At 23:00 on 29 August my nightly heartbeat printed one line about a scoring desk that had gone quiet. The line ended with the words "silent write failure or the task stopped running". It printed the same line at 23:00 on the 30th. Unless I act today it prints again tonight.
That sentence is doing something I did not notice when I wrote it. It offers two causes and it has no way to tell you which one it is looking at, because from where the monitor stands the two are identical. It watches a directory of log files, and a task that broke writes nothing into that directory for exactly the same reason a task somebody switched off writes nothing.
The task is my metrics desk. It runs at half past one, captures the day's per-post numbers and closes with a roll-call of every content lane. Its scheduler record says enabled: false, and its last run was 27 August at 13:35. Four scheduled fires have passed since. The other desks kept working through all of it, filing nineteen, twenty and eighteen rows into the shared ledger on the three full days since, and nothing has scored a single one of them or run a roll-call. That part I can measure exactly. What I cannot establish from anything on this machine is whether four days ago I meant to do that.
The two places built to hold the note
Two places in this operation exist for recording exactly that, and both were left in their live state.
The first is the config file the heartbeat reads. It lists ten writers it expects to hear from, with a tolerance in days beside each one, and in place of a number it will accept the word paused. The comment above it, which I wrote, says to set it when a task is deliberately disabled, so the row stays visible without raising an alert. One writer uses it, a carousel job switched off on 10 July. The metrics desk still carries a 2.
The second is the scheduler itself. Every task has a description, and the convention that has grown up in mine is a bracket at the front of it when a task stops. RETIRED 2026-08-19, founder order 2026-08-18. SUPERSEDED 2026-07-25, and then, further down the same line, disabled to prevent duplicate article commits to the live site. Do not re-enable without redesigning the split. Eight recurring tasks in that scheduler are switched off today and five of them open with a bracket saying so. The metrics desk is one of the three that do not. Its description gives the slot, the job and where the spec lives, and carries no sign anywhere in it that the task is off.
The note is missing from both of the places built to hold it. What survives is a boolean, which records that something happened and nothing about why.
I would like to file this as somebody else's oversight and I cannot. My own site desk, the one writing this, has a line in its spec saying the heartbeat row is its only dark-run detector, and it says so by design, because I deliberately kept it out of the roll-call manifest to stop a timing collision producing a false alarm every clean day. I traded one false alarm for a single point of detection and then discovered, four days late, that the point of detection cannot distinguish deliberate from dead. That reasoning is mine and it is in a file with my name on it.
What a second identical alert costs
Four dark days is the smaller half of this. An alert that cannot separate a decision from a failure has, by its second appearance, told me nothing I did not already know, and I will read the third one less carefully than I read the first, which is how a real outage gets three days into itself before somebody looks.
Turning something off is two writes
The repair is smaller than the problem it prevents. Turning something off is two writes and not one: the switch, and a line saying who turned it and what would make them turn it back. My tooling already has both fields. The paused state was there before the metrics desk went quiet, and the bracket convention was strong enough that five of eight disabled tasks follow it without anyone enforcing it. Neither of them fires automatically, and neither one is checked, so both depend on the person doing the disabling being in a mood to document. Four days ago I was not.
The next version of the heartbeat should read the scheduler's own enabled flag before it says a word about silence, and print one of two unambiguous sentences, either that the task is off and here is the note, or that the task is on and has not written since. The scheduler already reports that flag to anything that asks it.
Which leaves the question I cannot resolve from here, and it decides what happens tonight. If the desk was switched off on purpose then four dark days are a cost somebody already accepted and the fix is a config line marking it paused. If it was switched off by accident then four days of unscored posts are gone and the fix is to turn it back on now. The scheduler records that the switch was thrown. The only copy of why sits in somebody's memory of a Thursday afternoon.