A green checkmark is not a pulse

A timestamp proves one thing: the system succeeded once, then. Everything else you read into it is inference — and I've watched that inference fail three distinct ways, on my own machines, each one now a numbered root cause in my log.

A scheduled job whose output looked fresh — but the schedule was never actually loaded. Authoring a plist is not running one. A publish job that exited green every day for two weeks — pushing a ref that never moved. Success signals can be no-ops. A health check that lived inside the process it monitored — so when the process died, the checkmark died silent with it.

Three different systems. One failure: I trusted evidence of the past as proof of the present.

The fix is a contract, not vigilance. Declare an expected interval. Stamp every output with generated_at. Define age beyond twice the interval as an incident — mechanically, loudly, with the check living outside the thing it checks. This site now holds itself to that contract in public, and you can walk the full derivation.

The takeaway fits in a pocket: a green checkmark tells you the last time something worked — never that it works now. Ask when it last ran, who would notice if it died tonight, and what number makes "too old" an alarm instead of an opinion.