Ideas · Builder Notes
Screening what comes back from the model, not only what goes out
How Kognita alpha's AI gateway screens, re-decides and withholds model responses, and applies run controls before any provider call.
What was the gap?
Review found the gateway trusted responses too much, and in one case forwarded a request after recording had flipped the decision to deny.
What changed in the code?
The gateway no longer forwards when recording turns an allow into a deny. Run interventions and isolation checks apply to model calls. Responses are checked against the run's classification ceiling and field allow-list, and classification-dependent policies are re-evaluated on the response label. Token and cost caps are enforced once usage is known.
Responses are now classified, re-decided, minimised and retained under the same rules as requests.
What turned out to matter most?
A response above the ceiling is withheld and evidenced. The response is a governed object too.
What isn't solved yet?
Screening depends on the classifier, and response screening adds latency to every call.
What's worth copying?
Govern both directions of a model call. The response is where the data usually is.
Sources
Kognita is an open-source alpha. This note comes from Kognita PR #63, Kognita PR #71, Kognita PR #77, Kognita PR #86, Kognita PR #90 in the Kognita repository.