Why every automation we run only proposes — never applies

Our AI estate runs across two machines talking to five different backends. Before security-watch existed, we hit the failure mode it was built to catch: an authentication fix was applied and verified on one machine, and quietly never made it to the second. Nobody was lying about the system's state — nobody was checking both machines identically, on a schedule. It drifted, unauthenticated, for more than 12 days before anything caught it.

Security-watch runs one behavioral probe — make the real unauthenticated call, read the real HTTP status — identically against every machine in scope, without relying on anyone remembering to look. On our most recent run, that exact class of gap showed up again on schedule (the gateway's authentication was found open), and was fixed and re-verified the same day. 8 of 10 flagged findings that run were fixed and verified live; 2 were correctly left for a human — full-disk encryption, because enabling it needs physical console access nothing the automation runs with can reach, and a home-router firewall rule, because it needs the router's own admin UI, not anything scriptable from either machine.

Neither router-watch nor security-watch has write access to the config it reports on. Every finding is a dated proposal a human reads and applies by hand, then the automation re-verifies. That durable record — a human decision attached to every change — is the actual product. An automation that silently applied its own fixes would still be making the same changes, but it would stop being evidence anyone could audit later.