Knowledge Discovery for GA2 based on GA1

Introduction

This repository is using the Linked Open Data representation of the Global Asssessment 1 to identify new knowledge / literature produced after GA1 (2018).

Methods

We use snowballing to identify the literature. The key-papers are the publications which were used for the Backbround Mesages (BMs) as identified by the LOD representation of GA1.

Building

The work is split across four targets projects, defined in _targets.yaml. Each has its own script and its own store, so rendering a report can no longer reach a target that spends GPU time or API credit:

Project	Holds	Credentials
main	collection: LOD → refs → zotero → works → snowball → works_citing	API_openalex
factcheck	citing works → NLI (Phase 1) → LLM verification (Phase 2)	API_openrouter
training	key papers → NLI → LLM → training set → fine-tune	API_openrouter
reporting	everything that renders	none

targets::tar_make()                                    # main (the default project)

Sys.setenv(TAR_PROJECT = "factcheck"); targets::tar_make()
Sys.setenv(TAR_PROJECT = "training");  targets::tar_make()

# or address a project directly, without changing the environment:
targets::tar_make(script = "_targets_reporting.R", store = "_targets_reporting")
targets::tar_visnetwork()                              # dependency graph of the current project

Which named configuration each project uses — and which assessments it covers — comes from a purpose block in input/config.yaml (fact_checking:, training:), not from a global active: setting.

Two cautions worth knowing before a first run:

• The scoring projects need a live NLI pod for their pool health check, even for a pass that will skip every claim.
• factcheck's first run rebuilds its claim × premise cross-join from scratch.

See CLAUDE.md for build system details (credentials, system dependencies) and the full pipeline architecture, and TODO_PIPELINE_SPLIT.md for why the split is shaped this way.

Reports

• IPBES_Fact_Checker.html — the main fact-checker report (built by report_fact_checker in the reporting project: targets::tar_make(names = "report", script = "_targets_reporting.R", store = "_targets_reporting"))
• TD_targets.html — the targets pipeline architecture design doc (rendered from TD_targets.qmd)