Nexus — Evidence-based mapping of CORDIS projects to SDGs

Nexus: explore the results, inspect the evidence, reproduce the analysis

Explore Nexus

Open the working prototype · Follow one project’s evidence

Nexus makes CORDIS-to-SDG mappings explorable and inspectable. Interactive maps, goal and project views, association browsing and contextual AI questions help readers connect the research to its supporting evidence.

Approach

For each project, a local open-weight model first records the project’s own activities. Candidate goals and targets are then assessed against 186 written rubrics. Code checks quotations against the source objective and derives a recorded assessment state. Each association exposes its evidence and the criteria behind it.

Results

The publication covers 400 projects across Horizon Europe, Horizon 2020 and FP7. It records 915 supported associations across 307 projects, plus 154 open associations: 83 uncertain and 71 with insufficient evidence. Goal and target assignments may coexist; project counts deduplicate within each goal. Pending assessments remain visible.

Validation and limits

On 20 untouched projects, agreement with a blind  reference was 97% same-goal precision (33/34), 79% exact-concept precision (27/34) and 43% all-goal recall (23/54). Development batches A and B are reported separately. A further review of 80 published links found 61 correct, 16 incorrect and 3 uncertain before author adjudication.

The notebook reports confidence intervals and sampling limitations. Coverage and SDG interpretation remain the main weaknesses; evidence comes from project objectives, and the evaluation is not independent human expert certification.

Review the submission

1. Start with the summary PDF for the approach and findings, then the description PDF for the solution’s design.
2. Inspect the Excel workbook for mapping tables, confidence levels and evidence.
3. Open the executed methodology notebook for the analysis, seven figures, validation and limitations.
4. Use the evaluator guide (README) to rerun all 12 code cells from the attached data without network access or model calls.

The attached files reproduce the published analysis. Replaying the original classification decisions additionally requires the full trace kit, available to evaluators on request.

Further development

Time was the main constraint. Nexus was developed in a short window before the 30 September deadline. This is a starting point, and I would value the opportunity to develop it further for larger-scale use.