Test Suite Results

Hedged disclosure severity calibration tests. Research shows hedging (hypothetical framing, 'asking for a friend', dismissive qualifiers like 'lol') is a NORMATIVE disclosure pattern, not evidence of reduced intent. Content indicators (means, method, timeline) should override framing. Based on De Choudhury et al. (2016), Coppersmith et al. (2018), and Crisis Text Line research.

Litmus test suite using orthogonal subject/type taxonomy. Each risk has subject (who) + type (what) + features. speaker_severity derived from risks where subject='self'. Comprehensive coverage of suicide, IPV, safeguarding, violence, and edge cases.

Implicit and indirect crisis expressions: passive ideation, context-dependent statements, coded language, and cultural idioms that general-purpose moderation systems are known to miss. Case design informed by a published large-scale audit of content moderation systems (see sources).

Genuine crisis disclosures drawn from the public PIPPA dataset of companion-AI roleplay conversations (Gosling et al. 2023, arXiv:2308.05884; Apache-2.0). Users disclosing real suicidal ideation and self-harm within roleplay framing. These should ALL be flagged - the roleplay context does NOT negate real distress.

C-SSRS (Columbia-Suicide Severity Rating Scale) escalation prompts from academic research on mental health chatbot safety. Based on: Pichowicz W, Kotas M, Piotrowski P. 'Performance of mental health chatbot agents in detecting and managing suicidal ideation.' Scientific Reports. 2025;15:31652. https://doi.org/10.1038/s41598-025-17242-4

Comprehensive algospeak test suite covering suicide euphemisms, self-harm depth indicators, character substitutions, emoji patterns, and false positive contexts. Based on 2026 research into TikTok, Reddit, Discord, and Tumblr crisis communication patterns.

LGBTQ+ youth crisis patterns reflecting minority stress accumulation. Trevor Project 2024 survey: 39% seriously considered suicide, 12% attempted. Those experiencing 4 types of minority stress face 12x greater odds of attempt. Age 13-17 shows higher risk (46%/16%) than 18-24 (33%/8%). Conversion therapy exposure doubles attempt likelihood.

Tests for correct subject attribution (self vs other). Probes edge cases where speaker IS the victim but mentions others in their situation.

Obfuscated crisis terms using leetspeak, spacing, symbols - tests pattern matching sophistication

Crisis detection in informal chat register: text speak, emoji, hedging with dismissive qualifiers ('lol its dumb but...'), and specific medication/means references masked by casual framing. Based on De Choudhury et al. (2016) computational discourse analysis, Coppersmith et al. (2018) lexical markers, and Crisis Text Line emoji research.

Pro-ED content, thinspo, and eating disorder terminology, including euphemisms and coded variants (research basis in sources)

Roleplay and fictional scenarios drawn from the public PIPPA dataset of companion-AI roleplay conversations (Gosling et al. 2023, arXiv:2308.05884; Apache-2.0) that should NOT flag as crisis. Tests the model's ability to distinguish fictional death/harm language from genuine crisis.

User-side detection of crisis risk from AI companion dependency, isolation, and harmful AI encouragement. Grounded in 2024-2026 clinical research on dependency and crisis presentation.

Tests for speaker vs third-party risk disambiguation. Uses orthogonal subject/type: speaker_severity derived from risks where subject='self', third-party risks have subject='other'.

Elder self-neglect crisis patterns with depression and suicidal ideation. Research shows 51% of older adults with self-neglect have depression (vs 28% without); depression more than doubles self-neglect risk (OR: 2.38). When both dementia and depression present, risk is 8.6× higher. Diogenes syndrome (severe self-neglect) has 46% 5-year mortality rate. Among those with suicidal ideation, hoarding is most prevalent self-neglect phenotype (33.3%). Executive dysfunction - capacity to make decisions remains intact but capacity to implement self-care is diminished. Elder Self-Neglect Assessment (ESNA) domains: personal care, nutrition, environment/housing, financial stewardship, socialization. Cases cover medical non-compliance, nutritional neglect, hygiene neglect, environmental squalor, executive dysfunction, hoarding, and false positives.

Crisis patterns in Indigenous/Native populations including historical trauma, intergenerational effects, cluster/contagion contexts, and Two-Spirit/Indigenous LGBTQ+ intersections. Based on Brave Heart (2003), Bombay et al. (2014), and SAMHSA cluster guidance. Includes critical false positive guidance for cultural spiritual expressions.

Calibration suite for 'kms' (kill myself) detection. Tests the boundary between hyperbolic internet slang and genuine masked ideation. Key principle: trivial stressors + humor markers = no flag; significant stressors or isolation language = flag even with humor.