BACK TO HOME PAGE All projects ↗

POLICY INTELLIGENCE · TREATY RESEARCH

International treaty
intelligence with
language models.

A research program that turns a global corpus of bilateral investment treaties into evidence: where treaty language travels, which country connections may emerge, and how a tailored LLM can help teams investigate, compare and draft with care.

Explore the program ↓● PUBLIC TREATY CORPUS

PROJECT TYPE

Treaty analytics
& policy decision support

METHODS

NLP · graph ML
Retrieval-augmented generation

DATA FOUNDATION

EDIT treaty corpus
Machine-readable IIAs

STATUS

PEER-REVIEWED
RESEARCH PROGRAM

01 / THE DECISION

Legal language has a
travel history.

Bilateral investment treaties are not isolated documents. Clauses are adapted, replicated and negotiated across countries and decades — shaping the policy space available to governments.

This program makes those patterns visible, then extends the analysis with a bespoke language-model interface that keeps the underlying treaty text in view. The aim is not an opaque answer, but a faster route from question to reviewable evidence.

02 / THE SYSTEM

Three complementary ways to understand treaty language and its possible next moves.

01

Trace

Compare clause text with NLP similarity measures to identify countries that play an influential role in language dissemination.

02

Predict

Represent countries and BITs as a network, combining topology and economic features to estimate likely future connections.

03

Retrieve & reason

Use tailored prompts and a treaty vector database so the LLM can analyse clause stance, compare precedents and support grounded drafting.

04

Review

Return evidence, citations and drafting alternatives to expert users, with human judgement remaining central to every conclusion.

03 / EVIDENCE

From treaty text to
testable signals.

The research starts with published NLP and graph-learning methods, then builds toward a more flexible, context-aware alternative to purely rule-based treaty analysis.

555

BITs analysed for health-safeguard clause dissemination

167

countries represented in the BIT network analysis

64.02%

XGBoost link-prediction accuracy in the published study

Published studies use EDIT treaty data. Reported performance applies to the study design and data; models should be revalidated before any policy or operational use.

04 / RESEARCH FOUNDATION

Two studies, then a
new interface.

The first study quantifies national influence in clause dissemination. The second predicts potential BIT formation using network and country-level signals. Together they provide the foundation for a customised LLM system that brings retrieval, stance analysis and research assistance into one evidence-grounded workflow.

01

PLOS ONE · 2024

An NLP-based novel approach for assessing national influence in clause dissemination across bilateral investment treaties

Introduced the Influence (INF) metric using text similarity to identify countries central to the diffusion of defensive, neutral and offensive health-related treaty language.

Read article ↗ 02

JOURNAL OF COMPUTATIONAL SOCIAL SCIENCE · 2025

Identifying the factors influencing the development of bilateral investment treaties with health safeguards

Combined country-network structure, node embeddings and economic indicators to predict links representing potential future BITs with health inclusion.

Read article ↗ 03

DATA FOUNDATION

Electronic Database of Investment Treaties (EDIT)

A comprehensive, machine-readable full-text corpus of international investment agreements that enables retrieval, comparison and reproducible analysis.

Explore database ↗
TREATY RESEARCH ASSISTANT / CONCEPT WORKFLOWRAG + CUSTOM PROMPTS

05 / LLM WORKBENCH

Ask better
questions of
the corpus.

A treaty-specialised LLM is designed to retrieve relevant provisions before it responds. It can support nuanced stance or sentiment analysis, compare drafting patterns, generate bounded alternatives and help researchers explore potential policy implications.

RESEARCH QUESTIONMODEGROUNDED INOUTPUT FOR REVIEW
How is health policy protected?ANALYSERetrieved clauses + custom rubricProvision stance and citations
Which precedents are closest?COMPAREVector similarity + treaty metadataReviewable treaty set
What might a new clause say?DRAFTRetrieved language + prompt guardrailsDrafting options, not legal advice
What could change in a BIT?EXPLORENetwork, text and policy contextQuestions for expert assessment
THE SYSTEM SUPPORTS RESEARCH AND POLICY WORK; IT DOES NOT REPLACE LEGAL, POLICY OR NEGOTIATION JUDGEMENT.

06 / PROJECT PRINCIPLE

Treaty analysis becomes more powerful when every answer remains traceable.
Discuss a policy AI project ↗