Answer, Ask, or Escalate? Evaluating Public-Service Triage in Global South Languages

Public-service assistants are often judged as ifsuccess means answering correctly. In practice,that is not enough. A safe system should some-times answer, sometimes ask for one missingfact, and sometimes escalate the user to anofficial channel. We introduce a document-grounded triage benchmark with 90 revieweditems across Sinhala, Tamil, and Swahili, cov-ering civil-registration procedures in Sri Lankaand Tanzania through 15 language-specificservice cards. Models must choose Answer,Ask, or Escalate under fixed official evidence.Answer-only evaluation substantially overstatesreadiness: Qwen 2.5 3B gets answer-only score1.00 but only 0.344 triage-safe score, with 0.917harmful over-answer rate. A smaller Qwenmodel fails in the opposite direction and askson nearly all items. Claude Sonnet 4.6 reaches0.844 triage-safe score with much lower harmfulover-answering, which suggests the task issolvable but model-tier dependent. A 45-rowreviewed human-evaluation slice is consistentwith the automatic ranking. These resultssuggest that public-service evaluation shouldmeasure whether models know when to answer,ask, or escalate, not only whether they answerroutine cases. A public reproducibility packageis available at github.com