Project Requirements Specification

1. Project Overview

Background

The EU AI Act (enacted in 2024, fully enforced in 2026) classifies AI systems by risk level and mandates conformity assessments, post-deployment monitoring, and incident reporting for high-risk AI. Organizations must manage the entire lifecycle of AI models — from model registration, risk classification, deployment approval, compliance assessment, to incident response.

This system automates AI model governance within a single bounded context. It implements the core requirements of the EU AI Act (risk tier classification, conformity assessment, incident management) using DDD tactical patterns and the Functorium framework, serving as a full-stack DDD example that demonstrates practical patterns across the Domain/Application/Adapter layers.

Goals

1. Automate core EU AI Act requirements (risk tier classification, compliance assessment, incident management)
2. Implement independent lifecycles of 4 Aggregates and cross-domain rules using DDD tactical patterns
3. Demonstrate practical application of LanguageExt IO advanced features (Timeout, Retry, Fork, Bracket)
4. Automate OpenTelemetry 3-Pillar observability using Source Generators
5. Provide a full-stack DDD reference implementation of the Functorium framework

Target Users

Persona	Role	Core Goal
AI Governance Administrator	AI model registration, deployment approval, risk tier classification	Manage the entire lifecycle of organizational AI models and ensure regulatory compliance
Compliance Officer	Compliance assessment, incident investigation, audit reporting	Evaluate EU AI Act requirement compliance and respond promptly to incidents

Success Metrics (KPIs)

Metric Type	Metric	Target	Measurement Method
Leading	Model registration completion rate	> 95%	Success ratio relative to registration requests
Leading	Incident auto-quarantine response time	< 1 second	Time from Critical incident report to deployment quarantine
Lagging	Compliance assessment pass rate	> 80%	Passed ratio relative to total assessments
Lagging	Average unresolved incident resolution time	< 24 hours	Average time from Reported to Resolved

Technical Constraints

• .NET 10 / C# 14
• Functorium framework (DDD Building Blocks, LanguageExt IO, Source Generator)
• FastEndpoints (HTTP API)
• OpenTelemetry 3-Pillar (Metrics, Tracing, Logging)
• InMemory / Sqlite persistence switching

---

2. Non-Goals

Scope explicitly excluded from this project:

• Model training pipeline — Model training/fine-tuning is the responsibility of the ML platform; this system handles only governance of trained models
• A/B testing platform — Model performance comparison experiments are conducted on a separate experimentation platform; this system manages only regulatory compliance of deployed models
• Real-time dashboard — Model performance monitoring dashboards are provided by a separate observability platform; this system is responsible only up to creating incidents when drift thresholds are exceeded

---

3. Ubiquitous Language

English	Definition
AIModel	An AI/ML model registered and subject to governance
ModelName	Name of the AI model (100 characters or fewer)
ModelVersion	Model version in SemVer format
ModelPurpose	Description of the model’s intended use (500 characters or fewer)
RiskTier	EU AI Act-based 4-level classification: Minimal, Limited, High, Unacceptable
ModelDeployment	A deployment instance of an AI model in a production environment
DeploymentStatus	Current status of the deployment (Draft, PendingReview, Active, Quarantined, Decommissioned, Rejected)
DeploymentEnvironment	Target deployment environment: Staging, Production
EndpointUrl	Service endpoint of the deployed model
DriftThreshold	Model performance drift detection threshold (0.0~1.0)
ComplianceAssessment	Regulatory compliance assessment for a deployment
AssessmentCriterion	Individual criterion item of a compliance assessment
AssessmentScore	Composite assessment score in range 0~100; passing score is 70 or above
AssessmentStatus	Assessment progress status (Initiated, InProgress, Passed, Failed, RequiresRemediation)
CriterionResult	Individual criterion evaluation result: Pass, Fail, NotApplicable
ModelIncident	Accident/issue report related to an AI model
IncidentSeverity	Critical, High, Medium, Low
IncidentStatus	Incident progress status (Reported, Investigating, Resolved, Escalated)
IncidentDescription	Detailed description of the incident (2000 characters or fewer)
ResolutionNote	Incident resolution record (2000 characters or fewer)
RiskClassificationService	Risk tier classification based on model purpose keywords
DeploymentEligibilityService	Cross-Aggregate eligibility verification before deployment
HealthCheck	Health status check of a deployed model
DriftReport	Model performance drift monitoring result
ModelRegistry	External model metadata repository

---

4. User Stories

AI Governance Administrator

ID	Story	Priority
US-001	As an AI Governance Administrator, I want to register an AI model to understand the current state of AI models within the organization.	P0
US-002	As an AI Governance Administrator, I want to classify a model’s risk tier to determine the appropriate management level per EU AI Act regulations.	P0
US-003	As an AI Governance Administrator, I want to deploy a model to serve it in the production environment.	P0
US-004	As an AI Governance Administrator, I want to submit a deployment for review to get approval through eligibility verification.	P0
US-005	As an AI Governance Administrator, I want to quarantine a problematic deployment to prevent further damage.	P0
US-006	As an AI Governance Administrator, I want to search models and deployment status to manage them by risk tier and status filters.	P1

Compliance Officer

ID	Story	Priority
US-007	As a Compliance Officer, I want to initiate a compliance assessment to verify a deployment’s regulatory compliance.	P0
US-008	As a Compliance Officer, I want to report an incident to document the issue and prompt an immediate response.	P0
US-009	As a Compliance Officer, I want to investigate and resolve incidents to identify root causes and prevent recurrence.	P1
US-010	As a Compliance Officer, I want to search incident lists by severity/status to respond according to priority.	P1

---

5. Aggregate Candidates

Aggregate	Core Responsibility	State Transitions	Key Events
AIModel	Model registration, risk tier classification, archive/restore	(None, Soft Delete guard)	RegisteredEvent, RiskClassifiedEvent, ArchivedEvent
ModelDeployment	Deployment creation, state transitions, health check recording	Draft -> PendingReview -> Active -> Quarantined -> Decommissioned	CreatedEvent, ActivatedEvent, QuarantinedEvent
ComplianceAssessment	Assessment creation, criterion evaluation, score calculation, completion	Initiated -> InProgress -> Passed/Failed/RequiresRemediation	CreatedEvent, CriterionEvaluatedEvent, CompletedEvent
ModelIncident	Incident reporting, investigation, resolution, escalation	Reported -> Investigating -> Resolved/Escalated	ReportedEvent, ResolvedEvent

Aggregate Relationship Diagram

---

6. Business Rules

AIModel Rules

1. Model name must be 100 characters or fewer and must not be empty
2. Model version must be in SemVer format
3. Model purpose must be 500 characters or fewer and must not be empty
4. Risk tier is classified into 4 levels: Minimal, Limited, High, Unacceptable
5. Automatic risk tier classification based on model purpose keywords is supported
6. Archived models cannot be modified (Soft Delete guard)
7. Archive and restore operations are idempotent

ModelDeployment Rules

1. Endpoint URL must be a valid HTTP/HTTPS URL
2. Drift threshold must be in the range 0.0~1.0
3. Deployment status transitions follow only the defined transition map (6 states, 2 terminal states)
4. Health checks can be recorded

ComplianceAssessment Rules

1. Default assessment criteria of 3: Data Governance, Technical Documentation, Security Review
2. For High/Unacceptable tiers, an additional 3 criteria: Human Oversight, Bias Assessment, Transparency
3. For Unacceptable tier, an additional 1 criterion: Prohibition Review
4. All criteria must be evaluated before the assessment can be completed
5. The composite score is automatically calculated as the Pass ratio among applicable criteria (0~100)
6. 70 or above is Passed, 40~69 is RequiresRemediation, below 40 is Failed

ModelIncident Rules

1. Incident description must be 2000 characters or fewer
2. Critical/High severity incidents trigger automatic deployment quarantine
3. Incident status transitions follow only the defined transition map

Cross-Domain Rules

1. Risk Tier Classification: Risk tier is determined by analyzing model purpose keywords (RiskClassificationService)
2. Deployment Eligibility Verification: Prohibited tier check -> Compliance assessment check -> Unresolved incident check (DeploymentEligibilityService)
3. Automatic Assessment Initiation on Risk Tier Upgrade: RiskClassifiedEvent -> Create assessment for each active deployment (EventHandler)
4. Automatic Deployment Quarantine on Critical Incident: ReportedEvent -> Quarantine deployment (EventHandler)

---

7. Use Cases + Acceptance Criteria

Commands (Write)

Use Case	Input	Core Logic	Output	Priority
RegisterModelCommand	Name, Version, Purpose	VO composition -> Risk classification -> Model creation	ModelId	P0
ClassifyModelRiskCommand	ModelId, RiskTier	Model lookup -> Reclassify -> Update	—	P0
CreateDeploymentCommand	ModelId, Url, Env, Drift	VO composition -> Model verification -> Deployment creation	DeploymentId	P0
SubmitDeploymentForReviewCommand	DeploymentId	Deployment lookup -> Eligibility verification -> Submit	—	P0
ActivateDeploymentCommand	DeploymentId, AssessmentId	Deployment/Assessment lookup -> Pass verification -> Activate	—	P0
QuarantineDeploymentCommand	DeploymentId, Reason	Deployment lookup -> Quarantine	—	P0
InitiateAssessmentCommand	ModelId, DeploymentId	Model/Deployment lookup -> Assessment creation	AssessmentId	P0
ReportIncidentCommand	DeploymentId, Severity, Desc	VO composition -> Deployment lookup -> Incident creation	IncidentId	P0

RegisterModelCommand Acceptance Criteria

Success Scenario:

Given: A valid model name, SemVer version, and model purpose are provided
When:  An AI Governance Administrator registers a model
Then:  The model is created, the risk tier is automatically classified based on purpose keywords, and a ModelId is returned

Rejection Scenario:

Given: The model name is empty and the version is not in SemVer format
When:  An AI Governance Administrator registers a model
Then:  Both errors are returned simultaneously (ApplyT parallel validation)

SubmitDeploymentForReviewCommand Acceptance Criteria

Success Scenario:

Given: A deployment in Draft status exists, the model is not in a prohibited tier, the compliance assessment has passed, and there are no unresolved incidents
When:  An AI Governance Administrator submits for review
Then:  The deployment status transitions to PendingReview

Rejection Scenario (Prohibited Tier):

Given: A deployment referencing a model with Unacceptable risk tier exists
When:  An AI Governance Administrator submits for review
Then:  A ProhibitedModel error is returned and the deployment status remains unchanged

Rejection Scenario (Failed Compliance):

Given: A High risk tier model has no passed compliance assessment
When:  An AI Governance Administrator submits for review
Then:  A ComplianceAssessmentRequired error is returned

ReportIncidentCommand Acceptance Criteria

Success Scenario (Auto-Quarantine):

Given: A deployment in Active status exists
When:  A Compliance Officer reports a Critical severity incident
Then:  The incident is created in Reported status, and the event handler automatically quarantines the deployment

InitiateAssessmentCommand Acceptance Criteria

Success Scenario:

Given: A registered model and deployment exist
When:  A Compliance Officer initiates an assessment
Then:  Assessment criteria are automatically generated based on the risk tier (Minimal: 3, High: 6, Unacceptable: 7) and an AssessmentId is returned

Queries (Read)

Use Case	Input	Query Strategy	Output	Priority
GetModelByIdQuery	ModelId	IModelDetailQuery	Model detail (deployment/assessment/incident aggregation)	P0
SearchModelsQuery	RiskTier?, Page, Size	IAIModelQuery	Model list	P1
GetDeploymentByIdQuery	DeploymentId	IDeploymentDetailQuery	Deployment detail	P0
SearchDeploymentsQuery	Status?, Env?, Page, Size	IDeploymentQuery	Deployment list	P1
GetAssessmentByIdQuery	AssessmentId	IAssessmentRepository	Assessment detail (including criteria)	P0
GetIncidentByIdQuery	IncidentId	IIncidentRepository	Incident detail	P0
SearchIncidentsQuery	Severity?, Status?, Page, Size	IIncidentQuery	Incident list	P1

Event Handlers (Reactive)

Trigger Event	Action	Priority
ModelIncident.ReportedEvent	Auto-quarantine deployment on Critical/High severity (QuarantineDeploymentOnCriticalIncidentHandler)	P0
AIModel.RiskClassifiedEvent	Auto-create assessment for active deployments on High/Unacceptable upgrade (InitiateAssessmentOnRiskUpgradeHandler)	P0

QuarantineDeploymentOnCriticalIncidentHandler Acceptance Criteria

Given: A Critical severity incident is reported for a deployment in Active status
When:  The ReportedEvent is published
Then:  The event handler auto-quarantines the deployment, including the severity in the quarantine reason

InitiateAssessmentOnRiskUpgradeHandler Acceptance Criteria

Given: A model upgraded from Minimal to High risk tier has 2 Active deployments
When:  The RiskClassifiedEvent is published
Then:  The event handler creates a ComplianceAssessment for each active deployment (2 total)

---

8. Prohibited States

Prohibited State	Prevention Strategy	Functorium Pattern
Unacceptable tier model having active deployments	riskTier.IsProhibited check in DeploymentEligibilityService	Smart Enum domain property (RiskTier.IsProhibited) + Domain Service
Deployment transitioning directly from Draft to Active	Structurally blocked by DeploymentStatus transition map	Smart Enum + HashMap transition map (CanTransitionTo)
Assessment completed without all criteria evaluated	Complete() method returns Fin.Fail when unevaluated criteria exist	Aggregate Root guard method

---

9. Priority Summary

Priority	Criteria	Use Case Count	Notes
P0 (Required)	Cannot ship without it	12	MVP: Command 8 + Query 4 (ID lookup) + EventHandler 2
P1 (Important)	Competitiveness weakened without it	5	Search/Filter: Query 3 + Incident investigation/resolution
P2 (Optional)	Differentiating if present	—	EfCore/Dapper persistence, Prometheus dashboard

---

10. Timeline

Milestone	Scope	Target	Dependencies
Phase 1 (MVP)	P0 Use Cases + InMemory Persistence	Week 4	—
Phase 2	P1 Use Cases + Sqlite Persistence	Week 8	Phase 1 Complete
Phase 3	P2 + Observability Dashboard + Alert Rules	Week 12	Phase 2 Complete

---

11. Open Questions

ID	Question	Category	Blocking	Owner
Q-001	Should ComplianceAssessment criteria be dynamically loaded from an external configuration file?	engineering	Non-blocking	Architect
Q-002	Is integration with external notification systems (Slack, PagerDuty) needed for incident escalation?	product	Non-blocking	PM
Q-003	Should registration of Unacceptable tier models be blocked entirely, or is blocking only deployment sufficient?	legal	Non-blocking	Legal
Q-004	Is automatic incident creation needed when model drift threshold is exceeded?	engineering	Non-blocking	Architect

---

12. Next Steps

1. architecture-design — Project structure + infrastructure design
2. domain-develop — Detailed design + implementation of each Aggregate
3. application-develop — Use case implementation
4. adapter-develop — Persistence + API implementation
5. observability-develop — Observability design + implementation
6. test-develop — Test authoring
7. review — Code review + documentation