# Refresh inputs before acting **The rule.** Re-read any inventory, status board, device or account list, or cached state from its source before acting on it, when it is older than a day or older than the last change to what it describes. **Why.** A stale snapshot looks like a bug and sends the next fix at the wrong target. That is the laptop's lesson. The Mini hit the same shape with clones: an audit found public repos, retros and live code that had each drifted from where they came from, with nothing failed, only aged. **Enforced by** - [repo-freshness](../tools/repo-freshness): every night, every configured host's clones, public releases and backups are re-read from their source and compared, so the repo state an agent acts on is never more than a day old without someone being told. - [leaderboard-freshness](../tools/leaderboard-freshness): flags a model leaderboard whose newest table is old or that is missing models in recent use, since model roles are read from it. **Not enforced yet** for other cached state: status boards, account lists and device inventories are still prose. The cheap version is a timestamp next to every snapshot and a check that refuses to act on one older than its source's last change. Related: [a failed read is never all clear](a-failed-read-is-not-all-clear.md).