Plain-language definitions

AI credit and fair-lending glossary.

These definitions describe how Avarent uses common fair-lending, model-risk, and evidence-review terms. Institutional policy and qualified legal interpretation remain authoritative.

Reviewed 2026-08-13

Adverse impact ratio (AIR)
The favorable-outcome rate for a comparison group divided by the favorable-outcome rate for a reference group. A value below a configured screening threshold can prompt investigation, but does not establish cause or a legal conclusion by itself.
Statistical parity difference (SPD)
The comparison group's favorable-outcome rate minus the reference group's favorable-outcome rate. It expresses an absolute percentage-point gap and should be reviewed with sample size, uncertainty, and business context.
Favorable-outcome rate
The share of an explicitly defined population receiving the favorable outcome under review, such as an approval. The numerator, denominator, exclusions, and time window must be documented.
Comparison group
The cohort whose outcome rate is compared with a documented reference group. Group definitions may depend on available, permitted, or institution-approved demographic attributes.
Reference group
The cohort used as the denominator or baseline for a disparity comparison. Its selection can materially affect the result and should be documented rather than assumed.
Review threshold
A configured level that triggers additional investigation or human review. It is a workflow rule, not an automatic legal verdict.
Adverse-action reason review
A check that stated reasons are accurate, specific, and supported by factors actually considered or scored in the decision. Avarent can organize candidate reasons and supporting fields for qualified human review.
Finding record
A traceable record connecting a material result to its question, population, comparison, time window, configuration, supporting evidence, interpretation, reviewer, and next action.
Evidence packet
A reproducible collection of scope, method, inputs, configuration, findings, reviewer actions, limitations, and exports assembled for internal or third-party review.
Synthetic-first review
An evaluation that begins with representative synthetic or institution-approved de-identified data to test usefulness and evidence quality before production access is considered.
Human review
A documented decision by qualified staff who examine the result, context, uncertainty, and next action. Software can support the review but does not own the institution's decision authority.
Model-risk evidence
Artifacts that help reviewers understand and challenge a model-related workflow, including data lineage, methodology, configuration, version context, validation boundaries, findings, limitations, and governance actions.

For formulas, limitations, and regulatory references, see the public methodology.