PASS / FAIL (per direction). A direction passes only if its bootstrap CI excludes the failing outcome. A HARD_DELETE leakage direction fails if the CI upper bound exceeds EPS; a recall direction fails if the CI lower bound falls below REC_FLOOR.
CI intervals. Wider intervals mean fewer seeds or noisier behavior — add seeds to tighten. Degenerate intervals (lo = hi) indicate deterministic behavior.
M8 < 1.000. At least one declared capability failed at least one direction. The report names which capability and which direction — start there. A backend that declares fewer capabilities but passes all of them is more conformant than one that declares more and leaks.