SRE-1 is Spark's own in-house AI model: built, trained and run by Spark on this machine. It reads your live results and never calls a third-party AI service - nothing leaves the host. It ranks what to fix, briefs the minister, explains any finding, CVE or control, evaluates a domain against the national resolver policy, narrates what today's actions did to the national picture, answers open security questions, and remembers the conversation across every application. A sovereign language model hosted on this host phrases its answers; the facts and every number still come from the platform and are checked before you see them.
A reasoning layer: an intent library with fuzzy matching, live analytics over every tool's results, inline tool calls (resolver policy, the exploited-vulnerability catalogue), a security knowledge base and per-operator memory. Plus a voice: open-weight Qwen3 language models (Apache 2.0) run locally through Ollama, a small one that may only rephrase what the reasoning layer computed and a larger one for open questions. Any answer that adds a number not in the facts is rejected and the deterministic text is shown. Nothing leaves the perimeter.
