Empirical analysis: methodological alignment in EU-funded artificial intelligence research

Version 2.1.0 adds the results annex (separata.pdf): a self-contained, skimmable PDF presenting the results of the audit (the Marcus-criterion classification distribution, its breakdowns by research type, disciplinarity and year, and the reproduction instructions), extracted from the parent manuscript Hybrid AI or NLP-as-usual? Auditing Methodological Commitments in Publicly-Funded Research. The annex names the analysed project; the interpretation, the deliverable-level mechanism analysis and the procedural corrective are carried by the parent manuscript. This record is part of an empirical analysis of methodological alignment in EU-funded artificial intelligence research. The analysis evaluates whether the methodologies actually deployed in the research output that a publicly-funded artificial-intelligence research project produces align with the methodological commitments the project declared at its proposal level. The alignment evaluated in this record is the alignment between the methodological architecture of the publication corpus produced by the project and a reconstructed methodological criterion — the Marcus-tradition criterion of hybrid AI, which specifies architectural integration of symbolic and statistical components such that the statistical layer's outputs are constrained by symbolic structure (or vice versa), the architecture closes the explainability deficit pure-statistical baselines exhibit, and the system demonstrates compositional or causal reasoning beyond pattern-matching. The criterion is reconstructed from the canonical neurosymbolic literature (Marcus 2018, 2020; Garcez & Lamb 2023; Hitzler & Sarker 2022; Sarker et al. 2021; Hamilton et al. 2022) and operationalised across four axes — symbolic_component, integration_pattern, output_type, inspectability — that compose the aggregate marcus_score. In order to ensure the analysis stays at institutional level (avoiding personal attribution to researchers), individual publications are referenced only by an opaque identifier (pub_uid, P001–P075). Bibliographic information (titles, author names, venue names, DOIs) is not in any file in this record. Contents: scripts/marcus-analysis.R (R script; ggplot2-based; consumes the v1.0.0 classification table to produce rendered figures and the summary statistics), figures/fig-marcus-{distribution, by-research-type, by-disciplinarity, by-year}.{png,pdf} (four rendered plots), tables/headline-stats.md (top-line counts), summary-stats.md (long-form summary statistics: per-axis cross-tabulations + narrative reading). Reproducibility: with the v1.0.0 record's derived-tables/ folder downloaded into the v2.0.0 record's root, running Rscript scripts/marcus-analysis.R regenerates all rendered outputs. The R script uses only ggplot2 + scales and requires no network access. The classification protocol — the four axes (symbolic_component, integration_pattern, output_type, inspectability) composing the aggregate marcus_score — is documented in the companion v1.0.0 record (see relatedIdentifiers).