Skip to main content
This platform detects two kinds of drift and used to miss the one that matters most: An agent whose model version and tool manifest are both unchanged, but whose decisions have shifted, was invisible. That is exactly what a prompt edit or a new retrieval corpus produces without touching anything under change control.

​
Built from what is already there

No new instrumentation and nothing extra to send. The fingerprint is computed from decisions the ledger already holds: the mix of outcomes, which gates the agent tends to stop at, the shape of what it attempts. That matters for a reason beyond convenience. A behavioural signal an agent reports about itself is a signal an agent can be wrong about. This one is derived from what the platform observed, so a changed agent cannot describe itself as unchanged.

​
What a shift looks like

Not “this decision was wrong”. The platform does not have an opinion about correctness, and any product claiming to detect that from outside is selling a feeling. What it can say is that the distribution moved: an agent that refused at policy twice a week now does so forty times a day, or one that never reached review now reaches it constantly. Both are worth a human asking what changed, and neither requires knowing what the right answer was.

​
What to do with it

A behavioural shift is a prompt to investigate, not a refusal. The things that actually stop an action are the nine gates, and they are unaffected by this. The useful pairing is with certification: a certificate pins the provenance and the manifest an agent was cleared against, behaviour catches the change those two cannot see, and together they cover the ways an agent stops being the thing that was approved.