Abstract
<title>Abstract</title> <p>Background and objectives. Large language models (LLMs) may support decision-making in multidisciplinary tumor boards (MDTs), but rigorous evaluation is hindered by hallucinations and poor reproducibility, where identical inputs yield divergent recommendations. Sarcomas, with heterogeneous biology, rare subtypes, and reliance on multidisciplinary consensus, represent a demanding test case. We developed and validated a schema-enforced LLM framework producing reproducible, hallucination-resistant sarcoma tumor board recommendations as a basis for subsequent concordance studies against historical MDT decisions. Methods. We built an LLM-based tumor board simulator using Anthropic Claude Opus 4.5 (temperature 0.0) producing structured JSON output via tool-use enforcement against a Pydantic schema covering nine clinical domains and 21 controlled-vocabulary decision codes. The methodology was iteratively refined in three phases (free-text Markdown; schema-enforced JSON; semantic clarifications with conditional logic). The final framework (schema v4) was tested on 51 fully de-identified sarcoma cases from a single center (September 2020 – September 2025), each processed three times. The pre-specified primary endpoint was per-case decision-code reproducibility (proportion of codes identical across all three runs). Hallucinations were classified into five pre-specified categories. Results. Overall reproducibility was 93.9% (1,006 / 1,071 decision-code instances stable). Seventeen cases (33.3%) achieved 100% reproducibility across all 21 codes. Per-domain reproducibility ranged from 88.9% (radiotherapy) to 98.0% (imaging). Six of 21 codes (29%) reached 100% stability. The reconstruction decision was least stable (74.5%), reflecting clinically interchangeable alternatives. No structurally fabricated content occurred in any of the 154 runs under the final schema. Conclusions. Schema-enforced LLM output with explicit semantic and conditional rules achieved high reproducibility and prevented structurally fabricated hallucinations in sarcoma tumor board simulation. Residual variability clustered in clinically equivocal domains, supporting subsequent concordance studies against historical MDT decisions.</p>