EU AI Act, California SB 53 and ISO/IEC 42001 evidence
Ariana Nexus prepares the Afghan-language evidence that AI governance files are usually missing: what your model or system was tested on in Pashto and Dari, how it performed, where it fails, who reviewed it and how it is monitored. Each document is written to the EU AI Act, California SB 53 or ISO/IEC 42001 requirement it supports.
See what the evidence file containsAI governance documentation for language coverage is the written evidence of how an AI model or system was built, tested, limited and monitored in a specific language. For Afghan languages it records the Pashto and Dari training and test data, evaluation results by language and dialect, known failure modes, human oversight arrangements and incident history, in the structure that the EU AI Act, California SB 53 and ISO/IEC 42001 ask for.
From the records you already hold to one evidence file that three frameworks can read.
A model card says the model supports more than a hundred languages. The technical documentation reports accuracy on English benchmarks. The risk assessment lists multilingual performance as a single line. None of that tells a regulator, an auditor or a deployer what happens when a person writes to the system in Pashto or Dari.
That gap is now a documentation problem, not only a quality problem. California SB 53 requires frontier developers to publish the languages a model supports. The European Commission’s training-content template asks general-purpose AI providers to describe the linguistic characteristics of their data. The EU AI Act requires high-risk systems to be trained and tested on data that represents the people and settings they are used in, and to declare accuracy and known limitations. ISO/IEC 42001 asks for an assessment of the system’s impact on individuals and groups.
A management system can conform to a standard and still hold no evidence about an entire language community. Conformance is not coverage. This service supplies the coverage.
Ten controlled documents. Each one can be lifted into your technical documentation, framework, transparency report or management system without rewriting.
Delivered as controlled Word and PDF documents with version history, plus a machine-readable index (JSON and CSV) that imports into your document control or GRC system.
One body of evidence, three readers. The same Pashto and Dari test results answer a different question in each framework.
References are a navigation aid prepared on 19 September 2026 against the current legal texts. They are not legal advice. Your counsel decides which obligations apply to you.
Checked against the Official Journal and the California code on 19 September 2026.
Both are low-resource languages. That changes what can be assumed from English results, and what has to be shown directly.
More than forty years of war and displacement have placed Afghan communities in front of AI systems in asylum offices, hospitals, schools, job centres and social platforms across Europe, North America and Australia.
Six steps, each with a written output you can inspect before the next begins.
A first evidence file for Pashto and Dari on one system typically takes six to ten weeks from a signed scope. Updates for a new model version are faster, because the test sets, procedures and crosswalk already exist.
Every engagement is scoped and priced as a whole, with a named partner accountable for it. We do not take per-item review work or place unnamed reviewers in another firm’s pipeline.
Governance consultancies know the frameworks and cannot read the languages. Language vendors can read the languages and do not know what an auditor will sample. This firm was built to do both.
We do not build or sell AI models, governance software or certification. We do sell evaluation, red teaming and data services; where an evidence file reports on work we performed, the file says so.
All 24 appear in the coverage statement. We publish readiness tiers rather than a flat claim, because speaker availability differs by two orders of magnitude across this list.
Dari deliverables are Afghanistan Dari, not Iranian Persian. Turkmeni deliverables are Afghan Turkmeni, not the standard Turkmen of Turkmenistan. Hazaragi is a variety of Dari and is staffed separately because the lexical distance is large enough to change a result.
Ariana Nexus is led and operated by alumni and scholars of leading universities who are native speakers of the languages they assess. The people who read the Pashto and Dari output are the people who write the findings.
Hassan Ukasha oversees the firm’s operations and its AI governance documentation program. He is the accountable partner on every evidence file Ariana Nexus issues: he approves the scope, signs the independence and conflicts declaration, and is the point of escalation for your general counsel, notified body or certification auditor.
Not in those words. The Act does not list languages. It requires high-risk systems to use training, validation and test data that is relevant and sufficiently representative for the intended purpose and the setting of use (Article 10), to declare accuracy levels and known limitations (Articles 13 and 15), and to document testing in the Annex IV technical file. If Afghan-language speakers are within the intended purpose, the file needs evidence that covers them, or a stated limitation.
Before deploying a new or substantially modified frontier model, a frontier developer must publish a transparency report that includes the languages the model supports, its intended uses and its restrictions. Large frontier developers must also summarise their catastrophic-risk assessments and the role of third-party evaluators. SB 53 prohibits materially false or misleading statements about catastrophic risk and its management, which is one reason safety claims should be evidenced in every language the report lists. The law has applied since 1 January 2026.
ISO/IEC 42001 certifies the management system, not the model. An auditor samples documented information: the AI system impact assessment (clause 6.1.4 and Annex A.5), data quality and provenance records (A.7), verification and validation (A.6.2.4) and information for users (A.8.2). Where a system in scope is used in Pashto or Dari, those records need language-specific content. We supply it; your certification body audits it.
No. ISO/IEC 42001 is a voluntary management system standard and is not a harmonised standard under the EU AI Act. It gives a governance structure that much of the Act’s documentation can sit in, but the Act’s specific obligations still have to be met and evidenced separately.
Yes. In Common Crawl’s August 2026 crawl, 0.0043% of pages were identified as Pashto against 40.45% English, and Dari is not identified separately from Persian at all. Models see little Afghan-language data, safety training transfers unevenly and public benchmarks are scarce, so performance and safety claims cannot be assumed from English results. They have to be evidenced directly.
They can be disclosed, but they are weak evidence. A machine-translated test set carries the translation system’s errors and contains no Afghan names, places, institutions or idiom, so it measures translation quality as much as model quality. We use test items written by native speakers and state the method in the report so an assessor can judge its weight.
No. Ariana Nexus is not a law firm, a notified body or a certification body. We produce Afghan-language evidence and documentation and map it to the requirements it supports. Your counsel decides what the law requires of you; your certification body or notified body decides conformity. We work alongside them, never in place of them.
University-educated native speakers tested in the specific variety they review, led by partners and principals who are alumni and scholars of Cornell University, the University of Chicago, the University of British Columbia and Otto-von-Guericke University. No U.S. certification exists for Pashto or Dari, so every reviewer is assessed against the firm’s written standard and the result is recorded in the evaluator record. All work is done in-house. We do not use subcontractors or crowd platforms.
All 24 appear in the coverage statement, in three published readiness tiers. Pashto and Dari, including the Hazaragi variety, receive the full ten-document evidence file today. Uzbeki, Turkmeni, Balochi, Pashayi and the Nuristani group are evidenced on an agreed scope. The smallest languages, such as Ormuri, Parachi, Tirahi and the Pamir group, are scoped to verified speaker availability, and the file says “not evidenced” where that is the truth.
Often, yes. Under the EU AI Act, deployers of high-risk systems must use them according to the instructions, assign competent human oversight and, for public bodies and certain services, complete a fundamental rights impact assessment before first use. If your clients, patients or applicants include Afghan-language speakers, those steps need language-specific content. We also review a vendor’s language claims before you buy.
A gap assessment takes two to three weeks. A first full evidence file for Pashto and Dari on one system typically takes six to ten weeks from a signed scope, depending on how much evaluation already exists. Updates for a new model version are faster.
Work begins under a mutual non-disclosure agreement. Access is limited to the named engagement team, data stays in the environment agreed in the scope, and contributors are pseudonymised in deliverables. Ariana Nexus routes no data, documents or inquiries through channels controlled by the de facto authorities in Afghanistan. European engagements run under GDPR and UK GDPR terms.
Those services produce the measurements. This service turns measurements, ours or your own, into the governance record: coverage statements, data and risk files, oversight procedures and the crosswalk that auditors and regulators read. Many clients start here with a gap assessment, then commission the evaluation the gaps call for.
Obligations for general-purpose AI model providers have applied since 2 August 2025, and Commission enforcement powers over them began on 2 August 2026. Regulation (EU) 2026/1744 moved the high-risk dates: stand-alone Annex III systems must comply from 2 December 2027, and AI in Annex I regulated products from 2 August 2028.
Tell us which systems and language claims are in scope and which deadline you are working to. A partner will reply with the questions we need answered to scope a gap assessment.
Request a scoping conversationEngagements begin under a mutual non-disclosure agreement. Washington, D.C. Serving the United States, the European Union and clients worldwide.
Reviewed by the Ariana Nexus Cultural Compliance Bureau. Last updated 19 September 2026.
