Reproducing

Every number in the README and paper is generated by a script from files in results/.

make install
make test
make reproduce      # simulations, re-analyses, figures (no model needed)

Real-model runs need Ollama and about 15 GB of disk. They are resumable and recorded in a record/replay cache:

OLLAMA_NUM_PARALLEL=4 ollama serve &
python scripts/run_real.py --models qwen2.5-3b llama3.2-3b --runs 10 --workers 4

Seeds are fixed everywhere; configs/models.json lists the exact model tags. Raw trajectories and record/replay caches of the local-model run are not in git and are not currently published; only the aggregate results/real_main.json is stored, so the local-model numbers cannot be recomputed from raw runs. See EXPERIMENTS_AND_REPRODUCIBILITY.md in the repository root for what was reproduced, what was only partially verified and what was not rerun.