The model grades itself every morning at six

The fraud model ships Tuesday. By Friday you'd like to know whether it's still any good.

scoreboard.ts starts with // @schedule 0 6 * * *. At 06:00 UTC the exec service fires it. The script pulls yesterday's predictions and the actuals that arrived overnight from /hoody/databases/scoring.db, computes precision and recall, writes one row to a daily_eval table. If precision drops more than four points from the seven-day mean, it POSTs to /api/v1/notifications/notify and your phone buzzes before you're out of bed.

Every fire lands in fires.log; GET /api/v1/exec/schedules/history?scriptPath=... shows the last 100 runs when you want to know if the 6am job actually ran on the 14th.

No Airflow DAG, no MLflow server, no GitHub Action waiting on a cron trigger. A file, a comment, a table.

Model monitoring is a cron line above a SELECT.