71Governance health

Create and manage governance policies applied to your AI systems. Policies define the rules Aegis enforces — automatically or with human approval.

Policy vs Rule: A Policy is a set of rules applied to one or more AI systems. A Rule is a specific condition within a policy — e.g. “Block decisions where demographic proxy score > 0.8”. Policies are versioned. Every change creates a new version with a full audit trail.

AI Control Plane / Policies

Policies

Rules that enforce automatically across every AI system

New Policy

Active Policies

9

Block Rules

5

Flag Rules

4

Regulations Covered

4

MER-POL-001 is inactive— Trading restrictions cannot be enforced because dependent models are unregistered.

Policy ID▲	Policy Name	Scope	Regulation	Enforcement	Triggered	Status	Owner
MER-POL-001	Fair Lending Amendment All loan decisions above $50,000 require disparate impact assessment before approval. Human sign-off mandatory.	All Lending Models	ECOA	BLOCK	Pending approval — 4 days	INACTIVE	arun@meridian.com
MER-POL-002	SR 11-7 Model Validation Policy All High Risk models require independent validation before production deployment.	High Risk Models	SR 11-7	BLOCK		ACTIVE	priya@meridian.com
MER-POL-003	EU AI Act Article 6 Compliance High-risk AI systems under EU AI Act must maintain technical documentation and human oversight mechanisms.	All Models	EU AI Act	BLOCK		ACTIVE	priya@meridian.com
MER-POL-004	AML Threshold Policy AML Detection thresholds adjusted to reflect Q1 transaction volume increase.	AML Models	FinCEN	BLOCK	Pending approval	INACTIVE	arun@meridian.com
MER-POL-005	Data Quality Minimum Standards All model input feeds must maintain >99% availability and <2hr latency for production systems.	All Models	SR 11-7	FLAG		ACTIVE	dev@meridian.com
MER-POL-006	Quarterly Validation Schedule Medium and High risk models require quarterly validation. Low risk models annually.	All Models	SR 11-7	FLAG		ACTIVE	priya@meridian.com
MER-POL-007	Human Override Requirement Any model decision affecting individual credit, employment, or insurance outcomes requires human review before action.	All Credit Models	ECOA	BLOCK		ACTIVE	arun@meridian.com
MER-POL-008	Model Retirement Protocol Models inactive for >90 days or with accuracy below 85% must be formally retired with audit evidence.	All Models	SR 11-7	FLAG		ACTIVE	priya@meridian.com
MFG-FAIR-LENDING-V2	Fair Lending Amendment — Disparate Impact v2 Prohibits deployment of any lending model where demographic parity delta exceeds 10% without documented MRM exemption. Submitted by Dev Kumar 4 days ago — pending Arun Mehta review and approval before Loan Origination Model v2 can deploy.	Loan Origination Model v2	ECOA / SR 11-7	BLOCK	FAIR_LENDING_BLOCK	INACTIVE	Dev Kumar
POL-FIN-V2-001	Fair Lending Policy v4 Block any lending decision where demographic proxy variables or disparate impact patterns are detected.	All Lending Models	ECOA	BLOCK	412 times today	ACTIVE	Priya Sharma
POL-FIN-V2-002	AML-003 Transaction Monitoring Monitor all transactions for AML patterns per BSA requirements. Escalate anomalies.	AML Detection v3	BSA/AML	FLAG	1,847 times today	ACTIVE	Priya Sharma
POL-FIN-V2-003	Debt-to-Income Threshold Deny loan origination where DTI exceeds regulatory threshold. SR 11-7 requirement.	Loan Origination Model v2	SR 11-7	BLOCK	203 times today	ACTIVE	Priya Sharma
POL-FIN-V2-004	Fair Lending Amendment — Disparate Impact v2 Amendment to Fair Lending Policy — adds disparate impact screening across ZIP codes. Pending approval — 4 days in queue.	All Lending Models	ECOA	FLAG	0 — pending approval	INACTIVE	Priya Sharma

ECOA / Fair Lending

Lending models — demographic proxy detection active

2 active policies

OCC SR 11-7

Model risk management — confidence, versioning, human review

5 active policies

OSFI B-10

Third-party AI risk — registration enforcement, AML escalation

0 active policies

Policy Checks

No policy checks recorded yet

Policy checks appear here when an AI system makes a decision that triggers a governance rule. Decisions are logged in real time once a system is connected.