{% extends "base.html" %} {% block title %}AI Config · AICAudit{% endblock %} {% block page %}config{% endblock %} {% block content %}

AI configuration

Verdicts are evidence-grounded and fail-safe: unparseable or low-confidence answers become unverified, never auto-confirmed.

relay / 中转 openai claude openrouter ollama (local) mock (off)
How verdicts work
The LLM sees the full enclosing function, the file's imports, and the engine-traced taint path.
Injections get dataflow instructions; policy findings (secrets / weak crypto / random) get usage-context instructions.
confirmed false_positive unverified — findings are marked, never silently deleted.
{% endblock %}