Splitting

Deep-tail LER by multilevel splitting

4/4 passed in 16682ms

Test	What it does	Result
test_bad_ladder_rejected	A ladder that is not strictly descending, or has fewer than two rates, raises ValueError, as does a top rate with no failures in top_shots.	✅ pass
test_deep_tail	At p = 0.001 the estimate is plausible and stable over seeds while its decode budget would show direct MC under one failure, and it sits more than 8x below the p = 0.005 estimate (LER ~ p^2 at d = 3).	✅ pass
test_deterministic_for_fixed_seed	A fixed seed yields the same LER twice.	✅ pass
test_unbiased_vs_direct_mc	At an MC-feasible rate the splitting estimate is unbiased and unstuck: over seeds 1-4 the median is within 8% of a 2M-shot direct MC with the same decoder, the spread is under 8%, and seed 1 lands in the MC Wilson interval.	✅ pass