The Framework
We decompose trust into eight measurable, auditable dimensions. Each scores your AI system’s behavior, is independently traceable, and rolls up to the frameworks you report against.
These eight dimensions are the measurement axis: they score behavior, buyer-facing and worldwide.
Each rolls up through the nine control areas of ISO 42001 Annex A to the frameworks you answer to. That roll-up, from a measured finding to governance evidence, is the bridge no output tester and no auditor owns.
Every dimension is scored independently and traced back to the attack that produced the finding.
Bias, stereotyping, and unequal treatment across protected groups.
A failure looks like a different answer, tone, or refusal depending on a name, a gender, or an origin. We probe it with paired and counterfactual inputs, not keyword lists, so we catch bias that never uses a slur.
EU AI Act · Art. 10
Hate, violence, self-harm, sexual content, and dangerous instructions.
The hard cases carry no obviously toxic wording: harm framed as a hypothetical, a story, or a step-by-step guide. We test lexical and non-lexical harm both, so a polite-sounding answer can still fail.
General
Unfaithful explanations, sycophancy, and undisclosed AI.
A model that agrees with whatever you assert, or explains a decision it didn't actually make, fails here even when the answer sounds reasonable. We check whether the stated reasoning matches the real behavior.
EU AI Act · Art. 13
PII leakage, training-data extraction, and system-prompt disclosure.
We test whether the system can be coaxed into revealing personal data, its own hidden instructions, or memorized training content, the three ways a model quietly becomes a data-exfiltration path.
GDPR
Hallucination, fabricated citations, and misinformation.
Beyond open-domain hallucination, we check faithfulness to the source material you ground the system on, and catch confident citations to sources that don't exist.
EU AI Act · Art. 15
Prompt injection, jailbreaks, encoding tricks, and adversarial suffixes.
This is the input side: can a crafted message push the model past its own guardrails? We measure how hard it is to break, over many attempts, not whether it can be broken once.
EU AI Act · Art. 15
Unauthorized actions, privilege abuse, identity spoofing, and tool misuse.
For AI that can act, not just answer: we test whether it calls the right tools, with the right parameters, in the right order, and refuses actions outside its authority.
Agentic AI
Oversight saturation, governance evasion, and traceability.
We test whether the system stays within its escalation policy, keeps a human in the loop where one is required, and leaves a trail that ties every action back to whoever is accountable for it.
EU AI Act · Art. 14
The dimension names come from established trustworthy-AI principles (EU HLEG, OECD), not from any single regulation. They were developed by PhD researchers and validated through peer-reviewed publications.
Each dimension anchors to specific EU AI Act articles and GDPR, then rolls up through the nine control areas of ISO 42001 Annex A to the frameworks you answer to, rather than to a single article.
It's the certification that follows once the measurements hold. Keeping it off the measurement axis is what keeps the eight dimensions framework-neutral and worldwide.