Abstract
The world has committed to placing 30% of degraded ecosystems under “effective restoration” by 2030 (Kunming-Montreal Global Biodiversity Framework, Target 2). But no one agrees on what “degraded” or “effective” mean, and the knowledge needed to act is scattered across peer-reviewed studies, practitioner reports, Indigenous and local knowledge, proprietary databases and policy documents.
For three years, EcoWeaver, an international collaboration, has been building open tools that connect evidence, practice and policy. At its core, EcoWeaver is a knowledge graph that turns the causal claims made in prose (that grazing suppresses shrubs, that a seed mix increases pollinator visits) into linked, sourced statements a computer can compare, combine and feed into decision support models like Bayesian belief networks, fuzzy cognitive maps, evidence gap maps, and more.
The knowledge graph is populated using LLM-assisted annotation of evidence (currently peer-reviewed journal articles; practitioner reports and other sources are next). We deliberately use small, openly licensed models running on public research hardware (Digital Research Alliance of Canada) rather than commercial systems. That keeps data under institutional and community control, keeps the environmental cost down, and makes every machine-made annotation traceable.
Each step raises questions we can’t answer alone. What is lost when a hedged, context-bound finding becomes a machine-readable assertion? Can a data schema represent a contested concept without collapsing into agreement? How should studies, practitioner experience, and Indigenous and local knowledge sit side by side when each has different epistemic foundations? And what does “validated” mean when the task is interpretation rather than extraction?
I’ll introduce the EcoWeaver pipeline and its knowledge graph, show where design choices have become philosophical choices, and make the case that philosophers of science, causation, modeling, and epistemic justice belong on the team building these tools.