ARIA Model Card

ARIA is intended to help mid-market teams understand AI governance and compliance: the EU AI Act, GDPR, the UAE AI Charter, risk assessments, transparency obligations and documentation. It is a drafting and education aid, not a decision-maker. A qualified human reviews and signs off any output before it is relied upon.

ARIA sends your question, plus relevant context from your workspace, to Anthropic's Claude model with a compliance-focused system prompt. Claude is a large language model trained by Anthropic on a broad mix of licensed and publicly available text; AM8 does not train models on customer data. Responses are generated probabilistically from that training, not retrieved from an authoritative legal database.

ARIA can be confidently wrong. As a large language model it may hallucinate facts, cite regulations inaccurately, miss jurisdiction-specific nuances, or reflect gaps and biases in its training data. It has a knowledge cut-off and is not aware of the very latest regulatory guidance or your full legal context. It cannot access the internet in real time, cannot guarantee completeness, and must never be treated as a substitute for professional advice. Always verify material decisions with a qualified compliance professional or lawyer.

The main risks are over-reliance, inaccurate output, and use outside ARIA's domain. Safeguards: every AI response carries a persistent disclosure that it is AI-generated; generated compliance documents are watermarked DRAFT until a named human reviews them; and ARIA is bounded to the compliance domain. Users are reminded to verify and to seek professional advice.

ARIA will refuse or decline requests that fall outside its remit. In particular it does not provide legal advice (it redirects you to a qualified lawyer), declines topics that are out of scope of AI compliance, and refuses requests to produce harmful content or to bypass its guidelines. These refusals are by design.

We evaluate ARIA against a fixed golden dataset of compliance questions, under an evaluation protocol registered before the results are read, and we re-run it when the underlying model version changes. The dataset includes cases whose correct answer is a refusal, such as a request for legal advice or a question about a jurisdiction we do not cover, so declining is scored as success rather than as a miss.

We publish no accuracy or refusal-rate figure here. A number without the dataset, the interval and the rule-pack version it was measured against is not evidence, and a marketing page cannot carry those. Enterprise buyers and auditors can request the evaluation record under NDA.