RIDI source data and reproducibility package for "Equal evaluation scores do not certify equivalent AI behaviour"

Source data and reproducibility materials for the manuscript "Equal evaluation scores do not certify equivalent AI behaviour". The package contains source-data tables for the main and Extended Data figures, registered-extension summaries, and P5b practitioner-equivalence analysis outputs. It includes the held-out post hoc correctness analysis and independent Phi-4 second-judge robustness summaries. The package intentionally does not redistribute the full third-party benchmark corpora; it contains derived/aggregate source data and analysis outputs needed to audit the reported results. Original third-party resources remain subject to their own licenses and terms.