

Requirement	TealTiger Module	How
Human-in-the-loop for high-risk actions	TealApproval (REQUIRE_APPROVAL decision)	Pauses execution, waits for human sign-off
Ability to intervene and override	freeze() kill switch	Immediate agent termination, zero policy required
Real-time monitoring	observe() + GovernanceDashboard	Live event feed, cost tracking, anomaly alerts
Decision to not use or interrupt	freeze() + TealCircuit	Kill switch + circuit breakers for cascading failures
Human override evidence	TealApproval evidence (who approved, when, what they saw)	Full audit trail of human interventions

Evidence Artifacts: TEEC receipts with decision: REQUIRE_APPROVAL, approval evidence records, freeze-block audit events

---

Article 15 — Accuracy, Robustness, and Cybersecurity

Obligation: High-risk AI systems shall be designed and developed in such a way that they achieve an appropriate level of accuracy, robustness, and cybersecurity.

TealTiger Coverage: FULL

Note: Article 15 cybersecurity requirements deferred to December 2, 2027 under Digital Omnibus (ref: tagwercher.io/writing/eu-ai-act-article-15-deferral/). The obligation itself is unchanged; only the enforcement date moved.

Requirement	TealTiger Module	How
Robustness against errors/faults	TealCircuit (circuit breakers) + TealReliability (retry budgets, fallback chains)	Cascading failure prevention
Cybersecurity — adversarial inputs	TealGuard (prompt injection detection) + TealToolIntegrity (v1.4)	Deterministic defense against injection and tool poisoning
Cybersecurity — unauthorized access	TealGrant (per-identity scoped access) + freeze()	Fine-grained authorization without external token service
Cybersecurity — model manipulation	TealRegistry (definition drift detection) + TealDrift (behavioral monitoring)	Detect tool/model changes and behavioral anomalies
Resilience against agent loops	TealRunawayGuard (v1.4)	Loop detection, spending velocity, per-trace token budgets

Evidence Artifacts: TealCircuit state records, TealToolIntegrity violation events, TealDrift anomaly reports, TealRunawayGuard alerts

---

Article 17 — Quality Management System

Obligation: Providers of high-risk AI systems shall put in place a quality management system that ensures compliance with this Regulation.

TealTiger Coverage: PARTIAL

TealTiger provides the runtime enforcement and evidence layer of a quality management system. It does not define the organizational QMS (policies, procedures, personnel responsibilities).

Requirement	TealTiger Module	How
Systematic examination of known/foreseeable risks	TealDrift + Policy Test Harness	Continuous risk monitoring + automated policy testing
Techniques for testing and monitoring	Policy Test Harness + observe() baselines	JUnit XML export for CI/CD integration
Design of governance controls	TealEngine policies + TEEC evidence	Deterministic policy enforcement with verifiable receipts
Post-market monitoring	observe() + assess CLI + GovernanceDashboard	Continuous operational monitoring

Gap: Organizational QMS design (reporting structures, personnel roles, incident management procedures) requires complementary governance framework.

---

Article 19 — Quality Management System Documentation (Retention)

Obligation: Providers shall keep the documentation at the disposal of the national competent authorities for a period of 10 years or, where specific sectoral legislation provides for a longer retention period, for that longer period.

TealTiger Coverage: INFRASTRUCTURE SUPPORT

Requirement	TealTiger Module	How
Documentation retention	retention_until field in TEEC + TealAudit configurable policy	Receipts carry retention metadata; storage layer enforces

Note: The 10-year provider obligation applies to technical documentation (Art. 11), not necessarily to all operational logs. The interplay between Art. 12 log retention (unsettled, 6-month floor) and Art. 19 documentation retention (10 years) requires regulatory clarification. TealTiger supports configurable retention policies that can be adjusted once guidance is finalized.

---

Article 26 — Obligations of Deployers

Obligation: Deployers shall use high-risk AI systems in accordance with the instructions of use and implement appropriate technical and organisational measures.

TealTiger Coverage: PARTIAL (deployer responsibility)

TealTiger is the technical measure that deployers use to comply. It does not define organizational measures.

Requirement	TealTiger Module	How
Technical measures for oversight	observe() + TealEngine + freeze()	Zero-config to full governance progression
Monitor operation based on instructions	GovernanceDashboard + assess CLI	Continuous monitoring aligned to provider instructions
Record events for traceability	TealAudit (automatic)	Always-on event recording
Inform provider of serious incidents	TEEC events exportable to provider	Structured evidence for incident reporting

---

Open Questions for Reviewer

The following questions are for regulatory review:

1. Art. 12 retention period: Given the unsettled state (6-month minimum under Art. 19/26 vs. potential longer figures), what retention default would you recommend we ship in the EU AI Act policy pack? Is "6 months, configurable" a defensible default?

2. Input-output traceability depth: Art. 12 qualifies with "where technically feasible." For agentic AI systems processing thousands of requests per hour, is content-addressable hashing (storing the hash in the receipt, raw content in a separate customer-managed store) sufficient? Or do regulators expect inline content in the log?

3. prEN ISO/IEC 24970 timeline: Any visibility on when this standard will be finalized? Should we wait for it before publishing the compliance pack, or ship with a mapping note that says "will add ISO/IEC 24970 adapter when finalized"?

4. Multi-agent coverage: Art. 12 references "the AI system." For multi-agent deployments where several agents collaborate on a task, do you expect regulators to require per-agent logs, per-system aggregate logs, or both?

5. Annex I "safety component" narrowing: You mentioned watching the narrowing on the Annex I side. Is there any risk that agentic AI embedded in products (e.g., AI agent in a medical device workflow) gets re-classified under the narrowed definition? Relevant because TealTiger supports both standalone and product-embedded agent deployments.

---

Architecture Constraints

All TealTiger compliance modules operate under these non-negotiable constraints:

Constraint	Description
In-process	No external service dependency. Runs inside the application.
Less than 5ms overhead	Deterministic evaluation. No LLM in governance path.
Zero network hop	No proxy, no sidecar, no SaaS dependency.
Offline capable	Works in air-gapped environments with no internet.
Open source	Apache 2.0 license. Full source available for audit.

These constraints ensure that compliance infrastructure does not introduce additional attack surface, latency, or operational dependencies.

---

Next Steps

1. Reviewer feedback on this mapping (open questions above)
2. Alignment with prEN ISO/IEC 24970 once published
3. Draft EU AI Act policy pack implementation (TealTiger module configuration)
4. Formal receipt-to-Article-12 field mapping with test corpus
5. Publication as part of TealTiger v1.4 documentation

---

Document End

TealTiger — Open Source AI Agent Security Platform
Apache 2.0 License | github.com/agentguard-ai/tealtiger
NVIDIA Inception Member | OWASP Contributor

Category

Improvement

What problem does this solve?

No response

Potential Impact

No response

Implementation Thoughts

No response

Examples or References

No response