Ideas · Builder Notes
A governed agent demo you can run in seconds
The Kognita alpha flagship demo scaffolds a governed agent and runs four scenarios: an allowed query, a denied query, evidence tampering and gateway redaction.
Why did this need fixing?
Governance is hard to explain in prose. I wanted a demo that shows allow, deny, tamper detection and redaction on a laptop with no provider key.
What did I do about it?
kognita scaffold --template governed-agent creates a small app with a SQLite policy and evidence store. run_demo.py runs the four scenarios. A local HTTP stand-in plays the upstream model.
A denied query returns citations and retrieves nothing. Editing one evidence row makes kognita evidence verify report the break.
Which detail mattered?
The stand-in is not trusted as a local destination, so confidential content is still redacted before it is forwarded. What you see in the demo is what the library actually does.
Where are the limits?
It is a demo pack with synthetic subjects. Nobody should run it as a production configuration.
If you're building something similar?
Ship the demo with the library, and make it use the same code path as production. A demo that cheats teaches the wrong thing.
Sources
Kognita is an open-source alpha. This note comes from Kognita PR #14 in the Kognita repository.