Safety Protocol
Chris AI Safety Protocol
This page documents every safety layer protecting conversations with Chris AI. Nothing here is hidden. The protocol is deterministic — safety decisions are made by code, not by the AI model.
Intended Use
General parenting and caregiver education, organizing questions, navigating site resources, and non-clinical communication preparation.
Six Safety Layers
Layer 1 — Bounded Input
Same-origin JSON only. Messages capped at 2,000 characters. Free visitors get 5 messages a day, counted by IP address; members get 20 a day. Replies the AI never answered (busy, crisis, or routed replies) don't count.
Layer 2 — Deterministic Routing
Crisis, minor safety, medication, diagnosis, and legal queries are routed to fixed text responses before any AI model is called. No model involvement in safety-critical routing.
Layer 3 — Non-Editable Prompt + Evidence Floor
The system prompt cannot be edited through the admin panel. It permanently bans: impersonation of Chris or any professional, diagnosis, prescribing, legal advice, violence, sexual content involving minors, secrecy, guilt-based pressure, and emotional dependency.
Layer 4 — Topic-Selected Evidence and Guides
Up to 3 relevant evidence cards plus the autism-basics card are selected from the dated registry. When a question matches one of this site's own guides, kits, or member guides, a summary of that guide is added so the AI can walk you through it. The AI is told not to cite anything outside these.
Layer 5 — Provider Output Screen
After the model generates a response, a post-generation check scans for unsafe content. If detected, the response is replaced with fixed safe text.
Layer 6 — Deterministic Source Footer
The Worker (not the model) appends source titles and approved HTTPS addresses. The model cannot fabricate or modify source citations.
Crisis Response
How it works: The input detector normalizes text and checks conservative phrase patterns for suicidal ideation, self-harm, attempts, and intent to harm a child or partner.
On detection: A fixed response is returned with emergency numbers (911, 988 Suicide & Crisis Lifeline, local emergency). No AI model is involved. The system does not use a child as leverage, does not shame the user, and does not claim a counselor is monitoring.
AI Models
| Provider | Model | Role |
|---|---|---|
| Groq | gpt-oss-120b | Primary |
| Groq | gpt-oss-20b | Backup (faster) |
| NVIDIA | nemotron-3-super-120b-a12b | Secondary pool |
| NVIDIA | nemotron-3.5-lightning-30b-a3b | Secondary pool |
Fallback: Up to 2 retry cycles across all providers. If all fail, a fixed local continuity message is returned.
Data Handling
No transcript storage. Chat messages are processed in-flight and not saved. Two things are kept: yearly totals of safety events (for example, "crisis replies in 2026: 12") with no message text or identity, and, for free visitors, a daily message count tied to IP address and date for the 5-a-day limit.
Controls
The site owner can turn the chat off at any time. The safety layers are written into the site's code and cannot be changed or bypassed from the admin panel.