Each session carries three hidden severity signals in [0,1]: D (depression cues from text/voice), A (anxiety cues) and S (a rarer suicidality risk factor). The classifier gate combines them into a single risk score with a logistic model: R = σ(w_d·D + w_a·A + w_s·S − b), σ(x) = 1 / (1 + e^-x) Higher Detection sensitivity lowers the decision threshold τ, so more borderline sessions get flagged. Sessions with R below τ pass to a digital therapist (CBT-style chat, e.g. Woebot) for self-guided support; sessions above τ are either escalated straight to a human clinician, or — in Auto-respond mode — first triaged by the chatbot before a high-confidence case is still escalated.

• Depression / Anxiety signal — mean severity of the simulated patient population
• Detection sensitivity — how aggressively the classifier flags borderline cases
• Therapist mode — auto-respond via chatbot first, or escalate flagged cases only

Illustrative model, not a clinical diagnostic tool.