EvoRiri: Interpretable Evolutionary Search for Diagnosing Reward Design Failures in a Bangla Mental Health Assistant

Mental health conversational systems must balance empathy, risk-aware safety, and actionable structure, yet optimising hand-designed proxies for these qualities can expose reward design failures. We present EvoRiri, a template-based Bangla–Banglish–English mental health assistant that uses no language model. Its response policy is an eight-gene mixed continuous–discrete genome (empathy, safety and structure weights, a distress-gating gene, two reserved escalation thresholds, a template selector, and a response-part budget). A genetic algorithm (population 100, 20 generations, tournament selection, and uniform crossover, with 2000 evaluations per run) maximises a scalar fitness that weights rule-based empathy, safety and structure scores (0.40/0.40/0.15) with per-example safety penalties. On 2872 de-identified records rendered as synthetic prompts, the evolved genome raises structural completeness from 0.65 to 1.00 and fitness from 0.556 to 0.703 over a fixed reference genome, with comparable safety. A one-factor ablation attributes this gain to a single structure-weight threshold, and over 30 paired seeds, genetic and equal-budget random search are equivalent on validation prompts. Trace analysis exposed failures that aggregate fitness hid, including question-only collapse, empathy-marker stuffing, and response truncation. The contribution is a diagnostic case study of reward design on a synthetic single-turn corpus, not an optimisation advantage or clinical validation.