The rules that apply to you, in four markets

ISO/IEC FDIS 24029-3 tells you how to assess. It does not tell you who must be shown the result. These are the held frameworks that do, chosen by scan for the markets this buyer works in, in market order: United States, European Union, United Kingdom, Australia, with the international standards that sit underneath.

Fourteen controls, each stated in our words and cited to its code. Where a market's regime is not held, the page says so instead of stating it. Last updated: 2026-10-21.

International

The held frameworks that supply the management system, the risk process and the quality model. They are not a jurisdiction; they are the spine the evidence hangs on.

Code	Framework and control	What it asks, in our words	Evidence it expects
A.5.2	AI system impact assessment process ISO/IEC 42001:2023	The organisation has a way to assess what an AI system could mean for individuals, groups and society across its life cycle. The robustness result is one input to that assessment.	The impact assessment record, with the robustness result attached and dated.
A.6.2.4	AI system verification and validation ISO/IEC 42001:2023	The organisation writes down how the AI system will be verified and validated, and the criteria that decide when each measure is used. A statistical robustness test is one such measure.	The verification and validation plan, the criteria, and the run that applied them.
A.7.4	Quality of data for AI systems ISO/IEC 42001:2023	The organisation defines data quality requirements and checks that the data used to build and run the system meets them. The perturbation set and the test sample are data and need the same treatment.	Data quality requirements, the check results, and the provenance of the test data.
6.4.3.3	Assessment of likelihood ISO/IEC 23894:2023	The risk assessment estimates how likely each event and consequence is, states its uncertainty, and recognises that historical data can be scarce and that likelihood can change as the system learns or its context shifts.	The likelihood estimate, its uncertainty, and the evidence it rests on.
6.5.2	Selection of risk treatment options ISO/IEC 23894:2023	Treatment options are compared on what they gain against cost, effort and drawbacks, and can be combined: avoid, accept, remove the source, change likelihood, change consequences, share, or keep by informed decision.	The options considered, the comparison, the decision and the owner.
6.6	Monitoring and review ISO/IEC 23894:2023	Monitoring and review are planned in, compare the process with how the system behaves in use, pick up drift and changed context, and feed improvements back.	The monitoring plan, the drift signals, the review dates and the actions.
5.5	Robustness ISO/IEC 25059:2023	Robustness is a sub-characteristic of reliability: how far the system keeps its level of correctness under difficult conditions, including unseen, biased, adversarial or invalid inputs and outside interference.	The quality requirement, the measure and the result that shows the level kept.

United States

The NIST AI Risk Management Framework is the held federal framework for AI risk. Its inventory and measure functions are where robustness evidence lands.

Code	Framework and control	What it asks, in our words	Evidence it expects
AIRMF-GV-1.6	Inventory of AI systems NIST AI RMF 1.0	Mechanisms are in place to inventory AI systems, and the resource given to the inventory is proportionate to the risk the systems carry.	The inventory, the owner and the resourcing decision.
AIRMF-MS-2.1	Test and evaluation documentation NIST AI RMF 1.0	Test sets, metrics and the details of the tools used in test, evaluation, validation and verification are documented well enough for the evaluation to be repeated.	The test record: sets, metrics, tools, versions, seeds and results.
AIRMF-MS-2.3	Performance under deployment-like conditions NIST AI RMF 1.0	Performance or assurance criteria are measured qualitatively or quantitatively and demonstrated for conditions similar to deployment, with the measures documented.	The measurement, the deployment-like conditions and the documented result.

European Union

The EU AI Act is the held Union regime. Article 15 names robustness directly; Article 10 governs the data the assessment uses.

Code	Framework and control	What it asks, in our words	Evidence it expects
EUAI-Art.15	Accuracy, robustness and cybersecurity EU AI Act	High-risk AI systems are designed and developed to achieve an appropriate level of accuracy, robustness and cybersecurity, and to perform consistently in those respects through their life cycle, including resilience to errors, faults and attempts to alter use or output.	The conformity evidence, the robustness testing, and the consistency record.
EUAI-Art.10	Data and data governance EU AI Act	High-risk systems that train models use training, validation and testing data that meet the quality criteria, with data governance, bias examination, gap identification and statistically relevant datasets for the intended purpose.	The data governance record, the bias examination and the dataset description.

United Kingdom

The UK framework is non-statutory and applied by regulators. Principle 1 names safety, security and robustness, and points to testing and due diligence.

Code	Framework and control	What it asks, in our words	Evidence it expects
P1	Safety, security and robustness UK AI Regulation Framework	AI systems should work robustly, securely and safely across their life cycle, with risks continually identified, assessed and managed. Regulators apply the principle proportionately and may require testing or due diligence.	The risk assessment, the testing and the due diligence record a regulator can ask for.

Australia

The Australia AI Ethics Framework is the held national guidance. Principle 5 covers reliability and safety across the life cycle.

Code	Framework and control	What it asks, in our words	Evidence it expects
AUAIE-5	Reliability and safety Australia AI Ethics Framework	Across the life cycle an AI system should operate reliably in accordance with its intended purpose, be accurate and reproducible as appropriate, avoid unreasonable safety risk and be monitored and tested so it continues to meet its purpose.	The reliability and safety evidence, the monitoring record and the action on identified problems.

Questions to put to a supplier

When the neural network arrives inside a bought product, the evidence has to come from the supplier. Ask three things: which statistical method was used and on what sampling basis; which input conditions and corruptions the test covered; and whether the result is stable across data sources and time. A supplier that cannot answer all three has not done the assessment, whatever the datasheet says.

What is not held

The United States has no federal AI statute held for this subject. The EU AI Act and the UK framework are held as published policy and regulation; the Australian guidance is held as national ethics guidance. Where a jurisdiction's rule is not in the held set, this page names the regime and does not state its requirements.