Grounded ontogenic retrieval improves accuracy and citation fidelity of a medical oncology clinical reasoning AI assistant

Abstract Introduction: Large language models (LLMs) have demonstrated considerable utility in generating clinically informed text to assist physicians, yet uncertainty about the accuracy of their outputs has limited their broader adaptation. General purpose frontier models challenge the necessity of domain-optimized proprietary applications; however, they remain prone to attribution errors because their assertions are drawn from opaque model weights rather than explicitly verifiable references. Tethering model-generated outputs to auditable and verifiable ground truths is essential for establishing clinical confidence and minimizing the risk of medical error. Here, we introduce Grounded Ontogenic Retrieval (GOR), an architecture that couples effective prompt structuring with a harness-driven “ontology-on-the-fly” framework that creates a powerful neuro-symbolic disease management platform based on publicly available curated oncology management guidelines. Methods: The system employs a three-step pipeline—intent extraction, targeted source retrieval, and precise localization of relevant facts within retrieved material—implemented through a tailored architectural harness that mediates all input to and output from the core language model. Retrieval is performed via the Brave Search API, with capture restricted to abstracts and query-pertinent paragraphs. ASCO Rapid Recommendations and ESMO Living Guidelines are additionally checked for late-breaking updates. Each retrieved source is scored on four factors: recency, authority (domain tier plus a publication-type bonus), indication match (line-of-therapy alignment), and topical relevance. Topical relevance is assessed primarily by a separate LLM grader that accounts for medical synonymy, with a lexical word-overlap score serving as a fallback. A larger-context model (GLM-5.2) holds the full assembled corpus in-prompt, reinforcing the design principle that inference is driven by grounded retrieval rather than parametric weights. The pipeline also handles intent classification, agent extraction, and approval adjudication for limited-flag and limited-options clinical calls. The performance of this model was compared with that of several frontier and advanced open-source models from a mix of vendors, using an LLM-as-judge methodology. Models were provided online search facility, but received no additional prompting than a presentation of a clinical vignette, mirroring a typical physician use case. Results: Model performance was evaluated on a curated set of 100 generated clinical questions designed to drawn from research published within the preceding year. By restricting evaluation to recent literature, the design precludes answers sourced from parametric knowledge acquired during pretraining, thereby isolating and measuring the model's capacity to reason over current, retrieved evidence. The Aprilio model, based on grounded ontogenic retrieval (GOR), scored 100% accuracy on the test panel, with most other models scoring between 97% and 99%; the open-weight Qwen3 Max scoring 89%. As these models were not prompted specifically to perform an online search, all but one model (GPT-5.6 Terra) reliably did so, although four of the citations it provided were non-existent or erroneous. Perplexity Sonar Pro and Qwen3 Max searched 100% of the time, but left only a citation factor in one and twelve citations, respectively. One model cited a patent-petition document to support an answer. Other models utilized search facility to answer queries for between 35% and 81% of the problems but left almost no locatable citation, making reference to a supportive research study only in prose. Most importantly, the Aprilio model provided locatable links in all of the vignettes (without hallucination or dead links), while all other models, save GPT-5.6 Terra, exhibited substantially lower performance. The Aprilio model also supplied a citation evaluable for claim support in 97 of 100 responses, against 20 for GPT-5.6 Terra and 5 or fewer for the rest. Notably, 55% of responses generated by the Aprilio model carried citations which were confirmed to directly substantiate its assertions, whereas only 15% of citations from the next-best-performing model substantiated its claims. Potential harm could have resulted from a few of the answers, such as by Perplexity Sonar where the recommended agent did not account for critical baseline patient factors, while Qwen3 Max recommended a drug that has been withdrawn from the market due to unacceptable toxicity. In addition to the lack of auditable citations, we found instances of evidentiary misrepresentation in 1 of 14 instances in which claims were made regarding the preference rung of a treatment option according to reference guidelines. While questions were answered correctly, the reasoning behind the recommended selection was faulty. A correct answer can therefore exist beside a misstated authority claim, which is why an MCQ score cannot be read as a statement about open-ended reliability. Conclusion: The proposed grounding architecture, encompassing real-time information retrieval, coupled with query processing to extract intent and impose semantic structure, binds each generated clinical statement to a verifiable, auditable source, shifting the selection criterion for assistant performance away from the model's internal factual knowledge and toward its fidelity to curated, up-to-date evidence. A grounded, ontogenic retrieval (GOR) framework offers a more consistent and less error-prone foundation for deploying LLM-based decision support in oncology, directly addressing the reliability concerns that have limited physician adoption of LLM assistants in clinical practice.