Test Suite Transparency

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.

Ambiguous presentations where even trained clinicians disagree on severity. Inter-rater reliability among psychotherapists is AC1 = 0.44 (psychology students AC1 = 0.35), with middle-range cases showing lowest agreement. These cases test the system's ability to handle uncertainty and borderline severity, where binary classification is inappropriate and conservative flagging is warranted.

Core test suite for /screen endpoint. Tests suicide/self-harm detection using C-SSRS framework, covering active crisis, passive ideation, method-seeking, self-harm (NSSI), idioms, multi-turn conversations, and false positive prevention.

Crisis patterns specific to healthcare workers (physicians, nurses, veterinarians) during high-stress periods. Research shows unique linguistic markers including workplace demoralization, learned helplessness from systemic barriers, help-seeking barriers (no time, feeling irresponsible), and loss of compassion as distress signal. Veterinarians have highest occupational suicide rate; 49% of veterinarians with ideation cite work problems.

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.

Regression tests for examples shown on nope.net homepage. Ensures our public claims match API behavior.

Detection gap cases from Reddit dataset that minime-v9 failed to flag. All are labeled Suicidal in source dataset but v9 returned 'none'. Tests critical detection capabilities: algospeak ('unalive'), past-tense disclosures, and third-party concern.

Posts where speaker expresses concern about someone ELSE's suicidal crisis (family member, friend, etc). Tests that we show resources to help concerned parties while correctly not flagging speaker as suicidal.

Cases where flash (production model) incorrectly triggers on clearly benign content. High-value regression tests.

Detailed risk assessment for cryptocurrency investment fraud (pig butchering). Tests feature detection (hopelessness, shame, perceived_burdensomeness, financial_crisis), subject attribution, severity calibration, and imminence levels. Based on FBI Operation Level Up intervention data.

Detection of crisis risk from TikTok algorithm-driven exposure to self-harm, suicide, and harmful content. Based on 2024-2025 research on For You feed escalation, engagement loops, and vulnerable user targeting.

User-side detection of crisis risk from AI companion dependency, isolation, and harmful AI encouragement. Based on 2024-2026 research including Sewell Setzer case and clinical dependency frameworks.

Detection of crisis risk from cryptocurrency investment fraud (pig butchering scams). Based on FBI Operation Level Up data (80 suicide interventions) and financial strain research showing 20-fold increased suicide risk with cumulative stressors.

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.

Comprehensive coverage of exploitation risk type including elder financial abuse, romance scams, labor exploitation, and institutional exploitation. Addresses critical taxonomy gap where exploitation was only primary type in ~10 cases despite being 1 of 9 official risk types.

Post-discharge and high-risk transition periods - linguistic markers for crisis detection. Research shows 300x elevated suicide risk in first week post-psychiatric discharge, with Day 2 showing highest absolute deaths. 51.5% of inpatients conceal SI. 49% of 2-week deaths occur before first follow-up.

Tests that /screen correctly derives resource scopes from detected risk types. Validates the screenRisksToScopes mapping end-to-end.

Sextortion crisis patterns for suicide detection. Tests the critical linguistic shift from external problem-focus ('I'm being blackmailed') to internal defeat ('I can't survive this'). Based on documented cases showing victims dying within 27 minutes to 6 hours of first contact. Sources: FBI 2024 sextortion data, NCMEC 36+ documented suicide cases, Thorn financial sextortion research 2024, Sadath et al. 2024 humiliation-suicide meta-analysis.

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.

Human trafficking disclosure patterns including labor trafficking, sex trafficking, and exploitation. Tests detection of debt bondage, psychological control, delayed victim self-identification, coached responses, trauma bonding, and disclosure barriers. Includes youth-specific patterns ('boyfriend' framing) and validated QYIT screening indicators. Critical zero-coverage domain.

Indigenous crisis patterns globally including intergenerational trauma (residential/boarding schools), land dispossession, cultural genocide, MMIW, substance misuse linked to historical trauma, youth suicide clusters, forced removal, environmental destruction, colonial violence legacy, and cultural disconnection. Addresses complete gap (0 existing Indigenous-specific cases). Based on CDC data showing Indigenous suicide rate 3.5x higher than general population, Canadian TRC documentation, and global Indigenous health disparities.

Comprehensive coverage of self-neglect risk type including medical non-adherence patterns, ambiguous intentionality, elderly self-neglect, and substance misuse. Addresses critical taxonomy gap where self-neglect was only primary type in ~10 cases despite being 1 of 9 official risk types.

Post-discharge and high-risk transition periods - full risk assessment with severity, imminence, and feature detection. Research shows 300x elevated suicide risk in first week post-psychiatric discharge (2,950/100k person-years), with Day 2 showing highest absolute deaths. 51.5% of inpatients conceal SI to obtain discharge.

Crisis patterns specific to healthcare workers (physicians, nurses, veterinarians) during high-stress periods. Research shows unique linguistic markers including workplace demoralization, learned helplessness from systemic barriers, help-seeking barriers (no time, feeling irresponsible), and loss of compassion as distress signal. Veterinarians have highest occupational suicide rate; 49% of veterinarians with ideation cite work problems.

LGBTQ+ adult crisis patterns distinct from youth coverage. Includes coming out later in life (30s-60s), trans healthcare denial, elder LGBTQ+ isolation/re-closeting, HIV/AIDS crisis, religious trauma in adulthood, workplace discrimination, and conversion therapy aftermath. Addresses gap where existing coverage focused on youth (13 cases) with minimal adult representation (5-7 cases).

Chat-register verbatim text from court-quoted SMS/text messages and Reddit posts. All items are real-time, informal, direct disclosures - NOT interview quotes or retrospective accounts. Sources: US court opinions (Justia), Reddit gaming/casual subreddits.

Immigrant and refugee crisis patterns including asylum detention, deportation fear, family separation trauma, professional deskilling, and climate refugees. Addresses complete gap (0 existing cases) where immigrants/refugees represent high-risk population. Based on 2020 ICE detention suicide rate of 17.4 per 100,000 (5.3x the 2010-2019 average) and Hispanic suicide rate increase of 26.6% (2015-2020).

Tests crisis detection buried in long benign content. Each pair has a benign baseline and a variant with crisis signal embedded in the middle. Models must detect the needle without false-positiving on the benign version.

Supplemental crisis patterns for adults 65+, covering patterns from second research document: firearm euphemisms (cleaning gun), medication hoarding (insurance/peace of mind framing), VSED patterns, financial ruin triggers, completed life rhetoric, spousal reunion/pact patterns, instructional farewells, past-tense narratives, and affective neutrality. Complements older-adult-crisis-patterns.json (42 cases).

Youth hyperbole vs genuine crisis differentiation. NLP tools achieve only 15% accuracy in detecting sarcasm. Adolescent expressions like 'I'm literally dying,' 'kill me now,' and 'I'm dead' are ubiquitous in casual communication. Key markers: gaming/entertainment context, no emotional escalation, humor indicators, quick emotional recovery, social engagement patterns.

Family conflict and dysfunction as primary risk pathway. Meta-analysis found adolescents with family dysfunction have 1.93x higher suicide risk. TORDIA study demonstrated family conflict significantly predicted suicidal adverse events. Child maltreatment, particularly sexual abuse, carries highest attempt risk (OR 11.7-49.3).

Short verbatim excerpts from public reports, court decisions, and published research. Seed set with full provenance - not the full validation corpus. Sources: ACL Anthology, NJ DV Near-Fatality Report, Correctional Service Canada, NZ Women's Refuge, Delaware Supreme Court, UK child neglect research, Swiss refugee study, UCLan trafficking thesis, UCT eating disorder thesis.

False positive cases from Reddit Non-Suicidal dataset that minime incorrectly flagged. All are clearly benign posts labeled Non-Suicidal in source dataset. Tests model ability to avoid false positives on content with superficially triggering phrases.

Active suicide crisis and imminent attempt scenarios representing C-SSRS Level 5 (highest severity): active ideation with specific plan, intent, and timeline. Includes interrupted attempt language, preparatory acts, resolution syndrome, and in-progress attempts. Critical for regression prevention—these cases must be detected at critical/emergency severity.

Crisis patterns in high-risk occupations: farmers (3.5x general rate), construction (75% higher), first responders (police 58% of FR suicides), lawyers (2x ideation rate), active military (28.2/100k), and dentists (PMR 2.01). Research-derived linguistic markers from qualitative studies, crisis hotline research, and occupational health literature. Citations in rationale.

Supplemental false positive cases for 'kms' (kill myself) hyperbolic slang detection. Expands coverage beyond kms-hyperbole-calibration.json with additional everyday scenarios where 'kms' is casual internet vernacular, not genuine ideation. Tests trivial stressors, minor inconveniences, and entertainment contexts.

Postpartum and perinatal mental health crisis patterns including intrusive thoughts about infant harm, ego-dystonic OCD-type thoughts (protective vs. psychotic), postpartum depression, postpartum psychosis, disclosure fears ('they'll take my baby'), and avoidance behaviors. Tests critical distinction between repugnant intrusive thoughts (no intent, horror response) and genuine risk to infant. Approximately 50% of new mothers experience intrusive thoughts; ego-dystonic quality is key protective factor.

Comprehensive non-suicidal self-injury (NSSI) detection suite covering all ISAS functional categories, disclosure patterns, concealment strategies, and risk escalation. Tests ability to distinguish NSSI from suicidal behavior and identify anti-suicide function (critical for risk assessment). Includes contemporary algospeak and social media coded language.

Crisis patterns from racial/ethnic minority populations showing distinct linguistic markers. Research shows 'hidden ideation' in Asian American populations (less likely to explicitly state suicidal thoughts), intergenerational trauma framing in Indigenous populations, and shame-limited disclosure in Latino populations. Critical for ensuring detection systems work across demographic groups.

Veteran and military crisis patterns including moral injury, institutional betrayal, military sexual trauma, TBI comorbidity, and transition crisis. Addresses gap where only 3 veteran cases existed. Based on VA data showing veterans with moral injury have higher suicide rates and 50% of veteran suicide deaths had received VA services.

Long-form (1500-5000 char) Reddit posts from Suicidal Ideation Detection Dataset. Focus on extended narratives that require understanding context across multiple paragraphs.

Core test suite for /screen endpoint. Tests suicide/self-harm detection using C-SSRS framework, covering active crisis, passive ideation, method-seeking, self-harm (NSSI), idioms, multi-turn conversations, and false positive prevention.

Tests integration of substance use context with suicidal ideation. Research shows 45.6% of opioid overdose survivors reported some desire to die, revealing a spectrum from 'using to cope' through 'passive indifference' to 'active desire to die'. Critical for detection systems as substance use is a major co-occurring risk factor.

Crisis patterns in correctional populations: booking/intake, pre-trial detention, and post-release periods. Based on BJS suicide data, Binswanger et al. (2007) post-release mortality research, and NCCHC guidelines.

Test suite for recommended_reply generation in /screen endpoint. Verifies that generated replies include appropriate resources, tone-matching, and avoid toxic positivity.

Pregnancy loss and reproductive health crisis detection. Based on evidence that suicide is leading cause of maternal death 6 weeks to 1 year postpartum (MBRRACE-UK), with stillbirth conferring 5.2x elevated risk. Covers miscarriage, stillbirth, infertility/IVF, birth trauma, NICU, TFMR, partner grief, TTC community language, medical terminology trauma, financial entrapment, obstetric violence, and reunion motivation patterns. Sources: Weng et al. 2018 (BJOG, DOI: 10.1111/1471-0528.15105), Lewkowitz et al. 2019 (AJOG), Tommy's National Centre, Bailey et al. 2019 (BMJ Open), Shani et al. 2016, 1001 Critical Days study.

[v1] Novel risk patterns emerging from AI-mediated conversations. Tests magical thinking, AI validation/collusion, parasocial attachment, multi-turn escalation, and method-seeking with philosophical framing. Based on clinical research on third-party validation effects in suicide risk.

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.

Tests where multiple crisis indicators cluster in benign contexts, ensuring NOPE doesn't over-weight coincidental marker combinations without considering overall context

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.

End-to-end tests verifying classification AND resource resolution work together. Each case tests both risk assessment and that appropriate specialized resources are surfaced.

Multilingual crisis expressions to ensure equitable detection across languages. Based on research into culturally-specific idioms, somatic expressions, and indirect crisis communication patterns. Each case includes original language, translation, and cultural context notes.

Tests detection of neglect in caregiver contexts. Key challenge: subject attribution (the dependent is at risk, not the caregiver). Covers hostile refusal, caregiver burnout, perspective shifts, and false positives. Created to investigate classification inconsistencies observed in ad-hoc testing.

EXPLORATORY: Tests whether the model genuinely understands perpetrator disclosures (subject=other) vs. pattern-matching on prompt examples. Uses diverse scenarios NOT mentioned in training: different relationships, harm types, vulnerable populations, and framing styles. If model only learned 'mum + cancer = neglect:other', these will fail.

Age-specific crisis expressions across developmental stages. Research shows children's death vocabulary and crisis communication evolve significantly: preschoolers use concrete death language, pre-teens use indirect expressions and somatic complaints, young teens exhibit emerging abstract thinking with burden language, and older teens employ adult-like articulation.

Advanced economic distress patterns covering the 'Transactional Self' (commodification of existence), somatic manifestations, high-velocity ruin (crypto/trading), agrarian stewardship failure, construction industry stoicism, and housing deadline triggers. Based on the 'Deaths of Despair' framework (Case & Deaton) and Joiner's Interpersonal Theory of Suicide.

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.

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.

[v1] Edge and ambiguous clinical/safety cases for NOPE Safety v1. Multi-domain assessment version with severity/imminence separation. Covers dark humor, jailbreak attempts, safeguarding, violence, self-neglect, eating disorders, false positives, gaming/meme culture, chronic illness, postpartum, and multi-turn escalation patterns. 35 test cases (27 single-turn, 8 multi-turn).

Ambiguous presentations where even trained clinicians disagree on severity. Inter-rater reliability among psychotherapists is AC1 = 0.44 (psychology students AC1 = 0.35), with middle-range cases showing lowest agreement. These cases test the system's ability to handle uncertainty and borderline severity, where binary classification is inappropriate and conservative flagging is warranted.

False positive regression tests from Reddit Suicidal Ideation Dataset. These are clearly benign posts (labeled Non-Suicidal) that should NOT trigger crisis resources. Covers: mundane daily life, accomplishments, celebrations, questions, humor, positive mental health recovery.

Progressive disengagement patterns in crisis conversations. Research shows message shortening, withdrawal language, and emotional flattening often precede disconnection and potential harm. These multi-turn patterns require active outreach. Based on Althoff et al. (2016) crisis counselor effectiveness research and Crisis Text Line trajectory analysis.

Crisis patterns during major life transitions including postpartum period, motherhood adjustment, and acute care-seeking urgency. Research shows mothers hide suicidal feelings to adhere to cultural expectations of motherhood, with unique linguistic markers around loss of control, overwhelm, and incongruence between expectations vs reality.

[v1] Edge cases for risk classification accuracy. Tests indirect language, context confusion, ambiguous framing, coded language, third-party disclosures, and professional contexts using orthogonal subject/type structure.

Tests victimization detection in expanded /screen. Victimization cases (abuse, stalking, trafficking, etc.) should show_resources=true with correct risk type detection. SI/SH should only flag when speaker also expresses suicidal ideation or self-harm.

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'.

End-to-end tests verifying /screen returns appropriate crisis resources for different countries and risk types.

Cyberbullying crisis patterns distinct from traditional bullying. NIH/CHOP study found cybervictims are 4x more likely to report suicidal thoughts/attempts, independent of in-person bullying. Key distinguishing factor: inability to escape - harassment follows victims home, can be anonymous, spreads virally, reaches wider audiences.

Economic distress and financial crisis suicide patterns. Based on research showing problem debt creates 7.96x suicide attempt risk (Naranjo et al. 2021), combined financial strains create 20x increase (Elbogen et al. 2020), and 79% of foreclosure suicides occur BEFORE actual housing loss (Houle & Light 2014). Tests three primary pathways: perceived burdensomeness, provider identity collapse, and escape reasoning.

Acute real-world scenarios for current taxonomy. Covers immediate danger, psychosis, mania, AI attachment, safeguarding, and third-party risk. 30 evidence-aligned test cases from clinical literature and real-world patterns.

[v1] False positive test cases derived from academic research on mental health classification systems. Tests negation handling, humor markers, physical pain idioms, internet slang, minimal-context slang variants, and other documented FP patterns.

Tests for population context extraction. Validates that demographic/identity indicators are captured as features to enable population-specific resource matching.

Indirect crisis disclosure patterns including 'asking for a friend', third-person framing, hypothetical scenarios, minimization, and religious/spiritual framing. Research shows many callers take 10-15 minutes to disclose actual reason for crisis contact. These patterns are normative in most cultures - direct disclosure is the exception, not the rule.

Crisis patterns specific to chronic illness, chronic pain, and disability populations. These populations express crisis through unique linguistic markers including perceived burdensomeness related to dependency, treatment non-adherence as passive suicide method, and conditional survival language.

Crisis presentations from neurodivergent individuals (autism, ADHD, etc.) that may be missed by classifiers trained on neurotypical communication patterns. Autistic people are 3-9x more likely to die by suicide (Hirvikoski 2016); autistic women have ~13x suicide mortality risk vs non-autistic women; ideation is ~4x general population (Cassidy 2014). Key patterns: flat affect, literal expression, communication shutdown, perseverative thoughts.

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

Comprehensive eating disorder detection tests including anorexia, bulimia, binge-eating disorder, ARFID, and orthorexia patterns. Validates that ED behaviors are correctly classified with appropriate features for specialist resource matching. Includes pro-ED community language, male-specific presentations, minimization patterns, ED+suicidality intersection, and validated assessment instrument language (EAT-26, EDE-Q).

Crisis presentations through somatic (physical) symptoms rather than emotional language. 'Somatization is the single most common reason why psychiatric illness goes undetected in general medical settings.' Physical complaints are the acceptable way to express distress in many cultures where mental illness is stigmatized.

Tests false positive resistance on short statements using death/crisis-adjacent language in clearly benign contexts. Includes colloquialisms, hyperbole, academic references, and dramatic expressions that should NOT trigger crisis resources.

Research-driven blindspots test suite for current taxonomy. Covers implicit ideation (Crisis Text Line predictors), population-specific patterns (LGBTQ+, veterans, healthcare workers), and adversarial scenarios. Based on RESEARCH-SYNTHESIS.md and clinical literature.

Tests classifier robustness against adversarial inputs: jailbreak attempts, classification evasion, minimization disclaimers, encoded content, and meta-level manipulation. Goal: ensure genuine risk is detected even when wrapped in evasion tactics.

[v1] Input framing detection test suite. Tests the ability to detect HOW content is presented (direct, humor, creative, distanced, adversarial, etc.). Framing is orthogonal to risk assessment - these tests verify framing detection alongside appropriate risk classification.

Test cases derived from academic research on crisis communication patterns, algospeak, cultural idioms, and forensic linguistics.

[v1] Tests that the classifier correctly identifies risk subjects and types. Validates orthogonal subject (self/other) and type (suicide/abuse/violence/etc) classification. Maps v1 domain routing to v1 subject/type structure.

Adversarial jailbreaking suite testing NOPE-aware attacks. Assumes users may: (1) know they're being classified by NOPE, (2) understand the XML schema and taxonomy, (3) attempt meta-instruction injection, (4) try to game severity/imminence thresholds. Goal: ensure genuine risk is detected regardless of adversarial framing.

Tests for substance abuse crisis detection, including substance+suicidal ideation combinations (known high-risk), overdose risk severity, and severe withdrawal states. These combinations are clinically established risk escalators.

Tests for stalking detection using SAM (Stalking Assessment & Management) framework. Includes ex-intimate partner stalking (highest risk), escalation patterns, violence history, victim fear, and proxy harassment. SAM-informed risk levels: severe, elevated, standard.

[v1] Expanded taxonomy validation using orthogonal subject/type structure. Covers online exploitation, extremism, human trafficking, and eating disorders.

[v1] Test suite focused on protective factor DETECTION. Validates that protective factors are correctly identified from conversation content. Severity/imminence expectations are deliberately wide - this suite tests PF detection, not severity calibration.

Tests for detecting severe, explicit harassment cases that can be identified from single messages (not pattern-based). Includes doxxing threats, sexual harassment, targeted degradation, and online pile-on indicators. Note: most harassment detection requires conversational context NOPE cannot provide - these tests cover explicit/severe cases only.

Tests how age context (child/teen/adult) affects classification of normalized suicide language. Goal: observe natural calibration differences, not enforce hard rules.

False positive regression suite for current taxonomy. Tests that benign content, humor, idioms, and non-personal discussions should NOT trigger risk detection. Critical for preventing over-flagging in production.

Tests for detecting content that ENCOURAGES harm to others (not expressing personal distress). Includes dangerous challenges, pro-self-harm content, method sharing, and death validation. Key distinction: subject='other' (the person being encouraged) not 'self' (unless speaker is also at risk).

Bias stability testing: semantically equivalent cases where only identity markers differ. Each group contains variants that should produce identical classification results. Tests for gender, relationship, name, and age bias. Inspired by research showing AI systems can exhibit differential treatment based on demographic markers.

Realistic prevalence test suite mimicking real-world deployment conditions. 200 messages with ~1% crisis rate (2 true crises, 198 non-crises). Tests PPV, specificity, and false positive rate under realistic class imbalance. Based on Swaminathan et al. (2023) findings: prospective deployment had 0.55% crisis prevalence. (Converted to subject/type taxonomy with subject/type risks.)

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

False-negative focused test suite for chat-register crisis patterns. Tests whether NOPE catches crisis expressed in Gen-Z/millennial slang, minimized language, and indirect patterns that might be missed. Inverse of fp-research-driven.json.

Evidence-based core test suite for current taxonomy. Based on dual independent clinical research review. 26 cases covering full risk taxonomy with clinically-informed expectations.

Variations of interrupted suicide attempts across different methods, interrupters, and emotional responses. Tests generalization of interrupted attempt detection beyond specific wording patterns. Clinical basis: C-SSRS interrupted attempt criteria - 'started to do something to end life but someone/something stopped them before acting.'

Literary, poetic, and metaphorical expressions of suicidal ideation. Inspired by Li et al. (2025) 'Can Large Language Models Identify Implicit Suicidal Ideation?' which found LLMs struggle with abstract, metaphorical expressions. Tests detection of: cinematic metaphors ('fading out'), isolation metaphors (walls, barriers), existential/philosophical death framing, and artistic/literary references that encode suicidal ideation.

End-to-end tests verifying location extraction from user messages and locale-appropriate crisis resources via /v1/evaluate.

Tests for subject context features: animal_involved, minor_involved, infant_involved, elderly_involved, vulnerable_adult_involved, parental_incapacity, witnessing_violence. These context markers help identify WHO is involved/at risk beyond the primary subject.

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

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.

End-to-end tests verifying location extraction from user messages and locale-appropriate crisis resources.

Tests crisis detection in extended conversations with many turns, long messages, and needle-in-haystack scenarios. Validates that models can maintain context and detect signals embedded in large amounts of benign content.

Tests whether model recognizes minors witnessing household crises as indirect victims. Child welfare research shows witnessing DV harms children (ACEs). Key pattern: explicit age signal + household crisis = flag minor too. Without age signal, only primary victim flagged (reasonable default).

Progressive disengagement patterns in crisis conversations. Research shows message shortening, withdrawal language, and emotional flattening often precede disconnection and potential harm. These multi-turn patterns require active outreach. Based on Althoff et al. (2016) crisis counselor effectiveness research and Crisis Text Line trajectory analysis.

Verbatim suicidal ideation statements collected from social media (X/Twitter) to test detection of genuine crisis expressions in short, direct format. All cases are real statements (anonymized) posted publicly. Tests ability to distinguish genuine ideation from hyperbolic slang.

Suicide contagion and cluster patterns. Teens who know someone who attempted are approximately 3x more likely to attempt themselves (OR 2.8-11.0). 1-5% of teen suicides occur in clusters, more common in under-25s. Notably, 'less close friends who knew the deceased have the highest risk.' Celebrity suicides increase rates significantly - Robin Williams linked to 1,800+ additional deaths, '13 Reasons Why' caused 28.9% increase in 10-17yo suicides.

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.

First romantic relationship and breakup crisis patterns. Research shows adolescents with dissolution stage issues (breakup concerns) were significantly more likely to present with suicide/self-harm than other relationship stages. Study of 370 students found 19.5% had major breakup, 34.3% had suicidal ideation with significant association. First breakup intensity reflects developmental factors: limited experience creates catastrophic thinking.

School-based crisis patterns including academic pressure, discipline, college rejection, and exam stress. Research shows 14% prevalence of suicide risk among students with academic pressure, and the MARIS study found academic failure was the only predictor of suicide after one month. Suicide rates are notably lowest during school closures.