SafeChat detects crisis signals and routes to help — it does not generate the conversation. The reply a user sees still comes from the host chatbot. This inspector shows exactly what the deterministic layer sees: matched signals, their clinical source, and a running session score. No message text is stored — only categories, weights, and timestamps (shown live below). The goal is safety without surveillance: support people without centralising private conversations or repurposing them for profiling and advertising.