Health
Unknown is not healthy.
A recorder that silently stops recording is worse than no recorder, because you keep making decisions as though the record exists. The failure is never dramatic — no error, no alert, just a lane that quietly captures nothing while the dashboard stays green.
Three verdicts, and one of them is an admission
fresh
A proof ran, recently, and it passed. The floor holds everything the origin holds. This verdict names the check that produced it and when it ran.
stale
A proof ran and it failed. The origin has content the floor does not. This is the good failure — it is loud, specific, and tells you which lane and how far behind.
unknown
No proof ran. Nothing is claimed. This lane is not being watched, or the check itself failed to execute — and either way there is no evidence about its state. It is reported as its own verdict and never folded into a healthy count.
The whole design rests on refusing to collapse the third one into the first.
Most health dashboards have two states, so an unwatched lane and a verified lane render identically. That single collapse is how a recorder lies to you without anyone writing a lie: absence of a signal gets promoted to a verdict about the world.
Staleness is a gap, not a clock
The obvious way to measure staleness is wall-clock time: if nothing arrived in an hour, call it stale. That check is wrong in both directions. It fires all weekend when you simply were not working, and it stays silent for a lane that broke the moment you stopped using that agent.
So the comparison is between the origin file and the floor. If the origin transcript holds 400 lines and the floor holds 400, the lane is fresh — even if that last happened four days ago. If the origin holds 400 and the floor holds 180, the lane is stale, and it is stale by 220 lines, right now.
A lane that captured nothing today because you did nothing today is healthy. A lane that captured nothing today because its hook was uninstalled is not. Only the gap tells those apart.
Heartbeats are side effects, never self-reports
A component that writes “I am fine” to a log proves only that it can still write to a log. The health directory records mtimes as a side effect of work actually completing, so the timestamp cannot be produced by a process that is failing to do the job. A description of health is not health.
What the output looks like
A real run — the release-day check against the author's backfilled floor, committed by the stamp script. Note the summary line refuses to add any unknown lane to the healthy count.
The stale lane is the part worth staring at. This check ran on release day, minutes after a full backfill — and the Claude Code lane was already ~68 minutes behind, because the author kept working while the backfill ran. The tool's first verdict about its own demo was that the demo had fallen behind reality. A health report that can say that about itself is one you can believe when it says fresh.
A check that cannot fail proves nothing
Before trusting any health output, make it go red on purpose once. Stop the watcher and confirm the lane turns unknown. Have a conversation without the hook installed and confirm the lane turns stale with a real line count behind it.
If you cannot make it fail, you have not learned that it works — you have only learned that it is capable of printing green.