Role-Play Bench reproduction audit bundle

The evidence behind Sections 3.9, 3.10 and 4 and Appendix B of the Role-Play Bench reproduction report: the audit scripts, their results, and the frozen scores from which the reliability analysis can be recomputed without access to the platform. Nothing here contains a dialogue or a judge's rationale; scores, run configurations (without endpoint addresses) and session identifiers only. Times are UTC.

Check without running anything

shasum -a 256 -c SHA256SUMS      # in a copy of this directory

Recompute Section 4.2–4.4 from the frozen scores

mkdir -p work/scripts work/out && ln -s "$PWD/frozen/raw" work/raw
for f in common.py reliability.py; do sed "s#\${RPB_AUDIT_DIR}#$PWD/work#g" code/tools/experiments/rpb_audit_20261027/scripts/$f > work/scripts/$f; done
python3 -I work/scripts/reliability.py && diff -r work/out results/   # only the reliability files are produced

Log of this recomputation: logs/01-recompute.log. The inventory and the v6.0 timeline need the production database and the Cloudflare version list, which are not public; their outputs are in results/, and the scripts that produced them in code/.

Files