CSV, CSA, and AI Assurance Glossary

Term	Plain-language explanation	Practical example
Intended use	The job people rely on the system to perform	ERP blocks issue of quarantined material
Context of use	The circumstances in which the task is performed	Expert reviewers using source documents during protocol review
CSV	Establishing and maintaining evidence that a computerized system suits its intended use	Assessing and supporting an eQMS lifecycle
CSA	Software-assurance thinking focused on confidence in use through risk-informed activities	Selecting tests around meaningful production or quality failures
GxP	Shorthand covering different regulated good-practice areas	GMP and GCP have different contexts
Requirement	A defined need or expected behavior	An unauthorized user cannot close a CAPA
Risk	The assessed relationship between a possible failure and its consequences, using the applicable method	Wrong unit mapping could support an incorrect acceptance decision
Control	A measure that prevents, detects, or manages a failure	Blocking an unknown disposition code
Residual risk	Risk remaining after the assessed controls	An unresolved failure possibility accepted within defined conditions
Objective evidence	Information that supports an assessable conclusion	A traced transaction comparison with expected and actual outcomes
Traceability	A visible connection among use, requirement, risk, evidence, and decision	Requirement MAT-02 linked to interface test IF-T08
Scripted testing	Testing with predefined conditions, actions, and expected results	Expiry boundary check at a defined time
Exploratory testing	Learning, test design, and execution conducted together within a purpose	Investigating CAPA reassignment combinations
Negative test	A challenge to behavior that should be rejected or prevented	Attempting approval without permission
Boundary test	A check around a point where behavior changes	Just below, at, and above a specification limit
Regression testing	Checking relevant existing behavior after a change	Rechecking routing after a workflow-engine update
Supplier evidence	Evidence supplied by the product or service provider	Version-specific standard-feature test records
SaaS	Software delivered as an online service	A hosted eQMS with supplier-managed releases
eQMS	Electronic quality management system	Controlled documents, deviations, CAPA, and related workflows
LIMS	Laboratory information management system	Sample, test, result, and laboratory workflow management
MES	Manufacturing execution system	Managing and recording defined production execution activities
ERP	Enterprise resource planning system	Material, inventory, and business-process management
CAPA	Corrective and preventive action	Investigating and addressing a quality problem through a controlled process
Audit trail	Recorded history of relevant events and changes	Who changed a controlled value and when, with required context
Data lineage	The path from source information through transformations to use	PDF extraction, indexing, retrieval, and AI finding
ALCOA	Common data-integrity shorthand: attributable, legible, contemporaneous, original, accurate	A result linked to its creator, context, and trustworthy record
RAG	Retrieval-augmented generation: retrieved material is supplied to support generation	An assistant answers using selected approved procedure passages
Groundedness	The extent to which a claim is supported by its provided source	A cited passage actually supports the stated instruction
Hallucination	Generated content presented as factual without adequate support	An invented procedure requirement
False positive	A proposed finding that is not a real issue under the reference definition	Flagging an approved exception as a defect
False negative	A reference issue the system fails to identify	Missing an incorrect acceptance limit
Precision	Correct findings divided by all proposed findings	32 correct out of 38 proposed findings
Recall	Correctly identified reference issues divided by all reference issues	32 identified out of 40 reference issues
Evaluation leakage	Evaluation information influences development or otherwise undermines the intended independence	Tuning a prompt repeatedly on the final test cases
Drift	A change in data, behavior, or performance relevant to the use	New document formats lead to more extraction failures
Prompt injection	Untrusted content attempts to redirect the assistant’s behavior	A document tells the assistant to disclose restricted information
Agent	An AI-enabled application that can select or execute actions through tools	An assistant creating proposed investigation tasks
Human oversight	Defined human checks and decisions with suitable authority and capability	A qualified reviewer verifies sources and disposes findings
Reassessment	Revisiting a conclusion when relevant conditions change	Evaluating a new parser or source corpus