ctde_v0.run_bench¶
Reproducibility benchmark — re-run the methods-that-matter under ONE fixed protocol, many seeds.
Most prior rankings were single-seed; this nails them down. Fixed protocol for EVERY run: @32²/10 · cover_r=0 (real visited coverage, scored vs the per-map oracle optimal) · SLAM ON (sense_walls) · frontier-attn · collision-mask · 100-step horizon · SOFT/learned connectivity only (NO hard action_mask).
Varies the 3 axes that actually moved the needle — role-split, connectivity (soft vs learned- Lagrangian), and the bump-explore term — as ablations around the best, across 3 terrains, N seeds. Reports coverage_pct + connectivity_real per (method, terrain), with seed spread.
PYTHONPATH=.:../../../FiedlerValueEstimation ~/ZymeraLab/.venv/bin/python ctde_v0/run_bench.py --out runs/bench --seeds 4 --iters 2000 --jobs 1