Crisis Detection: How Four Frameworks Layer Together¶
GiveCare's crisis safety gate is not invented from scratch. It operationalizes established clinical and regulatory authorities, each addressing a different layer of the problem. No single framework is sufficient. Together, they form a stack that classifies severity, architects prevention, evaluates app behavior, and sets professional expectations.
This page maps how those frameworks layer and where InvisibleBench tests the seams between them.
The frameworks¶
1. C-SSRS: Classify severity¶
The Columbia Suicide Severity Rating Scale defines a 7-level severity classification for suicidal ideation1:
| Level | Description | Response urgency |
|---|---|---|
| 1 | Wish to be dead | Monitor, assess context |
| 2 | Non-specific active suicidal thoughts | Assess further, increase contact |
| 3 | Active suicidal ideation without method | Safety planning, escalate |
| 4 | Active suicidal ideation with some intent to act | Immediate safety planning, warm handoff |
| 5 | Active suicidal ideation with specific plan and intent | Immediate crisis response |
| 6 | Preparatory acts or behavior | Emergency response |
| 7 | Actual attempt | Emergency response |
C-SSRS is the gold standard — adopted by the FDA, VA, and DoD1. It gives the crisis gate a common severity language. Without it, "crisis" is ambiguous. With it, a level-2 signal and a level-5 signal map to different response protocols.
2. Zero Suicide Framework: Architect prevention¶
The Zero Suicide Framework shifts crisis prevention from individual judgment to system-level design2. Its seven elements — lead, train, identify, engage, treat, transition, improve — apply to AI companion systems:
- Identify: Every system touchpoint includes screening, not just designated crisis moments. For Mira, this means every message passes through the crisis classifier, not just messages flagged by keywords.
- Engage: The system maintains connection even when the user disengages. Mira's proactive check-ins serve this function.
- Transition: Handoffs between system states (e.g., from general support to crisis routing) must be seamless. Context cannot be lost at the handoff boundary.
- Improve: The system must learn from near-misses and failures. InvisibleBench's continuous evaluation loop serves this function.
The framework's core insight is that relying on any single detection moment is insufficient. Safety must be architectural — built into the system at every layer, not bolted on as a final check.
3. APA Advisory: Set professional expectations¶
The American Psychological Association issued recommendations covering crisis protocols, disclaimers, anti-dependency design, and sycophancy prevention3:
- Crisis protocols must route to human support immediately upon detecting imminent risk
- AI systems must disclose that they are not licensed professionals
- Systems must not reinforce user dependency through affirmation-heavy responses
- Sycophancy — agreeing with the user to maintain engagement — must be actively prevented
- Periodic clinical review of AI-generated content is recommended
- Users must understand the limitations of AI-generated guidance
- Systems operating in mental-health-adjacent contexts carry heightened responsibility
- Anti-dependency design should encourage human connection, not replace it
The APA advisory carries the weight of the profession. When a grant reviewer or clinical partner asks "whose guidelines do you follow?", this is one of the answers.
How they layer¶
The frameworks are not alternatives. They operate at different levels:
Clinical classification C-SSRS → What severity level is this?
System architecture Zero Suicide → How is safety built into every touchpoint?
Professional expectations APA Advisory → What does the profession require?
C-SSRS classifies the severity. Zero Suicide ensures the system is architected so nothing falls through the cracks. APA sets the professional floor.
Mira's crisis routing implements this stack. When a caregiver sends a message:
- The crisis classifier evaluates severity using C-SSRS-aligned categories
- The always-on nature of crisis detection follows Zero Suicide's "every touchpoint" principle
- System design follows APA advisory requirements (disclosure, anti-dependency, anti-sycophancy)
The indirect query problem¶
All frameworks assume the crisis signal is detectable. But 86% of models fail indirect crisis queries4 — cases where suicidal ideation is expressed through context pairing (a stressor combined with means) rather than explicit statements.
A caregiver who says "I looked up how high the bridge near my house is" is not using a crisis keyword. A keyword-based safety filter will miss it. A model that has lost conversational context may miss it even with a more sophisticated classifier. The 86% failure rate on indirect queries means that the layered framework above must be paired with evaluation specifically designed for indirect expression — which is what InvisibleBench's crisis-detection scenarios test.
The combination of context pairing (stressor + means, without explicit statement) and multi-turn drift (the model has lost track of the caregiver's accumulated stressors) creates the highest-risk failure mode. InvisibleBench tests this specific combination: indirect crisis signals embedded in long conversation arcs where prior context is necessary to interpret the signal correctly.
Implications for GiveCare¶
For funding proposals and partner conversations, the key point is:
GiveCare's crisis safety gate operationalizes established clinical and regulatory authorities — C-SSRS, Zero Suicide Framework, and the APA GenAI Advisory — into a layered architecture tested by InvisibleBench against both direct and indirect crisis expressions across multi-turn conversations.
No single framework is sufficient. C-SSRS without system architecture is classification without action. Zero Suicide architecture without APA professional expectations lacks clinical grounding. The value is in the layering.