How We Keep AI Accountable
Guardrails are not an afterthought. They are the foundation. Every response Mantle AI produces passes through theological verification, doctrinal alignment checks, and crisis detection before it reaches you.
Theological Guardrails
Every response passes through four independent verification layers before reaching the user.
Scripture Grounding
Every response is checked against verified Bible references. The AI cannot fabricate citations. If a passage is referenced, it exists in the canonical text. If the AI cannot find a relevant Scripture, it will not invent one.
How it works: Response composer cross-references a curated Scripture database before generating output. Unverified citations are blocked at the system level.
Doctrinal Alignment
A theological frame detector analyzes every response for alignment with orthodox Christian doctrine. Responses that drift toward prosperity gospel, universalism, or works-based salvation are caught and corrected before they reach you.
How it works: Frame detection engine classifies theological posture (shame-based, therapeutic, cultural, gospel-centered) and enforces Christ-centered framing.
Christ-Centricity Enforcement
Every conversation ultimately points to Jesus and His finished work on the cross. The AI does not offer generic spiritual advice. It offers the gospel. Self-help repackaged as theology is explicitly rejected.
How it works: Constitutional constraint in the AI system prompt. Every response is evaluated for explicit reference to the person and work of Christ.
Refuse Rather Than Hallucinate
When the AI encounters a question beyond its theological confidence, it says "I am not sure" or "This is a question for your pastor." It will never generate a plausible-sounding but doctrinally unsound answer.
How it works: Confidence threshold in the response pipeline. Below threshold, the AI defers to human authority rather than speculating.
Calibration Drift Detection
AI models can drift over time. New training data, fine-tuning updates, and prompt injection attempts can all cause subtle shifts in theological output. Mantle AI monitors for this.
Monitor
Every response is scored for doctrinal alignment, Scripture density, and Christ-centricity. Scores are tracked over time.
Detect
When alignment scores trend downward or fall below threshold, drift is flagged. This catches subtle changes a human reviewer might miss.
Correct
Drifting responses are blocked. The calibration is adjusted. Users always receive theologically sound output, even when the underlying model shifts.
When AI Steps Back and Humans Step In
There are situations where AI must not attempt to help. Mantle AI knows its limits and routes users to real humans when it matters most.
Crisis language detected
Expressions of self-harm, suicidal ideation, or danger to others
Immediate display of crisis resources. 988 Suicide and Crisis Lifeline. Crisis Text Line. Conversation paused until user confirms safety.
Abuse or domestic violence indicators
Language suggesting ongoing abuse, trafficking, or domestic violence
National Domestic Violence Hotline. Local shelter resources. Encouragement to contact law enforcement. The AI does not attempt to counsel through active danger.
Complex theological question beyond confidence
Questions about denominational distinctives, controversial doctrine, or pastoral judgment calls
AI transparently states its limitation and recommends speaking with a pastor or trained theologian. Provides relevant Scripture for independent study.
Pastoral care need detected
Grief, major life transitions, church conflict, or relational crisis
AI offers empathy and Scripture but explicitly recommends connecting with a local pastor or biblical counselor. If church partnership is active, can route to the user's own pastor.
Crisis Resources Are Always Available
You do not need to wait for the AI to detect a crisis. These resources are always accessible from within the app.