We built a research discovery pipeline that ingests papers from arXiv and Semantic Scholar, finds cross-domain connections, generates hypotheses with a multi-model ensemble, formally verifies them with Z3, then stress-tests survivors in adversarial debate.

The twist: we capture and display what each model said when critiquing. No single-model black box — you see GPT-4o, Claude, Gemini, and Grok arguing for and against the same hypothesis.

Example: [Distributed feedback control from microbial consortia enhances metabolic stability in Ginzburg-Landau cognition models](https://www.aegismind.app/discoveries/2af7c10d-18f8-42d5-8c9...). The hypothesis bridges synthetic biology and physics-of-cognition. The debate transcript shows Claude calling it "artificially stitched together" while Gemini finds it "a plausible theoretical synthesis." We surface both — and the evidence score (38% challenged) — instead of hiding the disagreement.

Pipeline: arXiv ingestion → cross-domain matching → multi-model hypothesis generation → Z3 theorem prover → adversarial debate → ranked discoveries. The whole thing runs autonomously; discoveries are published daily at [aegismind.app/discoveries](https://www.aegismind.app/discoveries).

We'd love feedback on the approach. Happy to answer questions about the architecture or the debate design.