Is every feed on time?
A feed is late when its newest run is more than 12 hours past when the next one was due, and failing past 24. NOAA's twice-daily learned-model runs are late past 24 hours and failing past 36; their GFS-started runs are listed separately below when they lag. The same document is published for machines at /data/status.json.
How this is measured
Every fifteen minutes desk/board.py reads each model's latest.json
and catalog.json from the network site that renders it, the same files that
site's explorer reads. The upstream sites write frames first, then the catalog, then
latest.json, so a run the Board can see is a run whose maps exist.
"Horizon" is the furthest forecast hour the catalog lists for 2 m temperature, against the hour a finished run of that model reaches from that cycle. HRES and the ensemble stop at 144 h on their 06 and 18Z runs, and that is counted as complete, not short.