Managing model inventory

Overview

The model inventory is where you register every AI model your organization uses, whether it's a third-party API, an open-source model you host, or something built in-house. Each model gets an approval status, an owner, and a record of changes over time.

The point is visibility. If you can't answer basic questions (what AI do we run? who owns it? what data does it touch? is it approved?) then you can't govern it. The inventory gives you that baseline.

Why bother?

• Regulatory requirement: The EU AI Act (Article 60) and ISO 42001 both require documented records of AI systems in use, especially high-risk ones
• Risk visibility: You can't assess risks for models you don't know about. The inventory surfaces everything for review.
• Accountability: Every model has an owner. When something goes wrong, there's a clear person responsible.
• Audit readiness: Auditors ask "show me all your AI systems." With an inventory, you can answer in seconds instead of weeks.

Accessing the model inventory

Open Model inventory from the sidebar (under Inventory). You get a table of all registered models with search, filters for status and provider, and summary cards showing counts by approval status at the top.

Registering a new model

To add a new AI model to your inventory, click the Add model button and provide the required information:

1. Provider: — The organization or service that provides the model (e.g., OpenAI, Anthropic, internal team)
2. Model name: — The specific model identifier (e.g., GPT-4, Claude 3, custom-classifier-v2)
3. Version: — The version number or release identifier
4. Approver: — The person responsible for approving this model for use

Model attributes

Beyond the basics, each model record supports these fields:

Approval status

Every model in the inventory has an approval status that controls whether it can be used in your organization:

• Pending: Model is awaiting review and approval before use
• Approved: Model has been reviewed and authorized for production use
• Restricted: Model is approved for limited use cases or specific projects only
• Blocked: Model is not authorized for use in the organization

Security assessment

Models can be flagged as having completed a security assessment. When enabled, you can attach security assessment documentation directly to the model record for easy reference during audits.

Linking evidence

You can attach evidence files to any model record. Typical examples:

• Model cards and technical documentation
• Vendor contracts and data processing agreements
• Security assessment reports
• Bias testing results and fairness evaluations
• Performance benchmarks and validation studies

MLFlow integration

If you use MLflow, VerifyWise can pull in model training metadata directly: training timestamps, parameters, metrics and lifecycle stage. This saves you from manually entering data that your ML platform already has.

Change history

VerifyWise automatically maintains a complete audit trail for every model in your inventory. Each change records:

• The field that was modified
• Previous and new values
• Who made the change
• When the change occurred

Auditors typically ask for this when reviewing how your organization tracks AI system changes over time.

Datasets

Datasets have their own page in the sidebar (Inventory → Datasets). You catalog the data used for training, validation, testing and production. Each dataset can be linked to models and use cases, so you can trace which data feeds which system.

Accessing datasets

Navigate to Datasets from the main sidebar under Inventory. The page shows all registered datasets in a searchable table with status summary cards at the top.

Adding a new dataset

To add a new dataset to your inventory, click the Add new dataset button and provide the required information:

1. Name: — A descriptive name for the dataset
2. Description: — Detailed explanation of what the dataset contains and its intended use
3. Version: — The version identifier for tracking dataset iterations
4. Owner: — The person or team responsible for maintaining the dataset
5. Type: — The purpose of the dataset (training, validation, testing, production, or reference)
6. Function: — The dataset's role in AI model development
7. Source: — Where the data originated from
8. Classification: — The sensitivity level of the data
9. Status: — The current lifecycle stage of the dataset
10. Status date: — When the current status was set

Dataset types

Datasets can be categorized by their purpose in the machine learning lifecycle:

• Training: Data used to train the model and learn patterns
• Validation: Data used to tune hyperparameters and prevent overfitting during training
• Testing: Data used to evaluate final model performance before deployment
• Production: Data that the deployed model processes in live environments
• Reference: Baseline or benchmark data used for comparison

Data classification

Each dataset should be classified according to its sensitivity level:

• Public: Data that can be freely shared without restrictions
• Internal: Data intended for use within the organization only
• Confidential: Sensitive data requiring access controls and handling procedures
• Restricted: Highly sensitive data with strict access limitations and regulatory requirements

Dataset status

Every dataset has a status indicating its current lifecycle stage:

• Draft: Dataset is being prepared or documented but not yet ready for use
• Active: Dataset is approved and currently in use for model development or production
• Deprecated: Dataset is no longer recommended for new use but may still be referenced by existing models
• Archived: Dataset is retained for historical purposes but not available for active use

Dataset attributes

Each dataset can include additional attributes to support governance and data quality:

Linking datasets to models

When creating or editing a dataset, you can link it to one or more models in your inventory. This traceability matters when a data quality issue surfaces: you can immediately see which models are affected.

Linking datasets to use cases

In addition to models, datasets can be linked to specific use cases (projects) in your organization. This helps maintain a clear view of which data supports which business applications, supporting both governance oversight and impact analysis.

Optional fields

Beyond the required fields, you can document additional metadata to enhance governance:

• License: The licensing terms governing data use (e.g., CC BY 4.0, MIT, proprietary)
• Format: The data format (e.g., CSV, JSON, Parquet)
• PII types: Specific types of personally identifiable information when PII is present