Scaling logic vulnerability discovery in high-value targets like Android and Chrome is difficult: traditional fuzzing is blind to non-crashing logic defects, while known LLM approaches fail on large-scale codebases due to context hallucination.In exploring LLMs for logic vulnerability discovery, we found it works best when combined with more verification and more tools. So we built a framework powered by "feedback-driven verification", which harvested 100+ Android and Chrome 0-days in just two months. We structured it with two parallel pipelines: an exploratory workflow modeling N-day patch primitives and generalizing them to verify 0-day variants, and a verification workflow auditing every phase through dedicated agents and toolchains. This step-by-step verification creates a closed-loop feedback mechanism, harnessing intermediate failures to dynamically correct logic and make finding vulnerabilities more efficient.We have reported 100+ 0-days to the Google VRP to date (many rated as high-severity), some of which expose severe threats to Web3 assets and enterprise security. To validate these threats, we chained seven of these high-severity flaws into the "Invisible Sovereign" kill chain: a non-root exploit chain that breaks Android's core trust and isolation boundaries on fully patched Pixel devices. It covertly exfiltrates Web3 mnemonics and payment QR codes, and enables persistent camera surveillance that invalidates Enterprise DPC policies.