What's new
v1.5.1 · 27 September 2026A clean-up release, in two parts. The archive had picked up the same broadcasts more than once under different names, so the transcripts were compared by content, the duplicates were removed, and the whole index was rebuilt from what was left. That left dozens of streams with no working video link, and the easy read was that they were all deleted — most weren't, so every one was checked by hand. Nothing new was added to the tool itself. Figures were measured on 27 September 2026.
46 duplicates removed since the last release.
3.2% fewer, so search has less repeated text to wade through.
Checked against the video's own duration, not just matched by name.
Most of the drop is streams that were never really missing.
One file per broadcast, and a link that's actually checked
With every broadcast in the archive once, a question is answered from one copy of the evidence instead of two, and a citation cannot point to a re-upload with the wrong date. And treating every unlinked stream as deleted would have quietly written off dozens that are actually fine: 622 of 660 links are now confirmed by checking the transcript length against the video's.
Compared by what was said
Compare what was said, not what the file is called
Many streams had been re-downloaded from other channels, so their file names carried the wrong date and time. The scanner ignores names and turns each transcript into a set of five-word phrases, then looks for pairs of streams that share a large part of them.
Catch cut-down re-uploads
The score is the share of the smaller stream that appears in the larger one, not the overlap of both. A trimmed or partial re-upload of a longer stream still scores high, where an ordinary overlap score would miss it.
Ignore stock phrases
Phrases that turn up in a large share of all streams (“thank you very much for the”) are left out, so two unrelated streams are not linked just because the host says the same things every week.
A person decides
The scan only reads files. It writes a review page with the two titles side by side, how much they overlap, their lengths and an Open in YouTube button for each, and nothing is deleted until someone has looked.
A fresh index
After the deletions the search index was thrown away and built again from the transcripts on disk, so nothing from a removed stream is left in it. The summaries already written were carried across for every stream whose transcript had not changed, and the few streams without one were summarised again.
Four checks before calling anything missing
Don't trust a missing link
A rollout on YouTube's side (SABR streaming) can make the scanning tool report a completely normal, public video as unplayable — and when that happens the video simply never gets written down, so it looks exactly like a deleted stream. 24 “missing” videos turned out to just be a scanning failure, confirmed by opening them directly and watching them play.
Check the date against a second source
A few streams were filed under the wrong day entirely — a transcript dated the 29th when the real broadcast, and its video, was the 30th. The archive's Plex export was used as an independent check: every transcript's date was compared against Plex's, and every mismatch found was corrected.
Open it and look
For anything the tools still disagreed on, the video was opened in a browser rather than trusted to a one-line error. That's the only way to tell a real copyright block, which names the claimant on the page, apart from a scanning failure, which shows nothing wrong at all because nothing is.
Only call it gone once there's nowhere left to look
Every stream here is Unlisted, so the channel scan is the only way any of them can be found — they don't turn up in search or on the channel page. A stream is only treated as truly missing once it's absent from a scan that accounts for the rollout above, and its transcript has been checked against the rest of the archive for a hidden duplicate under another name.
Most of the gap wasn't a real gap
24 links recovered from a scanning bug
Live, public videos a YouTube-side rollout was making yt-dlp report as unplayable. Confirmed by opening each one directly.
4 transcripts filed under the wrong date
Found by cross-checking every transcript's date against the archive's own Plex library export, an independent record of the same files. One example: a transcript dated the 29th, when Plex's copy of the same broadcast — and its video — were both dated the 30th.
12 streams confirmed genuinely blocked
A real copyright claim, checked by opening the video and reading the claimant's name off the page — not guessed from a scanning error.
2 stale “deleted” tags corrected
The archive's own library tags said a couple of streams were gone when the video actually still played fine. Fixed once confirmed.
Limits worth knowing about
No new features
Search, Ask, summaries and characters work exactly as they did in the last release. What changed is the archive underneath them, and how many streams link to a real video.
85 streams still have no video link
Most of the drop from 118 is streams that were never really missing. The confirmed-blocked ones account for some of what's left; the rest are absent everywhere this pass checked, including a cross-check against an independent, community-kept list of Francis's streams, which turned up a handful more worth a closer look.
This check can still be wrong
It relies on the YouTube channel scan being complete. If that scan misses a video for a new reason, a stream can look deleted when it isn't — which is exactly what happened at the start of this pass, until it was caught.