Problem

LLMs shouldn't get the final say on whether a DNA variant causes disease. So they argue, and a deterministic engine decides.

Approach

One agent argues the variant is pathogenic, another argues benign; TypeSafe cross-examines both; a jury revises a shared structured state over up to three rounds. A deterministic ACMG-style rule engine issues the final classification. Weave traces every step and ARIA compares runs to recommend better loop configurations. FastAPI backend, Nuxt 4 frontend, CLI demo, marimo research notebook.

Results

A working research prototype with a pytest suite — explicitly not for clinical use.