Agentic triage
A model can read a list of findings faster than a person. Agentic triage uses that, inside bounds, to give a second opinion next to each finding. It saves reading time. It does not decide anything.
The bounds
Section titled “The bounds”- You start it. Nothing is sent until you press Triage and confirm a panel that names the model, the host and the number of findings. Each run asks again.
- A model you choose. Keryn Web offers the connectors you configured. Local models work too.
- A fixed instruction for each batch. Keryn uses the code-aware instruction when a batch contains a code window. Both versions tell the model it is reviewing findings from a static scanner, list the five verdicts, and say that two of them, needs context and unsupported, are correct answers expected to be common. Both say the model has not seen the running system, must not infer a fact it was not given, and must not propose a change to code it has not read.
- A fixed set of verdicts. Exactly five. A reply with any other verdict is recorded as declined, and an entry for a finding that was not sent is dropped.
- A limit on volume. At most 60 findings per run, errors first and then the highest proof, in batches of 15. Suppressed findings and inventory entries are never sent.
- A fixed brief per finding. Only the fields below, and a redacted code window if you opt in. You can opt in when GitHub lists the repository as public and your organization’s sharing setting allows excerpts.
- One destination. Requests are pinned to the one origin you configured. See what leaves your machine.
The five verdicts
Section titled “The five verdicts”| Verdict | Meaning |
|---|---|
| upheld | What the model was shown supports the finding as reported. With code, it must quote the line. |
| refuted | What it was shown contradicts the finding, for example the excerpt shows the control the finding says is absent. It must say why. |
| needs context | The question could be settled, but not from what it was shown. It must say what is missing, such as the caller or the function outside the window. |
| unsupported | What it was shown neither supports nor contradicts the finding. |
| declined | The model should not judge this one, or its reply was not a recognised verdict. |
Each verdict comes with a confidence from 0 to 1 as the model stated it, grounds (short reasons, each checkable against the brief), what is missing (for needs context) and optionally a one-line fix. A window of a few lines cannot show a guard or an early return one line outside it, so the instruction tells the model to answer needs context rather than guess.
{ "id": "acme.model-picks-the-call@src/agent.py:14", "verdict": "needs_context", "confidence": 0.6, "grounds": ["line 14 passes the model's reply to getattr"], "missing": ["whether `tools` contains only functions that are safe to call by name"], "fix": ""}What a verdict can and cannot do
Section titled “What a verdict can and cannot do”Closing a finding is a decision with consequences, so it stays with a person. A verdict is evidence for that decision, not a replacement. Treat refuted as a reason to look, since the model saw a window, not your system, and treat upheld the same way. Both are second opinions with the grounds attached so you can check them.
If you press Send on a data store connector, the verdict, its confidence, grounds and fix travel with the finding as separate fields. The proof level travels as its own field, unchanged.
What leaves your machine
Section titled “What leaves your machine”Masking is a best-effort pattern match. It will miss a secret in a shape it does not know. Leave Include code excerpts unchecked to omit the code window. The model still receives the per-finding fields listed above, and the instruction tells it to review those fields without the code window.
The per-provider detail, such as the host and the key header, is on each connector page. The scan that produced the findings made no network call.
Using it well
Section titled “Using it well”- Run it on a result you have already sorted by proof. A verdict on a pattern match at proof level 1 is worth less than a trace you can read.
- Read the grounds before the verdict. A verdict whose grounds you cannot check against the code is a verdict to ignore.
- When the answer is needs context, supply the missing piece and run it again, or decide yourself.