# Nexus - evaluator guide

## Start here

1. Explore the working prototype: https://nexus.foaap.app/
2. Follow one project from its associations to its evidence: https://nexus.foaap.app/#/project/101137814/associations
3. Read Nexus-summary.pdf (one page) and Nexus-description.pdf (eight pages).
4. Inspect cordis-sdg-mapping.xlsx for the formatted mapping and supporting tables.
5. Open cordis-sdg-methodology.ipynb for the method, executed analysis, seven figures, validation and limitations.

## What each resource provides

- PDFs: approach, findings and explanation of the solution.
- SUMMARY.md and WHY-NEXUS.md: accessible text versions of the overview and rationale.
- cordis-sdg-mapping.csv: 1,069 association rows, with source evidence and exact assessment states (915 supported, 83 uncertain, 71 insufficient evidence).
- cordis-projects-summary.csv: all 400 projects, including provisional outcomes and pending assessments.
- Nine nexus-projects-*.ndjson parts plus nexus-map-metadata.json: the unchanged publication data used by the notebook.
- sdg-rubrics-2.4.8.json: the written criteria for SDG concepts.
- review-f-score.json: the separate 80-link review results.
- trace-verification.json: saved report from replaying the original decisions.
- checksums.json: original filenames, byte sizes and SHA-256 checksums for the submission resources.

## Reproduce the analysis

Download the notebook, both CSVs, all nine NDJSON parts and the four JSON inputs (map metadata, rubrics, review score and trace report) into one folder. Install Python with pandas, matplotlib and Jupyter, launch Jupyter from that folder, open the notebook and run all cells.

The portal can add suffixes such as `_0` to downloaded filenames. The notebook detects those variants and checks each data file against its embedded SHA-256; no manual renaming is required. A missing or changed file produces an explicit error naming the required input. The notebook creates a figures/ folder and makes no network requests or model calls. All 12 code cells were checked using the portal-style filenames on 23 October 2026.

The data is split solely to meet the portal's 5 MB per-file limit. The nine parts and metadata reconstruct the unchanged published map. The workbook and CSV values correspond to the live export data.

## Interpret the validation

Only batch C (20 projects) is the untouched evaluation of the frozen method. Its AI-reference results are 33/34 same-goal precision, 27/34 exact-concept precision and 23/54 all-goal recall. Batches A and B informed development. The separate review of 80 published links reported 61 correct, 16 incorrect and 3 uncertain before author adjudication. Confidence intervals, sampling limitations and adjudicated results appear in the notebook and saved scores. These are agreements with AI reviewers, not independent human expert certification.

## Verification scope and access

This notebook reproduces tables, figures and reported validation from the attached publication and saved scores. Reading trace-verification.json does not rerun the original decision verifier. That replay requires the frozen code and recorded answers in the private repository, which evaluators may request. It reproduces the recorded publication without running the classification model.

## Official sources

- Horizon Europe: https://data.europa.eu/data/datasets/cordis-eu-research-projects-under-horizon-europe-2021-2027
- Horizon 2020: https://data.europa.eu/data/datasets/cordish2020projects
- FP7: https://data.europa.eu/data/datasets/cordisfp7projects
- SDG taxonomy: https://data.europa.eu/sdg

## Submission record

Original entry received by email on 30 September 2026. The platform was reopened for registration on 23 October. This upload clarifies documentation and adapts packaging to the portal; it does not rerun classification or change the published project decisions.