Kanonik checks each draft on its own server before anyone is asked to approve it.
We call it the Verifier, and in the app it’s the safety check. Your AI can’t skip it.
If a draft falls short, the reasons go back to the AI so it can try again.
A draft gets up to three rounds before it reaches you.
- Kanonik
Fixed rules come first.
Are the required fields filled in? Does the draft point at records that actually exist? A second copy of a vendor you already have stops here as well. No model is involved.
- Kanonik’s model
Then a model reads it, the way a careful reviewer would.
It asks the questions you’d want answered about that kind of record. Does this control really address the requirement it claims to meet? Is this risk a real threat with a believable impact? It writes down its judgment, how confident it is and why.
- Kanonik
If it falls short, the AI gets it back.
The reasons go with it. The AI revises the draft and sends it again, up to three rounds in all, and your To decide list stays quiet while that happens. If a draft still fails after the third round, the AI stops and tells you which one and why.
- You
The draft reaches you with its history.
Open it, and a fold called How this was prepared tells you whether it needed more than one round, and what the check objected to each time. Passing the check doesn’t put anything on the record, because that’s still your decision to make.
The Verifier is one of five stops between your AI and your Record
Three rounds with the Verifier
Brightdesk Support is a vendor draft from Dana’s AI. Watch it go to Kanonik’s server, and watch what happens when the server finds something wrong.
Put a flaw into a fresh draft
Or let it fail
Your AI
Round 1 of 3
Drafting
Vendor
Brightdesk Support
Owner none
Awaiting approval
Kanonik’s server
Fixed rules
No model. Required fields, records that exist, a second copy of a vendor.
Waiting
Model review
A model reads it the way a careful reviewer would.
Waiting
Dana’s To decide 0
Nothing waiting for you
What each round found
Waiting for the draft to go
A sample, with sample data from Harbor Ledger, a made-up company. The reasons are the ones Kanonik’s checks give, shortened where you see [...]. The model’s reasons are sentences from the Verifier’s own test cases. Nothing here is saved.
Some decisions get no AI opinion at all.
Take a decision to live with a risk, which your record calls a risk acceptance. Two good security leads can make opposite calls on the same risk and both be right, because the answer turns on cost, timing and appetite that live nowhere in your compliance data. So there’s nothing for a model to compare your call against, however good the model is.
Exceptions and remediation decisions are the same kind of call. They get the fixed rules and a person, and the record says plainly that no model judged them.
A score on a call like that would look like a check without being one, and we won’t put one on your record.