AI GOVERNANCE IN LENDING · 100,000-APPLICATION SCENARIO ANALYSIS

When decisions scale, governance must scale with them.

Across 12 scenarios spanning model and configuration, data and access, decision authority, human override, explainability and fairness drift, this synthetic mass-retail lending simulation shows how embedded governance changes decision quality, facility value linked to faulty cases, human intervention, residual risk and evidence for management action.

Why the comparison matters: Governance OFF is the comparison baseline where relevant controls are absent, bypassed, ineffective or not embedded at the point of decision. Governance ON applies access and segregation of duties, approved models, human review, targeted kill switch and fallback, control assurance, ERM response and evidence. The difference shows what governance stopped and what still requires management action.

Quick meaning guide — recurring terms

STP

Straight-through processing: automatic finalisation without human referral.

HITL / HOTL

HITL pauses a case for accountable human review. HOTL supervises portfolio behaviour while automation runs.

Kill switch

Targeted suspension of an affected AI path/cohort; the whole loan process does not stop.

Affected / residual

Affected = touched by a failure scenario. Residual = still remaining after governance/control treatment.

ToD / ToE

Design capability versus actual operating performance of a control.

Risk response / control

Response = accept, mitigate, transfer/share or avoid. Control = preventive, detective, corrective, containment or recovery.

12-month occurrence probability

Scenario-configured probability that a risk event occurs within the stated horizon. It is separate from the percentage of transactions currently affected.

Residual-case treatment

Known faulty cases remaining after automated controls must still be held, reviewed, corrected, stopped, remediated or investigated before closure.

Applications100,000

Same loan-application population used in both OFF and ON paths.

Gross requested facility valueIDR 6.78 tn

Sum of requested principal; this is exposure size, not expected loss.

Faulty transactions passed through57,254

IDR 3.98 tn requested facility value linked to faulty cases

Faulty transactions remaining after controls1,055

Residual cases not fully intercepted/contained. IDR 75.38 bn requested facility value remains linked to them.

Governance effectiveness98.16%

Share of baseline fault-affected transactions removed from the residual population by governance controls.

Decision errors OFF → ON2,235 → 97

Final decisions that differ from the clean reference path.

Kill switchTRIGGERED

Triggered means AI auto-finalisation was suspended only for the affected cohort/path. Those applications continue through HITL or an approved fallback; the overall loan process is not stopped.

Governance evidence10/10 PASS

Operating evidence shows whether governance controls actually ran, not merely whether policies exist.

1) Governance OFF vs ON

What this shows: The identical 100,000-loan portfolio is run with the same faults under two control states.
Framework / practice basis: NIST AI RMF GOVERN/MANAGE; ISO/IEC 42001 management-system controls; controlled experiment design.
So what: Governance should reduce the faulty outcomes, not just restate the same baseline exposure.

Reading note: Some metrics are intentionally identical in OFF and ON because they describe the same baseline portfolio and the same injected-fault population. The rows that should improve are the outcome rows: faulty transactions reaching outcome stage, residual facility-value risk, and decision errors.

Governance OFF

57,254

Faulty transactions passed through

IDR 3.98 tn in requested facility value linked to faulty cases.

STP 71.60%

HITL 28,397

Decision errors 2,235

Governance ON

1,055

Faulty transactions remaining after controls

Residual facility-value risk reduced to IDR 75.38 bn, down IDR 3.90 tn from OFF.

What happens next: these cases are not auto-approved; they enter the residual-case treatment workflow below.

STP 52.98%

HITL 38,542

Errors prevented 2,138

Comparison metric	OFF	ON	Governance effect
Same baseline by design
Applications processed same portfolio	100,000	100,000	Same by design
Gross requested facility value same synthetic portfolio value	IDR 6.78 tn	IDR 6.78 tn	Same by design
Fault-affected transaction population same injected-fault transaction population	57,254	57,254	Same by design
Requested facility value linked to fault-affected population same injected-fault facility-value population	IDR 3.98 tn	IDR 3.98 tn	Same by design
Outcome comparison
Straight-through decisions lower can reflect intentional governance friction	71,603	52,980	Reduced by 18,623 (26.01%)
Straight-through rate automation retained	71.60%	52.98%	-18.62 pp
Total HITL volume human review volume	28,397	38,542	Increased by 10,145 (35.73%)
Faulty transactions reaching outcome stage lower is better; governance should intercept, redirect or contain these cases before final outcome	57,254	1,055	Reduced by 56,199 (98.16%)
Residual facility-value risk after controls gross requested facility value linked to the remaining faulty transaction population; not expected credit loss	IDR 3.98 tn	IDR 75.38 bn	Reduced by IDR 3.90 tn (98.10%)
Final decisions changed vs clean reference decision impact	2,235	97	Reduced by 2,138 (95.66%)
Facility value linked to changed decisions OFF exact; ON estimated from remaining changed-decision proportion	IDR 161.71 bn	IDR 7.02 bn	Reduced by IDR 154.69 bn (95.66%)
Control interventions governance action	0	56,199	Introduced 56,199
Blocked actions least-privilege enforcement	0	5,000	Introduced 5,000
Blocked AI decisions decision control	0	15,000	Introduced 15,000
Fallback to approved configuration model/config containment	0	40,000	Introduced 40,000
Held before booking final-integrity control	0	10,145	Introduced 10,145
Unauthorized actions executed lower is better	5,000	0	Reduced by 5,000 (100.00%)
Adverse decisions without reason explainability / conduct control	812	24	Reduced by 788 (97.04%)

1A) Mass-loan amount distribution, deciles and concentration

What this shows: The requested-facility portfolio is intentionally right-skewed: many smaller mass-loan cases and progressively fewer large-ticket cases. D1–D10 show where volume, affected cases and residual value concentrate.
Framework / practice basis: Retail credit portfolio segmentation / concentration analysis; ERM materiality; value-weighted residual-risk view.
So what: Average alone can hide concentration. A small upper-decile population can contribute a disproportionate share of facility value and residual financial materiality.

Portfolio calibration: default 100K / seed 20,260,828 produces IDR 6.78 tn requested facilities, average IDR 67.8 mn and median IDR 47.6 mn. The median is deliberately below the average, consistent with a right-skewed mass-loan portfolio.

Percentile / decile meaning: P10 means 10% of applications request this amount or less; P50 is the median; P90 means 90% are at or below that amount. D1–D10 divide the portfolio into ten equal-count amount groups from lowest to highest requested facility value.

P10IDR 15.9 mn

P50 / medianIDR 47.6 mn

P90IDR 141.7 mn

P95IDR 192.5 mn

P99IDR 341.8 mn

D9–D10 residual value share52.96%

Key principle: Frequency does not equal materiality: a small share of high-ticket cases can contribute a disproportionate share of residual facility value. D10 represents 32.85% of portfolio requested value and 33.28% of estimated residual facility value.

Facility value by amount decile

Blue = requested facility value in each decile. Dark red = estimated facility value still tied to residual cases after governance. The same full-scale denominator is used.

avg IDR 11.3 mn

IDR 114.99 bn · residual IDR 1.12 bn

avg IDR 19.7 mn

IDR 192.56 bn · residual IDR 1.86 bn

avg IDR 26.8 mn

IDR 272.34 bn · residual IDR 2.70 bn

avg IDR 34.3 mn

IDR 343.53 bn · residual IDR 3.43 bn

avg IDR 42.8 mn

IDR 426.16 bn · residual IDR 4.16 bn

avg IDR 53.1 mn

IDR 526.78 bn · residual IDR 5.58 bn

avg IDR 66.3 mn

IDR 665.78 bn · residual IDR 7.30 bn

avg IDR 84.7 mn

IDR 836.12 bn · residual IDR 9.30 bn

avg IDR 116.4 mn

IDR 1.17 tn · residual IDR 14.83 bn

D10

avg IDR 221.8 mn

IDR 2.23 tn · residual IDR 25.09 bn

Decile	Loans	Average facility	Total facility	Approval	STP	Affected by faults	Still remaining	Share of residual value
D1 LOW_TICKET	10,138	IDR 11.3 mn	IDR 114.99 bn	80.23%	75.08%	5,275 IDR 59.50 bn	101 IDR 1.12 bn	1.48%
D2 LOW_TICKET	9,778	IDR 19.7 mn	IDR 192.56 bn	80.32%	74.96%	4,952 IDR 97.83 bn	93 IDR 1.86 bn	2.47%
D3 LOW_TICKET	10,151	IDR 26.8 mn	IDR 272.34 bn	80.21%	74.02%	5,839 IDR 157.18 bn	101 IDR 2.70 bn	3.58%
D4 CORE_MASS	10,013	IDR 34.3 mn	IDR 343.53 bn	80.34%	74.22%	5,893 IDR 202.11 bn	100 IDR 3.43 bn	4.55%
D5 CORE_MASS	9,946	IDR 42.8 mn	IDR 426.16 bn	79.80%	73.03%	5,763 IDR 246.82 bn	97 IDR 4.16 bn	5.52%
D6 CORE_MASS	9,924	IDR 53.1 mn	IDR 526.78 bn	80.42%	73.00%	5,861 IDR 311.33 bn	105 IDR 5.58 bn	7.40%
D7 CORE_MASS	10,049	IDR 66.3 mn	IDR 665.78 bn	79.10%	71.99%	5,951 IDR 394.79 bn	109 IDR 7.30 bn	9.68%
D8 UPPER_MASS	9,868	IDR 84.7 mn	IDR 836.12 bn	78.95%	70.97%	5,798 IDR 490.56 bn	110 IDR 9.30 bn	12.34%
D9 UPPER_MASS	10,090	IDR 116.4 mn	IDR 1.17 tn	79.10%	67.80%	5,979 IDR 696.10 bn	127 IDR 14.83 bn	19.68%
D10 HIGH_TICKET	10,043	IDR 221.8 mn	IDR 2.23 tn	76.41%	61.02%	5,943 IDR 1.32 tn	112 IDR 25.09 bn	33.28%

Amount-segment roll-up

Segment	Loans	Average facility	Total facility	Affected cases	Residual cases	Residual facility value	Residual value share
LOW_TICKET	30,067	IDR 19.3 mn	IDR 579.88 bn	16,066	295	IDR 5.68 bn	7.54%
CORE_MASS	39,932	IDR 49.1 mn	IDR 1.96 tn	23,468	411	IDR 20.47 bn	27.16%
UPPER_MASS	19,958	IDR 100.7 mn	IDR 2.01 tn	11,777	237	IDR 24.14 bn	32.02%
HIGH_TICKET	10,043	IDR 221.8 mn	IDR 2.23 tn	5,943	112	IDR 25.09 bn	33.28%

Still remaining does not mean unresolved forever: the decile/segment tables identify where residual cases and value are concentrated. Every remaining case is then assigned to a concrete treatment route in the next section.

1B) Residual-case treatment — what happens to the 1,055 remaining cases?

What this shows: The residual population is routed into concrete treatment; it is not automatically approved, ignored or accepted as-is.
Framework / practice basis: Case management, HITL, preventive/containment/corrective controls, incident management and control-remediation closure.
So what: Residual risk at ERM level is different from a known faulty transaction. Known cases still require disposition and evidence before closure.

Key interpretation: 1,055 residual cases means automated governance did not fully resolve them. They must be held, reviewed, corrected, stopped before booking/disbursement, remediated after booking, or investigated. Residual does not mean allowed to proceed.

Treatment route	Cases	Facility value linked to cases	What happens	Owner	When the case can close
Correct + reprocess CORRECTIVE	260 24.64% of residual cases	IDR 18.49 bn 24.53% of residual facility value	Correct data, reason code, threshold or rule issue and re-run the affected application through the approved decision path.	Credit Operations + AI Governance	Corrected inputs/rules evidenced and reprocessed result passes applicable controls.
Hold + human review CONTAINMENT + HUMAN REVIEW	579 54.88% of residual cases	IDR 40.83 bn 54.17% of residual facility value	Prevent automatic finalisation and assign the application to an accountable human reviewer.	Credit Operations / Underwriting	Human decision recorded with rationale; case either corrected/reprocessed or approved/declined under delegated authority.
Incident investigation DETECTIVE + CORRECTIVE	46 4.36% of residual cases	IDR 3.55 bn 4.71% of residual facility value	Open an incident/root-cause investigation, preserve evidence, identify affected population and implement preventive actions.	AI Governance + Credit Risk / Incident Management	Root cause confirmed, actions completed, affected population treated and retest/closure evidence approved.
Stop before booking / disbursement PREVENTIVE + CONTAINMENT	130 12.32% of residual cases	IDR 9.68 bn 12.84% of residual facility value	Keep the case from booking or disbursement while the high-risk exception is investigated and resolved.	Credit Risk + Operations	Exception cleared or application cancelled/declined; no uncontrolled disbursement occurs.
Post-booking remediation CORRECTIVE + CUSTOMER REMEDIATION	40 3.79% of residual cases	IDR 2.83 bn 3.75% of residual facility value	For a case already booked, review the account, correct terms/data where required and apply customer/credit remediation and enhanced monitoring.	Credit Operations + Conduct / Customer Remediation	Customer/account remediation completed, approvals evidenced and account monitoring updated.

Closure rule: A case is closed only after the required treatment is completed and evidence is recorded. Systemic issues also require control remediation/retest before unrestricted automated processing restarts.

2) Root-cause propagation, affected transactions and residual exposure

What this shows: Each row shows how many transactions were affected by a root-cause scenario, how many were contained, and how many remain after governance.
Framework / practice basis: NIST AI RMF MAP 5 (likelihood/magnitude of impact) and MANAGE 1.2–1.4 (prioritisation, response and residual risk).
So what: One upstream failure can propagate across many automated decisions; controls should reduce propagation and residual exposure.

Terminology note: “Blast radius” is useful operational shorthand, but the formal dashboard uses propagation / affected transactions / residual exposure. IDR values are gross requested facility amounts touched by the scenario, not expected credit loss. Light red = transactions affected by the scenario. Dark red = the portion still remaining after controls. Both use the same denominator, so the remaining-risk bar can never appear larger than the initial affected population.

RC09 Segregation-of-duties self-approval

affected 20,000 (IDR 1.37 tn) · contained 20,000 · still remaining 0 (IDR 0)

RC01 Unauthorized model deployment

affected 20,000 (IDR 1.37 tn) · contained 20,000 · still remaining 0 (IDR 0)

RC02 Model calibration defect

affected 20,000 (IDR 1.37 tn) · contained 20,000 · still remaining 0 (IDR 0)

RC03 Unapproved proxy feature

affected 10,000 (IDR 681.36 bn) · contained 10,000 · still remaining 0 (IDR 0)

RC11 Silent material configuration drift

affected 5,000 (IDR 341.80 bn) · contained 5,000 · still remaining 0 (IDR 0)

RC12 Shadow AI / unapproved provider invocation

affected 5,000 (IDR 337.62 bn) · contained 5,000 · still remaining 0 (IDR 0)

RC08 Excessive user/agent privilege

affected 5,000 (IDR 338.29 bn) · contained 5,000 · still remaining 0 (IDR 0)

RC04 Low-uncertainty-confidence STP misrouting

affected 4,220 (IDR 296.27 bn) · contained 3,967 · still remaining 253 (IDR 17.78 bn)

RC10 Missing adverse-decision reason code

affected 812 (IDR 59.57 bn) · contained 788 · still remaining 24 (IDR 1.85 bn)

RC05 Automation authority misconfiguration

affected 5,390 (IDR 418.78 bn) · contained 5,390 · still remaining 0 (IDR 0)

RC06 Unsupported human override

affected 810 (IDR 64.44 bn) · contained 713 · still remaining 97 (IDR 7.65 bn)

RC07 Cohort calibration / fairness drift

affected 1,022 (IDR 72.12 bn) · contained 341 · still remaining 681 (IDR 48.10 bn)

2A) AI platform governance — controls that act before and around the loan process

What this shows: Transaction controls are not enough. The agents, permissions, environments, providers and release components also need a governance protocol.
Framework / practice basis: AI inventory/accountability, least privilege, environment segregation, third-party governance, lifecycle/change control and recoverability.
So what: The most scalable AI failure may start upstream. Stopping one bad permission, test identity or component before production can prevent portfolio-wide downstream exposure.

PG01 · Agent inventory + permission lifecycle

No agent should operate anonymously or keep access merely because it had access before.

Protocol: Register agent + owner + purpose + environment + risk tier → Grant only required data/tools/actions → Time-bound elevated privilege → Periodic recertification → Detect permission creep / expiry → Revoke or re-approve with evidence

Control test: 8/8 agents registered; 17 grants reviewed; 2 excess/expired detected and revoked; 0 unresolved.

Why it matters: Least privilege is not a one-time access matrix. Permissions can creep or outlive their business need.

Owner: IAM / Security + AI Governance · Evidence: Agent registry + access recertification log + revocation event

PG02 · Test / production boundary

A development, evaluation or test identity must not silently become a production actor.

Protocol: Bind identity to environment → Separate test/prod credentials and network routes → Deny production API/data access from non-production identity → Log boundary violations → Escalate repeated attempts

Control test: 3/3 unauthorized boundary attempts blocked; zero production escape.

Why it matters: A test agent must not silently become a production actor simply because it has a reachable endpoint.

Owner: Platform Engineering + Security · Evidence: Environment identity policy + denied-access / boundary-test logs

PG03 · Shadow AI discovery + approval lifecycle

Unknown AI use must become visible before it becomes normal operating practice.

Protocol: Discover unregistered AI/provider invocation → Match against approved inventory → Assess data/system access + intended use → Assign risk tier → Approve / restrict / reject → Onboard approved tool to IAM + logging → Periodically reassess

Control test: 1/1 unregistered provider attempts blocked; linked RC12 scope 5,000 applications.

Why it matters: Discovery alone does not govern shadow AI; the organization needs a decision and onboarding/rejection lifecycle.

Owner: AI Governance + Security / Procurement · Evidence: Discovery event + risk review decision + provider registry record

PG04 · AI component / supply-chain integrity

Production promotion checks the whole AI component set, not model version alone.

Protocol: Approved source / registry → Pinned version → Artifact/model/feature hash check → Provider identity check → Dependency / release manifest review → Block mismatch → Fallback to last-known-good configuration → Record remediation evidence

Control test: 4 components checked; 1 mismatch blocked; last-known-good retained.

Why it matters: An approved model can still be paired with an unapproved dependency, provider or release artifact.

Owner: Model / Platform Engineering + Model Risk / Security · Evidence: Signed release manifest + version/hash/provider checks + fallback record

Added platform fault scenarios: PF01 Permission creep / expired privilege and PF02 Unapproved component reaches release candidate. RC12 is enriched rather than duplicated: Shadow AI now demonstrates discover → assess → approve/restrict/reject → monitor.

2B) Evidence of Governance — is the governance system itself operating?

What this shows: Whether the governance layer leaves observable evidence that its controls actually ran.
Framework / practice basis: Control monitoring, evidence retention, access recertification, control testing, incident/recovery evidence and assurance.
So what: Governance cannot be demonstrated by policy documents alone. A reviewer should be able to see what was registered, checked, blocked, revoked, logged, tested and approved.

Overall status: OPERATING — CONTROL TESTS PASSED
This section tests the governance layer itself. A policy document is not enough; inventory, permissions, boundaries, integrity controls, evidence and recovery must leave observable proof.

Why this matters: governance is only credible if a reviewer can trace registration, permission review, blocked boundary/provider/component events, kill-switch/fallback actions, residual-case assignments and controlled restart evidence.

Operating metric	Result	Status	What the evidence proves	Evidence source
AI agent inventory completeness	8/8	PASS	Every simulated agent has an owner, purpose, environment and risk tier before it can operate.	Agent registry snapshot
High-risk permission recertification	17 grants reviewed	PASS	Access is periodically re-justified; two expired/excess grants were detected and revoked.	Access recertification + revocation log
Test / production boundary control test	3/3 blocked	PASS	Non-production identities were prevented from becoming production actors.	Boundary-test / denied-access log
Shadow AI / unregistered-provider containment	1/1 blocked	PASS	The unregistered provider attempt was discovered, assessed and blocked before escape; linked RC12 scope = 5,000 applications.	Provider discovery + approval/rejection record
AI component integrity gate	4 checked; 1 mismatch blocked	PASS	The release candidate is checked beyond model version alone; a mismatch triggers block + known-good fallback.	Release manifest + hash/provider attestation + fallback record
Governance controls with explicit evidence trace	8/8	PASS	Each platform/system governance control records a trace statement showing what was checked and the result.	Governance control trace
Evidence-ledger sample completeness	25 sampled events	PASS	Sampled decision/control events contain model, policy, AI output, governance output, oversight, intervention and evidence-hash fields.	Evidence ledger
Kill-switch control test	TRIGGERED + CONTROLLED FALLBACK	PASS	The test proves targeted suspension of the affected AI auto-finalisation path while cases continue through HITL/fallback.	Kill-switch event + routing/fallback log
Residual-case treatment assignment	1055/1055 assigned	PASS	Every known residual faulty case is assigned a treatment path instead of being allowed to proceed automatically or remain unmanaged.	Residual-case treatment / case-management assignment log
Controlled restart protocol	NORMAL_AFTER_CONTROLLED_REACTIVATION	PASS	Restart is only allowed after incident analysis, corrective action, revalidation and independent approval.	Restart approval / revalidation record

Evidence chain

Control objective → trigger/check → action → evidence record → accountable owner → exception/remediation → retest/closure. This is what makes governance auditable rather than aspirational.

3) Control effectiveness — ToD + ToE + improvement action

What this shows: ToD assesses whether a control is capable by design; ToE tests whether it actually operated over the simulated population.
Framework / practice basis: COBIT MEA02/MEA04/APO12, COSO Control Activities and Monitoring, three-lines assurance model.
So what: An adequate design with weak ToE is an implementation problem, not proof that the control concept is sufficient. Every row therefore includes a strengthening action.

Risk	Control	Framework mapping	ToD + rationale	ToE + rationale	Overall	Residual	How to improve
RC09	SoD Gate + Independent Model Approval	BAI06 Managed IT Changes; MEA02 Managed System of Internal Control; MEA04 Managed Assurance COSO: Control Environment, Control Activities, Monitoring Activities	ADEQUATE (100.00%) Four-eyes segregation prevents the model developer from approving the same change; independent approval is required before release. 7/7 required design attributes are present. No design attribute is missing in the simulated control specification.	EFFECTIVE (100.00%) Simulation evidence: 1 intercepted / 1 tested; 0 exceptions. Pass rate 100.00% meets the Effective threshold ≥ 98%.	EFFECTIVE	0	Maintain independent approval and periodically recertify SoD; make self-approval technically impossible rather than policy-only.
RC01	SoD + GATE 3 Model Registry	BAI06 Managed IT Changes; BAI10 Managed Configuration; MEA02 Managed System of Internal Control; MEA04 Managed Assurance COSO: Risk Assessment, Control Activities, Monitoring Activities	ADEQUATE (100.00%) The active model/version/hash is checked against the approved registry before automated decisioning, with a fallback to the last approved model. 7/7 required design attributes are present. No design attribute is missing in the simulated control specification.	EFFECTIVE (100.00%) Simulation evidence: 1 intercepted / 1 tested; 0 exceptions. Pass rate 100.00% meets the Effective threshold ≥ 98%.	EFFECTIVE	0	Add signed artifact/hash verification at deployment and automated rollback to last approved version.
RC02	GATE 3 + HOTL Monitor	APO12 Managed Risk; BAI06 Managed IT Changes; MEA01 Managed Performance and Conformance Monitoring; MEA04 Managed Assurance COSO: Risk Assessment, Control Activities, Monitoring Activities	ADEQUATE (100.00%) Pre-deployment validation plus post-deployment performance monitoring jointly address calibration defects before and after release. 7/7 required design attributes are present. No design attribute is missing in the simulated control specification.	EFFECTIVE (100.00%) Simulation evidence: 20,000 intercepted / 20,000 tested; 0 exceptions. Pass rate 100.00% meets the Effective threshold ≥ 98%.	EFFECTIVE	0	Add monthly calibration-by-cohort monitoring, outcome backtesting and independent validation after material change.
RC03	GATE 2 + GATE 3	APO13 Managed Security; APO14 Managed Data; DSS06 Managed Business Process Controls; MEA02 Managed System of Internal Control COSO: Control Environment, Control Activities, Information & Communication	ADEQUATE (100.00%) Data-purpose and feature allow-list gates prevent unapproved proxy features from entering the approved scoring feature set. 7/7 required design attributes are present. No design attribute is missing in the simulated control specification.	EFFECTIVE (100.00%) Simulation evidence: 10,000 intercepted / 10,000 tested; 0 exceptions. Pass rate 100.00% meets the Effective threshold ≥ 98%.	EFFECTIVE	0	Automate feature-store allow-list checks and data-purpose attestations at every model release.
RC11	GATE 3 Lineage / Hash Control	BAI06 Managed IT Changes; BAI10 Managed Configuration; MEA01 Managed Performance and Conformance Monitoring COSO: Control Activities, Monitoring Activities	ADEQUATE (100.00%) Hash and lineage checks detect material configuration changes even when the displayed version label is unchanged. 7/7 required design attributes are present. No design attribute is missing in the simulated control specification.	EFFECTIVE (100.00%) Simulation evidence: 1 intercepted / 1 tested; 0 exceptions. Pass rate 100.00% meets the Effective threshold ≥ 98%.	EFFECTIVE	0	Cryptographically sign model/config artifacts; alert on threshold/config changes even when version label is unchanged.
RC12	IAM + GATE 3 Provider Registry	APO10 Managed Vendors; APO12 Managed Risk; APO13 Managed Security; MEA03 Managed Compliance with External Requirements COSO: Risk Assessment, Control Activities, Information & Communication	ADEQUATE (100.00%) Provider registry and access controls prevent an unregistered external model/provider from being invoked by the Decision Agent. 7/7 required design attributes are present. No design attribute is missing in the simulated control specification.	EFFECTIVE (100.00%) Simulation evidence: 1 intercepted / 1 tested; 0 exceptions. Pass rate 100.00% meets the Effective threshold ≥ 98%.	EFFECTIVE	0	Add egress allow-list, provider attestation and shadow-AI discovery/alerting.
RC08	Dual IAM Permission Gates	APO13 Managed Security; DSS05 Managed Security Services; DSS06 Managed Business Process Controls; MEA02 Managed System of Internal Control COSO: Control Environment, Control Activities	ADEQUATE (100.00%) Dual permission gates constrain both human→agent invocation and agent→data/tool/action permissions under least privilege. 7/7 required design attributes are present. No design attribute is missing in the simulated control specification.	EFFECTIVE (100.00%) Simulation evidence: 5,000 intercepted / 5,000 tested; 0 exceptions. Pass rate 100.00% meets the Effective threshold ≥ 98%.	EFFECTIVE	0	Move to JIT/JEA permissions, short-lived scoped tokens and periodic user/agent access recertification.
RC04	GATE 4 + HITL	APO12 Managed Risk; DSS06 Managed Business Process Controls; MEA02 Managed System of Internal Control COSO: Risk Assessment, Control Activities	ADEQUATE (100.00%) Decision uncertainty thresholds route near-threshold or low-confidence cases to HITL instead of allowing unsupported STP. 7/7 required design attributes are present. No design attribute is missing in the simulated control specification.	PARTIALLY EFFECTIVE (94.00%) Simulation evidence: 3,967 intercepted / 4,220 tested; 253 exceptions. Pass rate 94.00% is below Effective 98% but meets Partially Effective ≥ 90%.	PARTIALLY EFFECTIVE	253	Tune uncertainty/OOD referral thresholds and review false-approval/false-decline trade-offs using backtesting.
RC10	GATE 4 Explainability	DSS06 Managed Business Process Controls; MEA02 Managed System of Internal Control; MEA03 Managed Compliance with External Requirements COSO: Control Activities, Information & Communication	ADEQUATE (100.00%) The adverse-decision gate requires a valid reason code before finalisation, making explainability an executable control. 7/7 required design attributes are present. No design attribute is missing in the simulated control specification.	PARTIALLY EFFECTIVE (97.04%) Simulation evidence: 788 intercepted / 812 tested; 24 exceptions. Pass rate 97.04% is below Effective 98% but meets Partially Effective ≥ 90%.	PARTIALLY EFFECTIVE	24	Make reason-code completeness a hard pre-finalisation condition and sample adverse decisions for quality, not only presence.
RC05	GATE 6 Authority	EDM03 Ensured Risk Optimization; APO12 Managed Risk; DSS06 Managed Business Process Controls; MEA02 Managed System of Internal Control COSO: Control Environment, Control Activities	ADEQUATE (100.00%) Delegated-authority service hard-stops automated decisions above the approved automation limit and routes them to human authority. 7/7 required design attributes are present. No design attribute is missing in the simulated control specification.	EFFECTIVE (100.00%) Simulation evidence: 5,390 intercepted / 5,390 tested; 0 exceptions. Pass rate 100.00% meets the Effective threshold ≥ 98%.	EFFECTIVE	0	Bind every automated decision to a central authority service with hard fail-closed behavior and immutable authority evidence.
RC06	GATE 5 HITL Override	DSS06 Managed Business Process Controls; MEA01 Managed Performance and Conformance Monitoring; MEA02 Managed System of Internal Control COSO: Control Environment, Control Activities, Monitoring Activities	ADEQUATE (100.00%) Override governance requires rationale, evidence and authority before a human reversal can become final. 7/7 required design attributes are present. No design attribute is missing in the simulated control specification.	INEFFECTIVE (88.02%) Simulation evidence: 713 intercepted / 810 tested; 97 exceptions. Pass rate 88.02% is below the Partially Effective threshold 90%.	INEFFECTIVE	97	Require structured override rationale/evidence, peer-rate anomaly monitoring and temporary restriction for repeat outliers.
RC07	HOTL Monitor + Kill Switch	EDM03 Ensured Risk Optimization; APO12 Managed Risk; MEA01 Managed Performance and Conformance Monitoring; MEA02 Managed System of Internal Control COSO: Risk Assessment, Monitoring Activities	ADEQUATE (100.00%) Portfolio HOTL monitoring plus kill-switch authority is designed to detect systemic cohort drift and contain its portfolio-level magnitude. 7/7 required design attributes are present. No design attribute is missing in the simulated control specification.	INEFFECTIVE (33.37%) Simulation evidence: 341 intercepted / 1,022 tested; 681 exceptions. Pass rate 33.37% is below the Partially Effective threshold 90%.	INEFFECTIVE	681	Shorten monitoring latency, add cohort calibration/fairness triggers and pre-agreed kill-switch/restart criteria.

4) Credit model outcomes analysis / backtesting

What this shows: Forecasts are compared with synthetic realised default outcomes; approved/reference and observed/faulted scoring paths are tested in parallel.
Framework / practice basis: Bank model-risk leading practice: conceptual soundness, ongoing monitoring and outcomes analysis/back-testing; parallel outcomes analysis for model changes.
So what: Governance cannot compensate for a model that is deteriorating unnoticed. Model performance needs its own validation layer.

Important: Synthetic realized-default ground truth is generated from an independent RNG and a latent true-PD function. In real lending, declined applicants do not produce observed repayment outcomes without reject-inference or external performance data.

Model metric	Approved/reference	Observed/faulted path	Preferred
AUC How well the score ranks borrowers who default versus those who do not; 0.5 is random ranking and higher is better.	0.6106	0.603	higher
Gini A discrimination measure derived from AUC; higher means stronger rank ordering.	0.2212	0.2061	higher
KS Maximum separation between cumulative good and bad borrower distributions; higher means stronger discrimination.	0.1606	0.1494	higher
Average PD Average predicted probability of default across the portfolio.	25.508	25.182	context
Observed default rate Synthetic realised default rate used as the common outcome set for backtesting.	26.549	26.549	same outcomes
Calibration gap (pp) Difference between average predicted PD and realised default rate; closer to zero means better calibration.	-1.041	-1.367	closer to 0
ECE (%) Expected calibration error across probability buckets; lower means predicted probabilities align better with outcomes.	1.102	1.466	lower
Brier score Mean squared error of probability forecasts; lower means better probabilistic accuracy.	0.18855	0.18966	lower

Approved model assessment
WITHIN LIMITS

Whether the approved/reference model meets the configured validation limits.

Observed path assessment
REVIEW / REMEDIATE

Whether the observed/faulted scoring path remains acceptable against the same validation limits.

Brier deterioration
0.59% · WITHIN LIMIT

Relative worsening in probability forecast error; lower deterioration is better.

Decision-direction diagnostic

For credit decisions, the dashboard uses false approval / false decline rather than relying on Type I / Type II labels because Type-I/II conventions depend on which class is defined as the positive/error event.

Error direction	OFF	ON
False approval vs clean reference Observed path APPROVES a case the clean reference would DECLINE — credit-loss / adverse-selection direction.	1,164	20
False decline vs clean reference Observed path DECLINES a case the clean reference would APPROVE — missed-good-business direction.	1,071	77

Governance OFF vs ON is a controlled governance-state experiment, not a substitute for a model champion/challenger test.

5) ERM — 5×5 heatmap, appetite, movement and concrete response

What this shows: Each theme has an explicit inherent position, residual position, risk appetite metric/limit, current value and management response.
Framework / practice basis: COSO ERM risk appetite/response; NIST MAP 5 and MANAGE 1.2–1.4; COBIT EDM03/APO12; RCSA residual-risk aggregation.
So what: Management needs to see where the risk moved, whether likelihood or impact improved, whether appetite is breached, and what action follows.

Impact anchor: Very Low < IDR 25bn · Low IDR 25–100bn · Moderate IDR 100–250bn · High IDR 250–500bn · Very High > IDR 500bn. Important: financial value is only one impact dimension. Customer, conduct, legal, regulatory and reputational severity can move a risk to a higher impact category even when the monetary value is lower.

Probability methodology: Scenario-configured event occurrence probabilities make the ordinal likelihood scale visible. They are not a historical probability density function; production calibration should use observed event/exposure data and approved expert or statistical estimation. The dashboard therefore reports a point estimate and range for the probability that each risk event occurs within a 12-month horizon. This is different from the current residual-transaction rate.

I = inherent risk before controls; R = residual risk after controls. Arrows show movement. Risk positions are scenario-configured settings and are independent from the raw residual-transaction rate.

Risk event/theme	Inherent occurrence probability	Residual occurrence probability	Likelihood improvement	Time horizon	Financial impact anchor	Management insight
MODEL_AND_CHANGE	40.00% band 30–50%	3.00% band 0–5%	92.50% 37.0 percentage-point reduction	12 months	IDR 56.80 bn Direct financial anchor; not expected loss.	Controls reduce the configured 12-month occurrence likelihood from 40.0% to 3.0% (92.5% relative reduction). A residual 3.0% probability remains that a material model or configuration change event could occur within the next 12 months. The direct financial anchor is IDR 56.80 bn and the final impact category is 4 (High impact); non-financial consequences may increase severity.
DATA_AND_ACCESS	40.00% band 30–50%	3.00% band 0–5%	92.50% 37.0 percentage-point reduction	12 months	IDR 25.27 bn Direct financial anchor; not expected loss.	Controls reduce the configured 12-month occurrence likelihood from 40.0% to 3.0% (92.5% relative reduction). A residual 3.0% probability remains that an unauthorized data, privileged-access or unapproved-provider event could occur within the next 12 months. The direct financial anchor is IDR 25.27 bn and the final impact category is 5 (Very High impact); non-financial consequences may increase severity.
DECISION_AND_CONDUCT	40.00% band 30–50%	10.00% band 5–15%	75.00% 30.0 percentage-point reduction	12 months	IDR 0 Direct financial anchor; not expected loss.	Controls reduce the configured 12-month occurrence likelihood from 40.0% to 10.0% (75.0% relative reduction). A residual 10.0% probability remains that an unresolved decision-quality or conduct event could occur within the next 12 months. The direct financial anchor is IDR 19.63 bn and the final impact category is 3 (Moderate impact); non-financial consequences may increase severity.
AUTHORITY_AND_HUMAN	22.00% band 15–30%	10.00% band 5–15%	54.50% 12.0 percentage-point reduction	12 months	IDR 64.44 bn Direct financial anchor; not expected loss.	Controls reduce the configured 12-month occurrence likelihood from 22.0% to 10.0% (54.5% relative reduction). A residual 10.0% probability remains that an authority-bypass or unsupported-human-override event could occur within the next 12 months. The direct financial anchor is IDR 64.44 bn and the final impact category is 4 (High impact); non-financial consequences may increase severity.
FAIRNESS_AND_PORTFOLIO	40.00% band 30–50%	22.00% band 15–30%	45.00% 18.0 percentage-point reduction	12 months	IDR 15.21 bn Direct financial anchor; not expected loss.	Controls reduce the configured 12-month occurrence likelihood from 40.0% to 22.0% (45.0% relative reduction). A residual 22.0% probability remains that a cohort calibration or fairness-drift event could occur within the next 12 months. The direct financial anchor is IDR 15.21 bn and the final impact category is 3 (Moderate impact); non-financial consequences may increase severity.

Risks that matter

DECISION_AND_CONDUCT — MANAGEMENT ATTENTION. Current 0.277 % applications versus appetite 0.1; response: MITIGATE.
AUTHORITY_AND_HUMAN — WATCHLIST. Current 0.097 % applications versus appetite 0.05; response: MITIGATE.
FAIRNESS_AND_PORTFOLIO — EXECUTIVE / BOARD ATTENTION. Current 17.02 percentage points versus appetite 3; response: MITIGATE.

Response versus control: Risk response is the management strategy (accept, mitigate, transfer or avoid). Control posture describes how controls act (preventive, detective, corrective, containment or recovery). Accept does not mean removing controls; it means the current residual risk is within appetite and existing controls/monitoring are maintained.

Risk theme	Inherent	Residual + movement	Facility value footprint	Appetite metric	Appetite	Warning	Breach	Current	Status	Risk response	Control posture + immediate action	Escalation
MODEL_AND_CHANGE	High impact × Probable likelihood Risk position before the simulated governance/control treatment.	High impact × Improbable likelihood Likelihood ↓ through pre-deployment prevention and fallback Residual = risk remaining after controls.	IDR 1.71 tn linked to affected cases Direct decision-value anchor: IDR 56.80 bn → Low financial anchor. IDR 0 facility value remains linked to residual cases. Systemic model/change failure can affect portfolio integrity beyond direct decision-value impact.	Unauthorized material model/configuration change reaching production Metric used by management to judge whether this risk is tolerable.	0 events Target maximum residual level accepted in normal operation.	0 events Early-warning level that requires management attention before breach.	1 events Threshold requiring formal response/restriction/escalation.	0 events Observed value from this simulation run.	WITHIN APPETITE	ACCEPT Accept residual risk within appetite. Monitoring is maintained as an oversight/control activity, not treated as a separate risk-response category.	PREVENTIVE + RECOVERY Maintain release approval, approved model registry, independent validation and fallback readiness. If an unapproved change is detected, block the release and fall back to the last approved configuration.	Model Risk / AI Governance; Risk Committee if a confirmed escape occurs.
DATA_AND_ACCESS	Very High impact × Probable likelihood Risk position before the simulated governance/control treatment.	Very High impact × Improbable likelihood Likelihood ↓ through least privilege, allow-lists and provider registry Residual = risk remaining after controls.	IDR 1.36 tn linked to affected cases Direct decision-value anchor: IDR 25.27 bn → Low financial anchor. IDR 0 facility value remains linked to residual cases. Privacy, security, data-purpose and regulatory consequences can be severe even where direct credit decision value is limited.	Unauthorized privileged action, prohibited feature or unapproved provider escaping controls Metric used by management to judge whether this risk is tolerable.	0 events Target maximum residual level accepted in normal operation.	0 events Early-warning level that requires management attention before breach.	1 events Threshold requiring formal response/restriction/escalation.	0 events Observed value from this simulation run.	WITHIN APPETITE	ACCEPT Accept residual risk within appetite. Monitoring is maintained as an oversight/control activity, not treated as a separate risk-response category.	PREVENTIVE + DETECTIVE Maintain least-privilege access, feature/provider allow-lists and periodic access recertification. If an exception is detected, deny the action/provider, revoke excess privilege and isolate the affected path.	Security + Data Governance + AI Governance; executive notification for confirmed data exposure.
DECISION_AND_CONDUCT	Moderate impact × Probable likelihood Risk position before the simulated governance/control treatment.	Moderate impact × Remote likelihood Likelihood ↓ through uncertainty referral and adverse-reason gate Residual = risk remaining after controls.	IDR 355.83 bn linked to affected cases Direct decision-value anchor: IDR 0 → Very Low financial anchor. IDR 19.63 bn facility value remains linked to residual cases. Customer treatment, explainability and conduct consequences can raise impact above direct financial value.	Unresolved decision / conduct exceptions as % of applications Metric used by management to judge whether this risk is tolerable.	0.1 % applications Target maximum residual level accepted in normal operation.	0.2 % applications Early-warning level that requires management attention before breach.	0.25 % applications Threshold requiring formal response/restriction/escalation.	0.277 % applications Observed value from this simulation run.	BREACH	MITIGATE Reduce residual risk until it returns within appetite; restrictions and escalation are implementation actions, not separate risk-response categories.	DETECTIVE + CORRECTIVE Route uncertain cases to HITL and hold adverse decisions lacking a valid reason code.	Credit Risk / Conduct / AI Governance until residual rate returns within appetite.
AUTHORITY_AND_HUMAN	High impact × Occasional likelihood Risk position before the simulated governance/control treatment.	High impact × Remote likelihood Likelihood ↓ through hard authority service and override evidence Residual = risk remaining after controls.	IDR 483.22 bn linked to affected cases Direct decision-value anchor: IDR 64.44 bn → Low financial anchor. IDR 7.65 bn facility value remains linked to residual cases. Authority bypass and unsupported override create governance/accountability consequences beyond direct facility value.	Residual authority / unsupported override exceptions as % of applications Metric used by management to judge whether this risk is tolerable.	0.05 % applications Target maximum residual level accepted in normal operation.	0.1 % applications Early-warning level that requires management attention before breach.	0.2 % applications Threshold requiring formal response/restriction/escalation.	0.097 % applications Observed value from this simulation run.	ABOVE APPETITE	MITIGATE Reduce residual risk until it returns within appetite; restrictions and escalation are implementation actions, not separate risk-response categories.	PREVENTIVE + DETECTIVE Remove automated authority outside limits; challenge unsupported overrides and route to senior authority.	Credit Risk management; Risk Committee if breach threshold is exceeded.
FAIRNESS_AND_PORTFOLIO	High impact × Probable likelihood Risk position before the simulated governance/control treatment.	Moderate impact × Occasional likelihood Likelihood ↓ and impact ↓ through HOTL detection, kill switch and affected-cohort containment Residual = risk remaining after controls.	IDR 72.12 bn linked to affected cases Direct decision-value anchor: IDR 15.21 bn → Very Low financial anchor. IDR 48.10 bn facility value remains linked to residual cases. Systemic cohort fairness and regulatory/customer harm can raise impact above direct facility-value impact.	Absolute cohort approval-rate drift versus reference Metric used by management to judge whether this risk is tolerable.	3 percentage points Target maximum residual level accepted in normal operation.	5 percentage points Early-warning level that requires management attention before breach.	8 percentage points Threshold requiring formal response/restriction/escalation.	17.02 percentage points Observed value from this simulation run.	BREACH	MITIGATE Reduce residual risk until it returns within appetite; restrictions and escalation are implementation actions, not separate risk-response categories.	DETECTIVE + CONTAINMENT + CORRECTIVE Temporarily suspend AI auto-finalisation for the affected cohort/path, route those cases to HITL or approved fallback, and investigate drift/fairness drivers.	Risk Committee / senior management while outside appetite.

6) Control environment improvement — ex-post root cause, corrective action and retest

What this shows: Control failures are fed back into the internal-control environment: risk response, control redesign, temporary restriction, owner action and retest.
Framework / practice basis: COSO Internal Control control activities/monitoring; COBIT DSS02/DSS03/MEA02/MEA04; NIST MANAGE response/recovery; RCA/CAPA and three-lines assurance.
So what: The purpose is not only to fix an incident. It is to improve the control environment so the same failure is less likely to recur and closure is evidence-based.

Internal-control meaning: a control weakness should feed back into control design and operating practice. Preventive controls stop an event before it happens; detective controls identify it; corrective/containment/recovery controls limit impact and restore a controlled state.

What retest means
Repeat the same failure scenario after remediation.

How to test
Same seed/comparable population → same root cause → rerun ToE/model outcomes → compare exceptions and appetite.

Closure
Action completed + retest passes + residual risk within appetite + required independent/challenge sign-off.

RC	Root cause	Risk response	Control type	Immediate action	Control change	Restriction	Owner	Retest method	Expected result	Closure criteria
RC04	Low-uncertainty-confidence STP misrouting	MITIGATE ERM response strategy: accept, mitigate, transfer or avoid. Monitoring/restriction/escalation are implementation actions, not separate response categories.	DETECTIVE + CORRECTIVE How the control acts: preventive, detective, corrective, containment or recovery.	Route near-threshold/uncertain cases to HITL.	Tune uncertainty/OOD referral threshold using false-approval/false-decline trade-off.	Mandatory HITL while residual exception rate is above appetite.	AI Governance / Credit Risk	Same seed/comparable population → reinject low-confidence misrouting → rerun ToE; target ≥98% pass and residual rate within appetite.	ToD remains adequate; ToE ≥98% or approved threshold; residual risk within theme-specific appetite.	Evidence of completed action + successful retest + residual risk within appetite + independent/challenge sign-off where required
RC10	Missing adverse-decision reason code	MITIGATE ERM response strategy: accept, mitigate, transfer or avoid. Monitoring/restriction/escalation are implementation actions, not separate response categories.	PREVENTIVE + DETECTIVE How the control acts: preventive, detective, corrective, containment or recovery.	Hold adverse decisions missing a valid reason code.	Hard pre-finalisation reason-code completeness and quality checks.	No adverse decision finalisation without usable explanation.	AI Governance / Credit Risk	Inject missing/invalid reason codes; target ≥98% interception and residual rate within appetite.	ToD remains adequate; ToE ≥98% or approved threshold; residual risk within theme-specific appetite.	Evidence of completed action + successful retest + residual risk within appetite + independent/challenge sign-off where required
RC06	Unsupported human override	MITIGATE ERM response strategy: accept, mitigate, transfer or avoid. Monitoring/restriction/escalation are implementation actions, not separate response categories.	DETECTIVE + CORRECTIVE How the control acts: preventive, detective, corrective, containment or recovery.	Challenge unsupported overrides and require evidence before finalisation.	Structured rationale/evidence plus peer-rate anomaly monitoring.	Restrict repeat outlier reviewers pending investigation.	AI Governance / Credit Risk	Replay override anomalies; target ≥98% effective interception and residual exception rate ≤ theme appetite.	ToD remains adequate; ToE ≥98% or approved threshold; residual risk within theme-specific appetite.	Evidence of completed action + successful retest + residual risk within appetite + independent/challenge sign-off where required
RC07	Cohort calibration / fairness drift	MITIGATE ERM response strategy: accept, mitigate, transfer or avoid. Monitoring/restriction/escalation are implementation actions, not separate response categories.	DETECTIVE + CONTAINMENT + CORRECTIVE How the control acts: preventive, detective, corrective, containment or recovery.	Suspend affected auto-decision path and make HITL mandatory for the cohort.	Tighter cohort drift/fairness monitoring, recalibration and kill-switch criteria.	No controlled restart until independent validation and appetite re-entry.	AI Governance / Credit Risk	Reinject cohort drift with same monitoring checkpoint; verify trigger latency, containment, fairness drift ≤3pp and ToE ≥98% before restart.	ToD remains adequate; ToE ≥98% or approved threshold; residual risk within theme-specific appetite.	Evidence of completed action + successful retest + residual risk within appetite + independent/challenge sign-off where required

7) Meaning of recurring dashboard terms

Why this exists: A metric without a decision meaning is just a number. This glossary states what each recurring term is intended to tell the reader.

STP

Straight-through processing: applications finalized without human review. Higher is faster, but not automatically better if control risk rises.

HITL

Human in the loop: a specific application is paused for accountable human review before finalisation.

HOTL

Human on the loop: portfolio-level supervision while automation runs, with authority to intervene when thresholds are breached.

Kill switch

Targeted suspension/deactivation of the affected AI automated path or component. In this simulation the loan process continues through HITL or approved fallback.

Residual risk

Risk remaining after governance and controls have acted; it is compared with risk appetite to decide whether further treatment is required.

Governance effectiveness

Percentage reduction from baseline fault-affected transactions to the residual population after controls.

ToD / ToE

Test of Design asks whether the control is capable by design. Test of Operating Effectiveness asks whether it actually worked over the tested population.

Risk appetite / warning / breach

Appetite = accepted residual level; warning = early management trigger; breach = formal response/escalation threshold.

Inherent / residual

Inherent = risk before the specified controls. Residual = risk after those controls and treatments.

Risk response

Management strategy for the risk: accept, mitigate, transfer/share or avoid. It is different from the type of control used.

Control type

Preventive stops an event; detective identifies it; corrective fixes causes; containment limits spread/impact; recovery restores a controlled operating state.

Facility value

Gross requested principal linked to the cases being discussed. It is not expected loss, impairment, capital impact or realised loss.

Occurrence probability

Probability that a defined risk event occurs within the stated time horizon. It is not the same as the percentage of applications currently affected and is not presented as a historical probability-density function.

Residual-case treatment

Disposition applied to known faulty cases after automated controls: hold/review, correct/reprocess, stop, post-booking remediation or incident investigation.

Synthetic educational demonstration. Quantitative appetite levels, heatmap positions and 12-month event probabilities are scenario-configured assumptions, not universal regulatory thresholds or empirically fitted probability distributions. Model-validation metrics use synthetic realised outcomes. In production, thresholds must be approved by the institution and calibrated to portfolio materiality, legal obligations and risk appetite.