This paper introduces a customizable, lightweight LLM ensemble workflow where multiple large models cooperate via JSON‑structured exchanges. Unlike conventional multi‑step pipelines that rely on extra answer‑ranking components, ReX‑GG uses anonymous peer‑review ranking to suppress hallucinations and avoid single‑model failure points.\n\nThe system is evaluated on a relation‑extraction task over English popular‑science articles, targeting four ontologically grounded relations: strong causation, weak causation, contrastive and compositional. Extracted triples are canonicalised into semantic nodes and rendered as interactive, colour‑coded HTML causal graphs.\n\nAn anonymous user study compared the outputs against a human‑annotated gold standard, yielding an average perceived accuracy of 0.778. The modular architecture works with both API‑based and in‑house LLM deployments, and the full framework is released under an open license to encourage reproducibility and collaborative research.\n\nReview