Abstract / Summary
MIRA is a research prototype of a fixed six-stage sequential triage pipeline (five LLM-backed agents plus a deterministic red-flag layer at the Triage stage) for WhatsApp-style symptom messages in English and romanized Hindi/Telugu. This corrected version supersedes v1.2-interim, which misdescribed the system. The evaluation is a 30-case synthetic set written by the first author, run locally on 2026-09-11 against hosted open-weight models (Groq and NVIDIA NIM; three Gemini calls) with synthetic text input; no Anthropic model, PostgreSQL, Docker, or learned transliteration model was used. Main verified finding: 17 of 30 cases (56.7%) stopped at an Intake completeness gate and never reached the Triage-stage red-flag layer, and none received a safety or referral message; this is a system-level observation, not an accuracy estimate. Clinical review was partial and not blind (Rater 1 is a co-author and the first author's mother). Tier agreement with usable Rater 1 assessments was 1/13 for explicit written tiers and 8/23 including 10 inferred from "agree"; the v1.2 14/29 scoring convention was never recorded and is retained only as a historical reconstruction, and v1.2's second-rater aggregates are withdrawn. Results are a retrospective selection of infrastructure-complete attempts and are not fully reproducible. MIRA is not clinically validated, not deployed, not evaluated on real patients, and not for clinical use. The archive contains code, the synthetic set, run outputs, structured logs, scoring scripts and derived tables; third-party and provider-generated material is described in THIRD_PARTY_NOTICE.md.