Algorithmic Supply Chain

The origin of all training, validation, and fine-tuning data. Training Data Lineage tracks the end-to-end origin, movement, and transformation history of all datasets. Key considerations include:

• Provenance: Verifying the source and legal right to use data.
• Consent: Ensuring data was collected with appropriate user permissions.
• Copyright: Scanning for intellectual property violations via a Copyright Infringement Scan.

Entities that pre-train large-scale models on massive datasets. They are subject to General Purpose AI Obligations under the EU AI Act. Critical transparency artifacts include:

• Foundation Model Transparency Report: Disclosing training data, compute resources, and capabilities.
• Model Card: A structured document detailing intended use, benchmarks, and limitations.

Providers of the software infrastructure used to build, test, and deploy AI. This includes MLOps platforms, vector databases, and orchestration tools. Risk is often concentrated in Hyperscaler Concentration Risk, the operational vulnerability from over-dependence on a single cloud provider. Interoperability Standards like ONNX are crucial to mitigate Vendor Lock-In Risk.

Specialists who adapt a foundation model to a specific domain using techniques like Parameter-Efficient Fine-Tuning (PEFT). They introduce new data and weights into the supply chain, creating a new Model Provenance trail. Their work must be captured in an AI Bill of Materials (AIBOM), a formal inventory of all components used.

Third-party auditors and internal red teams that stress-test models. They produce Red-Teaming Reports and measure Jailbreak Susceptibility. Key techniques include:

• Adversarial Robustness Benchmark: Testing resilience against evasion and poisoning.
• Dangerous Capability Benchmark: Measuring proficiency in domains like bioweapons design.

The final stage where the model is served to users. This layer introduces operational risks requiring Guardrail Configuration and Output Moderation APIs. A Kill Switch Mechanism must be in place for immediate shutdown during a critical failure. Post-Market Surveillance continuously monitors real-world performance and safety.