The assistant and its limits¶
The rule is simple: the engine decides; the assistant explains and drafts. No language model takes part in deciding which CVEs apply to a device, in which order to fix them, or what a device's status is. Those come from the deterministic engine and stored data.
Some screens offer help from the assistant (Claude, by Anthropic) for writing:
| Feature | Where | What it reads | What it never receives |
|---|---|---|---|
| Ask about your environment | Overview | Your stored checks, through the same read-only views as the screens | Nothing outside your organization |
| Explain this finding | CVE record | That CVE record's evidence, and optionally one of your devices' product and version | The device name |
| Digest summary | Changes | The digest groups: products, counts, CVEs, before and after values | Device names, site names |
| Report narrative | Reports | That report's saved figures | Device names, people, owners, maintenance windows |
| Change ticket draft | Fix first | The vendor's and government's numbered statements about the product and action | Customer data |
| Import suggestions | Import | One row's own cells, for rows the catalog could not identify | Other rows |
The guardrail¶
Every output is checked by deterministic rules before you see it:
- Every line must cite the facts it uses.
- Every number, number word, CVE identifier and version in a line must appear in the facts that line cites.
- Banned wording is refused: safety and compliance claims, speculation about attackers (unless the cited facts themselves say it), and the product's copy rules.
- A refused draft is sent back with the reason, at most twice. If it still fails, nothing is shown, and the screen says so.
For imports, a suggestion is re-run through the catalog resolver and shown with the resolver's verdict; accepting it writes the catalog product's own name, never the assistant's text. A person decides.
Turning it on and what is logged¶
The assistant is off unless your organization turns it on. Every call is logged with the user, the feature, the model, the scope of the input, the output and the token counts. Every output is labelled "written by the assistant (Claude, by Anthropic)" with what it read.
The device API uses no model¶
POST /api/v1/assets/correlate-cves, the Windows checks, fix plans, the resolver and the MCP tools run entirely on the deterministic engine. An assistant connected over MCP is a client of that engine: it calls the tools and writes its own answer from their results, and the tools tell it to repeat the honest assessment of each device.