We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
DRAFT-PATCH: a multi-pass large-language-model protocol for extracting monitoring data from the primary literature
Summary
Scientists have created an AI tool that automatically reads through thousands of research papers and pulls out data on contaminants like "forever chemicals" (PFAS), pesticides, and microplastics, a task that would take humans years to do by hand. This matters because it can quickly reveal gaps in our monitoring of harmful chemicals in the environment, helping researchers and regulators spot where dangerous contaminants might be slipping through the cracks undetected.
DRAFT-PATCH reads primary publications and writes structured monitoring records. A first pass drafts the records; up to four further passes re-read the source and correct omissions and formatting errors, which is what makes extraction reliable enough to run across thousands of papers. The protocol is schema-driven and not specific to PFAS: the same pipeline can be pointed at pharmaceuticals, pesticides, microplastics or any contaminant whose measurements are scattered through the literature. This repository holds the tool and every analysis built on its output for theaccompanying paper: - 01_extraction_draft_patch - the extraction tool itself: prompts, the multi-pass correction loop, a browser interface, Docker setup and a packaged benchmark.- 02_benchmark_scoring - scoring against hand-curated ground truth. Extracted rows are aligned one-to-one to ground truth by Hungarian matching on a weighted field score (PFAS type 0.10, location 0.20, date 0.20, concentration 0.30, unit 0.20); a row counts as a true positive at 0.70. Includes the alignment-threshold sensitivity analysis.- 03_meta_analysis - harmonisation and random-effects meta-analysis of co-measured pairs, yielding cross-media transfer factors. Entry point run_site_level_meta.py.- 04_occurrence_and_gaps - catchment construction, compositional analysis (CLR-PCA and non-negative matrix factorisation), source association, and the monitoring-coverage audit.- 05_figures - plotting scripts for the main and supplementary figures.