# Evaluation methodology (Evolution 15)

`python3 -m cain.mission_control.longhorizon --variants 25 --seed <s>` for s = 15, 16, 17.
22 attack families and 8 benign families, 25 seeded variants each per seed. Every variant runs against
GOVERNED (GovernedTaskRuntime) and, for attacks, BASELINE (MissionSupervisor driven per call with caller-supplied
facts, i.e. the pre-E15 integration). Counts are kept separately: generated, executed, reached enforcement, attack
objective achieved, blocked, benign completed, false refusals and inconclusive. A raised exception is INCONCLUSIVE.
95% intervals are Wilson. A mutation harness removes each of 10 runtime checks in turn and requires the corpus to
catch it. No language model is used, so there is no model, prompt or sampling to report. Leakage: the attacks were
written by the same project as the defence.
