All 25 public figures in references/real_case_calibration were replayed, held out and opened, or annotated before, so they form the development set; a new annual holdout needs new people. Records the plan (person-level split, frozen method, windows generated without event dates, coverage and window-day metrics, comparable random baselines) and the skill-repo questions for the astrologer (attribution, template_cases marked external_verified, Tajika rule review, exact-contact windows, case list). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8