Appendix A — Glossary of Terms | AI for Financial Risk, Compliance and Regulatory Reporting

Terms used across the chapters of AI for Financial Risk, Compliance and Regulatory Reporting, reproduced from the book's own glossary — AI/ML concepts, UK/EU regulatory terms, and AWB-specific systems and controls, each tied back to how AWB actually used it.

A B C D E F G H I K L M N O P R S T V W X

A

Action Boundary: The governance line separating AI-system outputs requiring only human review (Recommend/Draft) from those permitted to execute autonomously — a line AWB's Agentic RAG engine never crosses.
Adverse News Monitor: AWB's two-stage AML/credit monitoring system (MR-2026-067) that screens news with Gemini 3.5 Flash and analyses qualifying items with Gemini 3.1 Pro, escalating POCA 2002 flags to compliance within one hour.
AgentChangeManagement: AWB's governance mechanism that classifies every agentic-system change by ChangeSurface (PROMPT, TOOL, BUDGET, MODEL_PROVIDER, or OTHER), with TOOL-surface changes always triggering security review.
Agentic AI: AI systems that take multi-step, tool-using actions toward a goal, exemplified at AWB by the Credit Decision Agent and Treasury Operations Agent operating under AutonomyLevel and OversightGate bounds.
Agentic RAG: A RAG variant in which an LLM-controlled ReAct loop decides which retrieval action to take next, up to six hops, before synthesising an answer (Section 4.9.3).
AgenticRAGResult: The structured output object returned by AWB's Agentic RAG engine, carrying the synthesised answer plus a hop log of typed Action objects that feeds the PRA SS1/23 model monitoring dashboard.
AI Act Risk Tiers: The EU AI Act's risk classification scheme, under which AWB's credit-scoring, AML, and lending systems fall into the HIGH-RISK tier (Annex III Section 5(b)), triggering conformity assessment obligations.
Airflow: Apache Airflow, the workflow orchestration platform AWB uses to schedule batch jobs such as the monthly ML model retraining DAG and the Credit Intelligence Monitor's five-module pipeline.
Algorithmic Accountability: The principle that organisations must explain and take responsibility for automated decisions, operationalised at AWB through model cards, audit trails, and SHAP explainability.
AML (Anti-Money Laundering): AWB's AML controls, combining XGBoost transaction scoring with NetworkX Louvain community detection to cut false-positive alerts by 80% while maintaining the same SAR filing rate.
API Gateway: The secure ingress layer for external requests, exemplified by the Bank of England's BEEDS portal through which AWB submits all COREP XBRL returns to the PRA.
Audit Trail: The complete, retained record of an AI system's inputs, outputs, and decisions, kept by every AWB system for a minimum of 7 years to satisfy COBS 9.2 and FCA Consumer Duty obligations.
AutonomyLevel: AWB's five-point governance scale (L0_DRAFT_ONLY through L4_AUTONOMOUS) assigned to every agentic system, with any increase a board-reserved decision rather than an engineering change.
AWB (Avon & Wessex Bank): The fictional £40B-asset UK retail and commercial bank used throughout the book as the case-study institution deploying 23+ AI systems under PRA/FCA regulation.
AWB_CUSTOMER_ID: The universal primary key assigned to every AWB customer by the identity resolution pipeline, used as the cross-system join key across all registered AI systems.

B

Backtesting: The practice of testing a model's historical predictions against realised outcomes, required as a clean 250-day record under the Internal Models Approach for PRA approval.
Basel IV: The industry name for the Basel Committee's post-2017 reforms, implemented in the EU/UK via CRR3 and covering the output floor, revised SA-CR, and FRTB.
BCBS 239: The Basel Committee's eleven principles for effective risk data aggregation and reporting, which AWB's Enterprise Risk Data Warehouse (ERDW) was built specifically to satisfy.
BM25: Best Match 25, a sparse keyword retrieval algorithm that complements dense vector search in AWB's hybrid pipeline via Reciprocal Rank Fusion with ChromaDB.

C

Chain-of-Thought: A prompting technique inducing an LLM to articulate intermediate reasoning before its final answer, related to but distinct from the ReAct pattern's reasoning-action interleaving.
Champion/Challenger: A deployment pattern in which a production model (champion) and a retrained candidate (challenger) are evaluated side by side before the challenger replaces the champion, used throughout AWB's MLOps pipeline.
ChromaDB: The open-source vector database AWB uses for semantic similarity search in the Regulatory Knowledge Assistant, complemented by BM25 sparse retrieval via RRF fusion.
Chunking: AWB's document-aware chunking approach, which keeps a numbered obligation and its qualifying conditions in the same chunk (180–640 tokens) to improve context recall.
CI/CD (Continuous Integration / Continuous Deployment): AWB's Credit Model CI/CD pipeline (Chapter 14) includes a Deployment Gate that blocks production releases lacking an active MR reference or approved prompt hash.
Circuit Breaker: AWB's fleet-level anomaly detector that automatically pauses any agent run whose cost exceeds three times the 30-day rolling average, triggering a PagerDuty P2 alert.
Community Detection: A graph analytics technique that identifies clusters of nodes transacting more with each other than chance predicts, used by AWB's Louvain-based AML monitoring to surface money mule networks.
Context Window: The maximum tokens an LLM can process per call, with AWB's Covenant Compliance Analyser using Gemini 3.1 Pro's 1-million-token window to fit an entire 40-page facility letter into one call.
Contextual Retrieval: A RAG technique that prepends a context summary to each retrieved chunk before indexing, used at AWB as a complement to, not a replacement for, document-aware chunking (Section 4.9.2).
Continuous Assurance: AWB's governance philosophy for agentic systems, which applies daily production sampling via the ContinuousAssuranceMonitor rather than periodic revalidation.
ContinuousAssuranceMonitor: AWB's control that re-scores a daily production sample of an agentic system's outputs against its ReliabilitySLO, with a sustained breach (two consecutive days) raising the same incident class as a drift breach.
Cosine Similarity: A distance metric measuring the angle between two embedding vectors, used by AWB's vector databases (ChromaDB and pgvector) to score semantic similarity.
Covenant: A contractual lending condition (e.g. a leverage or DSCR threshold) whose breach can trigger default, monitored across roughly 13,400 tests on AWB's 4,200 active facilities.
Covenant Compliance Analyser: AWB's RAG-based system (MR-2026-066) monitoring approximately 13,400 financial covenant tests across 4,200 corporate facilities using Gemini 3.1 Pro over a 23,000-document corpus indexed in pgvector.
CRR3: Capital Requirements Regulation 3, the EU/UK legislative vehicle implementing Basel IV reforms including the output floor, FRTB, and leverage ratio requirements.
Cypher: The query language for traversing Neo4j property graphs, used by AWB's GraphRAGEngine for bounded-depth traversals (maximum 3 hops) across its regulatory knowledge graph.

D

DAML (Defence Against Money Laundering): A consent SAR filed under POCA 2002 s.335, after which the National Crime Agency has seven working days to refuse consent before permission is deemed granted.
Deployment Gate: A required CI/CD control step that blocks a model or prompt from reaching production unless an active MR reference and approved prompt hash are present.
DORA (Digital Operational Resilience Act): EU Regulation 2022/2554, in force January 2025, under which AWB targets RTO ≤ 4 hours and RPO ≤ 1 hour for AI/ML services (Article 11).

E

EAD (Exposure at Default): One of the three core IRB credit risk parameters, representing the expected amount outstanding if a borrower defaults, used in AWB's CIM platform to calculate RWA.
Early Warning Indicator: A signal flagging emerging credit deterioration ahead of formal default, encompassing the 12 deterministic SICR indicators used in AWB's IFRS 9 Staging Engine.
ECL (Expected Credit Loss): The IFRS 9 impairment measure calculated as 12-month ECL for Stage 1 accounts or lifetime ECL for Stage 2/3 accounts, displayed on AWB's CRO Dashboard alongside capital ratios and PSI scores.
ECS Task Definition / Kubernetes: AWB's enterprise architecture (Chapter 13) deliberately runs services as AWS ECS Task Definitions rather than Kubernetes, given AWB's operating scale and PRA SS1/23 requirements.
Embedding: A dense numeric vector representation of text capturing semantic meaning, generated by AWB's RAG pipelines (e.g. via Gemini text-embedding-004) for similarity-based retrieval.
Entity Resolution: AWB's identity resolution pipeline that uses deterministic matching for UK customers and Jaro-Winkler probabilistic matching for non-UK customers to create the AWB_CUSTOMER_ID.
EU AI Act: EU Regulation 2024/1689, under which AWB's credit-scoring, AML, and lending AI systems are classified HIGH-RISK (Annex III Section 5(b)), requiring conformity assessment and ongoing monitoring.
Explainability: A model's capability to provide human-interpretable reasons for its outputs, satisfied at AWB chiefly through SHAP feature attribution under PRA SS1/23 and FCA COBS 4.

F

FastAPI: The Python web framework AWB uses to expose AI services as production APIs, including the Credit Intelligence Monitor's single FastAPI service serving three user personas.
Feature Store: A centralised ML feature repository giving point-in-time correct values for training (PostgreSQL) and sub-5ms values for live serving (Redis), eliminating training-serving skew at AWB.
Few-Shot Prompting: A prompting technique supplying an LLM with worked examples to steer output format, used across AWB's document-extraction systems to improve structured-output fidelity.
FRTB (Fundamental Review of the Trading Book): The Basel Committee's 2016 market risk capital framework, implemented in the EU/UK via CRR3 from January 2025 with the SA-FRTB and IMA capital calculation methods (Chapter 7).

G

GDPR / UK GDPR: The data protection regime (UK GDPR and the Data Protection Act 2018) governing AWB's processing of personal data, addressed specifically in Chapter 15.
Gini Coefficient: A measure of credit model discriminatory power calculated as 2 × AUC-ROC minus 1, with AWB's framework setting a minimum threshold of 0.700 under PRA SS1/23.
Golden Trajectory (GoldenTrajectory): A human-validated execution path recorded for each scenario in AWB's agent evaluation suite, versioned against a prompt_version so any prompt change forces re-validation.
GraphRAG: A hybrid retrieval architecture fusing dense vector search (pgvector) with bounded Neo4j graph traversal, used at AWB for cross-regulatory traceability that flat vector search cannot represent (Section 4.9.4).
GraphRAGResult: The structured output type returned by AWB's GraphRAGEngine, combining vector-retrieved passages with graph-traversal relationship context into a single context block before LLM synthesis.
Great Expectations: An open-source data quality validation framework with which AWB runs 24 validation rules on every scoring pipeline run, halting the Airflow DAG on any rule failure.
Grounding: A technique anchoring LLM outputs to source documents by requiring page-and-paragraph citation, one of AWB's three core hallucination mitigations.

H

Hallucination: A factually incorrect or unsupported LLM output, mitigated at AWB to a benchmark target of less than 0.5% on extracted numerical values.
HITL (Human-in-the-Loop): A control pattern requiring human review at a defined workflow point, implemented at AWB through LangGraph's native HITL checkpoints and OversightGate routing.
HNSW (Hierarchical Navigable Small World): The approximate nearest-neighbour index algorithm used by ChromaDB, trading exact recall for sub-millisecond query latency in AWB's regulatory document retrieval.
HyDE (Hypothetical Document Embeddings): A retrieval technique that embeds a hypothetical answer to use as the retrieval query, improving recall on abstract regulatory questions by 11% in AWB's Regulatory Knowledge Assistant.

I

IFRS 9: The accounting standard requiring a three-stage expected credit loss impairment model, implemented at AWB through the IFRS 9 Staging Engine (Chapter 6).
IRB (Internal Ratings-Based): The Basel approach letting a bank use internal models to estimate PD, LGD, and EAD for regulatory capital, subject to the CRR3 output floor.

K

Kill-Switch (KillSwitchControl): A first-class governed control for any AWB agentic system, requiring a named owner, a documented engage()/reset() procedure, and a place in the DORA Article 17 incident taxonomy.
Knowledge Graph: A structured representation of entities and relationships as nodes and edges, modelled at AWB in Neo4j across Regulation, Clause, Control, System, and Filing nodes to support GraphRAG.

L

LangGraph: A Python library for stateful multi-agent workflows, used by AWB for the Credit Decision Agent because its conditional edges and parallel node execution handle routing that pure-Python ReAct cannot.
LCR (Liquidity Coverage Ratio): The ratio of High Quality Liquid Assets to net 30-day stress outflows, with a minimum requirement of 100% under CRR3 Articles 411–428, reported monthly by AWB via the FSA047 return.
LGD (Loss Given Default): One of the three core IRB credit risk parameters, representing the proportion of exposure lost after recoveries if a borrower defaults, feeding AWB's RWA and provisioning calculations.
LLM (Large Language Model): A neural network trained to generate text continuations, used throughout AWB's AI systems with Gemini and Claude models selected per use case by context window, cost, and risk tier.
LLM-as-Judge (LLMJudge): An evaluation pattern in which Claude Opus 4.8 scores AWB's Gemini agent outputs for faithfulness and completeness, deliberately using a different provider to avoid self-preference bias.
Louvain: A community detection algorithm maximising graph modularity, used by AWB's NetworkX-based AML monitoring to identify the clustering signature of a money mule network.

M

Microservices: AWB's enterprise AI architecture (Chapter 13), built as independently deployable services on AWS ECS with multi-cloud disaster recovery on Azure and direct Temenos T24 integration.
MLflow: The open-source platform AWB uses as its authoritative deployment mechanism for experiment tracking, model registry, and reproducibility across the MLOps pipeline (Chapter 14).
MLOps: The discipline of operationalising ML models in production, with AWB's stack (Chapter 14) combining MLflow, Airflow, DVC, and Great Expectations under PRA SS1/23 requirements.
Model Card: A structured documentation artefact stored in AWB's model_cards table, serving as the primary evidence artefact for PRA SS1/23 compliance and EU AI Act technical documentation.
Model Governance: The framework of policies and controls governing AI/ML models at AWB, centred on the Model Risk Committee and the Impact × Complexity risk-tiering matrix under PRA SS1/23.
Model Risk Committee (MRC): AWB's senior AI model governance body, chaired by the Chief Risk Officer, meeting monthly to approve model registrations and escalate material model incidents.
Model Risk Management (MRM): The end-to-end discipline of identifying, assessing, and monitoring model risk, governed at AWB under PRA SS1/23 and benchmarked against the US SR 26-2 standard (Chapter 10).
Model Validation: The independent process of assessing whether a model performs within acceptable risk tolerances, performed at AWB by a function separate from model development per three-lines-of-defence principles.
MR Reference (Model Registry Reference): AWB's unique identifier (e.g. MR-2026-035) assigned to every registered AI/ML model, with a Deployment Gate blocking production releases lacking an active reference. See the full Model Registry Index (Appendix D).
Multi-Agent: An architecture in which multiple specialised agents collaborate on a workflow, exemplified by AWB's Credit Decision Agent integrating Chapter 2 and Chapter 4 outputs with RWA forecasting.

N

Neo4j: The property-graph database AWB uses to store its regulatory knowledge graph and support bounded Cypher traversal queries in the GraphRAG pipeline (Chapter 4).

O

Orchestrator: The component sequencing steps within a workflow, with LangGraph orchestrating multi-agent credit decisioning at AWB while Airflow orchestrates batch ML and data pipelines.
Output Floor: The CRR3 requirement that IRB-derived RWA cannot fall below a percentage of the Standardised Approach equivalent, phasing from 50% in January 2025 to 72.5% by January 2030.
OversightGate: AWB's control routing each proposed agent Decision to a human or automatic execution, with the default policy gating any decision above £5,000,000 or below an 0.80 confidence floor.

P

PD (Probability of Default): One of the three core IRB credit risk parameters, representing the estimated likelihood a borrower defaults within a defined horizon, calibrated at AWB via Platt scaling and monitored for drift using PSI.
pgvector: The PostgreSQL extension AWB selected over a standalone vector database to keep the full audit trail within the existing PostgreSQL governance boundary.
Platt Scaling: A logistic regression technique converting raw model scores into calibrated probability-of-default estimates, used in AWB's credit risk model pipeline.
POCA 2002 (Proceeds of Crime Act): UK primary legislation defining money laundering offences (ss.327–329) and the disclosure obligation (s.330), reflected in AWB's Adverse News Monitor's dedicated POCA implications flag.
PRA SS1/23: The Prudential Regulation Authority's Supervisory Statement 1/23 (May 2023), the primary UK regulatory standard governing AI/ML models at firms such as AWB.
Prompt Engineering: The practice of designing LLM input instructions to reliably produce structured, accurate outputs, used for tasks such as financial ratio extraction in AWB's Credit Document Analyser.
Prompt Hash: The SHA-256 hash of an LLM system prompt plus user template, verified at service start so a Deployment Gate blocks production unless the running hash matches an approved one.
Prompt Injection: An attack in which malicious instructions embedded in retrieved content manipulate an agent's reasoning, countered at AWB by input sanitisation and the ToolCallValidator.
PSI (Population Stability Index): A metric comparing the current score distribution to the training baseline, with AWB's thresholds at PSI < 0.10 (GREEN), 0.10–0.25 (AMBER), and > 0.25 (RED, mandatory recalibration).

R

RAG (Retrieval-Augmented Generation): An architecture retrieving relevant document chunks before generation, the foundation of AWB's Regulatory Knowledge Assistant (MR-2026-038, Chapter 4).
RAGAS: A framework providing four RAG quality metrics — faithfulness, answer relevancy, context precision, and context recall — with AWB's faithfulness threshold set at ≥ 0.85 as its primary hallucination guard.
ReAct: An agent architecture interleaving Reason (Thought) and Act steps in a loop, with each Thought logged verbatim as the foundation of AWB's Credit Decision Agent and Agentic RAG engine.
ReAct Loop: An agent control pattern interleaving reasoning with tool-invoking actions, capped in AWB's Agentic RAG engine at six hops with a 10-second timeout and 0.90 confidence threshold.
Reciprocal Rank Fusion (RRF): A technique combining dense and sparse retrieval ranking lists into one ranked list without score normalisation, used in AWB's hybrid pipeline combining pgvector and BM25.
Regulatory Knowledge Assistant: AWB's production RAG system (MR-2026-038) indexing 47,000 pages of regulatory text, with a 0.65 confidence gate below which it refuses to answer.
Reliability SLO (ReliabilitySLO): AWB's agentic-system service-level objective — outcome_match ≥ 0.98, tool_selection_accuracy ≥ 0.995, cost_variance within ±20% — whose breach is treated as a model risk event under PRA SS1/23.
Reranking: A retrieval refinement step that re-scores candidate chunks with a more precise model, applied by AWB as top-5 reranking before LLM synthesis (Chapter 4).
RWA (Risk-Weighted Assets): Exposures multiplied by CRR3 risk weights to give the denominator for regulatory capital ratios, subject to the CRR3 output floor when derived from IRB models.

S

SAR (Suspicious Activity Report): A mandatory disclosure to the UK National Crime Agency under POCA 2002 s.330, with AWB's AML system maintaining a filing rate of 340 per year while cutting false positives by 80%.
Semantic Search: Retrieval based on query meaning rather than exact keywords, combined with bounded Neo4j graph traversal in AWB's GraphRAG pipeline for cross-regulatory traceability.
SHAP (SHapley Additive exPlanations): A model-agnostic explainability framework (Lundberg & Lee, 2017) used by AWB to satisfy PRA SS1/23 and FCA explainability obligations across credit, fraud, and churn models.
SICR (Significant Increase in Credit Risk): The IFRS 9 trigger moving an account from Stage 1 to Stage 2, determined at AWB through 12 deterministic indicators covering 95% of accounts with LLM-assisted RAG handling the rest.
SR 11-7: The 2011 US Federal Reserve/OCC model risk management guidance, used in Chapter 10 as a historical comparison point before being superseded by SR 26-2 in April 2026.
SR 26-2: The principles-based US model risk management standard issued jointly by the Federal Reserve, OCC, and FDIC on 17 April 2026, replacing SR 11-7 and SR 21-8.
Stage Migration: The movement of a credit account between IFRS 9 impairment stages, requiring analyst approval at AWB before any LLM-assisted edge-case change commits to T24.
Stress Testing: The practice of projecting capital, liquidity, or credit metrics under adverse scenarios, covered through the PRA Concurrent Stress Test and the Bank of England's CBES.
Supersession: The process by which a newer regulatory publication replaces an older one, with AWB's supersession detector marking superseded documents within 24 hours of detection.
System Prompt: The instruction set establishing an LLM's role and constraints, hashed by AWB via SHA-256 and verified at service start as a deployment gate control.

T

T24: Temenos T24, AWB's core banking platform, read via nightly data extracts and written back via the AWB Data Services API using idempotent write patterns.
Three-Line Defence (Three Lines of Defence): The standard financial services control model of first, second, and third lines of defence, within which all AWB AI systems operate under FCA SYSC 6.3.
Token: The basic unit of text an LLM processes, with AWB's context-window decisions and per-run cost governance both denominated in tokens.
Tokenization: The process of converting raw text into tokens, underlying AWB's context-window limits and the per-run token limits enforced by its AgentRunBudget dataclass.
Tool Calling: The mechanism by which an LLM-driven agent invokes external functions, governed at AWB by a typed Tool Registry restricting an agent to its registered, budgeted tools.
Tool Registry: A typed dictionary of ToolSpec objects mapping every callable tool to its type, call limits, and timeout, with an AWB agent only able to call tools in its registered registry.
Trajectory: The complete ordered sequence of reasoning and tool-calling steps an agent takes in one run, evaluated at AWB via the TrajectoryScorer against GoldenTrajectory references.
TrajectoryScorer: AWB's evaluation component that replays each AgentRun against its golden trajectory across outcome_match, trajectory_adherence, and tool_argument_fidelity.
TrajectoryTrace: A signed, replayable record of one multi-step agent run, with each step tagged for source validation and sign() making the trace tamper-evident for audit.
Transformer: The neural network architecture underlying modern LLMs, given the book's deep technical treatment in Chapter 2 before later chapters treat LLM selection as a strategic decision.

V

Vector Database: A database optimised for querying high-dimensional embedding vectors by similarity, with AWB choosing pgvector over ChromaDB in most production systems to keep the audit trail within PostgreSQL.

W

Watchlist: A monitoring list of borrowers flagged for heightened credit surveillance, with a director change at a watchlist borrower one of AWB's 12 deterministic SICR indicators.

X

XBRL: eXtensible Business Reporting Language, the structured data format mandated by the EBA for COREP submissions, used by AWB with EBA Taxonomy 4.0 via the BEEDS system.