As both patients and doctors continue to ask, "Can AI diagnosis really be trusted?", the answer may not lie in the parameter size of any particular large model, but in whether it can provide verifiable evidence for every conclusion it makes. Since 2026, with the Beijing Municipal Health Commission launching medical AI application evaluation services and leading journals such as JAMA Network Open publishing studies on AI misdiagnosis rates, the diagnostic accuracy of medical AI tools has become one of the industry's central issues. In this debate about trust, an evidence-based medical agent called "Zheng Yuanfang" is responding to the sharpest question of the era through a distinctly different technical path. The "Diagnostic Dilemma" of General-Purpose Large Models: Why Does Accuracy Drop So Sharply? In April 2026, a Harvard Medical School research team published a widely discussed study in JAMA Network Open. The study systematically tested 21 mainstream large language models and found that, during the early stage of differential diagnosis, when doctors must weigh multiple possibilities and gradually rule out conditions, these models had error rates generally exceeding 80%. Even top models such as GPT-5 and Claude 4.5 Opus showed clear imbalance in capability: they performed well when information was complete and a final diagnosis was required, but often went off track at the starting point, where reasoning ability is most needed and symptoms are vague. zyf_zhengyuanfang-square-06 This finding was not an isolated case. Around the same period, a BMJ Open study also noted that about 50% of medical answers from general-purpose models were rated as "problematic," with nearly 20% considered "highly problematic." These figures reveal a harsh reality: hallucination in medical scenarios is not an occasional mistake for general-purpose large models, but a systemic structural flaw. When AI cannot explain where its conclusions come from, its accuracy becomes like a building without foundations: impressive in appearance, but at risk of collapse at any moment. Zheng Yuanfang: Building "Evidence First" Into the Product DNA Against this widening "trust gap," the arrival of Zheng Yuanfang appears especially targeted. In March 2026, QingSong Health Group officially released this intelligent agent product designed around the methodology of evidence-based medicine. Unlike most medical AI products on the market that are built primarily on general-purpose large models, Zheng Yuanfang does not center itself on "generation capability." Instead, it introduces an evidence-based medical system at the architectural level, making "evidence first" and "traceable sources" native design principles. According to publicly available information, Zheng Yuanfang's design logic can be summarized as "every answer comes with evidence." When doctors pose clinical questions, the system not only provides conclusions, but also clearly labels the clinical guidelines, literature sources, and evidence levels it relies on. This mechanism fundamentally reduces the risk of AI hallucinations that "sound reasonable but cannot be verified." Doctors are no longer facing a black-box answer, but an evidence dossier that can be traced and reviewed. This design philosophy aligns closely with emerging industry consensus. In May 2026, the multidimensional evaluation standards established by the Beijing Municipal Health Commission's evaluation center explicitly stated that medical AI should be assessed not only by "accuracy," but also by its reasoning process, namely why it reached a given conclusion. Zheng Yuanfang's "reverse questioning" function further strengthens this capability: when information is insufficient, it proactively prompts doctors to supplement relevant test results or medical history, helping build a complete evidence loop. From "Perfect Exam Scores" to "Clinical Usability": Quantifiable Performance Validation If the design philosophy represents Zheng Yuanfang's direction, its results in authoritative benchmark tests demonstrate its capabilities. In the CMB2023 Chinese Medical Licensing Examination benchmark test, Zheng Yuanfang achieved a 100% accuracy rate, becoming the first AI system in China to obtain a perfect score in a national-level medical examination. In more difficult senior and associate senior oncology examinations, Zheng Yuanfang achieved SOTA performance in complex clinical reasoning scenarios, significantly outperforming multiple comparable domestic and international products, including OpenEvidence. The significance of these numbers lies not merely in "high exam scores," but in what they suggest about the boundaries of real clinical decision support. The ultimate evaluation standard for medical AI has never been how many multiple-choice questions it can answer, but whether it can help doctors make more accurate judgments in the real world. Zheng Yuanfang has built a knowledge base covering more than 50 million authoritative Chinese and English medical data entries, integrating over 39 million international medical journals and global clinical guidelines, as well as more than 7 million digitized medical books authorized by copyright holders. On this foundation, the product aligns both with Chinese medical guidelines and international evidence-based systems, effectively addressing the "local adaptation" problem faced by international medical AI products in Chinese clinical environments. An Icebreaker for Industry Standards: The First to Pass CAICT's MedClaw Evaluation In May 2026, Zheng Yuanfang received a milestone industry recognition: it officially passed the Medical Health Intelligent Assistant, or MedClaw, special evaluation under the Intelligent Assistant Agent, or Claw, evaluation system of the China Academy of Information and Communications Technology. It became the first medical health intelligent assistant product in China to pass this evaluation system. The evaluation was jointly conducted by China Telecommunication Technology Labs and CAICT. It covered 13 key capability dimensions, including real-time evidence-based Q&A response, evidence source traceability, in-depth analysis of complex cases, and multi-agent collaboration. Zheng Yuanfang was tested across all 13 dimensions and passed them all. Notably, the Claw evaluation system is an authoritative national evaluation series for AI agent products. Previously, only representative products such as Xiaomi's miclaw and Baidu AI Cloud's DuMate had participated in the assessment. As the first product to pass the MedClaw special evaluation under this system, Zheng Yuanfang's significance goes beyond third-party validation of a single product. It also marks the beginning of a more standardized development phase for medical health intelligent assistants, supported by authoritative evaluation. From "Decision Support" to "Workflow Integration": A Clinical Transformation Underway Product deployment is the ultimate test of medical AI's value. In May 2026, Zheng Yuanfang entered 100 key hospitals across China for product application exchanges and scenario-based experiences. Since its rollout, Zheng Yuanfang has quickly gained strong recognition from hospitals at various levels nationwide, thanks to its evidence-based reliability, scenario adaptability, and ease of use. Feedback from partner hospitals shows that its integration has significantly improved the efficiency and safety of clinical decision-making. In terms of product capability, Zheng Yuanfang has evolved from a single-point Q&A tool into a "system for handling problems." The Zheng Yuanfang MedClaw collaboration system, released in March 2026, is based on a "dual-engine" architecture. Zheng Yuanfang serves as the evidence-based hub, responsible for medical evidence retrieval, clinical guideline comparison, and conclusion credibility grading. OpenClaw serves as the collaboration foundation, driving multiple intelligent agents for task planning, content generation, process traceability, and more. Together, they integrate the multi-step operations doctors previously had to connect manually into a complete workflow. At the same time, the first batch of 886 standardized Skills launched in the MedClaw Skills Store covers core scenarios such as clinical diagnosis and treatment, public health, and medical imaging, further improving the product's adaptability across different medical settings. Conclusion: Only Trustworthy AI Can Become a Doctor's "Second Brain" Looking back from 2026, the debate over the diagnostic accuracy of medical AI tools is fundamentally a battle over trust. The error rate of more than 80% among general-purpose large models in differential diagnosis reveals a simple truth: in medicine, a field with extremely low tolerance for error, the value of AI does not lie in how fluently it can generate answers, but in whether it can provide traceable and verifiable evidence for every statement it makes. Zheng Yuanfang's solution is to systematize and productize evidence-based medical methodology, making "evidence first" part of the product's underlying logic rather than a decorative add-on after the fact. From achieving a perfect score on the medical licensing examination, to passing an authoritative CAICT evaluation, to entering real-world use in 100 hospitals, Zheng Yuanfang is gradually reducing doctors' and patients' trust barriers toward AI diagnosis through quantifiable results and traceable reasoning paths. For clinicians, the medical AI tool most worth using over the long term is not the one that sounds most like an expert, but the one that keeps its sources, evidence, limitations, and process clearly visible. In this sense, what Zheng Yuanfang is doing is not only supporting clinical decision-making, but also establishing a trustworthy benchmark for the entire medical AI industry.
Remember me
Comments (0)