Structured AI-Assisted Evidence Synthesis (SAES) v0.2.1-alpha · open protocol
A schema-first protocol designed to make each step of LLM evidence synthesis explicit and auditable: every coded field is paired with the verbatim text it came from, generation is kept separate from evaluation, and searches and screening decisions are logged, so that any consolidated claim can be traced back to the papers it draws on. Three stages: setup, extraction with audit, accumulation.
Run end to end on three corpora spanning medicine, agriculture and environmental impact; operated by hand and not yet benchmarked. An adversarial audit of the worked examples (September 2026) found that extraction quoted sources accurately where it could be checked, and that one consolidation had been written from the model's memory rather than the rows; the findings are recorded in the repo's KNOWN_ISSUES file. Next: benchmark against human coders, enforce provenance in code, run the audit in an independent model.
source: github.com/grahamprescott/structured-evidence-synthesis
Measurement rules that standard-setters adopted
Led development of the measurement framework of the Verra Nature Framework: a unit of biodiversity impact commensurable across every ecosystem type, calculated above a counterfactual baseline, and now Verra's published standard for nature credits. Lead author of the TNFD guidance on measuring changes to the state of nature, which corporates and financial institutions apply globally; wrote the tropical forest and intensive land-use biome guidance and project-managed the biome series. Joint first author of SBTN's biodiversity position paper.
Building a research programme in Myanmar
After Myanmar's democratic opening I built a collaboration from scratch with the Forest Department and WCS Myanmar, ran a horizon scan of forest-conservation issues with a large group of co-authors from government, NGOs and universities, and took the shortlist to a regional governor to ask which problem he most needed solved. Through fieldwork and economic modelling I calculated that the government's preferred interventions, enforcement and alternative livelihoods, could not plausibly deter informal gold mining, and redirected the policy question. source: Conservation Biology 2017, Conservation Science and Practice 2020