How confidence scoring works - Kanonik Help Center

Before any change reaches you for approval, Kanonik checks it and gives it a confidence score. The score is not a gate that decides things behind your back. It decides how a change is presented to you: ready for a quick approval, flagged for a closer look, or stopped because it is clearly wrong. You always make the final call, and this article explains how the score is worked out so the number is meaningful to you and so you can set the level of scrutiny you want.

What the score is made of

Every proposed change goes through two checks, and the results are combined into a single confidence figure.

A fit check. Kanonik measures how closely the change matches what it is meant to do. For a control being mapped to a requirement, for example, it compares the wording and meaning of the control against the requirement and produces a similarity score. A high score means they are clearly about the same thing; a low score means the connection is weak on wording alone.
A reviewer check. A separate AI reviewer reads the change in context and judges whether it is actually correct and well formed, and reports how confident it is. This is the step that understands meaning rather than just word overlap, so it can confirm a connection that is genuine even when the wording differs.

Alongside these, Kanonik runs quick structural checks (is there an owner, is the text present, is the framework known) that can raise small warnings.

These signals are weighted and combined into one confidence figure, shown as a percentage. The reviewer's judgment carries the most weight, the fit check supports it, and the structural warnings pull it down when something is off.

How the score decides what happens next

The combined confidence is compared against an approval bar.

At or above the bar: the change is ready for your one click. You still see the score and the reasoning, and you still approve it. Nothing is written until you do.
Below the bar: the change is not rejected, it is flagged for a closer look. This is the right outcome for a change that is plausible but that the automated checks were not sure about. You read it, and if you agree, you approve it. Your judgment is what settles it.
Clearly wrong: if the reviewer determines the change is incorrect, it is rejected before it reaches your queue, so you are not asked to approve something that does not hold up.

A low fit score on its own never blocks a change. A control can correctly address a requirement even when the wording does not line up neatly, and you are the one asserting that it does. A weak word-level match lowers the confidence and sends the change for review, but it does not veto your decision. Only the reviewer judging a change actually wrong will stop it.

Why the fit score is trustworthy

The fit check compares the change against every option in the framework, not just the handful that looked closest at first glance. So when you see a low similarity, it means the wording really is a weak match, not that the option simply did not surface in a short list. That keeps the number you see truthful: a low score is a real signal you can weigh, not an artifact of how the search was run.

Setting the level of scrutiny you want

Different organizations want different amounts of review. Kanonik lets your workspace set how much evidence and scrutiny each approval asks for, from a lighter touch for routine work to a stricter posture where more is surfaced and confirmed before anything is sealed. Raising the rigor means more changes are sent for a closer look rather than presented as ready; lowering it means more routine changes come straight to a quick approval. The confidence scoring itself does not change, what changes is where the bar sits and how much is put in front of you.

Whatever level you choose, the guarantees underneath stay the same. Every change is checked, every change carries a visible confidence score and the reasoning behind it, and nothing is written to your record until you approve it.

In short

The confidence score combines a wording-fit measure with an AI reviewer's judgment of correctness, shown as a percentage. High confidence comes to you as a quick approval, lower confidence comes as a flag for review, and a change the reviewer judges wrong is stopped before it reaches you. A weak fit score lowers confidence but never overrides your decision, and you can set how strict the review should be for your workspace.

More help