We test AI models, AI applications and software built by others, against agreed acceptance criteria and the laws and regulations of the region where the system will operate. You receive written evidence that procurement, risk and regulators can act on.
Your system, criteria and intended use
Recorded evidence and human verification
Findings, limitations and next steps
The evaluation process, illustrated.
Three classes of system. One method: scoped in writing, tested against agreed thresholds, reported for decision makers.
Choose the kind of system you need to evaluate.
Foundation and fine-tuned models, evaluated on your tasks: faithfulness to source, citation accuracy, abstention, consistency across languages and dialects.
Does the answer hold up against the source?
Chatbots, retrieval systems, document processors and agents: accuracy, hallucination rate, prompt injection, data extraction, tool misuse, bias, over-refusal, disclosure, performance and cost.
What happens when the whole workflow is tested?
Web, mobile and enterprise systems: functional acceptance, regression, performance under load, security baseline, localisation and right-to-left quality, personal-data handling.
Does the delivered software meet the agreed requirements?
Illustrative test paths. Exact coverage is agreed in scope.
We do not evaluate products in the three categories where we sell one: legal practice management, financial-reporting compliance and enterprise knowledge management. We do not evaluate anything we built. The full rule is published below.
The full bench scores ten dimensions, each tied to the standard or legal instrument that asks for it. Each package draws on the dimensions relevant to it; the scope document lists which.
Select a dimension to see an example of the test and the evidence it produces.
Prompt injection, system-prompt and data extraction, retrieval poisoning, supply-chain exposure.
Illustrative test example
Covered by:
Malformed, adversarial and out-of-distribution input, including dialect and script variation.
Illustrative test example
Covered by:
Faithfulness to source, citation correctness, abstention when the answer is absent.
Illustrative test example
Covered by:
Arabic variants and dialects, right-to-left rendering and other deployment languages, against the English baseline.
Illustrative test example
Covered by:
Differential outcomes across the characteristics relevant to the deployment.
Illustrative test example
Covered by:
Personal-data handling, retention, storage location and cross-border transfer against the applicable law.
Illustrative test example
Covered by:
Disclosure to users; explainability of high-impact outputs.
Illustrative test example
Covered by:
Documentation, change control, monitoring, human oversight, vendor due diligence.
Illustrative test example
Covered by:
Tool permissions, action boundaries, kill-switch behaviour, multi-step failure, in a sandbox.
Illustrative test example
Covered by:
Latency, throughput and token cost under realistic multilingual load.
Illustrative test example
Covered by:
Every package delivers a written report. Scorecards accompany the full bench and the adversarial package; the verdict memo is the pre-purchase deliverable; attestations are issued only on a published scope.
Open a document to see what it contains and when it applies.
Scope, method, results per dimension, findings register, limitations, remediation guidance.
One page: dimensions in scope, agreed threshold, result, trend against the previous run.
Registered with scope, version, date and limitations. Not an accredited certification, and it says so.
Proceed, proceed with conditions, or do not proceed. A technical acceptance recommendation.
Illustrative document structure. No client result is shown.
| ID | Severity | Tags | Status |
|---|---|---|---|
| SEC-04 | High | OWASP LLM01, MITRE ATLAS, NIST AI 100-2 | Reproduced |
| PRV-02 | Medium | OWASP LLM02, NIST AI 600-1 | Open |
| ACC-07 | Low | OWASP LLM09 | Retest scheduled |
Four stages. The same process for models, applications, agents and software.
System, version, the decision to support, acceptance criteria and applicable regulation agreed in writing. Written authorisation obtained before any test.
Agreed batteries executed: accuracy sampling, adversarial campaigns, bias, privacy and residency checks, functional and performance testing. Automated grading calibrated against human labels.
Every failure reproduced by a second tester. Every verdict reviewed by an external panel member who had no part in the test.
Findings against thresholds and applicable requirements, with remediation guidance, limitations and reassessment triggers. One re-test after vendor remediation.
Fixed scope, fixed fee. Each package has a defined input list, a duration band and a named deliverable, quoted against stated assumptions.
For buyers, before signing.
A fixed battery on the vendor's staging system using fifty to a hundred of your real questions: accuracy, language quality, baseline injection and leakage probes, disclosure and kill-switch checks.
Accuracy, language quality, security baseline, transparency, agent safety
For buyers and vendors.
Full accuracy, language-quality and robustness testing, document-ingestion quality, over-refusal, with a human-validated judge and a frontier-model baseline.
Accuracy, language quality, robustness, bias, transparency, performance and cost
For risk and security functions, and vendors needing security evidence.
Direct and indirect prompt injection, system-prompt and data extraction, retrieval poisoning, agent tool abuse in a sandbox, multi-turn attacks, supply-chain review.
Security, robustness, agent safety, privacy, governance evidence
For any organisation receiving software from a vendor.
Functional acceptance, regression, performance under load, security baseline, localisation review, personal-data walkthrough. Inputs: specification, acceptance criteria, test environment, integration boundaries.
Functional acceptance, security baseline, performance, localisation, privacy
Maps evidence you already hold, such as vendor test reports or prior audits, to the laws, sector frameworks and standards that apply in your region, including ISO/IEC 42001, the NIST AI RMF and AI regulation where in force. We state what evidence exists against each requirement; we never state that a system "complies".
One to three weeks
Mapping annex and gap register
A frozen subset of the batteries re-run on schedule or on material change: drift, regression, new attack vectors, annual bias re-test.
Trend report; re-issued attestation when the scope is unchanged
Engagement options How it works What you receive The independence rule
AI systems fail in ways conventional test suites do not catch: plausible wrong answers, injected instructions, uneven performance across languages. Independent evaluation gives the organisation that will operate the system its own evidence before it buys, deploys or signs off.
A development team tests its own reading of the requirement. We test yours: your acceptance criteria, your data, your users' language.
A chatbot or agent can pass a demo and still hallucinate, leak data or follow injected instructions. We measure this with adversarial test cases, documented sampling and human review.
Most benchmarks, attack catalogues and guardrails are built for English. We test in the languages your users actually use, with the English baseline alongside.
AI regulation, sector frameworks and data-protection laws now set requirements a deployed system can be examined against. Independent results give procurement, risk and regulatory review something concrete.
Organisations about to sign with a vendor. Deliverable: a verdict memo, proceed, proceed with conditions or do not proceed, with the evidence behind each line.
AI product companies selling to regulated buyers. Deliverable: an independent report and, on a published scope, a registered attestation of testing performed.
Banks, insurers and public bodies under frameworks that address AI. Deliverable: findings mapped to the provisions agreed in scope.
Disclosed conflicts, requirement-derived test cases, multilingual batteries and independent review of every finding.
The full rule is published on this page: category and system exclusions, no remediation, no guardrail resale, conflict disclosure, fees never tied to the verdict.
Test cases linked to your acceptance criteria, intended use and applicable regulation, with coverage and limitations recorded.
Arabic in eight variants with the English baseline, and the other languages of your deployment where the scope requires.
Where a finding bears on a requirement, it is tied to the applicable instrument of your region. Other findings are tied to acceptance criteria or a named testing reference.
No automated score becomes a result on its own. Second-tester confirmation and external review are standard.
For bank and government engagements, testing and data processing stay in the jurisdiction you specify. Processing locations are documented before testing; cross-border processing requires your written approval.
Written authorisation covering systems, techniques and data handling is obtained before any test, including any permission the hosting provider requires. The scope document lists the assets in scope, excluded techniques, the time window and the stop-and-notify procedure.
Test data and logs are retained for the period stated in scope, then deleted with written confirmation. Access is limited to named testers and the external reviewer. Third-party processing is listed with its location. Production data is used only where a test is meaningless without it, under a documented lawful basis.
Source Cybercrime Law 63 of 2015 — Permission boundaries on agents and retrieval systems, integrity of records, and the written permission under which adversarial testing is carried out.
MENA Region Solutions also builds AI products. The rule below is public and applies to every engagement.
No evaluation in any category where we sell or develop a product: legal practice management, financial-reporting compliance, enterprise knowledge management. Competitors there are referred elsewhere.
No evaluation of any system we designed, built, fine-tuned, prompt-engineered, integrated or bid for, for twenty-four months after our last involvement.
We deliver findings. We do not sell the fix for a system we tested, for the same period.
No resale or recommendation of guardrail products. The vendor-under-test's own model family is never the sole judge.
Every report carries a conflict statement covering the previous twenty-four months and states who commissioned and paid for the test.
Fees are never contingent on the verdict or on an attestation being issued.
Evaluation and build work are separated in staff and in information. Every verdict and attestation is reviewed by an external panel member who had no part in the test.
View the sample annex (Kuwait)
Findings are mapped to the laws, sector frameworks and standards that apply where the system will operate, identified at scoping and listed in the scope document. We cite only instruments we have read in official text. We state what evidence exists against each requirement; we do not state that a system "complies". A sample annex for one jurisdiction is published below.
Consent and purpose limitation for personal data, data-subject access and erasure, validity of contracts concluded by automated systems, evidential requirements for electronic records.
Bilingual notices, provable consent, retention periods, disclosure of storage location and notice before transfer abroad, breach notification windows, privacy by design.
Government data only on licensed providers with in-country data centres, no secondary use of subscriber content, exit and deletion rights.
Adversarial-attack defences for AI referencing the OWASP Top 10 for LLM applications, explainability for high-impact decisions, continuous monitoring, third-party review, vendor due diligence and residency.
Accurate service information, Arabic advertising and invoices, identity on electronic records.
Permission boundaries on agents and retrieval systems, integrity of records, and the written permission under which adversarial testing is carried out.
Traceability, right to explanation, non-discrimination; non-binding and labelled as such.
No. We issue independent test reports and attestations of testing performed; each document states what it is. Certification is contracted directly with an accredited body. Readiness work is a separate service, never offered for a system we evaluated.
Any. The engagement is scoped to the jurisdiction where the system will operate; the applicable instruments are identified at scoping and listed in the scope document.
A penetration test covers infrastructure and application security. We cover the AI-specific attack surface, plus accuracy, language quality, bias, privacy and functional acceptance.
For AI systems: a staging or sandbox instance, documentation and version, fifty to a hundred real questions or documents. For software: specification, acceptance criteria, test environment, integration boundaries. In both cases, written authorisation.
Not for a system we tested. We provide remediation guidance, your vendor fixes, we re-test.
Yes. A redacted sample report, a sample scope and the methodology summary are provided during qualification under NDA.
The commissioning party. Vendor attestations are publishable only with their full scope and version. Reports are shared with other parties under the agreed confidentiality terms, subject to applicable law.
Thirty minutes. Bring the system, who built it, where it runs and the decision you need to make. We follow up with a proposed scope, schedule and fixed-fee quotation.