Convergent assessment cannot be policed: capability corpus, modality characterisation and analysis code

Data and code supporting a residual-entropy bound on authorship attribution in assessment, and its consequences for academic integrity in quantitative disciplines. Contents: a capability corpus of 19 published measurements of generative AI attainment across 13 disciplines, each with a provenance note naming its primary source; a characterisation of 10 assessment modalities by residual entropy, the entropy of a response conditional on being correct; the three-regime classification (exposed, latent, detectable) derived by crossing the two; all analysis code; the generated figures and tables; and 37 unit tests. Associated with a manuscript submitted to Assessment & Evaluation in Higher Education.