Supplementary data and code for: "From hosing experiments to human impacts: a systematic evidence map of AMOC weakening"

Supplementary data and code for: "From hosing experiments to human impacts: a systematic evidence map of AMOC weakening" Author: Dimitri Defrance (CosmosClimae, Montpellier, France) Contact: [e-mail] Associated manuscript: submitted to Earth System Dynamics (2026) This repository contains the complete data, screening records, and code required to reproduce the systematic evidence map of the socio-ecological impacts of a weakened Atlantic Meridional Overturning Circulation (AMOC). It is released so that the search, screening, coding, and analysis can be independently reassessed. Overview of the workflow Web of Science search (4 Boolean queries, 15 Jan 2026) │ 2,672 records ▼ Deduplication + period restriction (2000–2025) │ 2,091 records ▼ Rule-based LLM screening (04_scripts/amoc_prisma_screener.py) │ Include / Exclude / Maybe ▼ Author adjudication of all "Maybe" records (full record) │ ▼ Inductive keyword pass (04_scripts/pass2_abstracts.py) │ ▼ Full-text reading + coding by author + DOI/bibkey resolution │ (04_scripts/build_final_corpus.py) ▼ Final corpus: 246 studies (03_corpus/AMOC_corpus_FINAL.csv) Folder structure 01_search/ Raw material of the literature search. SEARCH_QUERIES.txt — the four full Boolean search strings (master + three thematic queries) run on the Web of Science Core Collection on 15 January 2026, covering publications through 31 December 2025. (Fill in with the exact strings used.) Requête_Complete.csv — the raw records retrieved and retained after deduplication and restriction to 2000–2025 (2,091 records), as exported from Web of Science. DOI.txt — the list of DOIs of the studies carried through to the bibliography. 02_screening/ Screening records (the "screening log"). AMOC_screening_full.csv — all 2,091 screened records with the screening decision (Include / Exclude / Maybe) and the preliminary coded fields assigned by the rule-based LLM screener. AMOC_screening_INCLUDE.csv — records classified Include. AMOC_screening_MAYBE.csv — records classified Maybe; every one of these uncertain cases was subsequently adjudicated by the author on the full record. AMOC_pass2_keywords.csv — fine-grained, inductively extracted impact keywords per included abstract (second LLM pass). AMOC_pass2_clusters.txt — keyword frequencies and clustering used to organise the sectoral narrative. 03_corpus/ The final, frozen corpus and its statistics. AMOC_corpus_FINAL.csv — the 246 included studies, one row per DOI, with the full coding: bibliographic key, DOI, first author, year, AMOC condition (C1–C5), driving model, forcing magnitude, scenario, impact domain(s), region(s), consequences, and key idea. This is the core dataset. corpus_stats.json — all official counts (by condition, domain, region, period) reported in the manuscript. mapping_report.txt — the reproducibility report: exclusions with their categories, and the four author-identified studies added manually. 04_scripts/ Reproducible code. amoc_prisma_screener.py — Pass 1: the rule-based LLM screener. The system prompt embedded in this file is the coding dictionary: it defines the inclusion/exclusion rules, the AMOC-condition categories, the seven impact domains, their sub-impacts, the modelling-framework and region vocabularies, and the strict rules (e.g. palaeoclimate-only → exclude; ambiguous → Maybe). pass2_abstracts.py — Pass 2: the inductive keyword extraction over included abstracts. build_final_corpus.py — deterministic reconstruction of the final 246-study corpus from the screening output and the DOI→bibkey mapping, applying the documented exclusions and de-duplication. build_bibtex_from_dois.py — rebuilds a clean BibTeX file from the DOI list via Crossref. 05_bibliography/ AMOC_references_FINAL.bib — the reference database for the 246 studies plus background citations. doi_to_key.json — mapping from DOI to citation key. AMOC representation categories (C1–C5) Code Definition C1 Idealised freshwater hosing or imposed shutdown C2 Scenario-emergent weakening without explicit meltwater forcing C3 Abrupt emergent collapse (no idealised hosing) C4 Explicit ice-sheet meltwater forcing C5 No self-contained AMOC simulation (C5a: quantitative impact/IAM on an externally prescribed AMOC trajectory; C5b: conceptual, review, perspective, adaptation framework) Impact domains: base climate, climate extremes, ocean biogeochemistry, biodiversity, agriculture and food security, human health, society. Method notes and limitations Screening used a documented, rule-based large-language-model screener (Claude Sonnet 4, claude-sonnet-4-20250514) called through the Anthropic API against the fixed system prompt in amoc_prisma_screener.py. A second inductive pass used claude-sonnet-4-6. All records the screener flagged Maybe were adjudicated by the author on the full record, and every study retained for full-text assessment was read in full by the author, who verified the coding and checked every quantitative value against the source. As a single-author evidence map, no inter-rater reliability statistic (e.g. Cohen's κ) was computed. The full screening and coding record is released here so that reliability can be independently reassessed. Four studies meeting the inclusion criteria were known to the author and added manually because the automated queries did not retrieve them; all four apply explicit ice-sheet meltwater forcing (C4). A sensitivity analysis (reported in the manuscript) shows that excluding them leaves the qualitative conclusions unchanged. Licence Data and code released under CC-BY 4.0 / MIT (as applicable). If you use this material, please cite the associated manuscript and this repository.