DocTrust

Bangladesh's Election Commission has received 6.57 million applications to correct national ID records, and 1.36 million were rejected. When a person's documents disagree—a name transliterated two ways, a birth year off by one, digits in Bangla on one form and Latin on another—their bank account, their SIM, their life stops until a human reviewer untangles it. DocTrust is a reviewer agent built for exactly that job. It cross-examines a document set—national ID, birth certificate, bank form—in Bangla, English, and Hindi, and answers one question: do these documents tell a coherent story, and what evidence is missing? Muse Spark reads every field across scripts and digit systems, reading each one several ways and estimating its own uncertainty. When reads disagree, an investigation planner—the same Spark—chooses a step: re-read, zoom, or call SAM 3.1 to segment whatever blocks the view, including lifting an ink stamp to re-read the print underneath it. Deterministic code, never the model, then compares the evidence and issues the verdict: Consistent · Discrepancy: needs human review · Unreadable: request recapture. It never auto-rejects—every flag points to evidence a human can check. Across 100 measured synthetic document sets in two independently seeded batches, zero false flags on genuine files and zero silent passes—every failure ended as a human review or a recapture request. On a separate 30-set multi-persona batch, transliteration variance in father names still flags 7 of 30 genuine sets; the full evaluation, including where the agent does not beat a single careful read, is in the repo's EVALUATION.md. 191 automated tests. Live demo: https://rafiimam.github.io/DocTrust/demo/