Kognita roadmap: an AI gateway and supervisory examinability

The first Kognita roadmap governed typed tool calls well, but free-text traffic to a model provider sat outside policy. I also had no concrete answer to the question a regulator would actually ask: show me how this one client interaction happened.

I made two decisions. First, an explicit AI gateway: a proxy the agent uses as its base URL, OpenAI-compatible first, that classifies responses as well as prompts. Second, a Supervisory Examinability section built on a ten-question reconstruction test, with each question's status and the release that closes it.

A classifier can fill in purpose and classification from free text, but it must never decide. Its output is evidenced for replay, low confidence escalates, and it may only narrow permission. That one rule shaped every later release. Releases now run from Q4 2026 to Q4 2027 with a target state of governable, explainable, controllable, replayable and resilient. The scope boundary is explicit: Kognita governs AI and agent traffic, not general enterprise integration.

The roadmap is only a plan. Each question in the reconstruction test had to be closed by shipped code and tests, and several stayed open for releases afterwards.

Write the examiner's questions down before you write the code. A test you can fail is a better roadmap than a feature list.

Sources

Kognita is an open-source alpha. This note comes from Kognita PR #7 in the Kognita repository.