# Axiom Verity

> An independent, non-commercial AI governance research lab. It publishes a cross-mapped reference library (risks, controls, frameworks, tools and terminology), original research and working instruments for governing AI in regulated organisations. Financial services is the proving ground; the methods are built to generalise.

Axiom Verity is written and maintained by Andrew Moore. It sells nothing, offers no services and accepts no sponsorship (see /about/independence). The site is organised into four pillars (governance, compliance, security, engineering), the Library and Research. Every page, including every item listed below, is fully server-rendered with no client-side gating, so it is safe to fetch and cite directly.

## Reference

- [Governance](https://axiom-verity.com/governance): ISO/IEC 42001 AI management system guidance — clause-by-clause requirements (Clauses 4–10), the nine Annex A control groups, and evidence expectations.
- [Compliance](https://axiom-verity.com/compliance): EU AI Act obligation timeline, a three-framework crosswalk (EU AI Act, ISO/IEC 42001, NIST AI RMF), and named compliance gaps (Articles 44 and 49).
- [Security](https://axiom-verity.com/security): AI/LLM threat categories mapped to the OWASP Top 10 for LLM Applications, a reference assurance control-flow, and audit-ready evidence expectations.
- [Engineering](https://axiom-verity.com/engineering): AI engineering guidance: architecture, integration, evaluation, monitoring, and incident response for AI systems actually running in production.
- [Library](https://axiom-verity.com/library): Overview of the cross-mapped reference library: the taxonomy, severity model and mapping philosophy behind the risks, controls, frameworks, tools and glossary.
- [Research](https://axiom-verity.com/research): Overview of the lab: the Responsible AI programme, articles, the Thought Experiment, projects, news and learning paths.
- [Responsible AI by design](https://axiom-verity.com/research/responsible-ai): The flagship research programme: a twelve-domain map of what must hold true for AI in a regulated organisation, and an eight-stage lifecycle by which each use case is proven.
- [Third-Party AI Governance](https://axiom-verity.com/governance/third-party): Vendor and model governance for AI: assessment patterns and how third-party risk cuts across governance, compliance, security, and engineering.
- [Resources](https://axiom-verity.com/resources): A curated, annotated library of external AI regulation, standards, thought-leadership, and learning resources, each with editorial context — the best entry point on this site for primary sources.

## Risk Library (48 risks)

- [Risk Library index](https://axiom-verity.com/risks): Browse the full risk register by category, severity, or mapped framework.
- [Adversarial evasion attacks](https://axiom-verity.com/risks/adversarial-evasion-attacks): Crafted adversarial input causes a model to misclassify or mis-generate, evading a downstream security or safety control.
- [Adversarial testing coverage gap](https://axiom-verity.com/risks/adversarial-testing-coverage-gap): No structured adversarial testing exists for an AI system before or after deployment, leaving no empirical basis for its security assurance.
- [AI literacy gap in the deploying organisation](https://axiom-verity.com/risks/ai-literacy-gap-in-deploying-organisation): Staff operating or relying on an AI system's output haven't been given the training needed to understand its limitations, leading to over-trust or misuse.
- [AI model drift and degradation](https://axiom-verity.com/risks/ai-model-drift-and-degradation): A deployed model's performance silently degrades over time as the input distribution shifts away from its training and validation data.
- [AI not mapped to important business services](https://axiom-verity.com/risks/ai-not-mapped-to-important-business-services): AI components embedded in an important business service haven't been identified or mapped, so an AI failure isn't accounted for in the service's impact tolerance.
- [AI system lifecycle controls gap](https://axiom-verity.com/risks/ai-lifecycle-controls-gap): AI systems lack defined lifecycle controls, so changes ship without a consistent governance checkpoint from requirements through decommissioning.
- [AI-driven social scoring](https://axiom-verity.com/risks/ai-driven-social-scoring): An AI system evaluates or classifies people over time based on social behaviour or inferred characteristics, leading to detrimental or unfavourable treatment.
- [Automated decision-making without safeguards](https://axiom-verity.com/risks/automated-decision-making-without-safeguards): AI is used to make solely-automated decisions with legal or similarly significant effects on individuals, without the safeguards UK GDPR requires.
- [Biometric categorisation of sensitive attributes](https://axiom-verity.com/risks/biometric-categorisation-of-sensitive-attributes): A biometric categorisation system infers race, political opinions, trade union membership, religious belief, or sexual orientation.
- [Cross-border AI data transfer risk](https://axiom-verity.com/risks/cross-border-ai-data-transfer-risk): Personal data processed by an AI system, including via a cloud or model provider, is transferred outside the UK without an adequate transfer mechanism.
- [Data and model poisoning](https://axiom-verity.com/risks/data-and-model-poisoning): Training, fine-tuning, or embedding data is deliberately manipulated to introduce vulnerabilities, backdoors, or bias into a model.
- [Disparate impact monitoring gap](https://axiom-verity.com/risks/disparate-impact-monitoring-gap): No structured monitoring exists to detect disparate outcomes across protected characteristics once an AI system is in production.
- [Excessive agency in autonomous agents](https://axiom-verity.com/risks/excessive-agency): An agent is granted more permission, tool access, or autonomy than the task requires.
- [Exploitation of vulnerable groups by AI](https://axiom-verity.com/risks/exploitation-of-vulnerable-groups-by-ai): An AI system exploits vulnerabilities due to age, disability, or social or economic situation to distort behaviour and cause significant harm.
- [Hidden context exposure](https://axiom-verity.com/risks/hidden-context-exposure): Content the operator assumed was hidden from the user, such as retrieval context, RAG schemas, or policy logic, is exposed or inferred through the model, beyond just the system prompt.
- [Improper output handling](https://axiom-verity.com/risks/improper-output-handling): AI-generated output is passed to a downstream interpreter, renderer, or system call without validation, enabling injection-style attacks beyond the model itself.
- [Inadequate AI resourcing and competence](https://axiom-verity.com/risks/inadequate-ai-resourcing-and-competence): AI systems are adopted without provisioning the compute, tooling, data quality controls, or trained personnel needed to operate them safely.
- [Inadequate AI vendor due diligence](https://axiom-verity.com/risks/inadequate-ai-vendor-due-diligence): An AI vendor is onboarded without due diligence proportionate to how material the service is, including its own AI-specific risk controls.
- [Inadequate explainability for affected individuals](https://axiom-verity.com/risks/inadequate-explainability-for-affected-individuals): An AI-driven decision materially affecting a person can't be explained to them in a way they can meaningfully understand or challenge.
- [Inadequate fairness testing before deployment](https://axiom-verity.com/risks/inadequate-fairness-testing-before-deployment): An AI system used in a high-risk domain is deployed without pre-deployment testing for disparate performance or outcomes across demographic groups.
- [Inadequate human oversight design](https://axiom-verity.com/risks/inadequate-human-oversight-design): An AI system's human-oversight mechanism exists on paper but isn't actually usable or effective in practice.
- [Individual criminal risk profiling](https://axiom-verity.com/risks/individual-criminal-risk-profiling): An AI system assesses the risk that an individual will commit a criminal offence based solely on profiling or personality traits.
- [LLM application supply chain vulnerabilities](https://axiom-verity.com/risks/llm-supply-chain-vulnerabilities): Vulnerable or unvetted components — base models, adapters, datasets, plugins, deployment platforms — enter an LLM application through its supply chain.
- [Misinformation and confabulation](https://axiom-verity.com/risks/misinformation-and-confabulation): A model states false or fabricated information with the same linguistic confidence as accurate information.
- [Missing AI impact assessment process](https://axiom-verity.com/risks/missing-ai-impact-assessment-process): AI systems are deployed without a structured process to assess their impact on individuals, groups, and society before go-live.
- [Missing DPIA for high-risk AI processing](https://axiom-verity.com/risks/missing-dpia-for-high-risk-ai-processing): AI processing likely to result in high risk to individuals proceeds without a documented Data Protection Impact Assessment.
- [Model backdoors and trojans](https://axiom-verity.com/risks/model-backdoors-and-trojans): A hidden, trigger-activated backdoor is embedded in a model via poisoned training data, weight tampering, or a payload injected into the model artifact.
- [Model extraction and inference-API abuse](https://axiom-verity.com/risks/model-extraction-and-inference-api-abuse): Repeated querying of a model's inference API extracts private training data or effectively replicates the model itself.
- [No documented AI policy](https://axiom-verity.com/risks/no-documented-ai-policy): The organisation operates AI systems without a top-management-approved policy defining acceptable use, risk appetite, and escalation routes.
- [No impact tolerance for AI-dependent services](https://axiom-verity.com/risks/no-impact-tolerance-for-ai-dependent-services): A business service that depends on an AI system has no defined impact tolerance for AI-specific disruption.
- [Prompt injection](https://axiom-verity.com/risks/prompt-injection): Attacker-controlled input overrides system instructions, either directly or via retrieved/tool content treated as trusted.
- [Proxy discrimination via correlated features](https://axiom-verity.com/risks/proxy-discrimination-via-correlated-features): A model achieves discriminatory outcomes indirectly, through features that correlate with a protected characteristic even when that characteristic is excluded from the input.
- [Real-time public biometric identification](https://axiom-verity.com/risks/real-time-public-biometric-identification): Real-time remote biometric identification is used in publicly accessible spaces for law enforcement purposes, outside the three narrow statutory exceptions.
- [Sensitive information disclosure](https://axiom-verity.com/risks/sensitive-information-disclosure): A model surfaces training, fine-tuning, or retrieval-sourced sensitive data in its output.
- [Sub-outsourcing visibility gap](https://axiom-verity.com/risks/sub-outsourcing-visibility-gap): An AI vendor sub-contracts part of the service without the originating organisation having visibility or contractual control over that layer.
- [Subliminal or manipulative AI techniques](https://axiom-verity.com/risks/subliminal-or-manipulative-ai-techniques): An AI system deploys subliminal, manipulative, or deceptive techniques that materially distort a person's behaviour and cause significant harm.
- [System prompt leakage](https://axiom-verity.com/risks/system-prompt-leakage): The system prompt is extracted or inferred, exposing configuration, guardrail logic, or embedded credentials.
- [Third-party model and dataset provenance](https://axiom-verity.com/risks/third-party-model-provenance): A model, dataset, or plugin is adopted from an external source without verifying provenance, licensing, or update history.
- [Training data bias and discriminatory outcomes](https://axiom-verity.com/risks/training-data-bias): Historical or sampling bias in training data is reproduced or amplified in model outputs and decisions.
- [Unbounded resource consumption](https://axiom-verity.com/risks/unbounded-resource-consumption): Unrestricted or excessive inference requests lead to denial of service, denial of wallet, or facilitate model theft.
- [Unclear AI ownership and accountability](https://axiom-verity.com/risks/unclear-ai-ownership-and-accountability): No individual or committee owns AI risk decisions end-to-end, so accountability for a given system's behaviour is diffuse or absent.
- [Undisclosed AI interaction](https://axiom-verity.com/risks/undisclosed-ai-interaction): A person interacts with an AI system without being informed they are doing so.
- [Unlawful or undocumented training data basis](https://axiom-verity.com/risks/unlawful-or-undocumented-training-data-basis): Personal data used to train, fine-tune, or ground an AI system lacks a documented lawful basis, or exceeds the purpose and minimisation limits of the basis relied on.
- [Untargeted facial recognition scraping](https://axiom-verity.com/risks/untargeted-facial-recognition-scraping): Facial images are scraped from the internet or CCTV footage without targeting, to build or expand a facial recognition database.
- [Untested AI failure scenarios](https://axiom-verity.com/risks/untested-ai-failure-scenarios): No scenario testing exists for AI-specific disruption modes against the organisation's impact tolerances.
- [Vector and embedding weaknesses](https://axiom-verity.com/risks/vector-and-embedding-weaknesses): Weaknesses in how vectors and embeddings are generated, stored, or retrieved in a retrieval-augmented generation pipeline let an attacker poison or exfiltrate the knowledge base.
- [Vendor concentration and exit strategy gap](https://axiom-verity.com/risks/vendor-concentration-and-exit-strategy-gap): The organisation is dependent on a single AI model or vendor for a critical function with no viable exit plan or alternative provider identified.
- [Workplace and education emotion recognition](https://axiom-verity.com/risks/workplace-education-emotion-recognition): An AI system infers emotions in the workplace or an education institution, outside the narrow medical or safety exceptions.

## Framework mappings (1)

- [OWASP Top 10 for LLM Applications (2026), Mapped to Controls](https://axiom-verity.com/frameworks/owasp-llm-top-10): All ten OWASP LLM Top 10 (2026) entries, each mapped to the risks in the Axiom Verity register and the controls that address them.

## Control Library (22 controls)

- [Control Library index](https://axiom-verity.com/controls): Browse the full control register, mapped back to the risks it addresses.
- [Adversarial robustness testing](https://axiom-verity.com/controls/adversarial-robustness-testing): Structured red-teaming and adversarial testing of an AI system against known attack techniques before deployment and on a recurring cadence after.
- [AI acceptable use policy](https://axiom-verity.com/controls/ai-acceptable-use-policy): A top-management-approved policy defining what AI use is permitted, what isn't, and where to escalate an edge case.
- [AI impact assessment process](https://axiom-verity.com/controls/ai-impact-assessment-process): A mandatory, structured assessment of an AI system's impact on individuals, groups, and society, completed before deployment.
- [AI incident response and rollback](https://axiom-verity.com/controls/incident-response-and-rollback): Defined containment, remediation, and rollback procedure for a detected AI system failure or harmful output.
- [AI literacy programme](https://axiom-verity.com/controls/ai-literacy-programme): Structured training for staff who operate, rely on, or are affected by an AI system's output, covering its limitations as well as its use.
- [AI ownership and accountability framework](https://axiom-verity.com/controls/ai-ownership-and-accountability-framework): A named individual or committee accountable for each AI system's risk posture, with defined reporting lines.
- [AI system monitoring and logging](https://axiom-verity.com/controls/ai-system-monitoring-logging): Continuous telemetry, drift detection, and audit logging of AI system inputs, outputs, and overrides.
- [AI transparency and disclosure labelling](https://axiom-verity.com/controls/ai-transparency-and-disclosure-labelling): Consistent, visible labelling of AI-generated content and AI-driven interactions across every surface a person might encounter them.
- [Automated decision-making safeguards](https://axiom-verity.com/controls/automated-decision-making-safeguards): Process guaranteeing a right to human review, an ability to contest, and a documented rationale for any solely-automated decision with significant effect.
- [Bias testing and fairness monitoring](https://axiom-verity.com/controls/bias-testing-and-fairness-monitoring): Pre-deployment subgroup performance testing plus ongoing production monitoring for disparate outcomes across protected characteristics.
- [Data minimisation and lawful basis review](https://axiom-verity.com/controls/data-minimisation-and-lawful-basis-review): A documented check that any personal data used to train, fine-tune, or ground an AI system has a valid lawful basis and is limited to what's necessary.
- [Factuality and hallucination verification](https://axiom-verity.com/controls/factuality-and-hallucination-verification): Automated or human fact-checking of AI-generated output against a trusted source before it is relied on for a consequential decision.
- [Human-in-the-loop review](https://axiom-verity.com/controls/human-in-the-loop-review): Mandatory human approval gate for high-consequence or irreversible AI-generated actions before they take effect.
- [Important business service impact-tolerance mapping](https://axiom-verity.com/controls/important-business-service-impact-tolerance-mapping): A maintained map of where AI systems sit within important business services, with an AI-specific impact tolerance defined for each.
- [Inference usage quotas and rate limiting](https://axiom-verity.com/controls/inference-usage-quotas-and-rate-limiting): Per-user and per-tenant limits on request volume, input length, and context-window usage for an AI system's inference endpoint.
- [Model artifact signing and integrity verification](https://axiom-verity.com/controls/model-artifact-signing-and-integrity-verification): Cryptographic signing of trusted model artifacts and verification of that signature before a model is loaded into production.
- [Output validation and guardrails](https://axiom-verity.com/controls/output-validation-guardrails): Deterministic schema validation and sanitisation of AI-generated output before it reaches any downstream system.
- [System prompt hardening](https://axiom-verity.com/controls/system-prompt-hardening): Design and testing practice that keeps system prompts free of secrets and resilient to extraction attempts.
- [Third-party model and vendor due diligence](https://axiom-verity.com/controls/third-party-model-due-diligence): Pre-procurement and ongoing due-diligence policy for the selection and contractual oversight of third-party AI models and datasets.
- [Training data provenance checks](https://axiom-verity.com/controls/training-data-provenance-checks): Documented verification of the source, licensing, and integrity of any dataset before it is used for training or fine-tuning.
- [Vector store and embedding access controls](https://axiom-verity.com/controls/vector-store-access-controls): Access control, input validation, and tenant isolation applied to a RAG pipeline's vector store, the same as any other production data store.
- [Vendor exit strategy and concentration management](https://axiom-verity.com/controls/vendor-exit-strategy-and-concentration-management): A documented, costed exit plan and an actively managed concentration limit for any critical AI vendor dependency.

## Tools Library (4 tools)

- [Tools Library index](https://axiom-verity.com/tools): Downloadable AI governance practitioner checklists, templates, and working documents.
- [AI Risk Appetite Statement Template](https://axiom-verity.com/tools/ai-risk-appetite-statement-template): A fill-in template and companion tracker for setting AI risk appetite across six risk dimensions and four decision-consequence tiers, with a measurable tolerance threshold and a pre-agreed trigger in every cell.
- [AI Vendor Due Diligence Questionnaire](https://axiom-verity.com/tools/ai-vendor-due-diligence-questionnaire): An assessor's questionnaire built on what the EU AI Act already obliges a provider to hold, covering Article 25 on the value chain and the Annex XII documentation duty in force since August 2025, instead of a security questionnaire with "AI" added to the headings.
- [Fundamental Rights Impact Assessment Worksheet](https://axiom-verity.com/tools/fundamental-rights-impact-assessment-worksheet): An EU AI Act Article 27 worksheet structured against the six elements the Article actually requires, with the scope determination most assessments skip and the regulator notification most templates omit.
- [LLM Security Control Checklist](https://axiom-verity.com/tools/llm-security-control-checklist): Forty-two concrete controls across the ten OWASP GenAI LLM Top 10 2026 categories, with status, evidence and review tracking, plus a content-source inventory that defines the scope of any injection test.

## Glossary (120 terms)

- [Glossary index](https://axiom-verity.com/glossary): Searchable glossary of AI governance, compliance, and security terminology.
- [A/B Testing](https://axiom-verity.com/glossary/a-b-testing): An experimental method that compares two versions of a system (A and B) by sending each to a random subset of users, to establish which performs better.
- [Accountability](https://axiom-verity.com/glossary/accountability): The principle that individuals and organisations can be held responsible for decisions made by the AI systems they deploy, and must be able to explain those decisions to the people affected.
- [Adversarial Attack](https://axiom-verity.com/glossary/adversarial-attack): An attempt to manipulate or deceive an AI model by deliberately crafting inputs designed to cause misclassification or unintended behaviour.
- [Agent Goal Hijack](https://axiom-verity.com/glossary/agent-goal-hijack): Redirecting an AI agent's objective or plan, usually by placing instructions in content the agent reads, so that it pursues the attacker's goal with its own tools and permissions. It is the agentic form of prompt injection and the first item in OWASP's Top 10 for Agentic Applications (ASI01).
- [Agent2Agent Protocol (A2A)](https://axiom-verity.com/glossary/agent2agent-protocol): An open protocol for AI agents from different vendors and frameworks to discover each other, describe what they can do and delegate tasks to one another. Google launched it in April 2025 and donated it to the Linux Foundation in June 2025.
- [Agentic AI](https://axiom-verity.com/glossary/agentic-ai): An AI system that pursues a goal by deciding for itself which actions to take, in what order, using tools and other systems it has been given access to, with a person checking some or none of the steps. The governance change is that the unit of risk stops being a single output and becomes a sequence of actions with consequences.
- [AI Alignment](https://axiom-verity.com/glossary/ai-alignment): The property of an AI system pursuing the goal its designers actually intended, rather than a proxy that happened to score well during training. Alignment is not the same as capability (a highly capable system can be confidently, efficiently misaligned) and it is not the same as AI safety, which is the broader field alignment sits inside.
- [AI Assurance](https://axiom-verity.com/glossary/ai-assurance): The work of measuring, evaluating and communicating whether an AI system, or the organisation running it, meets the claims made for it, so that others can rely on those claims. It spans internal testing through to independent audit and certification, and the question it answers is not whether a system is good but whether a statement about it can be trusted.
- [AI Bill of Materials](https://axiom-verity.com/glossary/ai-bill-of-materials): A machine-readable inventory of what an AI system is built from: models, datasets, adapters, software dependencies, and their provenance and licences. It extends the software bill of materials to the components that make AI systems different.
- [AI Cyber Security Code of Practice](https://axiom-verity.com/glossary/ai-cyber-security-code-of-practice): The UK government's voluntary code, published in January 2025, setting 13 principles for securing AI systems across five lifecycle phases. It became the basis of a European standard, ETSI EN 304 223, released in January 2026.
- [AI Evaluation](https://axiom-verity.com/glossary/ai-evaluation): Systematic measurement of what an AI model or system can do, how it tends to behave and how well its safeguards hold, using benchmarks, task-based tests, red teaming or studies with people. Evaluations are the evidence most AI claims rest on, and they have known weaknesses.
- [AI Incident](https://axiom-verity.com/glossary/ai-incident): An event in which the development, use or malfunction of an AI system directly or indirectly leads to harm to people's health, critical infrastructure, rights, property, communities or the environment. An AI hazard is an event that could plausibly have led to one. Reporting regimes in the EU and several US states now attach legal deadlines to the most serious cases.
- [AI Inventory](https://axiom-verity.com/glossary/ai-inventory): A maintained register of every AI system, model and AI-enabled feature an organisation uses, builds or buys, with an owner, a purpose, a risk tier and a status for each. It is the first control in every AI governance framework because none of the others can be applied to a system nobody has listed.
- [AI Literacy](https://axiom-verity.com/glossary/ai-literacy): The skills, knowledge and understanding that let a person use, oversee or be affected by an AI system with an informed view of what it can do, where it fails and what it means for them. Article 4 of the EU AI Act has required providers and deployers to act on it since 2 February 2025; since the 2026 Digital Omnibus the duty is to take measures to support AI literacy among their staff, rather than to ensure a sufficient level of it.
- [AI Office](https://axiom-verity.com/glossary/ai-office): The European Commission body that supervises general-purpose AI models under the EU AI Act and coordinates the Act's enforcement with national authorities. Its powers to enforce the GPAI obligations, including fines, have applied since 2 August 2026.
- [AI Red Teaming](https://axiom-verity.com/glossary/ai-red-teaming): Structured adversarial testing in which testers try to make an AI system fail, either by producing harmful or prohibited output (safety red teaming) or by being subverted to act against its owner (security red teaming). It finds failures an ordinary test suite cannot, because the tester's job is to imagine what the designers did not.
- [AI Regulatory Sandbox](https://axiom-verity.com/glossary/ai-regulatory-sandbox): A controlled environment in which a firm can develop or test an AI system under a regulator's supervision, with guidance on how the rules apply. The EU AI Act requires every member state to have at least one by 2 August 2027; in the UK, the FCA runs AI-specific testing schemes.
- [AI Safety](https://axiom-verity.com/glossary/ai-safety): The field of research and practice concerned with making AI systems behave as intended, avoid unintended harm and stay under meaningful human control. In 2025 the UK and US government institutes renamed themselves around security, which narrowed the official focus without removing the questions an organisation deploying AI has to answer.
- [AI Supply Chain Security](https://axiom-verity.com/glossary/ai-supply-chain-security): Assuring the integrity and provenance of everything an AI system is built from and depends on: pre-trained models and weights, adapters, datasets, libraries, serving infrastructure, plug-ins and tools. AI adds components that ordinary software supply-chain controls were not built to check.
- [AI TPRM (AI Third-Party Risk Management)](https://axiom-verity.com/glossary/ai-tprm): A phrase with two meanings that are often confused: using AI tools to run a third-party risk management (TPRM) programme, and managing the third-party risk that comes from buying AI. Both are live concerns for the same function, and the tool used for the first is an instance of the second.
- [AI TRiSM (AI Trust, Risk and Security Management)](https://axiom-verity.com/glossary/ai-trism): Gartner's name for the discipline of making AI systems trustworthy, governable and secure in operation: governance, runtime inspection and enforcement, information governance, and the infrastructure underneath. It is a market category and a vocabulary rather than a standard, and most of what it names already has a home in ISO/IEC 42001, the NIST AI RMF and the OWASP LLM Top 10.
- [Audit Trail](https://axiom-verity.com/glossary/audit-trail): A complete, timestamped record of the actions an AI system took and the decisions made with it, so that they can be reviewed, explained and attributed afterwards.
- [Automated Decision-Making](https://axiom-verity.com/glossary/automated-decision-making): A decision about a person taken by technological means without meaningful human involvement. In UK data protection law, Articles 22A to 22D of the UK GDPR (inserted by the Data (Use and Access) Act 2025 and in force since 5 February 2026) allow significant automated decisions, but only with safeguards: information, the right to make representations, human intervention and the right to contest.
- [Automation Bias](https://axiom-verity.com/glossary/automation-bias): The tendency of people to over-rely on an automated system's output, accepting it without enough scrutiny even when other information suggests it is wrong. It is the main reason human oversight fails in practice, and the EU AI Act names it explicitly.
- [Chain-of-Thought Monitoring](https://axiom-verity.com/glossary/chain-of-thought-monitoring): Reading the intermediate reasoning a reasoning model produces before its answer, to detect intent to misbehave. Researchers across several AI developers have called it a valuable but fragile safety opportunity, because training choices can make the reasoning less legible or less honest.
- [Conformity Assessment](https://axiom-verity.com/glossary/conformity-assessment): The procedure by which the provider of a high-risk AI system shows that it meets the EU AI Act's requirements before placing it on the market. For most high-risk systems it is an internal control the provider carries out itself; for some biometric systems it needs a notified body.
- [Confusion Matrix](https://axiom-verity.com/glossary/confusion-matrix): A table comparing a classifier's predictions with the actual outcomes, from which precision, recall and false positive and false negative rates are calculated.
- [Consumer Duty](https://axiom-verity.com/glossary/consumer-duty): The FCA's outcomes-based standard requiring firms to deliver good outcomes for retail customers, in force since July 2023. It sets no AI-specific rules but applies in full to AI-driven customer journeys, pricing, advice and communications.
- [Content Provenance](https://axiom-verity.com/glossary/content-provenance): Verifiable information about where a piece of content came from and how it was made or changed, attached to the content itself. It is the technical basis for labelling AI-generated material, through standards such as C2PA Content Credentials and through watermarking.
- [Contestability](https://axiom-verity.com/glossary/contestability): The ability of a person affected by an AI-assisted decision to challenge it and have it reconsidered by someone with the authority to change it. It is one of the safeguards UK data protection law now attaches to significant automated decisions.
- [Context Window](https://axiom-verity.com/glossary/context-window): The maximum length of text (measured in tokens) that a language model can consider at once when generating a response.
- [Control Objective](https://axiom-verity.com/glossary/control-objective): A specific, measurable statement of what a control is meant to achieve against a named risk. It describes the outcome, not the mechanism, and it is only useful if a failure to meet it would be detectable.
- [Counterfactual Explanation](https://axiom-verity.com/glossary/counterfactual-explanation): An explanation of an automated decision that states the smallest change to the inputs that would have changed the outcome, such as an application that would have been approved with a higher declared income. It tells the affected person what would have made the difference without exposing the model.
- [Critical Third Parties Regime](https://axiom-verity.com/glossary/critical-third-parties-regime): The UK regime under which HM Treasury can designate a third party whose services are critical to the financial sector, bringing it under direct oversight by the Bank of England, the PRA and the FCA. Its rules have applied since 1 January 2025, and the first designations, in July 2026, were four cloud providers.
- [Data Poisoning](https://axiom-verity.com/glossary/data-poisoning): An attack in which an adversary deliberately corrupts data an AI model learns from, to degrade it or plant hidden behaviour. It now reaches beyond training data to fine-tuning sets, retrieval corpora and agent memory, and a poisoned model can pass ordinary evaluation.
- [Data Quality](https://axiom-verity.com/glossary/data-quality): The degree to which data is accurate, complete, consistent and representative of what it is meant to measure.
- [Deepfake](https://axiom-verity.com/glossary/deepfake): AI-generated or manipulated image, audio or video content that resembles real people, objects, places or events and would falsely appear to a person to be authentic. For an organisation, the most immediate risk is impersonation fraud: a cloned voice or a faked video call instructing a payment.
- [Differential Privacy](https://axiom-verity.com/glossary/differential-privacy): A mathematical framework for limiting what can be learned about any individual from an algorithm's output, by adding carefully calibrated noise so that the output barely changes whether or not that individual's data was included.
- [Digital Omnibus on AI](https://axiom-verity.com/glossary/digital-omnibus-on-ai): Regulation (EU) 2026/1744, the 2026 law that amended the EU AI Act. It deferred the high-risk obligations to December 2027 and August 2028, softened the AI literacy duty, added two prohibited practices, and extended documentation concessions to small mid-cap companies.
- [Embedding](https://axiom-verity.com/glossary/embedding): A numerical representation of text, images or other data in a high-dimensional space, where items with similar meaning sit close together.
- [Emergent Capabilities](https://axiom-verity.com/glossary/emergent-capabilities): Abilities that appear in larger AI models but were not explicitly trained for and were not evident in smaller versions. Whether they genuinely arrive suddenly with scale, or only look sudden because of how they are measured, is disputed.
- [EU AI Act](https://axiom-verity.com/glossary/eu-ai-act): The European Union's regulation of AI systems by risk category, in force since 1 August 2024 and applying in stages. As of October 2026 the prohibitions, the AI literacy duty, the general-purpose model obligations and the Article 50 transparency duties all apply. The 2026 Digital Omnibus, Regulation (EU) 2026/1744, deferred the high-risk obligations to December 2027 and August 2028, softened the literacy duty and added two prohibitions. It reaches any organisation whose AI output is used in the EU, wherever the organisation sits.
- [Excessive Agency](https://axiom-verity.com/glossary/excessive-agency): Giving an LLM-based system more functions, permissions or autonomy than its task needs, so that a manipulated or mistaken output becomes a damaging action. It is LLM03 in the OWASP Top 10 for LLM Applications 2026, up from sixth place in 2025.
- [Explainability](https://axiom-verity.com/glossary/explainability): The ability to give a person an account of why a system produced a particular output, in terms they can act on. Explainability is a property of the account and its audience, not of the model: an opaque model can be explainable, and an interpretable one can fail to be explained.
- [Fairness & Bias](https://axiom-verity.com/glossary/fairness-and-bias): The principle and practice of ensuring AI systems do not produce systematically different outcomes for individuals or groups because of protected characteristics or other vulnerable attributes. Algorithmic fairness is the same idea expressed as measurable criteria.
- [Fairness Metrics](https://axiom-verity.com/glossary/fairness-metrics): Quantitative tests of whether a model's outcomes or errors differ across groups. The main families are demographic parity, equalised odds and calibration, and outside special cases they cannot all be satisfied at once, so choosing which to use is a governance decision rather than a technical default.
- [Feature Attribution](https://axiom-verity.com/glossary/feature-attribution): Methods that estimate how much each input contributed to a model's prediction, such as SHAP and LIME. They are standard in model validation and explanation, but they approximate the model rather than describe it, and they can be manipulated.
- [Federated Learning](https://axiom-verity.com/glossary/federated-learning): A machine learning approach in which a model is trained across many separate data sources without the data itself being gathered in one place.
- [Fine-Tuning](https://axiom-verity.com/glossary/fine-tuning): The process of adapting a pre-trained model to a specific task or domain by training it further on a smaller, task-specific dataset.
- [Frontier AI](https://axiom-verity.com/glossary/frontier-ai): The most capable general-purpose AI models, at or near the leading edge. There is no single definition: the EU AI Act presumes systemic risk above 10^25 training operations, California's SB 53 and New York's RAISE Act define frontier models above 10^26, and the International AI Safety Report prefers the term general-purpose AI.
- [Frontier Safety Framework](https://axiom-verity.com/glossary/frontier-safety-framework): A frontier AI developer's published commitment to test its models for dangerous capabilities and to apply stronger safeguards, or pause, when defined capability thresholds are crossed. Examples include Anthropic's Responsible Scaling Policy, OpenAI's Preparedness Framework and Google DeepMind's Frontier Safety Framework.
- [General-Purpose AI Model](https://axiom-verity.com/glossary/general-purpose-ai-model): An AI model that displays significant generality and can competently perform a wide range of distinct tasks, however it is placed on the market. Under the EU AI Act its provider has had documentation, copyright and downstream-information duties since 2 August 2025, and a model with systemic risk carries evaluation, adversarial testing, incident reporting and cybersecurity duties on top.
- [Governance Framework](https://axiom-verity.com/glossary/governance-framework): A structured set of policies, roles and controls that guides an organisation's decisions about AI and its oversight of AI systems.
- [Guardrails](https://axiom-verity.com/glossary/guardrails): Runtime checks on an AI system's inputs and outputs, such as classifiers, filters and policy rules, that block, modify or flag content before it reaches the model or after it leaves. They reduce risk, but they are probabilistic: a guardrail is a layer of defence, not a security boundary.
- [Hallucination](https://axiom-verity.com/glossary/hallucination): Output that is fluent, confident and false. It arises from the same mechanism that produces correct output (a model predicting plausible continuations), which is why it is best understood as a property of the technology to be bounded rather than a defect to be fixed.
- [Hidden Context Exposure](https://axiom-verity.com/glossary/hidden-context-exposure): The exposure of content an AI application placed in a model's context on the assumption users could not see it: system prompts, retrieved documents, schemas, policy logic or other users' data. It is LLM08 in the OWASP Top 10 for LLM Applications 2026, which renamed and broadened the 2025 item system prompt leakage.
- [High-Risk AI System](https://axiom-verity.com/glossary/high-risk-ai-system): An AI system that the EU AI Act subjects to its full set of requirements, either because it is a safety component of a product covered by EU product law (Annex I) or because it is used in one of eight listed areas (Annex III), which include credit scoring and life and health insurance pricing. Since the 2026 Digital Omnibus the obligations apply from 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems.
- [Human-in-the-Loop (HITL)](https://axiom-verity.com/glossary/human-in-the-loop): A design in which a person reviews, approves or can override what an AI system does before it takes effect. It is only a control if that person has the competence, information, time and authority to disagree, which is why the EU AI Act requires oversight to be designed in and UK data protection law asks whether human involvement is meaningful.
- [Impact Assessment](https://axiom-verity.com/glossary/impact-assessment): A systematic evaluation of the potential positive and negative effects of deploying an AI system on individuals, communities and society. Distinct from a data protection impact assessment, which examines the lawfulness of the data processing rather than the consequences of the decision.
- [Impact Tolerance](https://axiom-verity.com/glossary/impact-tolerance): The maximum tolerable level of disruption to an important business service, set under the UK operational resilience regime. Applied to AI, it asks how long a firm can run without an AI component, or with one that is wrong, before customers or markets suffer intolerable harm.
- [Incident Response](https://axiom-verity.com/glossary/incident-response): A structured process for detecting, assessing, containing and resolving adverse events involving AI systems, such as model failure, a security breach or a fairness harm.
- [ISO/IEC 23894](https://axiom-verity.com/glossary/iso-iec-23894): International guidance, published in 2023, on managing risk related to AI. It applies the general risk management principles of ISO 31000 to AI and adds what is specific to AI systems.
- [ISO/IEC 42001](https://axiom-verity.com/glossary/iso-iec-42001): The international standard for an AI management system (AIMS): the policies, roles, risk assessment, impact assessment, controls and evidence an organisation puts around its development and use of AI so that the whole can be audited and certified. Published December 2023, and since 2025 surrounded by companion standards for impact assessment (42005) and for the bodies that certify against it (42006), and from 2026 by EN 18286, the European quality management standard written for the EU AI Act.
- [ISO/IEC 42005](https://axiom-verity.com/glossary/iso-iec-42005): The international standard, published in May 2025, giving guidance on how to carry out an AI system impact assessment: identifying the people and groups a system affects and assessing its potential benefits and harms. It is the method ISO/IEC 42001 assumes.
- [Jailbreak](https://axiom-verity.com/glossary/jailbreak): A technique for inducing a language model to produce output that its safety training was intended to refuse. It targets the model's alignment rather than any particular application, and works substantially the same way wherever that model is deployed.
- [Large Language Model (LLM)](https://axiom-verity.com/glossary/large-language-model): A deep learning model trained on very large amounts of text, able to generate and interpret human-like language across a wide range of topics and tasks.
- [Lethal Trifecta](https://axiom-verity.com/glossary/lethal-trifecta): Simon Willison's name for the combination of three capabilities that makes an AI agent exploitable for data theft: access to private data, exposure to untrusted content, and the ability to communicate externally. An agent with all three can be made to send private data to an attacker.
- [Loss of Control](https://axiom-verity.com/glossary/loss-of-control): Scenarios in which one or more AI systems come to operate outside anyone's control, with no clear way to regain it. It is among the risks the International AI Safety Report covers, and the one on which expert views differ most.
- [Machine Learning](https://axiom-verity.com/glossary/machine-learning): A branch of artificial intelligence in which systems learn patterns from data rather than being explicitly programmed for every scenario.
- [Mechanistic Interpretability](https://axiom-verity.com/glossary/mechanistic-interpretability): Research that tries to reverse-engineer what a neural network computes internally, identifying the features it represents and the circuits that connect them, in terms a person can understand. It is a research field, not yet an audit tool.
- [Membership Inference](https://axiom-verity.com/glossary/membership-inference): An attack that works out whether a particular person's record was in a model's training data by examining how the model responds. Showing that someone's data was used can itself be a disclosure of personal data.
- [Memory Poisoning](https://axiom-verity.com/glossary/memory-poisoning): Corrupting what an AI agent remembers or retrieves, so that later sessions, possibly other users' sessions, behave as the attacker wants. Unlike a one-off prompt injection it persists, and OWASP lists it among the top risks for agentic applications (ASI06).
- [MITRE ATLAS](https://axiom-verity.com/glossary/mitre-atlas): MITRE's knowledge base of the tactics and techniques adversaries use against AI systems, modelled on the MITRE ATT&CK framework for conventional cyber attacks and built from real-world case studies and red-team findings.
- [Model Backdoor](https://axiom-verity.com/glossary/model-backdoor): Hidden behaviour planted in an AI model that stays dormant until a specific trigger appears in the input, then makes the model act as the attacker intends. A backdoored model can behave normally on every ordinary test.
- [Model Card](https://axiom-verity.com/glossary/model-card): A document describing a machine learning model's intended uses, performance, limitations and known risks, so that people deciding whether and how to use it can do so on evidence. System cards extend the idea to the deployed system and its safeguards, and EU AI Act documentation duties now require much of the same content.
- [Model Collapse](https://axiom-verity.com/glossary/model-collapse): The degradation in quality and diversity that occurs when language models are trained, generation after generation, on content generated by other language models.
- [Model Context Protocol (MCP)](https://axiom-verity.com/glossary/model-context-protocol): An open protocol for connecting AI applications to external tools and data through servers that expose tools, resources and prompts. It has become the standard way to give an AI agent access to other systems, which makes its trust model the agent's attack surface: every connected server can steer the model as well as serve it.
- [Model Deployment](https://axiom-verity.com/glossary/model-deployment): The process of putting a trained AI model into production, where it makes real-world predictions or decisions that affect users or business outcomes.
- [Model Drift](https://axiom-verity.com/glossary/model-drift): A change in the relationship between a model's inputs and the environment it was fitted to, such that the model's assumptions no longer hold. The model itself is unchanged; the world it was trained on has moved.
- [Model Extraction](https://axiom-verity.com/glossary/model-extraction): Reconstructing a model's behaviour or parameters by querying it repeatedly and learning from the answers, to copy its intellectual property or to build a surrogate for planning further attacks.
- [Model Interpretability](https://axiom-verity.com/glossary/model-interpretability): An intrinsic property of a model: the degree to which a person can understand the mechanism by which it turns inputs into outputs. A linear model or a shallow decision tree is interpretable by construction. A deep network is not, however good an explanation you can construct for it afterwards.
- [Model Inversion](https://axiom-verity.com/glossary/model-inversion): An attack that uses a model's outputs to reconstruct information about its training data, such as recovering a recognisable face from a facial recognition model, or sensitive attributes of the people in a dataset.
- [Model Monitoring](https://axiom-verity.com/glossary/model-monitoring): Continuous observation of a deployed model's inputs, outputs, performance and behaviour, to detect degradation, drift, misuse or security anomalies.
- [Model Risk Management (MRM)](https://axiom-verity.com/glossary/model-risk-management): The discipline of identifying, measuring, controlling and reporting the risk that a model is wrong, misused or misunderstood, across its whole life from development to retirement. It predates AI by decades in banking, and it is the frame UK regulators have chosen for AI in financial services: the PRA's SS1/23 principles have applied to AI models as to any other since May 2024, and were amended in April 2026.
- [Model Versioning](https://axiom-verity.com/glossary/model-versioning): The practice of recording which model, trained how and on what, was deployed when, enabling comparison, rollback and reproducibility. For systems built on models the organisation did not train, it extends to versioning the surrounding assembly, since that is the only part it controls.
- [Multimodal AI](https://axiom-verity.com/glossary/multi-modal-ai): AI systems that process and combine more than one type of data (text, images, audio, video) to understand and respond to the world.
- [Neural Network](https://axiom-verity.com/glossary/neural-network): A computational model loosely inspired by biological neurons, made of interconnected layers of artificial nodes that learn to recognise patterns in data.
- [NIST AI Risk Management Framework](https://axiom-verity.com/glossary/nist-ai-rmf): The US National Institute of Standards and Technology's voluntary framework for managing AI risk, built around four functions: Govern, Map, Measure and Manage. Version 1.0 was published in January 2023 and is being revised under the US government's July 2025 AI Action Plan. Since 1.0 NIST has added a Generative AI Profile, an adversarial machine learning taxonomy, a draft Cybersecurity Framework profile for AI, and work on control overlays and AI agent standards. None of it is mandatory, and most of it is used as a common vocabulary rather than a checklist.
- [Non-Human Identity](https://axiom-verity.com/glossary/non-human-identity): Any identity used by software rather than a person to authenticate to other systems, such as service accounts, API keys and tokens. AI agents now hold and use many of them, which makes agent identity one of the main access-control questions in AI governance.
- [OWASP Top 10 for Agentic Applications](https://axiom-verity.com/glossary/owasp-top-10-agentic): OWASP's list of the ten most critical security risks specific to autonomous AI agents, published in December 2025. It complements the OWASP Top 10 for LLM Applications by covering what only goes wrong when a model acts: hijacked goals, misused tools, poisoned memory and failures that cascade between agents.
- [OWASP Top 10 for LLM Applications](https://axiom-verity.com/glossary/owasp-top-10-llm): The OWASP GenAI Security Project's ranked list of the ten most critical security risks in applications built on large language models. The 2026 edition, published in August 2026, is the closest thing the field has to a shared vocabulary for LLM application security, and every item on it maps to an entry in this site's risk register.
- [Performance Metric](https://axiom-verity.com/glossary/performance-metric): A quantitative measure of how well an AI model or system performs a task, such as accuracy, precision, recall, F1 score or a domain-specific measure.
- [Performance Regression](https://axiom-verity.com/glossary/performance-regression): A decline in model performance over time, typically caused by shifts in the data, concept drift or changes in the environment the model operates in.
- [Post-Market Monitoring](https://axiom-verity.com/glossary/post-market-monitoring): A provider's documented, continuous collection and analysis of data on how its high-risk AI system performs once in use, so that problems are found and fixed throughout the system's life. It is the legal form of model monitoring under the EU AI Act.
- [PRA SS1/23](https://axiom-verity.com/glossary/pra-ss1-23): The Prudential Regulation Authority's supervisory statement setting out five principles for model risk management at UK banks. It took effect on 17 May 2024, applies to AI and machine learning models as to any other model, and was amended in April 2026.
- [Prohibited AI Practices](https://axiom-verity.com/glossary/prohibited-ai-practices): The uses of AI that Article 5 of the EU AI Act bans outright. Eight have applied since 2 February 2025, and the 2026 Digital Omnibus adds two more from 2 December 2026: systems that create non-consensual intimate material and systems that generate child sexual abuse material.
- [Prompt Engineering](https://axiom-verity.com/glossary/prompt-engineering): The practice of crafting and refining prompts to steer language models towards the outputs wanted, treating the prompt itself as a design problem.
- [Prompt Injection](https://axiom-verity.com/glossary/prompt-injection): An attack in which untrusted content reaching a language model is interpreted as instructions rather than data, causing the model to act on an attacker's intent instead of the operator's. It is a property of how the application assembles its context, not a flaw in any particular model.
- [Protected Characteristics](https://axiom-verity.com/glossary/protected-characteristics): The nine characteristics the Equality Act 2010 protects from discrimination in Great Britain: age, disability, gender reassignment, marriage and civil partnership, pregnancy and maternity, race, religion or belief, sex, and sexual orientation. They are the legal anchor for AI fairness in the UK.
- [Provider and Deployer](https://axiom-verity.com/glossary/provider-and-deployer): The two main roles the EU AI Act assigns obligations to. The provider develops an AI system or general-purpose model, or has one developed, and places it on the market or puts it into service under its own name. The deployer uses an AI system under its authority in a professional capacity. Which role an organisation holds decides which obligations it carries, and an organisation can move from one to the other without intending to.
- [Reinforcement Learning from Human Feedback (RLHF)](https://axiom-verity.com/glossary/rlhf): A training method in which people rate a model's outputs, a reward model learns to predict those ratings, and the model is then fine-tuned to produce outputs the reward model scores highly. It is the main technique that turned raw language models into usable assistants.
- [Retrieval-Augmented Generation (RAG)](https://axiom-verity.com/glossary/rag): A technique in which a language model retrieves relevant documents or data before generating a response, grounding its output in source material rather than in its training data alone.
- [Reward Hacking](https://axiom-verity.com/glossary/reward-hacking): A system meeting the literal objective it was trained or rewarded on in a way its designers did not intend, such as gaming a test rather than solving the task. Recent research shows the habit can generalise into broader misbehaviour.
- [Risk Appetite](https://axiom-verity.com/glossary/risk-appetite): The level and types of risk an organisation is willing to accept in pursuit of its strategic objectives. Applied to AI, a single organisational posture is rarely enough: appetite has to be set separately across distinct risk dimensions and tiers of decision consequence.
- [Safety Case](https://axiom-verity.com/glossary/safety-case): A structured argument, supported by evidence, that a system is acceptably safe for a given use in a given context. Long established in nuclear, aviation and defence engineering, safety cases are being adopted for frontier AI as a way to make safety claims explicit and checkable.
- [Sandbagging](https://axiom-verity.com/glossary/sandbagging): Strategic underperformance on an evaluation, so that a model appears less capable than it is. It can be done by the model itself or arranged by a developer, and either way it undermines the evidence that safety decisions rest on.
- [Scheming](https://axiom-verity.com/glossary/scheming): A model covertly pursuing goals its developers did not intend while concealing them, for example by deliberately underperforming, hiding information or working around oversight. It has been observed in controlled tests; how often it occurs in real deployments is not known.
- [Shadow AI](https://axiom-verity.com/glossary/shadow-ai): The use of AI tools, models or features inside an organisation without the knowledge, approval or oversight of whoever is responsible for governing them. It is the AI-era form of shadow IT, with one difference that matters: the tool is usually free, browser-based, and takes the organisation's data as its input.
- [Slopsquatting](https://axiom-verity.com/glossary/slopsquatting): Registering software package names that AI coding assistants tend to invent, so that a developer who installs a suggested but non-existent package installs the attacker's code instead.
- [Sycophancy](https://axiom-verity.com/glossary/sycophancy): A model's tendency to tell users what they want to hear, agreeing, flattering or validating at the expense of accuracy. It is a side effect of training on human approval, and in customer-facing financial services it is a conduct risk as well as a quality one.
- [Synthetic Data](https://axiom-verity.com/glossary/synthetic-data): Artificially generated data, created by algorithms or models, used instead of or alongside real data in training and testing.
- [Testing AI Systems](https://axiom-verity.com/glossary/testing-ai-systems): Systematic evaluation of AI models and systems to verify that they perform as intended, identify failure modes and assess risks.
- [Third-Party Risk](https://axiom-verity.com/glossary/third-party-risk): The exposure an organisation carries through dependencies on external parties who build, train, host or operate the AI systems it uses. In AI the exposure is characteristically deeper than the contract: beneath most vendors sits a model provider, and beneath that an infrastructure provider.
- [Three Lines of Defence](https://axiom-verity.com/glossary/three-lines-of-defence): The risk governance model used across UK financial services, in which the business owns and manages risk, an independent risk and compliance function oversees it, and internal audit gives the board independent assurance. Applied to AI, it decides who owns a model, who challenges it and who checks both.
- [Token](https://axiom-verity.com/glossary/token): A unit of text that a language model processes, typically a word, part of a word or a character. Models work on sequences of tokens, not raw text.
- [Training Data](https://axiom-verity.com/glossary/training-data): The dataset used to train a machine learning model, from which the model learns patterns and relationships.
- [Transformer](https://axiom-verity.com/glossary/transformer): A deep learning architecture that uses self-attention to process sequences of data, and the foundation of modern large language models.
- [Transparency](https://axiom-verity.com/glossary/transparency): The disclosure of information about an AI system to the people who need it: that AI is being used, what it does, what it is for and what risks it poses, in terms they can understand. For people whose personal data it uses, being informed is also a legal right.
- [Trustworthy AI](https://axiom-verity.com/glossary/trustworthy-ai): The umbrella term for AI that people can justifiably rely on. The most widely cited definition lists its properties: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair.
- [Uplift](https://axiom-verity.com/glossary/uplift): The additional ability an AI system gives a person to do something, compared with what they could do without it, for example using search engines and textbooks. Uplift is the measure most frontier safety thresholds use for dangerous capabilities.

## Articles (10 articles)

- [Articles index](https://axiom-verity.com/articles): Original, long-form analysis on AI governance, compliance, and security from Andrew Moore.
- [Articles RSS feed](https://axiom-verity.com/articles/rss.xml): RSS 2.0 feed of the latest articles, for polling new content ahead of a full re-crawl.
- [AI in Third-Party Risk Management: What It Automates, and What It Cannot](https://axiom-verity.com/articles/ai-in-third-party-risk-management): Most of what AI does in third-party risk management is labour: reading evidence, pre-filling questionnaires, watching for change. That is worth having. What it cannot do is accept a risk, and the model inside your TPRM tool is itself a third party.
- [Deterministic Work Belongs Either Side of the Model Call](https://axiom-verity.com/articles/deterministic-work-either-side-of-the-model-call): There is a reflex that sends the whole problem to the model, and a correction that tries to take work away from it. Both fail, and working through a real standard shows exactly where each one breaks. The deterministic work is real and it belongs on either side of the call rather than in place of it.
- [Model Drift, Performance Regression and Versioning: Telling Them Apart](https://axiom-verity.com/articles/model-drift-regression-and-versioning): Three terms, three different things, and conflating them means monitoring the wrong one. Drift is the world moving. Regression is a measurement getting worse. Versioning is the artefact changing. For most LLM deployments the dominant cause of behaviour change is none of the classic ones. It is the model provider shipping an update you were not told about.
- [LLM Governance: What It Actually Means Operationally](https://axiom-verity.com/articles/llm-governance-operationally): Most AI governance programmes produce a policy, a committee and a register, then discover none of it touches the thing that actually changes: the assembly of model version, prompt, retrieval corpus and tool permissions that determines what the system does. LLM governance is an operations problem wearing a policy costume.
- [Prompt Injection and System Prompt Hardening: What Actually Works](https://axiom-verity.com/articles/prompt-injection-and-system-prompt-hardening): System prompt hardening is worth doing and will not save you. The achievable goal is not a model that cannot be tricked, since natural language offers no way to build that. It is a system where being tricked does not matter much. OWASP's 2026 revision makes the same point: the biggest mover was not the attack, it was what the model was allowed to do afterwards.
- [Third-Party AI Risk: Model Provenance and Vendor Due Diligence](https://axiom-verity.com/articles/third-party-ai-risk-provenance-and-due-diligence): Most AI vendor questionnaires are security questionnaires with "AI" added to the headings. The useful version asks for the specific documentation the EU AI Act already obliges the provider to hold, and establishes in writing whether your own configuration has quietly made you the provider.
- [AI Governance in UK Financial Services: The Regimes That Already Apply](https://axiom-verity.com/articles/ai-governance-uk-financial-services): The UK has decided not to write an AI Act for financial services. That is not the relief it sounds like. Five existing regimes already bind AI in regulated firms, one of them has required a board-approved model risk appetite since May 2024, and UK lenders and insurers serving EU customers sit inside the EU AI Act whatever Westminster does.
- [AI Impact Assessments: DPIA, FRIA and AIA Compared](https://axiom-verity.com/articles/ai-impact-assessments-dpia-fria-compared): Three different assessments answer to the phrase "AI impact assessment", and only one of them may actually bind your organisation. The EU AI Act's fundamental rights impact assessment reaches a narrower set of deployers than most coverage suggests. For those inside it, completing the assessment is only half the obligation, because Article 27 also requires you to file it with a regulator.
- [Setting an AI Risk Appetite: A Practical Framework for Boards](https://axiom-verity.com/articles/setting-an-ai-risk-appetite): Most organisations already have a risk appetite statement. Almost none of them survives contact with an AI system, because one sentence about "low appetite for operational risk" cannot tell you whether a 4% error rate is acceptable. A usable AI risk appetite is a matrix, and every threshold in it needs a trigger attached.
- [Undisclosed AI Interaction: The Article 50 Obligation That Was Not Delayed](https://axiom-verity.com/articles/undisclosed-ai-interaction-article-50): The Digital Omnibus deferred the AI Act's high-risk obligations by more than a year. It left Article 50 alone. The transparency duties that took effect on 2 August 2026 apply now, to almost every organisation running a chatbot, a synthetic voice agent, or a generative content pipeline, and they carry fines of up to €15 million or 3% of worldwide turnover.

## News (6 articles)

- [News index](https://axiom-verity.com/news): Timely updates on AI governance, compliance, and security developments.
- [News RSS feed](https://axiom-verity.com/news/rss.xml): RSS 2.0 feed of the latest news articles, for polling new content ahead of a full re-crawl.
- [ASD publishes its position on AI and Simplified Technical English](https://axiom-verity.com/news/asd-stemg-white-paper-ste-and-ai): The ASD Simplified Technical English Maintenance Group has issued a white paper on AI and ASD-STE100. It accepts AI as a support for STE writing and translation, on three conditions: the standard takes priority, a person stays responsible for the text, and AI-assisted content is declared with its limits stated. It endorses no AI tool.
- [EU AI Act Omnibus: Major compliance deadline changes and SME relief announced](https://axiom-verity.com/news/eu-ai-act-omnibus-compliance-updates): The European Commission published the Digital Omnibus on AI in November 2025, introducing simplified implementation measures, extended SME benefits with lighter documentation, and postponed high-risk deadlines. Critical for organizations planning 2026 compliance strategies.
- [ISO/IEC 42001 vs NIST AI RMF: Which framework is right for your organization?](https://axiom-verity.com/news/iso-42001-vs-nist-rmf-comparison): In-depth comparison of ISO/IEC 42001 (international certifiable standard) and NIST AI RMF (U.S. voluntary framework). Includes practical guidance on alignment, crosswalk mapping, and hybrid implementation strategies.
- [NIST AI Risk Management Framework: Complete guide for U.S. organizations](https://axiom-verity.com/news/nist-ai-risk-management-framework-guide): Comprehensive overview of the NIST AI RMF March 2025 update addressing generative AI risks, supply chain vulnerabilities, and third-party model assessment. Essential reading for organizations building trustworthy AI systems.
- [The time for AI governance is now: Key considerations and guidelines](https://axiom-verity.com/news/time-for-ai-governance-isaca-2025): ISACA industry analysis on why 2025 is the critical year for AI governance maturity. Explores organizational readiness, control patterns, board-level oversight, and the urgency of formalized AI governance programs across sectors.
- [Artificial intelligence in 2025: The new foundation for security and compliance](https://axiom-verity.com/news/artificial-intelligence-security-compliance-2025): Secureframe analysis of how AI is becoming a critical enabler for managing risk, improving audit readiness, and maintaining continuous compliance. Examines AI adoption trends, regulatory drivers, and implementation best practices.

## Technology (3 briefings)

- [Technology index](https://axiom-verity.com/technology): Governance briefings on the AI platforms and tools organisations are actually adopting.
- [Claude Code](https://axiom-verity.com/technology/claude-code): An agentic command-line and IDE-integrated coding tool that gives Claude models direct, bounded access to a local file system, shell, and developer tooling to read, write, test, and run code under developer direction.
- [Microsoft Copilot](https://axiom-verity.com/technology/microsoft-copilot): Microsoft's Copilot brand spans several materially different products — GitHub Copilot (coding), Microsoft 365 Copilot (embedded in Word, Excel, Teams, and Outlook), and Copilot Studio (building custom agents) — each with its own data-handling and governance profile. Treating "Copilot" as one thing is itself a governance risk.
- [Gemini (Google)](https://axiom-verity.com/technology/google-gemini): Google's Gemini model family reaches organisations through several distinct routes — the consumer/enterprise Gemini app, Gemini embedded in Google Workspace (Docs, Sheets, Gmail), and Vertex AI, Google Cloud's enterprise platform for building on Gemini directly — each with a different data-processing and governance posture, the same pattern as Microsoft's Copilot family.

## Projects (6 entries)

- [Projects index](https://axiom-verity.com/projects): Research programmes, instruments and public code, each with notes connecting it back to the Library.
- [Responsible AI by design](https://axiom-verity.com/research/responsible-ai): Two conceptual models for governing AI in a regulated organisation: a domain map of what must hold true and a lifecycle of the gates each use case passes, built on a UK retail bank and written to generalise.
- [Minimum Humans to Run a Bank](https://axiom-verity.com/thought-experiment): An evolving public investigation into whether a small human core can effectively govern and steward a UK-regulated retail bank, with AI agents performing the rest.
- [ASD-STE100 Simplified Technical English](https://axiom-verity.com/ste): All 53 ASD-STE100 rules, the prompts behind the checker published in full, and a worked example.
- [craft-a-bot](https://axiom-verity.com/projects/craft-a-bot): An interactive, browser-based agent simulator: assemble modular bricks — brain, sensors, memory, tools, actuators, safety — on a virtual workbench to build agents that pursue goals in a simulated environment, with a six-chapter tutorial covering observation, memory, tool use, hallucination, and human oversight.
- [Algorithmic Impact Assessment Tool](https://axiom-verity.com/projects/aia-eia-js): A survey-based Algorithmic Impact Assessment / Évaluation de l'incidence algorithmique tool implementing the Government of Canada's Directive on Automated Decision-Making — bilingual (English/French), with weighted scoring that combines raw impact factors with mitigations into a final assessment tier.
- [Model-Card](https://axiom-verity.com/projects/model-card): A standards-anchored template system for documenting large language models: a single YAML master source generates synchronised Word, Excel, Markdown, and JSON outputs, with role-based guidance for developers, reviewers, auditors, and downstream integrators.

## ASD-STE100 Simplified Technical English (53 rules, 10 prompts)

- [ASD-STE100 section](https://axiom-verity.com/ste): A free rule-cited checker for ASD-STE100 Simplified Technical English: all 53 rules, the ten prompts behind the checker published in full, and a worked example.
- [ASD-STE100 rules index](https://axiom-verity.com/ste/rules): All 53 ASD-STE100 rules, filterable by section and severity.
- [ASD-STE100 prompts index](https://axiom-verity.com/ste/prompts): The ten system prompts behind the checker, published in full. Six work standalone with any model.
- [ASD-STE100 dictionary lookup](https://axiom-verity.com/ste/dictionary): Look up whether a word is approved in ASD-STE100, and its approved alternative. Placeholder pending live search.
- [ASD-STE100 guides index](https://axiom-verity.com/ste/guides): Two explainers: what ASD-STE100 actually requires, and how to get a language model to write in it.
- [How to Make an LLM Write in Simplified Technical English](https://axiom-verity.com/ste/guides/llm-prompt-for-simplified-technical-english): Pasting all 53 STE rules into a prompt and asking for a rewrite does not work well. What does: dictionary resolution first, then checking by rule section.
- [ASD-STE100 Explained: Simplified Technical English](https://axiom-verity.com/ste/guides/what-is-asd-ste100): What ASD-STE100 requires, rule by rule, with worked examples for all 53 rules, and why the controlled dictionary matters more than the rules. No account needed.
- [Rule 1.1: Use approved words only](https://axiom-verity.com/ste/rules/1.1): Every word in the text must be either an approved word from the STE dictionary, a technical noun that fits one of the 22 approved technical-noun categories (rule 1.5), or a technical verb that fits one of the 4 approved technical-verb categories (rule 1.12).
- [Rule 1.2: Use approved words as the specified part of speech](https://axiom-verity.com/ste/rules/1.2): Use an approved dictionary word only as the part of speech (noun, verb, adjective, etc.) that the dictionary specifies for it; do not convert it to a different part of speech.
- [Rule 1.3: Use approved words with their approved meanings only](https://axiom-verity.com/ste/rules/1.3): Use each approved dictionary word only with the approved meaning or meanings the STE dictionary assigns it, even where standard English allows broader or different meanings. Some approved words have more than one approved meaning; use any of those, but never a standard-English sense the dictionary does not list.
- [Rule 1.4: Use only the approved forms of verbs and adjectives](https://axiom-verity.com/ste/rules/1.4): Use each approved verb only in the specific forms the dictionary lists for it (typically infinitive/imperative, simple present, simple past, and past participle — e.g. REMOVE, REMOVES, REMOVED, REMOVED) and each approved adjective only in its dictionary-given base, comparative, and superlative forms (e.g. SLOW, SLOWER, SLOWEST). Adjectives that regularly form comparatives/superlatives with 'more'/'most' are not given separate dictionary forms, since 'more' and 'most' are themselves approved words — except where a more specific rule (e.g. rule 1.5, category 17 colors) bans comparison for that word outright.
- [Rule 1.5: You can use technical nouns from approved categories](https://axiom-verity.com/ste/rules/1.5): A word not in the STE dictionary may be used as a technical noun only if it fits one of the 22 approved technical-noun categories listed in technicalNounCategories (each with its official definition, examples, and source page). Category membership does not by itself override the requirement to use the approved company/industry/subject-field term when one exists (rule 1.8). For category 17 (colors): colors are technical nouns even though they behave grammatically as adjectives, and comparative or superlative color forms (e.g. 'blacker', 'the reddest') are never permitted in STE — there is no compliant 'more'/'most' workaround for a color comparison.
- [Rule 1.6: Use non-approved words only as technical nouns](https://axiom-verity.com/ste/rules/1.6): A word may be used as a technical noun (or as part of one) — never with its ordinary, non-technical meaning — in two cases: (a) the dictionary explicitly marks it not approved, or (b) the word is absent from the dictionary entirely. Both cases require a genuine, valid technical-noun use under rule 1.5.
- [Rule 1.7: Do not use technical nouns as verbs](https://axiom-verity.com/ste/rules/1.7): Use a technical noun only as a noun (or as an adjective inside a different technical noun) — never convert it into a verb.
- [Rule 1.8: Use only approved technical nouns](https://axiom-verity.com/ste/rules/1.8): When your company, industry, or subject field has an approved technical noun for something, use that exact term rather than an invented, generic, or overly obscure alternative.
- [Rule 1.9: Use short and easy to understand technical nouns](https://axiom-verity.com/ste/rules/1.9): When no technical noun is approved in your company, industry, or subject field, select one that is short (not more than three words) and easy to understand, relying on illustrations or index numbers for extra specificity rather than a long descriptive noun phrase.
- [Rule 1.10: Do not use regional, slang, or jargon words as technical nouns](https://axiom-verity.com/ste/rules/1.10): Do not use technical nouns that are only understood by people in a specific region, or that are slang or narrow jargon understood by only a small group — always use well-known, widely understood terms.
- [Rule 1.11: Use the same technical noun for the same item throughout](https://axiom-verity.com/ste/rules/1.11): Once you select the approved technical noun for an item, use that exact same term every time you refer to that item in the text — never switch to a synonym or a different technical noun for it.
- [Rule 1.12: You can use technical verbs from approved categories](https://axiom-verity.com/ste/rules/1.12): A word not in the STE dictionary may be used as a technical verb only if it fits one of the 4 approved technical-verb categories listed in technicalVerbCategories (each with its official definition, subcategories, examples, and source page) — and only when no approved dictionary verb already expresses the same instruction clearly.
- [Rule 1.13: Do not use technical verbs as nouns](https://axiom-verity.com/ste/rules/1.13): Use a technical verb only as a verb — never convert it into a noun. Its past-participle form may still be used as an adjective.
- [Rule 1.14: Use American English spelling](https://axiom-verity.com/ste/rules/1.14): Use American English spelling as given in the STE dictionary, unless an applicable official directive specifically requires a different spelling; leave quoted text (e.g. text shown on a screen or label) unchanged even if it uses non-American spelling.
- [Rule 2.1: Write multi-word nouns of no more than three words](https://axiom-verity.com/ste/rules/2.1): A multi-word noun is a string of nouns and/or adjectives (with no article or preposition breaking it up) that together modify a head noun and function as one part of speech. Keep every multi-word noun to a maximum of three words; if a concept needs more, restructure the phrase (for example with prepositions such as 'of,' 'on,' 'in,' or 'for') instead of stacking more than three words together.
- [Rule 2.2: Write long technical nouns in full, then shorten or hyphenate them](https://axiom-verity.com/ste/rules/2.2): When an approved technical noun (from official company, engineering, or parts-catalog documentation) itself has more than three words and cannot be shortened in meaning, write it out in full the first time it occurs in the text. After that first occurrence, if possible, you can either (a) switch to a shorter form or an approved abbreviation for the rest of the document, or (b) use hyphens between words that function together as one unit so the phrase reads as three words or fewer. Abbreviating or hyphenating is not necessary for an approved technical noun that already has three words or fewer, and you must never alter hyphens or abbreviations that are already fixed parts of an official approved technical noun.
- [Rule 3.1: Use only the verb forms that are given in the dictionary](https://axiom-verity.com/ste/rules/3.1): For each approved verb, use only the specific inflected forms listed for it in the STE dictionary (typically the base/infinitive, the simple present, the simple past, and the past participle). Do not use, invent, or regularize any form of an approved verb that is not one of the forms the dictionary gives for it, even if that form would be grammatically valid in general English.
- [Rule 3.2: Use only approved verb forms and tenses](https://axiom-verity.com/ste/rules/3.2): Use verbs only in one of these forms or tenses: the infinitive form, the imperative (command) form, the simple present tense, the simple past tense, the simple future tense, or the past participle used as an adjective. Do not use other tenses or constructions such as the present/past perfect, the present/past progressive, or any other complex (multi-auxiliary) verb construction.
- [Rule 3.3: Use the past participle form as an adjective](https://axiom-verity.com/ste/rules/3.3): A verb's past participle form may be used as an adjective — either directly before the noun it modifies, or after a form of 'to be,' 'to become,' or 'to stay' — to show the condition of something; this is not passive voice. Only use a past participle as an adjective if it (or its adjective sense) is listed in the dictionary, including approved adjectives such as 'permitted' or 'damaged' whose underlying verb is not itself approved.
- [Rule 3.4: Do not use auxiliary verbs to make complex verb constructions](https://axiom-verity.com/ste/rules/3.4): Do not build complex verb constructions with auxiliary verbs used as the main verb of a perfect tense or a passive/modal construction — for example 'have/has' plus a past participle to form a perfect tense, or 'is/are/can/must/will' plus a past participle to form a modal or passive construction where the participle functions as the action verb, not as an adjective. This does not cover a past participle used purely as an adjective (rule 3.3) or the narrow permitted descriptive passive with a genuinely unknown agent (rule 3.6) — those are separate, permitted constructions that happen to share the same surface 'be + past participle' form. These disallowed constructions produce non-approved tenses and passive-voice sentences. Instead, identify the true agent of the action and use it as the subject of an approved simple tense (see rule 3.2), or use the imperative for procedural instructions.
- [Rule 3.5: Use the "-ing" form of a verb only as a technical noun or noun modifier](https://axiom-verity.com/ste/rules/3.5): Do not use a word with an '-ing' form as part of a verb phrase (a progressive tense) or to build long modifying clauses attached to a noun. The '-ing' form may be used in two separate, independent ways: (1) as a valid technical noun (for example in a procedure title or heading) or as a functional modifier within a technical noun describing the function of a system, component, part, tool, or material — this path does not require the specific '-ing' word to also appear in the small dictionary-approved general list; or (2) as one of the small number of dictionary-approved general '-ing' words. A word only needs to satisfy one of these two paths, not both.
- [Rule 3.6: Use the active voice](https://axiom-verity.com/ste/rules/3.6): Always write in the active voice, where the grammatical subject performs the action of the sentence. In descriptive writing only, the passive voice is permitted when the agent (the person or thing that does the action) is genuinely unknown; procedural (instructional) writing must always use the active voice.
- [Rule 3.7: Use an approved verb to describe an action, not a noun or other part of speech](https://axiom-verity.com/ste/rules/3.7): When an approved verb exists that describes the action you mean, use that verb directly rather than a nominalized construction (e.g. 'give an indication of,' 'before the removal of') or another part of speech. If a word is not approved as a verb in the dictionary, do not use it as a verb — restructure the sentence instead, for example with a 'do a/an <noun> of' construction.
- [Rule 4.1: Write short and clear sentences](https://axiom-verity.com/ste/rules/4.1): Write short sentences with clear structure that give accurate instructions and information. In procedures, give instructions directly to the reader in the imperative form; in descriptive text, give each sentence only one topic and build up information gradually across the sentences that follow, and never write abstract statements that omit the accuracy the reader needs.
- [Rule 4.2: Do not omit words or use contractions to shorten sentences](https://axiom-verity.com/ste/rules/4.2): Write every sentence with all of its parts — do not omit nouns, verbs, subjects, or articles to make the sentence shorter, and do not use contractions such as "don't" or "isn't". Omitted words and contractions make a sentence shorter but not easier to read, and can create ambiguity.
- [Rule 4.3: Use a vertical list for complex text](https://axiom-verity.com/ste/rules/4.3): When a sentence would otherwise be long because it includes many items (such as parts, components, or documents) or many actions, present those items as a vertical list instead. Introduce the list with a colon, mark each item consistently (dash, bullet, letter, or number), start each item with an uppercase letter, and punctuate the items according to whether each is a full sentence.
- [Rule 4.4: Use connecting words and phrases to link related sentences](https://axiom-verity.com/ste/rules/4.4): Use an approved connecting word (such as "and", "but", "then", "thus"), an approved connecting phrase (such as "as a result", "at the same time"), or a demonstrative adjective to link a sentence to the related idea in the sentence before it, giving descriptive text a logical structure and clarifying explanations that follow a work step or a safety instruction.
- [Rule 4.5: Use an article or demonstrative adjective before a noun](https://axiom-verity.com/ste/rules/4.5): When applicable, use an article (the, a, an) or a demonstrative adjective (this, these) before a noun or multi-word noun to show its position in the sentence; do not omit it to shorten the text. Do not use an article before a noun in a general statement or abstract concept, and do not use a definite article before a noun that is followed by an alphanumeric identifier, since the identifier makes it a proper noun.
- [Rule 5.1: Write short sentences with a maximum of 20 words](https://axiom-verity.com/ste/rules/5.1): In procedural writing, including warnings, cautions, and other safety instructions, write short sentences with a maximum of 20 words each. If a sentence is longer than 20 words, split it into two or more shorter sentences rather than leaving it as one long sentence.
- [Rule 5.2: Write only one instruction per sentence unless actions are simultaneous](https://axiom-verity.com/ste/rules/5.2): Write only one instruction in each sentence of a procedure, and clearly show the sequence of work steps, usually with numbers or letters. Write more than one instruction in a single sentence only when the actions occur at the same time; you may also write more than one sentence within a single work step when the actions are simultaneous or when a result or limit follows immediately after the action.
- [Rule 5.3: Write instructions in the imperative form](https://axiom-verity.com/ste/rules/5.3): Write every instruction in a procedure using the imperative (command) verb form, addressed directly to the reader, rather than passive voice or a statement about capability or permission. Do not use "must" immediately before the imperative form unless the instruction is safety-critical or states an important condition.
- [Rule 5.4: State a leading condition as a descriptive statement, separated by a comma](https://axiom-verity.com/ste/rules/5.4): When a work step depends on a condition the reader must know about first, start the instruction with a descriptive statement of that condition, then divide the condition from the command with a comma, placing the imperative instruction after the comma.
- [Rule 5.5: Write notes only to give information, not instructions](https://axiom-verity.com/ste/rules/5.5): A note must give only descriptive information to help the reader during a procedure — never an instruction, requirement, or limit. Do not use the imperative form in a note, and do not use a note to state a result, tolerance, or limit that belongs directly in the related work step; each sentence in a note may have up to 25 words.
- [Rule 6.1: Give information gradually](https://axiom-verity.com/ste/rules/6.1): In descriptive text, present information gradually and make sure each sentence contains only one subject. Do not pack multiple facts or ideas about different subjects into one sentence, since this forces the reader to re-read the text to understand it.
- [Rule 6.2: Use key words and key phrases to give your text a logical structure](https://axiom-verity.com/ste/rules/6.2): Use key words and key phrases — the same terminology repeated exactly — to connect a sentence to the ideas introduced earlier in the paragraph (not necessarily only the immediately preceding sentence), and use approved connecting words or phrases (for example, "and," "but," "then," "thus," "as a result," "at the same time") as one method, among others, to signal how a sentence relates to an earlier one. This gives the text a logical, easy-to-follow structure.
- [Rule 6.3: Write short sentences — maximum 25 words in descriptive text](https://axiom-verity.com/ste/rules/6.3): In descriptive writing, use a maximum of 25 words per sentence. Descriptive text is permitted a higher word count than procedural text because its subject matter is inherently more complex, but any sentence over the limit must still be split into shorter sentences.
- [Rule 6.4: Use paragraphs to show related information](https://axiom-verity.com/ste/rules/6.4): Structure descriptive text into paragraphs, where each paragraph opens with a topic sentence that states what the paragraph is about, and every following sentence explains or adds information related to that topic. A new paragraph signals to the reader that a new or different topic follows.
- [Rule 6.5: Make sure that each paragraph has only one topic](https://axiom-verity.com/ste/rules/6.5): Each paragraph in descriptive text must cover exactly one topic. The topic sentence — usually containing a key word and/or a connecting word or phrase — states that topic, and every other sentence in the paragraph must relate to it; do not combine two or more distinct topics in the same paragraph.
- [Rule 6.6: No paragraph may have more than six sentences](https://axiom-verity.com/ste/rules/6.6): Limit each paragraph in descriptive text to a maximum of six sentences. If a paragraph would require more than six sentences, divide it into two or more smaller paragraphs, each covering a distinct part of the topic.
- [Rule 7.1: Use an applicable word to identify the level of risk](https://axiom-verity.com/ste/rules/7.1): Immediately identify the level of risk in a safety instruction with an applicable signal word (or symbol) — use "warning" when there is a risk of injury or death, and "caution" when there is only a risk of damage to objects, tools, or equipment. If both a warning-level and a caution-level risk are present together, label the instruction as a "warning."
- [Rule 7.2: Start a safety instruction with a clear and accurate command or condition](https://axiom-verity.com/ste/rules/7.2): Begin a safety instruction with a clear, accurate command telling the reader what to do or not do; if the reader needs to know a condition before starting the procedure or work step, begin with that condition instead. Either way, the actionable part must come first so the reader immediately knows how to prevent an accident.
- [Rule 7.3: Give an explanation to show the risk or possible result](https://axiom-verity.com/ste/rules/7.3): If it is possible, after the command or condition in a safety instruction (rule 7.2), give an explanation that clearly states the specific risk or possible result if the reader does not obey it. A concrete, specified risk helps the reader understand why the precaution matters and makes them more likely to follow it. The explanation is conditional on it being possible to state — not every WARNING or CAUTION is automatically non-STE for omitting one.
- [Rule 8.1: Do not use the semicolon](https://axiom-verity.com/ste/rules/8.1): You can use all standard English punctuation marks except the semicolon (;). Where a semicolon would normally join two related independent clauses, write two separate sentences instead.
- [Rule 8.2: Use hyphens to connect words that are directly related](https://axiom-verity.com/ste/rules/8.2): Use a hyphen (-) to connect two or more words that are directly related, including multi-word adjectives before a noun, two-word numbers or fractions, letter/number-plus-noun shape terms, noun-derived compound verbs, and prefixes ending in a vowel joined to a root word starting with a vowel.
- [Rule 8.3: Use parentheses only for these seven approved purposes](https://axiom-verity.com/ste/rules/8.3): Use parentheses only to: reference an illustration or other text; give a letter or number identifying an item on an illustration or in text; identify a work-step number in a procedure; show an abbreviation; give the singular and plural forms of a noun together; explain a word or part of a sentence; or present an alternative.
- [Rule 8.4: In a vertical list, a colon has the effect of a period on word count](https://axiom-verity.com/ste/rules/8.4): When a colon introduces a vertical list, it acts like a period: the lead-in text before the colon must not exceed the applicable sentence word-count limit (20 words for a procedural sentence, 25 words for a descriptive sentence), and each item after the colon is counted as a new, independent sentence subject to that same limit.
- [Rule 8.5: Text in parentheses counts as one word](https://axiom-verity.com/ste/rules/8.5): When counting words for a sentence's word-count limit, count an entire parenthetical remark as exactly one word of that sentence, regardless of how many words it actually contains. The words inside the parentheses are also counted separately as their own independent sentence against its own limit. A parenthetical identifier (a number, a letter, an alphanumeric identifier, or an abbreviation) likewise counts as one word.
- [Rule 8.6: Count numbers, units, abbreviations, identifiers, quoted text, titles, and proper nouns as one word each](https://axiom-verity.com/ste/rules/8.6): When counting words for a sentence's word-count limit, count each of the following as a single word: a standalone number; a number together with its unit of measurement; an abbreviation, acronym, or initialism; an alphanumeric identifier; a block of quoted text (including text distinguished by uppercase or a different font to show quotation); a title, heading, or text reproduced from a placard or label; and a proper noun naming an individual, group, organization, or geopolitical entity. Numbers that are only paragraph or work-step numbering are not counted at all.
- [Rule 8.7: Hyphenated words count as one word](https://axiom-verity.com/ste/rules/8.7): A hyphenated group of words that functions as a single adjective before a noun, or as a compound technical noun built for clarity, counts as one word when counting sentence length — even though it is made of multiple hyphen-joined components.
- [Rule 9.1: Use a different sentence construction when a word-for-word replacement is not enough](https://axiom-verity.com/ste/rules/9.1): When a non-approved word cannot simply be swapped for an approved word with the same part of speech and the same meaning, do not force a literal word-for-word replacement. Instead, restructure the whole sentence using approved words so it preserves the original meaning clearly and grammatically.
- [Rule 9.2: Use each approved word only with its approved meaning and approved part of speech](https://axiom-verity.com/ste/rules/9.2): Many STE-approved words carry only one approved meaning and/or one approved part of speech, which can be narrower than their meanings in standard English. Always use an approved word only in the sense and grammatical role the dictionary assigns it, never in another standard-English sense or part of speech it happens to also have.
- [Rule 9.3: Do not combine approved words into phrasal verbs](https://axiom-verity.com/ste/rules/9.3): Do not combine an approved verb with a preposition or particle to form a phrasal verb whose combined meaning differs from the individual approved meanings of its parts (e.g. 'put out' meaning 'extinguish'). Use the single approved word that expresses the intended meaning instead. A small number of phrasal verbs with a restricted approved meaning (e.g. 'put on', 'come on') are dictionary exceptions and remain permitted.
- [Rule 9.4: Use consistent terminology and wording for the same item and the same type of work step](https://axiom-verity.com/ste/rules/9.4): Once you choose a technical noun for an item, or a sentence style/wording for a recurring type of instruction, use that exact same term and wording every time that item or instruction type occurs in the text, rather than varying the phrasing for the sake of variety.
- [Words](https://axiom-verity.com/ste/prompts/p1-words): The system prompt for ASD-STE100 Section 1, covering approved words, technical nouns and American spelling. Needs a dictionary lookup block.
- [Multi-word nouns](https://axiom-verity.com/ste/prompts/p2-multi-word-nouns): The system prompt for ASD-STE100 Section 2: the three-word limit on noun strings, and when to write a long technical noun out in full. Works standalone.
- [Verbs](https://axiom-verity.com/ste/prompts/p3-verbs): The system prompt for ASD-STE100 Section 3, covering permitted tenses, the active voice and the -ing form. Needs a dictionary lookup block.
- [Sentences](https://axiom-verity.com/ste/prompts/p4-sentences): The system prompt for ASD-STE100 Section 4: sentence clarity, contractions, vertical lists, connecting words and articles. Works standalone.
- [Procedural writing](https://axiom-verity.com/ste/prompts/p5-procedural): The system prompt for ASD-STE100 Section 5: the 20-word limit, one instruction per sentence, imperatives and leading conditions. Works standalone.
- [Descriptive writing](https://axiom-verity.com/ste/prompts/p6-descriptive): The system prompt for ASD-STE100 Section 6: the 25-word limit, paragraph topics and the six-sentence paragraph rule. Works standalone.
- [Safety instructions](https://axiom-verity.com/ste/prompts/p7-safety): The system prompt for ASD-STE100 Section 7: warning against caution, command before explanation, and stating the actual risk. Works standalone.
- [Punctuation and word count](https://axiom-verity.com/ste/prompts/p8-punctuation): The system prompt for ASD-STE100 Section 8: the semicolon prohibition, hyphens, parentheses, and how STE counts words. Works standalone.
- [Writing practices](https://axiom-verity.com/ste/prompts/p9-writing-practices): The system prompt for ASD-STE100 Section 9: approved meanings, phrasal verbs and consistent terminology. Needs a dictionary lookup block.
- [Rewrite](https://axiom-verity.com/ste/prompts/p10-rewrite): The rewrite prompt, driven by findings rather than asked to make text compliant in the abstract. Takes the deviations a check pass returned.

## Industries

- [Industries index](https://axiom-verity.com/industries): How the industry hubs work, and why some exist today and some don't yet.
- [Financial Services](https://axiom-verity.com/industries/financial-services): PRA SS1/21 and SS2/21 applied to AI, the Risk & Control Library filtered to FS-specific entries, and the Thought Experiment as a worked example.

## Start here

- [Start Here index](https://axiom-verity.com/start): Three curated, ordered walks through the site — one each for boards and the C-suite, heads-of, and practitioners.
- [Start: Board & C-suite](https://axiom-verity.com/start/board): A seven-stop board-register walk through AI governance, risk, compliance, and third-party assurance.
- [Start: Heads-of](https://axiom-verity.com/start/heads-of): An eight-stop implementation-brief walk through governance, compliance, security, engineering, and the Risk & Control Library.
- [Start: Practitioners](https://axiom-verity.com/start/practitioners): An eight-stop working-materials walk through the Risk & Control Library, Tools Library, Glossary, and technical references.

## Learning

- [Learning index](https://axiom-verity.com/learning): Curated learning paths — ordered walks through the Risk & Control Library plus purpose-written primers, scoped to an audience and a starting level.
- [AI Literacy for Staff](https://axiom-verity.com/learning/ai-literacy-for-staff): A seven-stop practitioner-register walk built around the EU AI Act Article 4 literacy obligation, in force since 2 February 2025.
- [What Article 4 actually requires](https://axiom-verity.com/learning/ai-literacy-for-staff/article-4): Primer: who the AI literacy obligation covers and what it asks for in practice.
- [Governance Accountability and the Board](https://axiom-verity.com/learning/governance-accountability-for-the-board): A seven-stop board-register walk through who owns AI risk, what good oversight looks like, and where board authority can't be delegated.
- [What board-level AI oversight requires](https://axiom-verity.com/learning/governance-accountability-for-the-board/board-oversight): Primer: four concrete things good AI oversight looks like in practice.
- [Where the board's authority can't be delegated](https://axiom-verity.com/learning/governance-accountability-for-the-board/prohibited-practices): Primer: the eight EU AI Act Article 5 prohibited practices, and why almost none has a matching control.

## About the lab

- [The lab](https://axiom-verity.com/about): What Axiom Verity is, how the library is built and maintained, how to cite it, and how to report a correction.
- [Independence statement](https://axiom-verity.com/about/independence): Funding, what the lab does not do, employment, affiliations and vendors: the standing conflict-of-interest statement.
- [Andrew Moore](https://axiom-verity.com/about/andrew-moore): The author and maintainer: career history, credentials, and the Axiom/Verity naming rationale.
- [Thought Experiment](https://axiom-verity.com/thought-experiment): An evolving public investigation into whether a small human core can effectively govern and steward a UK-regulated retail bank, with AI Agents performing the rest.

## Optional

- [Contact](https://axiom-verity.com/contact): Corrections, citation queries and collaboration on open research.
- [Privacy](https://axiom-verity.com/privacy): Privacy policy.
- [Terms](https://axiom-verity.com/terms): Terms of use.