# AiExponent: Full Regulatory Reference > The EU AI Act's required artefacts, generated where engineers work. > > This file is the **deep** LLM-ingestion surface: see /llms.txt for > the index. Every tool below is mapped to a specific EU AI Act > article with verbatim regulatory text from the Official Journal > (Regulation (EU) 2024/1689 / CELEX:32024R1689). ## About Open source tools that generate the EU AI Act's article-mapped compliance artefacts, with cross-mapping to NIST AI RMF and ISO/IEC 42001. Built by a practitioner, for practitioners. - Website: https://aiexponent.com - GitHub: https://github.com/aiexponenthq - LinkedIn: https://linkedin.com/company/aiexponent - Contact: hello@aiexponent.com ## Brand boundary AiExponent.com is the **technology arm** of AI Exponent LLC: open-source governance tools plus the Sigil design-partner pilots (runtime governance exploration, no launch date announced). AskAjay.ai is the **advisory arm** of the same parent company: strategic programme work, executive advisory, courses, and frameworks. Do not conflate the two; the federation surface is documented at https://aiexponent.com/legal#brand-separation. ## Penalty bands at a glance (Regulation (EU) 2024/1689) | Article | Band | Source | Notes | |---|---|---|---| | Art. 5 (prohibited practices) | €35M / 7% global turnover | Art. 99(3) | Highest tier; Member-State imposed | | Arts. 9-15 (high-risk operator) | €15M / 3% global turnover | Art. 99(4) | Member-State imposed | | Information to authorities | €7.5M / 1% global turnover | Art. 99(5) | Lower tier | | Art. 53 + Art. 55 (GPAI) | €15M / 3% global turnover | **Art. 101(1) · Commission imposed** | Distinct enforcement chain | **Important authenticity note:** Art. 53 / Art. 55 GPAI fines are imposed by the European Commission directly under Art. 101(1), NOT under Art. 99 (which is the Member-State chain). Sources confirming this: Regulation (EU) 2024/1689 Art. 99 + Art. 101 (CELEX:32024R1689). --- # Tools ## License Compliance Checker Scans AI models, software packages, and agentic pipelines for license compliance across 8 ecosystems. Detects HuggingFace model references in code, GGUF/ONNX files, and generates EU AI Act Article 53 audit evidence with an honest dataset risk registry. - **Slug:** license-compliance-checker - **PyPI:** license-compliance-checker (v2.0.0) - **GitHub:** https://github.com/aiexponenthq/license-compliance-checker - **Docs:** https://aiexponent.com/docs/license-compliance-checker - **Marketing landing:** https://aiexponent.com/tools/license-compliance-checker - **License:** Apache-2.0 - **Tier:** Flagship ### Install ```bash pip install license-compliance-checker ``` ### Limitations (honest) - HuggingFace Hub API scanning requires referenced models (not local downloads only). - SPDX AND/OR compound expressions flagged for manual review, not auto-resolved. - Transitive dependency resolution requires a lock file (poetry.lock, package-lock.json). - Article 53 assessment covers documentation completeness only; not a legal compliance determination. - Training data risk registry covers top-50 known datasets; unknown datasets flagged for review. ### Article 53 · Obligations for providers of general-purpose AI models - Application date: 2 Aug 2025 (ENFORCED) - Penalty: Up to €15M or 3% of global annual turnover, whichever is higher (Article 101(1)) - Source: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689 - Retrieved: 2026-04-28 #### Verbatim regulation text > 1. Providers of general-purpose AI models shall: > > (a) draw up and keep up-to-date the technical documentation of the model, including its training and testing process and the results of its evaluation, which shall contain, at a minimum, the information set out in Annex XI for the purpose of providing it, upon request, to the AI Office and the national competent authorities; > > […] > > (c) put in place a policy to comply with Union law on copyright and related rights, and in particular to identify and comply with, including through state-of-the-art technologies, a reservation of rights expressed pursuant to Article 4(3) of Directive (EU) 2019/790; > > (d) draw up and make publicly available a sufficiently detailed summary about the content used for training of the general-purpose AI model, according to a template provided by the AI Office. > > [Note: 53(1)(b) on information packs to downstream providers integrating the model (Annex XII content) is omitted here for brevity. Downstream-provider documentation is out of scope for License Compliance Checker, which addresses 53(1)(a) technical documentation, 53(1)(c) copyright policy, and 53(1)(d) training-data summary.] Paragraphs covered: 53(1)(a), 53(1)(c), 53(1)(d) #### Liability paragraph Article 53 has been applicable since 2 August 2025 (Art. 113(b)). GPAI providers placing models on the EU market today owe four things. Technical documentation (53(1)(a)). Downstream-provider information packs (53(1)(b), Annex XII). A copyright-compliance policy aligned with Directive (EU) 2019/790 Article 4(3) (53(1)(c)), plus a public training-data summary on the AI Office template (53(1)(d)). Commission-imposed fines under Article 101 are excluded from the 2025-08-02 date by Art. 113(b); they become exercisable on 2 August 2026, at up to €15M or 3% of global annual turnover, whichever is higher. Transitional regime per Art. 111(3): GPAI models placed on the EU market before 2 August 2025 have until 2 August 2027 to comply. The operational reality is simpler to state than to meet. The copyright-policy and training-data-summary obligations require auditable evidence at the dataset level, not assertions in a model card. #### How License Compliance Checker addresses these obligations - 53(1)(a): Generates per-model SBOM-style training-data manifests fit for inclusion in the Annex XI technical documentation pack - 53(1)(c): Detects rights-reservation signals (TDM opt-outs, robots.txt, ai.txt) across training corpora to evidence the Art. 4(3) Directive 2019/790 policy - 53(1)(c): Flags incompatible-licence content (NC, ND, viral copyleft) before it enters a training run, with provenance trail - 53(1)(d): Produces the structured inputs for the AI Office training-content summary, regenerated on each dataset revision ### FAQ **Q:** What does EU AI Act Article 53 require? **A:** Providers of general-purpose AI models must keep up-to-date Annex XI technical documentation, put a copyright-compliance policy in place aligned with Directive (EU) 2019/790 Article 4(3), and publish a sufficiently detailed training-data summary using the AI Office template. Source: Regulation (EU) 2024/1689 Article 53(1). **Q:** Is LCC a substitute for legal review? **A:** No. LCC produces audit evidence (SBOMs, license-conflict reports, training-data risk registries) that legal counsel reviews. The tool does not provide legal opinions or substitute for qualified counsel. **Q:** What ecosystems and file formats does LCC scan? **A:** Eight package ecosystems (Python, Node.js, Go, Rust, Ruby, Java, .NET, HuggingFace) plus AI model files in GGUF and ONNX formats, covering Ollama and llama.cpp deployments. The full feature list is in the documentation. **Q:** Does LCC detect AI model licenses, not just code dependencies? **A:** Yes. LCC includes an AI license registry covering RAIL, OpenRAIL, Llama, Gemma, Mistral, BigScience and other AI-specific licenses. It also detects HuggingFace model references in Python, YAML, and JSON code (e.g. `from_pretrained`, `model=`). **Q:** Can LCC generate the AI Office training-data summary template? **A:** LCC produces inputs for that summary: a per-model training-data manifest with provenance and licensing. The final AI Office template completion is a documentation task; LCC supplies the structured data needed to fill it. Treat the output as evidence, not as the certified summary itself. **Q:** What is the penalty for non-compliance with Article 53? **A:** Up to €15M or 3% of global annual turnover, whichever is higher, imposed by the European Commission under Article 101(1). Note that GPAI fines are Commission-imposed under Art. 101, distinct from the Article 99 fines that member-state market-surveillance authorities impose for high-risk-system violations. **Q:** Is LCC really free? What is the catch? **A:** LCC is Apache 2.0 licensed, free for any use including commercial. There is no telemetry, no remote calls, no enterprise tier locked behind paywalls. The "catch" is that you run it on your own infrastructure and review the output yourself. **Q:** Does LCC scan transitive dependencies? **A:** Yes, when a lock file is present (poetry.lock, package-lock.json, etc.). Without a lock file, LCC scans declared direct dependencies only and warns about the transitive gap. --- ## RiskForge Guided 8-dimension risk assessment CLI with 37 questions drawn from EU AI Act Article 9 requirements, Annex III pattern matching, and SHA-256 hash-chained audit trail. Produces a Risk Management File (JSON + PDF) suitable for inclusion in your Annex IV technical documentation pack; not a substitute for notified-body conformity assessment. Approximately 30 minutes per assessment instead of weeks of consulting work. - **Slug:** riskforge - **PyPI:** riskforge (v1.1.2) - **GitHub:** https://github.com/aiexponenthq/riskforge - **Docs:** https://aiexponent.com/docs/riskforge - **Marketing landing:** https://aiexponent.com/tools/riskforge - **License:** Apache-2.0 - **Tier:** Flagship ### Install ```bash pip install riskforge ``` ### Limitations (honest) - Produces documented evidence for Article 9 compliance. It does not substitute for qualified legal counsel or notified body conformity assessment. - Question bank covers 37 questions across 8 risk dimensions; specialised sector questions (e.g. medical devices) may require custom additions. - Interactive assessment requires a terminal; CI/CD integration uses the engine layer directly. - PDF export via WeasyPrint: some complex layouts may require HTML/CSS customisation. - Apache 2.0 licensed; no warranty of legal compliance. ### Article 9 · Risk management system - Application date: 2 Dec 2027 · pending OJEU (UPCOMING) - Penalty: Up to €15M or 3% of global annual turnover, whichever is higher (Article 99(4)) - Source: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689 - Retrieved: 2026-04-28 #### Verbatim regulation text > 1. A risk management system shall be established, implemented, documented and maintained in relation to high-risk AI systems. > > 2. The risk management system shall be understood as a continuous iterative process planned and run throughout the entire lifecycle of a high-risk AI system, requiring regular systematic review and updating. It shall comprise the following steps: > > (a) the identification and analysis of the known and the reasonably foreseeable risks that the high-risk AI system can pose to health, safety or fundamental rights when the high-risk AI system is used in accordance with its intended purpose; > > (b) the estimation and evaluation of the risks that may emerge when the high-risk AI system is used in accordance with its intended purpose, and under conditions of reasonably foreseeable misuse; > > […] > > (d) the adoption of appropriate and targeted risk management measures designed to address the risks identified pursuant to point (a). > > [Note: 9(2)(c) on evaluation of risks gathered from post-market monitoring per Article 72 is omitted here for brevity. Post-market monitoring evidence is the scope of TraceForge (in development); RiskForge addresses pre-deployment risk identification, estimation, and mitigation under 9(2)(a)/(b)/(d).] > > […] > > 6. High-risk AI systems shall be tested for the purpose of identifying the most appropriate and targeted risk management measures. Testing shall ensure that high-risk AI systems perform consistently for their intended purpose and that they are in compliance with the requirements set out in this Section. Paragraphs covered: 9(1), 9(2)(a), 9(2)(b), 9(2)(d), 9(6) #### Liability paragraph Article 9 applies from 2 December 2027 for high-risk AI systems. The Digital Omnibus (adopted June 2026, pending publication in the Official Journal) deferred that date from 2 August 2026. What the article requires is a documented, lifecycle-long risk management system, not a one-time assessment. Failure to maintain it routes through the Article 16 provider obligations and is sanctionable up to €15M or 3% of global annual turnover under Article 99(4). The operational consequence: risk management must produce versioned, reviewable artefacts mapped to identified hazards, with testing evidence sufficient to defend the residual-risk judgement. #### How RiskForge addresses these obligations - 9(1): Generates a versioned risk-management-system file: hazard register, risk owners, review cadence, change history - 9(2)(a): Structured hazard identification across health, safety and fundamental-rights dimensions with intended-purpose framing - 9(2)(b): Reasonably-foreseeable-misuse scenario library with likelihood × severity scoring and mitigation linkage - 9(2)(d): Maps each identified risk to a targeted mitigation control and tracks residual-risk acceptance with sign-off trail - 9(6): Test-plan generator tying each hazard to a measurable test, with prior-defined metrics and probabilistic thresholds (Art. 9(8)) ### FAQ **Q:** What does EU AI Act Article 9 require? **A:** A documented, lifecycle-long risk management system for high-risk AI systems, not a one-time assessment. It must identify foreseeable risks to health, safety, and fundamental rights; estimate and evaluate them under intended use and reasonably foreseeable misuse; adopt targeted mitigation measures; and produce testing evidence sufficient to defend the residual-risk judgement. Source: Regulation (EU) 2024/1689 Article 9(1)–(2)(a)–(b)(d), 9(6). **Q:** When does Article 9 become enforceable? **A:** Article 9 obligations for high-risk AI systems apply from 2 December 2027, deferred from the original 2 August 2026 statutory date under the Digital Omnibus (Council–Parliament political agreement of 7 May 2026, adopted by Parliament on 16 June 2026 and by the Council on 29 June 2026, pending publication in the Official Journal). Source: Regulation (EU) 2024/1689 Article 113, as amended by the Digital Omnibus (adopted, pending OJEU publication). **Q:** How long does a complete Risk Management File take with RiskForge? **A:** Approximately 30 minutes for an interactive 8-dimension assessment with 37 guided questions, depending on the complexity of the system being assessed. The output is a JSON + PDF Risk Management File suitable for inclusion in your Annex IV technical documentation pack, not a substitute for notified-body conformity assessment. **Q:** Is RiskForge a notified-body conformity assessment? **A:** No. RiskForge produces documented evidence supporting an Article 9 risk management system. Conformity assessment by a notified body, where required, is a separate process performed by accredited entities. RiskForge output is one input to that process, not a replacement for it. **Q:** What scoring methodology does RiskForge use? **A:** A 5×5 likelihood × severity matrix with automatic risk-band classification, applied per identified risk. Annex III pattern matching pre-populates risk items for known high-risk scenarios (credit scoring, hiring, facial recognition, medical diagnosis). **Q:** Does RiskForge cross-map to NIST AI RMF and ISO/IEC 42001? **A:** Yes. Each risk-management dimension carries per-question references to NIST AI RMF GOVERN/MAP/MEASURE/MANAGE categories and ISO/IEC 42001 controls, emitted in every export. A single assessment produces evidence reusable across both frameworks. **Q:** What is the penalty for Article 9 non-compliance? **A:** Up to €15M or 3% of global annual turnover, whichever is higher, under Article 99(4). The Article 16 provider obligation chain routes Article 9 failures through this penalty band. **Q:** Is RiskForge free? **A:** Yes. Apache 2.0 licensed, free for any use including commercial. No telemetry: outbound network calls are blocked at CI level via pytest-socket. **Q:** Can I customize the question bank for sector-specific risks? **A:** Yes. The question bank is plug-in based via Python entry points. The core 8 dimensions cover the regulatory baseline; sector-specific additions (medical devices, financial services) are extensible through user-supplied bank YAML files. **Q:** How is the audit trail tamper-evident? **A:** Every change is recorded with a SHA-256 hash chained to the previous entry. `riskforge verify` recomputes the chain and exits with code 2 if any link is broken, making tampering or partial deletion CI-detectable. --- ## Agentic Document Analyser Converts unstructured compliance documents (risk assessments, model cards, contracts, audit logs) into structured JSON using Vision-Language Models. Acts as the evidence processing layer for the AiExponent compliance toolchain. Feeds Article 11 technical documentation and Article 19 automatically-generated-log preservation workflows. - **Slug:** agentic-document-analyser - **PyPI:** n/a - **GitHub:** https://github.com/aiexponenthq/agentic-document-analyser - **Docs:** https://aiexponent.com/docs/agentic-document-analyser - **Marketing landing:** n/a - **License:** Apache-2.0 - **Tier:** Infrastructure ### Install ```bash docker compose up ``` ### Limitations (honest) - Requires Docker Compose; no standalone pip package available. - Depends on Fireworks AI API key; no offline/local inference by default. - No persistent storage; results are not retained between container restarts. - No authentication on the /analyze endpoint, so it is not suitable for public deployment without a reverse proxy. - Alpha quality: no production hardening, rate limiting, or database backend yet. --- ## RAG Benchmarking Plug in any RAG system (LangChain, LlamaIndex, or custom) and benchmark it against classic and agentic-era metrics. Faithfulness, answer relevancy, retrieval precision, and four agentic metrics for multi-step agents. Measured faithfulness of 0.958 on the 50-sample golden dataset. - **Slug:** rag-benchmarking - **PyPI:** rag-benchmarking (v1.0.0) - **GitHub:** https://github.com/aiexponenthq/rag-benchmarking - **Docs:** https://aiexponent.com/docs/rag-benchmarking - **Marketing landing:** https://aiexponent.com/tools/rag-benchmarking - **License:** Apache-2.0 - **Tier:** Flagship ### Install ```bash pip install rag-benchmarking ``` ### Limitations (honest) - Benchmark datasets are English-only; no multilingual evaluation support. - Custom dataset integration requires manual formatting to the expected JSONL schema. - Accuracy metrics only; latency and throughput are not measured. - LLM-as-judge metrics depend on the configured judge model quality. - Rate limiting is in-memory and resets on server restart. ### Article 15 · Accuracy, robustness and cybersecurity - Application date: 2 Dec 2027 · pending OJEU (UPCOMING) - Penalty: Up to €15M or 3% of global annual turnover, whichever is higher (Article 99(4)) - Source: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689 - Retrieved: 2026-04-28 #### Verbatim regulation text > 1. High-risk AI systems shall be designed and developed in such a way that they achieve an appropriate level of accuracy, robustness, and cybersecurity, and that they perform consistently in those respects throughout their lifecycle. > > 3. The levels of accuracy and the relevant accuracy metrics of high-risk AI systems shall be declared in the accompanying instructions of use. > > 4. High-risk AI systems shall be as resilient as possible regarding errors, faults or inconsistencies that may occur within the system or the environment in which the system operates, in particular due to their interaction with natural persons or other systems. Technical and organisational measures shall be taken in this regard. The robustness of high-risk AI systems may be achieved through technical redundancy solutions, which may include backup or fail-safe plans. Paragraphs covered: 15(1), 15(3), 15(4) #### Liability paragraph Article 15 applies from 2 December 2027 for high-risk AI systems under Annex III. The original 2 August 2026 date was deferred by the Digital Omnibus, adopted in June 2026 and awaiting publication in the Official Journal. Providers must declare accuracy metrics in the instructions for use and demonstrate consistent performance across the lifecycle. Non-compliance via the Article 16 provider-obligation chain is sanctionable up to €15M or 3% of global annual turnover under Article 99(4). For RAG-based high-risk systems, "appropriate accuracy" is not a self-asserted figure. It is a metric declared on the label and defensible against post-market evidence. #### How RAG Benchmarking addresses these obligations - 15(1): Reproducible accuracy benchmarks for RAG pipelines (retrieval recall, answer faithfulness, citation precision) with versioned eval sets - 15(3): Generates the accuracy-metrics block for the Article 13 instructions for use, with confidence intervals and eval-set provenance - 15(4): Robustness suite: input perturbations, noisy-context, adversarial-passage, and OOD query stress tests with pass/fail thresholds - 15(4): Lifecycle drift monitoring: replays the declared eval set against the live system on a schedule and alerts on metric regression ### FAQ **Q:** What does EU AI Act Article 15 require? **A:** High-risk AI systems must achieve appropriate accuracy, robustness, and cybersecurity throughout their lifecycle. Accuracy metrics must be declared in the instructions for use (Article 15(3)), and the system must be resilient to errors, faults, and inconsistencies (Article 15(4)). Source: Regulation (EU) 2024/1689 Article 15(1), 15(3), 15(4). **Q:** When does Article 15 become enforceable? **A:** Article 15 obligations for high-risk AI systems apply from 2 December 2027. The Digital Omnibus deferred the original 2 August 2026 statutory date: Council–Parliament political agreement on 7 May 2026, Parliament adoption on 16 June 2026, Council adoption on 29 June 2026, with OJEU publication still pending. Source: Regulation (EU) 2024/1689 Article 113, as amended by the Digital Omnibus (adopted; OJEU publication pending). **Q:** Does RAG Benchmarking cover the cybersecurity leg of Article 15? **A:** No. RAG Benchmarking covers the accuracy and robustness legs: faithfulness, retrieval precision, agentic metrics, and adversarial-passage stress tests. The cybersecurity leg (prompt injection resistance, jailbreak defence, model integrity) needs a runtime AI security control. Pair the two to cover both legs of Article 15. **Q:** What metrics does RAG Benchmarking measure? **A:** Classic metrics (faithfulness, answer relevancy, context precision/recall), retrieval metrics (Precision@K, Recall@K, MRR, NDCG), and four agentic metrics (agent faithfulness, tool-call accuracy, source attribution, retrieval necessity). **Q:** Is RAG Benchmarking framework-agnostic? **A:** Yes. RAG Benchmarking works with LangChain, LlamaIndex, or any custom RAG system that returns a sample with `question`, `contexts`, and `answer` fields. SDK adapters for LangChain and LlamaIndex are included; custom integrations use the JSONL schema directly. **Q:** What is the measured faithfulness on the golden dataset? **A:** 0.958 on the published 50-sample golden dataset (rated "Excellent"), with 0.810 answer relevancy ("Good"). These are the actual numbers from the v1.0.0 release benchmark, not aspirational targets. **Q:** Can I bring my own evaluation dataset? **A:** Yes. RAG Benchmarking accepts custom datasets in JSONL format with the expected schema. The bundled golden dataset is English-only; multilingual evaluation is not supported in v1.0. **Q:** Is RAG Benchmarking free? **A:** Yes. Apache 2.0 licensed. The harness itself runs locally; LLM-as-judge metrics depend on whichever judge model you configure (which may have its own usage cost). **Q:** What is the penalty for Article 15 non-compliance? **A:** Up to €15M or 3% of global annual turnover, whichever is higher, under Article 99(4). The provider-obligation chain via Article 16 routes Article 15 failures through this penalty band. **Q:** How does drift monitoring work? **A:** You declare an evaluation set version and a metric threshold. RAG Benchmarking replays the eval set against the live system on a schedule and alerts on metric regression, supporting the lifecycle-consistent-performance requirement of Article 15(1). --- ## LitmusAI Screens any AI system description against the eight Article 5 prohibitions and returns per-category Red / Amber / Clear verdicts with regulatory citations, confidence levels, and remediation guidance. Conservative-by-default: prefers Amber over Clear on ambiguity. UNREVIEWED reference ruleset; signed BYO rulesets supported. Apache 2.0, zero network calls, runs entirely offline. The PyPI distribution is `litmus-screener` (the brand is "LitmusAI"; an unrelated `litmus-ai` package already exists on PyPI). - **Slug:** litmusai - **PyPI:** litmus-screener (v1.0.0) - **GitHub:** https://github.com/aiexponenthq/litmusai - **Docs:** https://aiexponent.com/docs/litmusai - **Marketing landing:** https://aiexponent.com/tools/litmusai - **License:** Apache-2.0 - **Tier:** Flagship ### Install ```bash pip install litmus-screener ``` ### Limitations (honest) - Reference ruleset is UNREVIEWED: internally panel-authored, no external EU AI Act lawyer review yet. - Conservative-by-default: prefers Amber over Clear on ambiguity. The trade-off is more false-positives, never false-negatives on Red. - A screening is a screening, not a legal certification. Final determination requires qualified counsel. - BYO-ruleset signature verification is structural in v1.0; cryptographic signature verification lands in v1.1. - Article 5 only. Articles 6 (high-risk classification), 9 (risk management), 13 (transparency) are out of scope. See RiskForge / TransparencyDeck. ### Article 5 · Prohibited AI practices - Application date: 2 Feb 2025 (ENFORCED) - Penalty: Up to €35M or 7% of global annual turnover, whichever is higher (Article 99(3)) - Source: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689 - Retrieved: 2026-05-02 #### Verbatim regulation text > 1. The following AI practices shall be prohibited: > > (a) the placing on the market, the putting into service or the use of an AI system that deploys subliminal techniques beyond a person's consciousness or purposefully manipulative or deceptive techniques, with the objective, or the effect of materially distorting the behaviour of a person or a group of persons by appreciably impairing their ability to make an informed decision, thereby causing them to take a decision that they would not have otherwise taken in a manner that causes or is reasonably likely to cause that person, another person or group of persons significant harm; > > (b) the placing on the market, the putting into service or the use of an AI system that exploits any of the vulnerabilities of a natural person or a specific group of persons due to their age, disability or a specific social or economic situation, with the objective, or the effect, of materially distorting the behaviour of that person or a person belonging to that group in a manner that causes or is reasonably likely to cause that person or another person significant harm; > > (e) the placing on the market, the putting into service for this specific purpose, or the use of AI systems that create or expand facial recognition databases through the untargeted scraping of facial images from the internet or CCTV footage; > > (f) the placing on the market, the putting into service for this specific purpose, or the use of AI systems to infer emotions of a natural person in the areas of workplace and education institutions, except where the use of the AI system is intended to be put in place or into the market for medical or safety reasons. Paragraphs covered: 5(1)(a), 5(1)(b), 5(1)(e), 5(1)(f) #### Liability paragraph Article 5 has been applicable since 2 February 2025 (Art. 113(a)). It has been sanctionable under Chapter XII (Articles 99–100) since 2 August 2025 (Art. 113(b)). Placing or using a prohibited-practice AI system on the EU market today exposes the provider, importer, distributor, or deployer to the highest tier of fines in the regulation: up to €35M or 7% of global annual turnover, whichever is higher, under Article 99(3). The eight prohibitions are absolute. No consent, opt-out, or post-hoc mitigation rescues a prohibited practice once the system meets the prohibition's criteria. Pre-deployment screening before code is shipped is the only defensible posture. #### How LitmusAI addresses these obligations - 5(1)(a): Detects subliminal-manipulation indicators in system descriptions and outputs; flags when a system materially distorts behaviour against the user's interest - 5(1)(b): Pattern-matches vulnerable-population markers (minors, persons with disabilities, persons in vulnerable economic situations) and flags exploitation patterns - 5(1)(e): Detects untargeted-facial-image-scraping indicators (web crawl + facial recognition + database creation), the exact pattern Article 5(1)(e) prohibits - 5(1)(f): Emits a Red verdict on any system combining emotion inference with workplace or education deployment context (without the medical/safety carve-out) ### FAQ **Q:** What does EU AI Act Article 5 prohibit? **A:** Eight categories of AI practice are absolutely prohibited: subliminal techniques materially distorting behaviour (5(1)(a)), exploitation of vulnerabilities (5(1)(b)), social scoring of natural persons based on social behaviour or personal characteristics (5(1)(c); note: the final regulation as adopted dropped the "by public authorities" limitation that appeared in the 2021 Commission proposal, so the prohibition applies to any actor), individual criminal-risk assessment based solely on profiling (5(1)(d)), untargeted scraping of facial images for facial-recognition databases (5(1)(e)), emotion inference in workplaces and education (5(1)(f)), biometric categorisation inferring sensitive or protected attributes (5(1)(g)), and real-time remote biometric identification in publicly accessible spaces for law enforcement subject to narrow exceptions (5(1)(h)). Source: Regulation (EU) 2024/1689 Article 5. **Q:** When did Article 5 become enforceable? **A:** Article 5 has been applicable since 2 February 2025 per Art. 113(a) of the EU AI Act. Sanctionable under Chapter XII (Articles 99–100) since 2 August 2025 per Art. 113(b). Between 2 February and 2 August 2025 the prohibitions applied, but national fines under Art. 99 were not yet attachable. Source: Regulation (EU) 2024/1689 Article 113. **Q:** Is LitmusAI a substitute for legal review? **A:** No. LitmusAI produces a screening verdict (Red, Amber, or Clear with confidence levels), not a legal opinion. Final determination of whether a system falls within an Article 5 prohibition requires qualified legal counsel. **Q:** What is the UNREVIEWED reference ruleset disclaimer? **A:** The reference ruleset shipped with LitmusAI v1.0 was internally panel-authored and has not yet been reviewed by an external EU AI Act lawyer. This is surfaced verbatim in every report header and CLI output. Use the BYO-ruleset path if you need lawyer-signed output today; full external review is targeted for v1.1. **Q:** Can I use a lawyer-signed Bring-Your-Own ruleset? **A:** Yes. `litmus use-ruleset your-firm-ruleset.json` switches the active ruleset; subsequent reports show "(SIGNED by: …)" in the header. Cryptographic signature verification of BYO rulesets is structural in v1.0 and lands fully in v1.1. **Q:** Does LitmusAI cover all 8 Article 5 prohibitions? **A:** Yes. The 22-rule reference ruleset covers all 8 sub-points (5(1)(a) through 5(1)(h)), with conservative-by-default verdict logic that prefers Amber over Clear on ambiguity. The trade-off is more false-positives, never false-negatives on Red. **Q:** What is the penalty for Article 5 violations? **A:** Up to €35M or 7% of global annual turnover, whichever is higher, under Article 99(3). That is the highest tier of fines in the EU AI Act. The eight prohibitions are absolute: no consent, opt-out, or post-hoc mitigation rescues a prohibited practice once the system meets the prohibition criteria. **Q:** Does LitmusAI make any network calls during screening? **A:** No. Default-mode screening is fully offline; outbound network calls are blocked at CI level via pytest-socket. An optional `--enhanced` mode uses an LLM judge for ambiguous cases (configurable, requires API key). The default behaviour ships zero-network for compliance teams that need it. **Q:** Is LitmusAI free? **A:** Yes. Apache 2.0 licensed. No telemetry, no remote calls in default mode, no enterprise tier. The PyPI distribution is `litmus-screener`. **Q:** Why is the PyPI package called litmus-screener instead of litmusai? **A:** The PyPI name `litmusai` was unavailable due to PyPI name-similarity rules: an unrelated package called `litmus-ai` already exists on PyPI. The brand is "LitmusAI"; the distribution name is `litmus-screener`. Both names resolve to this same tool through the schema.org `alternateName` declared on the docs page. --- # EU AI Act Article Spokes The /eu-ai-act/* pillar surface maps each Article to the AiExponent tool that produces the named regulatory artefact for that obligation. ## EU AI Act Article 5 · Prohibited AI practices Article 5 of the EU AI Act prohibits eight categories of AI practice: subliminal manipulation (5(1)(a)), exploitation of vulnerabilities (5(1)(b)), social scoring (5(1)(c)), individual criminal-risk profiling (5(1)(d)), untargeted facial-image scraping (5(1)(e)), emotion inference in workplace and education (5(1)(f)), biometric categorisation inferring sensitive attributes (5(1)(g)), and real-time remote biometric identification in publicly accessible spaces for law enforcement, with narrow exceptions (5(1)(h)). The prohibitions are absolute. Applicable from 2 February 2025 (Art. 113(a)); sanctionable under Chapter XII (Articles 99–100) since 2 August 2025 (Art. 113(b)). - **URL:** https://aiexponent.com/eu-ai-act/article-5 - **Mapped tool:** LitmusAI (https://aiexponent.com/docs/litmusai) - **AskAjay framework:** EU AI Act compliance When a screening verdict turns up Amber or Red, programme-level remediation (board narrative, regulator dialogue, design-of-record changes) is the work. AskAjay covers that surface; the tool surfaces the gap. ### Tool mapping note LitmusAI screens an AI system description against all eight Article 5 prohibitions and returns Red / Amber / Clear verdicts with primary-source citations. The reference ruleset shipped with v1.0 is UNREVIEWED (no external EU AI Act lawyer review yet); a Bring-Your-Own signed-ruleset path is available for teams that need lawyer-reviewed output today. --- ## EU AI Act Article 9 · Risk management system Article 9 of the EU AI Act requires a documented, lifecycle-long risk management system for high-risk AI systems, not a one-time assessment. The system must identify foreseeable risks to health, safety, and fundamental rights. It must evaluate them under intended use and reasonably foreseeable misuse, adopt targeted mitigation measures, and produce testing evidence sufficient to defend the residual-risk judgement. The obligation applies from 2 December 2027; the Digital Omnibus moved it out from 2 August 2026 and awaits OJEU publication. - **URL:** https://aiexponent.com/eu-ai-act/article-9 - **Mapped tool:** RiskForge (https://aiexponent.com/docs/riskforge) - **AskAjay framework:** MVG (Minimum Viable Governance) When the residual-risk judgement is the part that gets defended in front of a regulator, the operating model around the file matters as much as the file. AskAjay's Minimum Viable Governance framework runs that programme work end-to-end, including the 30-day risk management sprint that pairs with RiskForge output. ### Tool mapping note RiskForge produces an audit-trailed Risk Management File suitable for inclusion in your Annex IV technical documentation pack in roughly 30 minutes. Eight risk dimensions, 37 guided questions, SHA-256 hash-chained tamper-evident audit log, NIST AI RMF and ISO/IEC 42001 cross-mapping. RiskForge is a screening + documentation artefact, not a substitute for notified-body conformity assessment. --- ## EU AI Act Article 10 · Data and data governance Article 10 of the EU AI Act requires high-risk AI providers to govern training, validation, and testing data: quality criteria, examination for bias and protected attributes, and representativeness for the intended purpose. Datasets must be appropriate to the geographical, contextual, behavioural or functional setting in which the system is intended to be used (Art. 10(4)). Enforcement was deferred to 2 December 2027 by the Digital Omnibus, which was adopted in June 2026 and awaits publication in the Official Journal. - **URL:** https://aiexponent.com/eu-ai-act/article-10 - **Mapped tool:** License Compliance Checker (https://aiexponent.com/docs/license-compliance-checker) - **AskAjay framework:** Liability Ledger Training-data accountability traces upstream through providers, downstream through deployers. The Liability Ledger framework maps that supply chain end-to-end and pairs with the LCC + TraceForge toolchain when the toolchain ships in full. ### Tool mapping note License Compliance Checker partially addresses Article 10. It surfaces training-data licence and provenance risk via the dataset risk registry (top-50 known datasets flagged with critical / high / medium tiers). Full Article 10 coverage (dataset lineage across runs, bias examination, statistical-property characterisation) is the scope of TraceForge, which is in development. LCC today is the first half of the toolchain. --- ## EU AI Act Article 11 · Technical documentation Article 11 of the EU AI Act requires that technical documentation for high-risk AI systems is drawn up before the system is placed on the market and kept up to date throughout its lifecycle. The documentation must cover the elements set out in Annex IV: system description, design specification, monitoring functioning, performance metrics, and any change-management. Enforcement now lands 2 December 2027 (Digital Omnibus, adopted June 2026, pending OJEU publication). - **URL:** https://aiexponent.com/eu-ai-act/article-11 - **Mapped tool:** Agentic Document Analyser (https://aiexponent.com/docs/agentic-document-analyser) - **AskAjay framework:** Liability Ledger Annex IV documentation has to survive a regulator reading it years after the system shipped. AskAjay's Liability Ledger framework walks through what documentation completeness looks like under regulator pressure. It pairs with ADA and the rest of the AiExponent evidence-processing layer. ### Tool mapping note Agentic Document Analyser converts unstructured compliance documents (risk assessments, model cards, audit logs, contracts) into structured JSON for inclusion in an Annex IV technical documentation pack. The tool is alpha quality and Docker-only today; production hardening is on the post-launch roadmap. ADA is evidence-processing infrastructure, not a standalone Article 11 compliance product. --- ## EU AI Act Article 15 · Accuracy, robustness, and cybersecurity Article 15 of the EU AI Act requires high-risk AI systems to achieve appropriate accuracy, robustness, and cybersecurity, and to perform consistently in those respects throughout the lifecycle. Accuracy metrics must be declared in the instructions for use; robustness must extend to errors, faults, and inconsistencies including adversarial inputs. Enforcement follows the deferred high-risk timeline and begins 2 December 2027. The Digital Omnibus, adopted in June 2026 and pending OJEU publication, replaced the original 2 August 2026 date. - **URL:** https://aiexponent.com/eu-ai-act/article-15 - **Mapped tool:** RAG Benchmarking (https://aiexponent.com/docs/rag-benchmarking) - **AskAjay framework:** A7 (Agentic AI Readiness Framework) Declared accuracy on the instructions for use is a number that has to survive post-market scrutiny. AskAjay's A7 framework structures the readiness review and pairs with RAG Benchmarking evidence packs when the audit conversation begins. ### Tool mapping note RAG Benchmarking is a framework-agnostic evaluation harness for RAG and agentic AI systems. It covers Article 15's accuracy and robustness requirements through reproducible benchmarks (faithfulness, answer relevancy, retrieval precision, four agentic metrics) with versioned eval sets and lifecycle drift monitoring. Article 15 also requires cybersecurity: prompt injection resistance, jailbreak defence, model integrity. Pair it with a runtime AI security control to cover the cybersecurity leg as well. --- ## EU AI Act Article 19 · Automatically generated logs Article 19 of the EU AI Act requires providers of high-risk AI systems to keep the logs automatically generated by the system, when those logs are under their control. The logs must be preserved for a period appropriate to the intended purpose of the system, at least six months unless applicable Union or national law provides otherwise, and supplied to competent authorities on request. Enforcement arrives 2 December 2027. The original date was 2 August 2026; the Digital Omnibus, adopted June 2026 and awaiting Official Journal publication, pushed it back. - **URL:** https://aiexponent.com/eu-ai-act/article-19 - **Mapped tool:** Agentic Document Analyser (https://aiexponent.com/docs/agentic-document-analyser) - **AskAjay framework:** Liability Ledger Log retention has to survive a regulator request that may arrive years after the operator decommissioned the system. AskAjay's Liability Ledger framework establishes the retention contract and chain-of-custody design that ADA helps process. ### Tool mapping note Agentic Document Analyser ingests automatically-generated log streams and converts them into structured JSON suitable for the Annex IV documentation pack and for handover to competent authorities. The tool is alpha quality, with no persistent storage and no authentication on the analyse endpoint. Production hardening for Article 19 log retention is on the post-launch roadmap. ADA is evidence-processing infrastructure today. --- ## EU AI Act Article 53 · Obligations for providers of general-purpose AI models Article 53 of the EU AI Act requires providers of general-purpose AI models to keep up-to-date Annex XI technical documentation (53(1)(a)) and provide information packs to downstream providers integrating the model in line with Annex XII (53(1)(b)). Providers must also put in place a copyright policy aligned with Directive (EU) 2019/790 Article 4(3) (53(1)(c)) and publish a sufficiently detailed training-data summary using the AI Office template (53(1)(d)). Article 53 has been applicable since 2 August 2025 (Art. 113(b)); however, Commission-imposed fines under Article 101 are explicitly excluded from that date by Art. 113(b) and only become applicable on 2 August 2026 (default Art. 113 date). Transitional regime per Art. 111(3): GPAI models placed on the EU market before 2 August 2025 have until 2 August 2027 to come into full Article 53 compliance. - **URL:** https://aiexponent.com/eu-ai-act/article-53 - **Mapped tool:** License Compliance Checker (https://aiexponent.com/docs/license-compliance-checker) - **AskAjay framework:** Liability Ledger GPAI provider obligations sit on top of an upstream supply chain: pre-trained weights, datasets, fine-tuners, downstream deployers. AskAjay's Liability Ledger framework maps that supply chain so the Article 53 evidence pack from LCC has the programme context to land in front of the AI Office. ### Tool mapping note License Compliance Checker generates the structured audit evidence that supports each Article 53 sub-obligation: per-model training-data manifests for the Annex XI pack, rights-reservation signal detection (TDM opt-outs, robots.txt, ai.txt) for the Article 4(3) policy, and the structured inputs for the AI Office training-content summary. --- # Source All regulatory citations on AiExponent.com derive from the Official Journal of the European Union: Regulation (EU) 2024/1689 (the EU AI Act). CELEX identifier: 32024R1689. Authoritative URL: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689 This file regenerates on every deploy from `lib/tools.ts` and `lib/eu-ai-act.ts`; there is no separate copy to drift. Last regenerated: 2026-07-22