What it does: When a rule starts applying, a watch records what the previous week measured, waits out a window, and compares. If LCP or INP is worse by more than the threshold it reverts the rule — through the same revert_rule() the operator’s own button uses — writes the reason auto-watch onto the rollback event, and e-mails the owner what was applied, what the numbers were, that it has been undone and how to put it back. How to access: VelocityForge > Assets, in the watch panel (the Pro add-on must be installed). Notes:

• Defaults: a 7-day baseline, a 24-hour window (settable 6–72), a 10 % regression threshold, and a sample floor of 30 measurements. The verdicts are kept, reverted and inconclusive.
• Six refusals, and they matter more than the feature: no licence, no beacon (nothing would ever land in the window), fewer measurements than the floor, Safe Mode on (somebody is already standing at a broken site), a protected handle or a rule the operator pinned, and one open watch per rule with one verdict per watch — the verdict is captured by an UPDATE whose WHERE still names the expected status, so a retried job and the sweeper cannot revert twice or send a second e-mail.
• The baseline and the result are stored, not recomputed: retention prunes page_profiles, and a verdict that recalculated its own baseline six weeks later would be reading a different set of rows.
• A sweeper on the queue finishes watches that lost their job on a host without wp-cron — a stuck watch counting down “closes in −312 h” is worse than any verdict.
• The free plugin carries the table (wp_velocityforge_rule_watches) and the two announcements (velocityforge/rule_created, velocityforge/rule_enabled) and nothing else: every decision is in the add-on. Disabling a rule announces nothing on purpose, or each automatic revert would open the window its own verdict just closed.
• Nothing watches the AI advisor’s plan, because there is no button that applies one.