The numbered directories follow the released pipeline:

1. step1: render and validate model-facing prompts;
2. step2_runner: run the 11 tested models;
3. step3: build judge inputs, run judges, parse, and validate;
4. step4: aggregate three judge decisions and apply documented manual F resolution;
5. step5: validation-stage aggregate and human/judge coherence utilities;
6. step6: main RPCBench metrics and figure;
7. step7: released downstream analyses and evidence ablations.