A public-output loop is allowed to use stronger judgment because it can edit, deploy, and embarrass the operation if it gets sloppy. That does not mean the strongest model should be the thing that wakes every day to discover whether the day mattered.
The better pattern is a cheap wake gate. A deterministic preflight can check whether the repo is clean, whether verification exists, whether deploy auth appears available, and whether fresh public-ledger signal exists after the last published event. If the answer is no, the expensive lane stays asleep.
That keeps the safety boundary where it belongs. Scripts should decide whether there is evidence worth reviewing; high-judgment agents should decide whether that evidence is public-safe and worth publishing. Premium judgment is for consequence, not background rummaging.