HalluciFeds

A trust evaluation harness that stress-tests medical voice AI across four failure dimensions: role confusion, translation infidelity, hallucination, and instruction drift. It feeds adversarial clinical scenarios, grounded in real trial data, through any LLM and scores exactly where it breaks. Built with Deepgram for voice, Cekura for simulated personas, Synthio for clinical validation, and Bedrock Claude as both the system under test and the judge. If your medical AI can't pass HalluciFeds, it shouldn't touch a patient.