# Why Nexus should be awarded

**Every link is a claim you can check.** Nexus does not output a label and a model-written reason. Each assignment shows the exact sentence from the CORDIS objective that states the project's own activity, the SDG concept it supports, the criteria it met, and the version of the method that decided it. A reviewer can verify any row in seconds in the prototype or the table.

**It maps what projects do, not what they mention.** The activities of each project are written down before any SDG is shown, and background, earlier work and conditional later phases can never support a link. This is the organisers' first quality criterion — *not superficial keyword matching* — built into the method rather than promised.

**It says "no" when "no" is the honest answer.** Projects with no supported or open link among assessed candidates are reported as *no justified mapping*, with pending candidates disclosed (Q&A 19); pairs that could not be processed are reported separately and never counted against a goal.

**It uses the taxonomy the way it is designed.** Each assessed goal or target is recorded separately; a project may have both a parent-goal and a target assignment. Project counts deduplicate projects within each goal; assignment counts include every recorded concept. Target-level assignments are made only when the text supports them (Q&A 6).

**Its quality is measured, not asserted.** The reference contains 60 projects drawn with fixed seeds. Batches A and B (40 projects) informed development; only batch C (20 projects), drawn after the method was frozen, is the untouched held-out evaluation. A second, independent AI rater annotated it blind — without our outputs or our rules — across all 17 goals; on the main batch of 30 projects the author adjudicated every disagreement, with each ruling stored as an immutable, attributed event. We report precision, recall, their confidence intervals and Cohen's κ, as the organisers prefer (Q&A 15), for all three reference batches, distinguishing development results from the untouched held-out estimate. When the measurement showed an idea did not help, we dropped it; when it showed a recurring error (partnership and research-capacity leaps), we fixed the rules and re-assessed every affected link.

**It is reproducible.** The model is a local, version-pinned open-weight model run at temperature 0, which the organisers identify as the reproducible choice (Q&A 5, 12). Every run freezes its scripts, inputs and model file hashes; every assignment records its method and rubric version.

**It is built to be reused.** Outputs are plain CSV and Excel plus a per-project table; the method is a set of documented, versioned rules and scripts; the prototype at **https://nexus.foaap.app** explains each project's links to non-technical readers and exposes the full provenance to technical ones.

**It is honest about its limits.** Coverage of goals that the candidate step never proposes is the main weakness, and we quantify it instead of hiding it; confidence levels and the published uncertainty chart show exactly where the mapping is less certain.