摘要

Biomedical analysis often requires researchers to translate published methods into executable code while reconciling incompatible repositories, incomplete resources, local data, and task-specific evaluation. We present PathoWish Evidence Loop, a research workbench that connects a host-selected language model to a versioned evidence and execution contract. The model reads selected literature and source material, develops independent project modules, chooses dependencies for the actual task, and reviews observed outputs. The workbench binds these actions to approved plans, input hashes, artifact dependencies, execution records, and corrective child versions. Independent branches continue when another branch lacks a prerequisite, and substitutions are recorded for post-run inspection. We evaluate the implementation through release-bound software checks and eight deliberately small public-data cases spanning pathology, radiological imaging, and multi-omics. The cases produced editable code, quantitative outputs, figures, and consolidated reports, with retained records of failed attempts and subsequent repairs. A held-out virtual-staining field achieved SSIM 0.6511; a tissue-segmentation example yielded macro-IoU 0.2185 when all classes with nonempty unions were counted; and a three-variable survival model achieved test concordance 0.6437. These heterogeneous outcomes illustrate why operational completion and scientific quality require separate records. The contribution is an inspectable workflow for method adaptation, execution, and revision, supported by case-level reproducibility evidence and explicit evaluation scopes.

案例与证据范围

论文整理 2026 年 10 月 15 日保存的八个公开数据最小案例，科学结果对应插件版本 0.4.16+client.20261015170000。每个案例保留独立评价口径，涵盖 IHC 定量、CT 测量、虚拟染色、细胞核与组织区域分割、预后预测、预定位 CT 病灶分割和多组学整合。方法条件、失败修正和第三方图像来源随文说明。

交互回看真实结果 · 按教程复现与继续修改

引用本版本

Luo, R. (2026). PathoWish Evidence Loop: Evidence-Guided, Host-Model-Driven Biomedical Research Workflows. Author preprint, 16 October 2026. Luoxiaoluo Tech (Beijing) Co., Ltd. https://lxlmedx.site/research

当前为作者公开的技术论文，尚未投稿 arXiv，未分配 arXiv 编号或 DOI。论文中第三方图像的来源与许可见图注及附录。

文件校验

下载文件为最终排版的英文 PDF，可直接公开获取。SHA-256：