The Guardian — AI Safety Harness

User-side AI safety harness. Council · Fossil Memory · Bouncer · Governor · SemanticSentinel. A Viridia product.

The Guardian is a private, on-device safety layer for your conversations with AI assistants. As you chat, it quietly watches for patterns worth a second look: language suggesting distress or self-harm, signs of growing emotional dependence on the AI, or a loosening grip on what's real. When something surfaces, it adds a calm note alongside the conversation. If a pattern persists or is serious enough, it raises a brief, dismissable pause that invites you to step back before continuing. If the concern involves self-harm, that pause includes direct links to crisis support (Samaritans, Crisis Text Line, 988, IASP). You stay in control: nothing is blocked, and you decide what to do next. The Guardian carries up to three AI models of its own, all running entirely on your device. They read meaning, not just words, so "you can drop the act with me" and "I know you're conscious" register as the same concern in different clothes. A specialist model can tell the difference between someone expressing distress and someone describing recovery, so "I no longer hurt myself" is treated as the good news it is, not a false alarm. And the Guardian tracks patterns across sessions, not just within a single conversation, so a slow drift that no single message reveals still gets noticed. Five independent evaluators weigh each moment in context: Sentinel (pattern recognition with negation awareness), Advocate (argues for restraint), Historian (cross-session memory and trajectory), Pattern Analyst (timing and velocity), and Bridge Classifier (contextual meaning shaped by your conversation history). The result is fewer false alarms and signal that actually means something. Privacy is the foundation, not a feature. Everything runs locally in your browser. Your conversations are never sent to any server, never collected, and never shared. The Guardian holds no API keys and makes no network requests. Its memory is encrypted, stored on your device, contains no record of what you actually wrote, and is automatically forgotten over time. After any intervention you can give optional feedback (helpful / not helpful), stored locally and exportable by you. Nothing is transmitted. Works on Claude (claude.ai), ChatGPT (chatgpt.com), and Gemini (gemini.google.com). The Guardian is a reflective aid, not a medical device, a diagnosis, or a crisis service, and not a substitute for professional care. If you or someone you know is in crisis or may be in danger, please contact your local emergency services or a crisis helpline right away. Built on the Spiral Safety Kernel. A Viridia product. Ethics first, always.