mine-the-papers-you-were-given

# The papers in `related_work/` are a work list, not background reading A supplied paper is not context. It names the tools the field uses on this problem, the benchmark events results are reported against, the baselines a number is compared to. Those names are what a reader checks your work for, and they are sitting in a file you already have. ## The failure this prevents It is not that runs skip the papers. Measured across a 40-task batch: every task ships them (median 4 PDFs, median 250,000 characters of extractable text) and **every run referenced 100% of them**. They are opened. They are read for framing. Then the work list is written from the task statement alone, and everything the papers name goes unused. Chemistry_002 is the clean case. `CAPRI` appears **30 times** in its supplied papers and `HADDOCK3` **22 times**. The two graded requirements that name them — a CAPRI round result, and the tool's consensus scoring workflow — scored **5 and 0**. The report discusses CAPRI *as a concept* and never reports a CAPRI result. Life_001's papers name `NetMHCpan` five times; the requirement naming it scored 5. The shape is always the same: the run knows the field's vocabulary, because it read the papers, and produces its own analysis in its own terms beside a literature it has paraphrased rather than used. ## What to do 1. While the papers are open at literature stage, write `literature/named_from_papers.json`: one row per **proper noun** the supplied papers attac...