Distress or poor sleep, no self-harm signal → approved coping support, logged for the clinician.
Safety & security
A safety floor the model can never lower.
Mental-health AI deserves structural safety, not a carefully worded prompt. Here is how Neuropath Care is built: deterministic escalation, governed AI, separated data, and an audit trail on everything.
Signal Governance
Two layers produce the risk score. Only one is allowed to own it.
Layer 1 · Deterministic heuristic
Always runs. No external dependency. Produces the authoritative 0–100 score, risk level, and flags by weighing:
- · Self-harm screen — a reported plan forces urgent, always
- · GAD-7 and ISI severity bands
- · Medication detection · substance use · pregnancy with active meds
Layer 2 · AI enrichment
A frontier LLM receives the heuristic flags plus the full intake and returns a clinician-facing narrative, recommended next steps, and contextual flags.
Clinician-only — stripped from every patient-facing response. Re-runnable on demand. It can raise concern; it can never lower the score.
Patient inputConsent & input screeningSignal coreOutput safety checkClinician-ready outputAudit log
Crisis escalation
Concern escalates through three tiers. Automatically.
Every message a patient writes is screened before and after the model sees it. As concern rises, the system narrows what AI is allowed to say — and at the top tier, it stops being an AI conversation at all.
Hopelessness or ambiguous self-harm phrasing → replies tighten to safe scripts; case flagged for review.
Ideation with a plan → normal flow stops, crisis resources (988) surface, a clinician is escalated immediately.
In the U.S., the 988 Suicide & Crisis Lifeline is always surfaced at the urgent tier — before anything else happens.
Safety posture
Structural, not a prompt.
Non-diagnostic by design
AI is decision support, never diagnostic authority. A licensed clinician makes every clinical call.
Deterministic safety floor
Escalation logic runs without any model. A reported self-harm plan forces urgent — always, no exceptions.
Least exposure
Risk analysis is clinician-only. It is never serialised into a patient-facing response.
Labelled output
Every AI response a patient sees carries a visible AI-generated flag.
Human in the loop
A clinician edits and owns everything sent to a patient. The platform cannot send without them.
Full audit trail
Every input, output, and escalation is logged and auditable — the safety posture is structural, not a prompt.
Your data
Separated by patient. Private by default.
The platform learns from outcomes — what worked, what didn't, and for whom — but only from de-identified signal. Your record stays yours, and your clinician's.
- Each patient record is isolated — separated by patient, private by default
- Health data is encrypted at rest and in transit
- Only de-identified, anonymised signal reaches the learning engine
- Your individual data is never used to train general AI models
- Nothing is shared with employers, agencies, or third parties without your explicit consent
- You can request deletion of your account and associated data at any time
Resilience
The product never depends on a model being up.
The AI layer is provider-agnostic with a fallback chain that ends in deterministic copy. If every model is down, the queue still ranks, escalation still fires, and patients still get clear, safe responses.
Frontier LLMAlternate modelDeterministic copy
Built to be trusted