Answers and explanations · Keep evidence useful as the application changes

1. Distance is 0.25: two 0.25 share changes, multiplied by one half. It measures category-mix difference, not correct outputs. It cannot detect a changed refund rule with identical input shares, a semantic failure within a category, or an unobserved new language. The 0.15 review threshold is illustrative, not a statistical significance boundary.
2. Both distributions are 70% routine and 30% boundary, so distance is zero. Volume changed tenfold; relative mix did not. Traffic volume is a separate operational signal and should not be inferred from this distance.

import opslab
print(opslab.total_variation({"routine": 7, "boundary": 3},
                             {"routine": 70, "boundary": 30}))
print(opslab.maintenance(opslab.load()["changes"], "2026-10-21"))

3. All three open items become overdue by the 21st. The closed item stays omitted. The function recalculates a returned list; it changes no file, task status, clock, reminder, or live schedule.
4. Example: fictional domain reviewer interprets the new policy; identifies affected guidance, route labels, and rubric anchors; records old and new contracts; creates a reviewed dataset/grader version; re-scores comparable artifacts or proposes authorized new observations where outputs must change. Completion evidence includes the interpretation, version note, manifest, and replay. This is proposed local planning, not permission to apply the rule to real users.

Self-check: an investigation trigger is not proof of degradation, and a due date is not execution authority. A good plan names the assumption that became stale and the evidence that would refresh it.