Sampling

run -sample 0.2 trades accuracy for time: it runs a random ~20% of the mutants, selected by a hash of their ID and -seed, and reports the rest SKIPPED, so they count towards no score and the score is computed over the sample (Gopinath et al., ICSE 2015/2016, TSE 2017: operator selection and stratified sampling are at best marginally better than uniform random sampling, and a ~10% sample loses under a point of score). totals.sampled records the fraction actually kept.

Mutant IDs are content hashes, and the hash covers the seed, so the selection needs no bookkeeping: with the same -seed, every shard of a run and every -previous run keeps the same mutants. minimize takes the kills as they are — a kill of a kept mutant is still a kill — but the requirements the sample left out may have changed the verdicts, so a redundant test is weaker under sampling.