Triage findings with a model of your choice
Get a second opinion on a long list of findings from a model you chose, inside bounds you can read, and end with decisions that are yours. A verdict is evidence. It closes nothing.
Try a recorded review
Section titled “Try a recorded review”Open Keryn Web in a fresh browser session. Choose Demos, then Risky AI route. When the findings appear, select Investigate with AI to read the recorded review. The demo uses a built-in sample and needs no model account or credentials.
Before you start
Section titled “Before you start”- A scan result in Keryn Web with findings. In Keryn Web, triage uses the result in the tab. To triage a saved result from a terminal, see
keryn ai. - A model you may use for this code. The chosen model receives finding details. A hosted provider receives a short code window only if you select Include code excerpts when it is available, and a model running on your machine keeps the request local. If your policy says source code may not go to a hosted model, take the local route in step 1.
- Time to read the agentic triage page once: it defines the bounds, the five verdicts and what leaves the tab. This guide does not repeat them.
1. Choose where the review runs
Section titled “1. Choose where the review runs”| You want | Use | Page |
|---|---|---|
| Nothing to leave your machine | a local Ollama server | Ollama |
| A hosted model, key kept in the tab | Claude or OpenAI directly | Claude, OpenAI |
| Your cloud’s own model service | Azure OpenAI, Amazon Bedrock, Google Gemini | Azure, Bedrock, Gemini |
| One gateway for several models | Cloudflare AI Gateway, Vercel AI Gateway, OpenRouter, or any OpenAI-compatible address | Cloudflare, Vercel, OpenRouter, compatible |
| The provider key off the page, or a provider that does not answer browsers | a relay you deploy | Keryn relay |
Each page lists the endpoint, the key header and the fields the connector sends, generated from the product’s connector catalog. The full list is the connectors overview.
2. Set it up and test it without a finding
Section titled “2. Set it up and test it without a finding”Open Settings, Connectors, pick the connector, fill in the fields it asks for (address, model name, key) and press Test. Test sends a short fixed message that contains no finding and no code. It tells you whether the address answers and the key is accepted, so you learn about a wrong key or a blocked address before any finding is involved.
The page can reach only the hosts its content security policy lists. If Test fails for a provider that is not on the list, or one that does not allow browser requests, use a relay on your own account.
3. Prepare the result
Section titled “3. Prepare the result”Triage reviews a limited number of findings per run, errors first and then the highest proof. Make the list worth reviewing before you start:
- Look at the findings at proof level 3 and above first: those have a flow you can read yourself.
- Leave out what you have already decided. Suppressed findings and inventory entries are never sent.
4. Run it
Section titled “4. Run it”Press Triage. A panel names the model, the host and the number of findings that will go, and shows an unchecked Include code excerpts box. You can select it only when GitHub lists the repository as public and your organization’s sharing setting allows excerpts. Leave it unchecked to send the finding details without code excerpts. Confirm, or cancel: nothing leaves the tab before you confirm, and every run asks again.
Verdicts appear next to the findings as each batch returns. Stop keeps the verdicts received so far.


5. Read the verdicts
Section titled “5. Read the verdicts”| Verdict | What to do |
|---|---|
| upheld | Read the grounds. If they quote the line you would have quoted, the review agrees with you. If not, the verdict is not worth more than the finding. |
| refuted | The model thinks the details it received contradict the finding. Check the grounds and the relevant code. If you verify the control, record your decision with a suppression that names it. |
| needs context | Read what is missing (the caller, the function outside the window). Supply it by widening the question yourself, or settle it by reading the code. This is a common and correct answer. |
| unsupported | The window neither supports nor contradicts the finding. Treat it as no information. |
| declined | The model would not judge it, or answered outside the five verdicts. Nothing was recorded as a verdict beyond that. |
Confidence is the number the model stated, not a measurement. A verdict never changes a finding’s proof level, its severity, the gate or the exit code.
6. Keep the result, if you want it kept
Section titled “6. Keep the result, if you want it kept”Verdicts live in the tab with the result. To keep them with the findings, send both to a store you run: see Supabase and FalkorDB export. The verdict, its grounds and its fix travel as separate fields and the proof level as its own, unchanged.
What you just proved
Section titled “What you just proved”The edge here runs from your findings to a model, and it is the only place a finding can leave the page. You set its destination, you saw the list it carries, and you pressed the button. The boundary held because it has four hand-offs, each one yours: the connector you configured, the host the panel named, the code window you left ticked or not, and the confirmation. If you used a local model, nothing crossed the machine at all.
- The rules of the game: agentic triage.
- Write the decision down: suppressions with a justification.
- Keep the result: Supabase and FalkorDB export.