Particle.news
Download on the App Store

Cisco Releases Open-Source Toolkit and Constitution to Verify AI Model Lineage

The effort sets a technical and definitional baseline for tracing open models through the AI supply chain.

Overview

  • Cisco's Model Provenance Kit, released Thursday, fingerprints transformer weights to tell if one model descends from another.
  • The companion Model Provenance Constitution defines lineage as a proven chain of weight derivation, sets a strict evidence bar, and defaults to no link when proof is unclear.
  • The toolkit uses a two-stage process that screens architecture first and then blends five weight-based signals into a single identity score.
  • It ships with compare and scan modes, open code on GitHub, and an initial fingerprint set for about 150 base models on Hugging Face.
  • Cisco reports 96.4% accuracy on a 111-pair test with strong precision and recall, while noting four misses after extreme architectural changes that push the method to its limits.