Cadence

1. Read every <item>, then drop trailers and bonus episodes, items with an unreadable pubDate, and anything dated more than two days in the future.
2. Sort ascending, then collapse episodes sharing a calendar day into one release event — otherwise a migrated back catalogue injects dozens of fake zero-day gaps.
3. Diff consecutive releases into a list of gaps in fractional days. Every statistic below comes from that list.
4. Take the median, never the mean. One long break moves a mean by weeks and a median by nothing.
5. Set the hiatus threshold from the show's own rhythm: max(3 × median, median + 4 × 1.4826 × MAD, 28 days). The multiple scales from daily to monthly shows, the MAD term protects genuinely irregular shows, and the 28-day floor stops a late weekly episode counting as a break.
6. Compute the median over four windows — the last 730 days, every release the feed still carries, and the 730 and 365 days ending at the last episode. The last-episode windows show the rhythm a show had while it was active, even if it has since gone quiet; when a show is younger than the window, it starts at the first episode and is flagged. Breaks stay in the data unless you exclude them; they are listed under each strip either way.
7. Compare days since the last episode against the expected gap: within 1.5× is on schedule, 3× is late, past the hiatus threshold is a break, and beyond that is dormant.

Two caveats the numbers cannot fix. Most hosts truncate the feed to the newest 50–300 items, so an all-time figure describes only what is still published — flagged on the card when the item count looks capped. And when a show has changed pace, the all-time median is pulled toward whichever era published more episodes, which is always the busier one. That is what the two-year window is for: when it disagrees with all-time, trust the two-year figure and read the difference as drift.