Behrens–Fisher programme: simulation data and code for Paper 1 (directional failure of small-sample intervals for a difference of means), version 2.0

Data, code and documentation for the paper "Which way do small-sample confidence intervals for a difference of means fail? Coverage, directional error and the signed skewness balance" (J. G. L. Verheyden). Five simulation stages are archived separately and never pooled:

• Stage 1: baseline, 256 settings, pooled t vs Welch;
• Stage 2: robustness, 384 settings;
• Stage 3: intervals, 120 settings, B = 10,000, R = 1,999;
• Stage 4: paired resampling sensitivity, R = 1,999 vs 9,999;
• Stage 5: confirmatory, 147 settings, 10,000 datasets and 1,999 resamples per setting, run under a predeclared plan.

Stages 1–3 are exploratory and Stage 4 is a sensitivity analysis. The Stage 5 pilot (B = 500) is included only as a code check and is not evidence.

Also included:

• all Colab notebooks (01–05);
• the shared simulation core bfcore.py, used by later papers in the programme, with a test that reproduces Notebook 05 counts exactly;
• an exploratory check of the first-order directional law outside the study grid;
• an automated check that every number in the manuscript traces to a result file;
• the predeclared plan, results note, manuscript version 2 and SHA-256 checksums.

The estimand throughout is δ = μ1 − μ2 (arithmetic means).