Long‑horizon deep search demands LLM agents to iteratively decompose complex information needs, gather evidence from external sources, and maintain a reliable research state under multiple constraints. Early mistakes tend to be amplified along the search trajectory: agents may anchor on an incorrect candidate, collect only partially supporting evidence, overlook contradictions, and accumulate errors that misguide subsequent reasoning.\ \ We introduce FALCA, a falsification‑aware framework that formulates deep search as a propose‑falsify process. FALCA adopts a collaborative Searcher‑Falsifier architecture: the Searcher proposes candidate answers or sub‑answers, while an independent Falsifier actively seeks counter‑evidence, checks unmet constraints, identifies conflicts, and produces concise falsification reports to revise the research state.\ \ To strengthen this loop, we propose falsification‑aware training, which combines SFT warm‑up on falsification trajectories with agentic reinforcement learning based on parallel rollouts and hierarchical credit assignment across research stages, state transitions, and actions.\ \ Experiments on five search benchmarks demonstrate that FALCA improves long‑horizon deep search performance over backbone and ablation models, while reducing confirmation bias and error accumulation through more effective falsification and research‑state revision.\ \ Review