APPENDIX J · GLOSSARY · 72 TERMS

The words, without the jargon.

Shared definitions keep the quality conversation precise across leadership, engineering, suppliers, digital teams, reviewers and assessors.

72 of 72 terms shown

Acceptance criterion: A condition agreed before evaluation that determines whether the intended use is acceptable.
Agreement study: A structured comparison of reviewer-to-reviewer and reviewer-to-system findings on the same case set.
AI-touched artefact: A quality record whose creation, comparison or recommendation was supported by an AI capability.
Anonymous Playbook: A browser-resident record of saved instrument results that requires no account to create.
APQP: Advanced Product Quality Planning: a structured lifecycle for turning requirements into controlled product and process evidence.
Attestation: A signed record of the exact configuration, evidence and human decision behind a controlled AI-supported output.
Automation bias: The tendency to accept a machine output because it is confident or fluent, not because it was verified.
Baseline: The current manual measure of time, cost, findings and missed findings against which improvement is judged.
Change control: A recorded process for approving, testing and documenting a modification to a controlled capability.
Configuration drift: A material change in model, prompt, retrieval, data, rules or use that can change task performance.
Coverage statement: A bounded account of what evidence was checked, what was not checked and what remains unresolved.
Critical characteristic: A product feature whose variation has a direct, significant effect on safety, fit, function or compliance.
Data readiness: The extent to which evidence is available, machine-readable, traceable and permission-mapped for a use case.
Desirability: Whether an intended user genuinely wants the capability enough to change their current behaviour.
Escalation path: The defined route by which a reviewer sends an uncertain or high-consequence finding to a more senior decision.
Evidence class: A classification — public, internal, confidential or restricted — that determines permitted processing and retention.
Evidence graph: A connected model of requirements, characteristics, controls, records and lessons used to trace one decision.
Facilitator separation: The rule that a person teaching or facilitating a cohort does not also grade that cohort's assessment.
Failure mode: A specific way a process, product or control can fail to deliver its intended function.
False finding: A reported issue that does not correspond to an actual defect or non-conformity in the evidence.
Feasibility: Whether a use case can be built, evaluated and operated within a reasonable evidence and engineering effort.
Fluency: The grammatical and stylistic smoothness of generated text, which is unrelated to its factual accuracy.
Ground truth: The agreed correct answer for a reference case, established independently of the tool being evaluated.
Golden set: A versioned reference set of cases with established ground truth used to evaluate an intended use.
Hallucination: Fluent output that is unsupported by the available evidence or incorrectly represents a source.
Human decision boundary: The point where a named competent person accepts, rejects, escalates or records a quality decision.
Incident record: A logged account of a failure, override or unexpected result from a live AI-supported capability.
Independent approval: Sign-off by a person other than the one who built or operated the capability being approved.
Intended use: The specific task, users, evidence, output and decision context a capability is approved to support.
Irreversibility gate: A rule that escalates the control level automatically whenever a consequence cannot be undone.
Known-correct case: A test case whose expected result has been established and agreed before the tool is run against it.
Level A: Assistive use such as drafting, searching or summarising with no direct influence on a product decision.
Level B: Controlled decision support that influences quality actions or APQP evidence and needs evaluation, traceability and monitoring.
Level C: High-consequence use that can materially affect safety, release, contractual conformity or major financial exposure.
Manual baseline: See baseline: the measured current-state performance of a task before any AI support is introduced.
Maturity ladder: A five-rung model of organisational AI-in-quality capability, each rung requiring supporting evidence.
Metadata: Context such as revision, part, project, supplier, source location and lifecycle phase that makes evidence usable.
Missed finding: An actual defect or non-conformity present in the evidence that the tool failed to report.
Model drift: A gradual change in a model's real-world performance as the underlying process or data distribution changes.
No-finding case: A reference case where the correct result is that no supported finding exists within the stated coverage.
Non-conformity: A documented instance where a product, process or record does not meet a specified requirement.
Override: A reviewer's documented decision to reject or change a machine-generated finding or recommendation.
Portfolio map: A view that positions candidate use cases by risk, verification cost and organisational readiness.
Prohibited use: A decision category, listed in advance, that a capability is never permitted to make on its own.
Qualification file: The consolidated evidence — intended use, configuration, results, limitations — supporting a release decision.
Reference set: See golden set: cases used to test a capability against agreed, known-correct outcomes.
Release verdict: A recorded decision — stop, improve, restrict, pilot or release — supported by qualification evidence.
Retrieval: The step in which a system selects source evidence to inform a generated answer, prior to generation itself.
Revalidation trigger: A defined material change or event that requires the intended use and acceptance evidence to be reviewed again.
Reversibility: Whether a decision's consequence can be undone without material harm if it later proves wrong.
Risk dimension: One of eight factors — consequence, reversibility, autonomy, scale, evidence, novelty, oversight, regulation — scored to classify a use case.
Root-cause analysis: A structured investigation to identify the underlying cause of a non-conformity or failure.
Scope creep: The gradual expansion of an approved capability's use beyond its original intended-use statement.
Seeded case: A deliberately inserted known-difficult or known-failing case used to test whether a tool fails safely.
Signature test: A check of whether a proposed machine action falls inside the list of decisions that must remain human.
Source authority: The standing of a document as the current, approved version of a requirement or specification.
Stale evidence: Evidence whose revision no longer matches the current approved version of the underlying requirement.
Survivability: Whether a deployed capability remains understandable, operable and revalidatable after the original builder leaves.
Thirty-second lab: An exercise comparing a vague finding with an evidence-rich finding to test verification speed.
Traceability: The ability to inspect the exact source, location, revision and configuration behind an output.
Use-case charter: A one-page record of the problem, owner, baseline, evidence, output and human decision for a candidate use case.
Use drift: Quiet expansion of an approved capability into a different destination, decision or consequence level.
Validation: The process of testing whether a capability performs its intended task on real, representative evidence.
Verification ratio: Verification time divided by production time for a proposed AI-supported task.
APQP gate: A defined checkpoint in the product lifecycle where evidence and open actions are reviewed before proceeding.
Corrective action: A defined step taken to eliminate the cause of an identified non-conformity, distinct from a workaround.
Durability class: A rating of how well a controlled capability's evidence and configuration will survive staff and process change.
Edition: A dated, versioned publication of the guideline; historical editions remain accessible at their own citation URL.
Language-pair evaluation: A distinct check of a capability's performance when source and output language differ from the primary evaluated language.
Reviewer agreement: The rate at which independent reviewers reach the same conclusion on the same evidence.
Signature list: The organisation-maintained register of decisions that require a named human signature and cannot be automated.
Time-to-verify: The measured time a competent reviewer needs to confirm or reject a machine-generated finding.