Skip to content

Feature Request: EU AI Act compliance metadata for unified ML framework #28938

@shotwellj

Description

@shotwellj

Summary

With the EU AI Act enforcement deadline on August 2, 2026, and Unify being based in London, Ivy falls within scope of the Act. As a framework that unifies multiple ML backends, Ivy has a unique opportunity: compliance metadata could flow through the framework layer, giving every downstream project compliance visibility regardless of which backend they use.

What this could look like

  • Art. 11 (Documentation): Expanded docstrings and type hints (currently at 24% and 45% respectively), auto-generated system cards for multi-backend deployments
    • Art. 12 (Record-Keeping): Framework-level logging that captures which backend handled each operation, with structured audit trails
      • Art. 14 (Human Oversight): Backend-agnostic budget controls and execution timeouts
        • Art. 15 (Security): Input validation at the framework boundary before dispatching to backends

Context

I ran Ivy through AIR Blackbox, an open-source EU AI Act compliance scanner (Apache 2.0). 1,473 Python files scanned, 7/45 checks passing (16%).

You can run it yourselves:

pip install air-blackbox
air-blackbox comply --scan . --no-llm --format table --verbose

Everything runs locally, no data leaves your machine.

Why this matters

As a framework layer, Ivy can surface compliance metadata that individual backends don't provide. This is a strong differentiator: teams that use Ivy get compliance observability across all their ML backends automatically.

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions