Paired-marking dataset and analysis code: UK and German grading of identical scripts

This deposit contains the anonymised paired-marking dataset and the Python analysis code supporting the manuscript "When a 75% feels low: compensation, thresholds and grading practice in UK and German higher education" (submitted to Assessment & Evaluation in Higher Education). Twenty-eight third-year undergraduate scripts from a single Behavioural Finance module (Edinburgh Napier University, academic year 2014/15) were marked independently and blindly by two experienced lecturers: the UK module leader at ENU and a German professor of finance at the partner Fachhochschule (Frankfurt University of Applied Sciences). Each marker applied the assessment culture, grade descriptors and threshold conventions of their own national system. Section A of the examination (multiple-choice, marked mechanically) is excluded from the deposit; only the two subjectively-marked components — Section B (short questions) and Section C (essay) — together with the constructed final mark are included. The dataset reveals a systematic pattern that is the central empirical contribution of the article: UK and German markers diverge substantially on the essay component (mean UK 52.05%, mean German 73.66%; Cohen's d = 0.93, large effect) but produce essentially identical short-question marks (d = 0.06). Despite the mean-level divergence on essays, paired Pearson correlations are uniformly strong across components (r = 0.690 for essays, 0.747 for short questions, 0.761 for the final mark), indicating that the two markers agree on the ranking of students even where they disagree about absolute level. A non-intercept regression yields an essay-conversion factor of 0.674 and a short-question factor of 0.968. The deposit comprises three files: paired_marking_dataset.csv (28 paired observations across eight variables), analysis.py (a Python script reproducing every numerical value in Section 5 of the manuscript), and README.md (explaining the variables, the consent and ethics provisions, the licence, and how to reproduce the analysis). Both the dataset and the code are released under a Creative Commons Attribution 4.0 International licence.