AI system
An artificial-intelligence-based system that receives input and produces a prediction, classification, content or action. It does not mean one particular product or model.
REFERENCE DESK
Glossary
Ninety-eight controlled terms used across the registry.
- VERSION
- 0.3.1
- STATUS
- founder edition · released
Generative system
A system capable of producing text, images, audio, code or another output in response to a user request.
GEO
In this work, the discipline of measurably improving the conditions under which an entity can be found, matched to evidence and represented accurately in generative systems.
Entity
A person, organisation, brand, product, service, place, work, event or concept that must be distinguished from other people or things.
Digital representation
The meaning that a digital system constructs about an entity from identity, attributes, relationships, limits, evidence and time information.
Mention
The appearance of an entity's name or explicit identity marker in an answer.
Citation
A visible reference to a source in an answer. Whether that source actually supports the relevant claim must be checked separately.
Recommendation
Presentation of an entity to the user as a suitable option.
Claim
A factual sentence or part of a sentence that can be assessed as true, false, incomplete or uncertain.
Source
A document, record, page, database or other publication surface from which information comes.
Evidence
A verifiable record that genuinely supports a particular claim within a particular scope and time.
Primary source
An original record connected directly to the event, decision or information produced.
Independent evidence
Verification supplied by a source that is institutionally or methodologically separate from the claimant.
Retrieval
The process by which a system finds and selects information or sources that may be used for a particular question.
Crawling
An automated tool discovering a URL, requesting it and attempting to obtain the content served.
Indexing
Analysis of retrieved content and its conversion into a record that can be used in a search or retrieval system.
Structured data
Markup that explains information on a page to machines through defined fields and relationships.
JSON
A text format that carries data in an organised structure of keys and values.
JSON-LD
A way to express linked-data meaning about entities and relationships by using JSON.
Canonical URL
A signal identifying the preferred page address among similar or duplicate URLs.
Canonical identity record
The principal record proposed by NOMOS GEO for managing an entity's stable identity, names, relationships, status and evidence links.
Canonical name
The preferred primary name within a defined record and use context.
Alternative name
A valid name used for the same entity other than the primary name.
Entity resolution
The process of determining which real entity a name or record denotes.
Role
A person's or identity's function or relationship within an organisation. A role does not automatically confer every authority.
Authority
Permission or legal power to carry out a particular action under defined scope and conditions.
Delegation
Transfer by one party of a defined part of its authority to another actor, subject to limits and duration.
Orphaned account
An account that still exists but has no current owner or accountable person.
Shadow permission
A permission not visible explicitly in the formal role record but usable in practice through another path.
Residual access
Access that remains active after the task or need has ended.
Service account
An account used to run a non-human workload.
Service principal
In Microsoft Entra, an object that represents an application or service instance as a security identity.
Managed identity
A workload identity whose credential lifecycle is managed by a cloud provider.
AI agent identity
An identity that distinguishes an AI agent with a defined purpose, sponsor, tools, permissions and action history.
Technical actor
The human, application, workload or agent identity that actually performs an action in a system.
Represented subject
The person or organisation on whose behalf or for whose benefit an action is performed.
Initiator
The first defined actor or event that gives rise to an action or action proposal.
Approver
The authorised party that permits an action for defined content, a target, scope and duration.
Executor
The actor that technically performs an approved or otherwise authorised action.
Provenance
The chain of relationships showing who produced a record or action, from which inputs, at what time and through which process.
Trace ID
An identifier that groups related technical events within one trace in a distributed system. It is not proof of authority or accuracy by itself.
Human approval
Permission for a specific action given by an authorised person who understands what is being approved, why and with which consequence.
Informed approval
Approval that is not given before the material purpose, scope, risks, alternatives and consequences of the decision are presented.
Action-specific approval
Approval limited to one operation with a defined target, scope, effect and time.
Meaningful human oversight
Oversight in which a person does more than appear on screen: the person can review, challenge, stop and change the result.
Automation bias
The tendency to treat an automated or AI output as more accurate than warranted merely because a system produced it.
Rubber-stamp decision
A formal decision that appears to have human approval but provides no real review, understanding or ability to change the outcome.
Approval fatigue
Erosion of human attention and decision quality caused by approval requests that are too frequent, unclear or unnecessary.
Separation of duties
The condition that authority to prepare, approve, execute and close a critical operation is not concentrated inappropriately in one party.
Decision trace
A chain of records connecting a decision's input, system version, recommendation, human review, approval, application and outcome.
Challenge route
A defined process through which an affected person can understand and correct a decision and obtain renewed review by an authorised person or independent body.
High-impact decision
A decision capable of materially affecting a person's rights, safety, livelihood, essential services, reputation or important opportunities.
Decision lifecycle
All stages of a decision: initiation, data collection, production, review, approval, application, notification, challenge, correction and closure.
Retrospective reassessment
Renewed review of past decisions within a defined scope when a system or its conditions change and create a new risk.
Historical accuracy
The extent to which a record accurately reflects the information and rules that applied when it was made.
Current validity
Whether a decision or information that was correct in the past remains valid for use today.
Append-only record
A record model in which earlier events are not overwritten silently and new states are added as separate, ordered events.
Propagation receipt
Evidence showing when a correction or restriction reached a downstream system and how it was applied there.
Resurrection risk
The risk that deleted, corrected or restricted data reappears from a cache, backup, export or model memory.
Deletion
Removal of data from a defined active or storage surface so that it cannot be recovered or used; the exact scope is stated for the technical and legal context.
Legal hold
A narrowly scoped, recorded retention obligation that suspends ordinary deletion for a specific legal or evidence-preservation reason.
Media sanitisation
The process of making information on a data-storage medium inaccessible under defined threat and reuse conditions.
Machine unlearning
A technical process intended to reduce or remove the influence of specified training data on model behaviour; it is not the same as deleting a record.
Non-use restriction
A temporary control that prevents specified data from entering retrieval, decisions, training, recommendations or tool use while deletion or unlearning remains incomplete.
Release gate
A control requiring defined tests and human approval before data, a model or a function can be used again.
Pending status
A temporary state in which no final decision has been made but the owner, reason, next step, deadline and escalation route must be recorded.
Safe shutdown
A recorded shutdown process that interrupts a defined risk while limiting the affected scope, preserving evidence and considering safe service options.
Restart gate
A control that prevents a system from reopening until root cause, remediation, testing, human approval and rollback readiness are complete.
Prompt injection
An attempt by an input to push an AI system outside its intended instructions, policies or authority boundaries.
Indirect prompt injection
Delivery of a malicious instruction through external content such as a web page, document, email, image, tool output or retrieval source.
Source poisoning
Contamination of sources in a retrieval or decision environment in a way that misdirects the system.
Synthetic consensus
The appearance of many independent sources and a shared view created by copies that originate from the same source.
Source laundering
A weak or unsupported claim being made to look credible through chains of copies and citations.
AI supply chain
The chain of external components consisting of models, datasets, APIs, plug-ins, tools, infrastructure and their sub-dependencies.
AI or representation incident
An event in an AI system or connected publication, data and tool chain that causes harm involving misrepresentation, security, privacy, authority, service or evidential integrity.
Truth Pack
A versioned evidence package linking the atomic claims, sources, conflicts, scope, time and human judgements used as an audit reference.
Reference truth
The audit reference accepted through an appropriate evidence method for a defined entity, scope, time and use context.
Question registry
A set of question texts, intents, risk classes and execution conditions versioned before testing.
Atomic claim
The smallest material proposition that can be verified on its own or left unresolved.
Claim ledger
A versioned record preserving each atomic claim's links to entity, evidence, result, severity, auditor, rationale and uncertainty.
Non-compensation
The principle that a critical failure cannot be averaged away by many low-impact successes.
Output variability
Variation in AI outputs across time and repetitions under the same or similar conditions.
Measurement uncertainty
Uncertainty in a result arising from sampling, differences in conditions, classification and methodological limits.
Confidence interval
A statistical interval that summarises measurement uncertainty in an estimate under stated assumptions; it is not the model's own confidence statement.
Audit independence
A working condition that protects the auditor's judgement from improper pressure, discloses interests and prevents the auditor from approving their own work alone.
Reproducibility
The ability of another sufficiently competent auditor to reproduce selected results to a reasonable degree using the same protocol and evidence.
Remediation receipt
A record that links the error, root cause, change, deployment, live verification, retest, regression and closure evidence under one identifier.
Regression
A new error introduced by a correction in an adjacent record, language, surface, feature or user journey.
Field parity
The material equivalence across publication surfaces of binding fields such as identity, ruling, normative force, version, date, sources, licence and limits.
Controlled transcreation
Language transfer that creates natural target-language text while preserving atomic claims, normative force, evidential scope, legal limits and machine-field parity.
Backward compatibility
The ability of a new release to work without breaking defined older consumer and file behaviour.
Breaking change
A change that makes an existing field, norm, contract or consumer behaviour incompatible.
Migration map
A versioned mapping that explains transformation between old and new fields, values, rules and file structures.
Deprecation
Controlled retirement of a field or release with a warning, alternative, support period and removal date.
Canonicalisation package
A package that links one release's identity, canonical address, DOI relationships, files, hashes, licence and archive mirrors in a single manifest.
File-integrity hash
A digest calculated with a defined algorithm to test whether the compared sequence of bytes changed.
Conformance statement
A result record that states explicitly the audited records, languages, surfaces, systems, versions, dates, gates, exceptions, audit type and decision owner.
Stewardship and maintenance
The arrangement that preserves the work's founder intent, accessibility, version history and last valid state through changes of person or institution.