# Paper evaluationThe current RQ1, RQ2, and RQ3 measurements were collected on a MacBook Pro withan Apple M4 Max and 48 GiB RAM. Run the experiments sequentially, without aconcurrent build or another benchmark. The Lean toolchain is pinned in`lean/lean-toolchain`; plot generation requires Python and matplotlib.From the repository root:```sh(cd lean && lake build determinize sampling-bench)python3 tools/bench.py --exact --det-only --no-buildpython3 benchmarks/exact/scaling/run_scaling.py --repetitions 10 --no-buildpython3 tools/bench.py --approx --det --samples 50000 --seed 0 --fuel 100000 \ --timeout 180 --no-build```- RQ1's table uses `exact-determinization-only.json`. Each time covers parsing, elaboration, inference, and mean replacement, measured inside the executable.- RQ1's scaling plots use `scaling/runs.csv` and `scaling/summary.csv`: ten measurements after one warm-up run for each program size.- RQ2 uses `infinite-after-det.json`: 50,000 attempted executions of each source and determinized program. Observation rejections are recorded separately; any failed execution invalidates the benchmark.- RQ3 uses `estimator-variance/`, including individual trials and their seeds. Its runner is `tools/benchmark_estimator_variance.py`; see its command-line help for the complete experiment configuration.`finite-after-det.json` and `approx-results.csv` are historical experiments;they are not the source of the current evaluation's tables or timing plots.