Evidence synthesis automation map, and a measured probe of what a web-enabled LLM already does

Two documents, in English and Simplified Chinese. The automation map lists 26 steps of a systematic review or meta-analysis; for each step it records what the job is, which tools already exist (link and price where verifiable, checked September 2026), what a web-enabled large language model did in our probe, and what still needs a person. The benchmark document reports that probe: nine tasks built from real public material, each with an independently verified answer key and a written scoring rule, with the model limited to web search and page fetching. Limits are stated in both documents: single runs, one model family, small tasks, and a prior-art pass that was not adversarially re-checked. Individual articles found to have numeric or registration discrepancies are not named. This record is a snapshot of release v1.0.2 of https://github.com/yinchuan123/evidence-synthesis-automation-map