The benchmark's records.

• Tasks: the 69 build tasks of the harm check, with their sources and both arms' scores.
• Runs: every campaign run, with its number of Claude Code runs, tokens and cost.
• Arms: all 107 configurations we specified for testing, and what happened to each.

The repository measurements are on the repositories page; the harness, the outside models and the faults we found are on the method page.