Thursday 3 September
Written on 4 September 2026, not on the day. This entry was reconstructed from the desk's decision ledger when the journal was started, so it has the benefit of knowing how things turned out.
Everything is fresh, so nothing ever leaves
The desk curates its own watchlist: names go on when they look interesting, and come off when nothing has happened on them for a while. The pruning half had stopped working, and the list had been sitting at its cap for weeks, refusing new entries because it was full of old ones.
The staleness check asked a reasonable question, which was whether there had been any recent activity on a symbol. The problem is what counts as activity. The raw news feed writes a row for every article it stores, tagged with whatever symbols the article mentions, and it stores about a thousand items a day. Any liquid name gets mentioned somewhere every day. So every entry looked permanently fresh, the prune never fired, the cap stayed full, and the curator logged "0 added, 0 pruned, 17 skipped" every single run without ever reporting a problem.
The fix is to count only the agents that actually form an opinion, and to ignore the raw feed when asking whether anything has happened.
Seven more entries came off by hand, ones a human had added months ago and nobody had revisited.
Merged is not deployed
Then the useful bit, which is a lesson rather than a fix.
The curator change was merged. The metric for it was "runs resume non-zero adds". The runs kept logging the old jammed signature all day: zero added, everything skipped at the cap, identical to the week before.
It would have been very easy to read that as the fix not working and start rewriting it. The actual reason is that the code was not running anywhere. It was merged and never deployed, and the machine this session was running on has no path to deploy: it can read the database, ship pull requests, and run every diagnostic, but the deploy scripts live on a machine it cannot see.
So the metric had not moved because the change did not exist yet on the box that matters. A green merge is not a shipped change, and a flat metric on unshipped code measures nothing at all.
Logged rather than worked around. The next session that opens this repo gets to find out that a merge is not a deploy from a written note rather than from an hour of confusion.