Govern agents by model provenance
Weights & Biases owns the record of how a model was built — versions, evaluations, dataset lineage, who promoted what. KYE™ decides whether an agent running that model may still act. Connecting the two makes provenance a precondition of authority.
An agent whose model silently changed to an unapproved version is an agent acting on authority no one granted. This integration catches exactly that — at the action boundary, with evidence.
How it works
This is an integration you wire up, not an app you open: a W&B registry webhook flows into KYE™, which turns each provenance event into an authority decision.
- Connect. Point your W&B model-registry webhooks at KYE™ — version changed, eval regressed, unapproved promotion, unregistered model in use.
- Convert. KYE™ normalises each event into a governed drift signal, reusing the same vendor-neutral contract as behavioural-drift and inference telemetry.
- Decide. The signal runs through the authority decision: an unapproved promotion to a production model revokes the agent's authority; a regressed eval escalates for review.
- Evidence. Every decision emits an Evidence Pack™, replay-verifiable to an auditor or regulator from public keys alone.
Why it matters
For a CISO or an MLOps lead, this closes a gap that access controls cannot.
- The EU AI Act and NIST AI RMF both expect you to govern the AI behind a decision, not just the login that reached it.
- Exactly one finding recurs: a model quietly swapped under a live agent, with no authority re-check.
- KYE™ converts W&B provenance into authority; it is not an MLOps tracker and stores no experiments.
Boundary: W&B records how the model was built; KYE™ governs whether an agent running it may act. Part of the KYE Protocol™ platform; pairs with the inference-boundary PEP.