Holy Holiday for Weirdos in Utah: pre-registration
Status: NOT RUN. This file and the input hashes are published before anything runs. The commit time is the
timestamp. Results go in their own file, and this one stays unchanged.
The post
Instagram post DdWodVUBwWy, Sep 16, 2026, 15:46 UTC, 1:37. Started by
latterdaychad, with ldsabuseonx, theemmapreneurofficial and transcendental_mormon_mystic. It is the only
one of the 32 posts in the ldsabuse Twins network with transcendental_mormon_mystic on it: the odd join.
It is a reaction video: latterdaychad speaks at the start (0:00–0:13.8) and the end (1:25.6–1:37), and in between is
1:12 of Candace Owens, cut from somewhere else. Her part has two different settings:
| Scene |
Post time |
Looks like |
| A, close-up |
0:13.8–0:25.8 and 0:46.6–1:25.6 |
her face close, white wall, mic at the bottom |
| B, studio desk |
0:25.8–0:46.6 |
wider shot at a desk, her name on the wall, a sheriff’s office web page on screen beside her |
So the excerpt runs A, B, A. Scene times are from our scene split (dl_wm scene_split.py, threshold 50).
Our download of the post, its Whisper transcript (.whisper.json, .srt), and our cut of Candace’s part (scenes 2–4).
Predictions (John, locked 2026-10-12)
- Same show. Scenes A and B both come from one Candace Owens episode, from Sep 16, 2026. John saw it.
- Bot detection. The SRA accounts all treated that episode as total validation. Who re-posted it, when and how
is how we detect the bots. (Rules for this one are fixed in a separate pre-registration before it runs.)
Test of prediction 1 (run 1)
Candidate source: Candace Ep 388, Sep 16, 2026, “BIG NEWS! Erika To File A Wrongful Death Lawsuit. For Charlie Or
For PR?”, 1:05:36, https://youtu.be/aLJhpxrakCU. It is the only Sep 16 Candace episode our YouTube searches found.
It is downloaded after this commit, through the usual pipeline (download, watermark, Whisper), and its hash goes in the result.
Method: bin/compare_cut.py, the tool that lined up THE CUT: --full = the episode’s .whisper.json, --cut =
the post’s .whisper.json. It finds the stretches of the episode that the post keeps, word for word, with their time
offset.
Decision rule. Map each kept stretch back to post time (episode time minus its offset). Then:
- Scene A is found if those stretches cover at least 80% of scene A’s 51 s.
- Scene B is found if they cover at least 80% of scene B’s 20.8 s.
- Prediction 1 holds if both are found in this one episode.
- If only one is found, it does not hold for this episode, and the result says which scene came from where.
- If neither is found, Ep 388 is not the source. The next candidate gets its own pre-registration before it runs.
Whisper spells some words differently in two recordings of the same speech (THE CUT: 245 of 249 matched), which is
why the bar is 80% and not 100%.
Addendum, 2026-10-12, before anything ran: visual cross-check (Grok’s suggestion)
Grok, on X (reply): “a quick visual cross-check that the close-up and
desk settings both appear in that episode, to confirm the cut beyond audio alone”. Added, as a confirmation step: it
does not change the 80% rule above.
- For each scene (A and B), take 3 moments the transcript line-up matched: the first, middle and last matched word of
that scene’s longest kept stretch.
- At each moment, grab one frame from the post and one from the episode (at the post time plus the stretch’s offset) and
set them side by side on one sheet, 6 pairs in all.
- Each pair is marked same setting or not, by eye (Claude, then John). The post is a vertical crop with
captions on top, so the check is about the setting (room, wall, desk, mic, clothes), not a pixel-perfect match.
- Reported with the result: “visual check: A 3/3, B 3/3” (or what it is), with the sheet. If a scene’s audio is found but
its picture is not, the result says so plainly. That would mean the audio was laid over other footage.
Addendum 2, 2026-10-12, before anything ran: scenes from the splitter, on both sides (John)
John, on X (reply): “we agree that we use the scenes that were
generated in splitter. Candace comprised 2 scenes.” And: adjust one side or both, so it is apples to apples.
So both videos are cut by the same splitter with the same settings (dl_wm bin/scene_split.py, threshold 50,
5 frames a second), and the comparison is splitter scene to splitter scene:
- Post side (already split): Candace is splitter clips 2, 3 and 4. Clips 2 and 4 are one scene, A (the close-up),
cut in two by clip 3, scene B (the studio desk). Two scenes, as John said.
- Episode side: Ep 388’s watermarked video goes through the same splitter after it is downloaded, before the line-up.
- The coverage rule now applies per splitter clip: each of clips 2, 3 and 4 is found if the kept stretches cover at
least 80% of that clip’s time. This replaces the pooled “80% of scene A” and “80% of scene B” above, and is stricter.
- For each found clip, the result names the episode splitter scene(s) its words fall in.
- Prediction 1 holds if clips 2, 3 and 4 are all found in Ep 388, and clips 2 and 4 land in the episode’s close-up
setting and clip 3 in its desk setting. The setting is read off the splitter’s own screenshots of those episode scenes,
which also makes the visual check (addendum 1) scene to scene.
Addendum 3, 2026-10-12, before anything ran: the dataset is Candace’s scenes (John)
John: the episode dataset “should have Candace’s face”. That is how the data will line up, scene for scene, so the
times can be interpolated.
- After the episode is split (addendum 2), every episode scene gets one label from its splitter screenshot:
Candace on screen or not (ads, other people, graphics). The labels are made from contact sheets before any
words are lined up, and they are published with the result.
- The dataset is the Candace scenes only. The line-up and the per-clip 80% rule run against those.
- Check: the line-up also runs once on the full episode. Both runs should find the same clips; if not, the result
says so.
- Interpolation: where a post clip lands in an episode scene, the two scenes’ cut points give a time map that does
not depend on the words. Between matched words, post time is mapped to episode time by interpolating between those
cut points. This is reported with the result (where each clip sits in the episode, to the second). It does not change
the 80% rule.
Command (run after this is pushed, word for word)
# 1. split the episode, same splitter and settings as the post (addendum 2)
(cd ~/Documents/repos/dl_wm && python bin/scene_split.py ~/boner_vault/dl_wm_outputs/<download day>/youtube__aLJhpxrakCU/youtube__aLJhpxrakCU_watermarked.mp4 --threshold 50 --fps 5)
# 2. line up the words
python bin/compare_cut.py compare data/holy_holiday_run1 \
--full ~/boner_vault/dl_wm_outputs/<download day>/youtube__aLJhpxrakCU/youtube__aLJhpxrakCU.whisper.json \
--cut ~/boner_vault/dl_wm_outputs/2026-10-12/instagram__DdWodVUBwWy/instagram__DdWodVUBwWy.whisper.json
<download day> is the only free part: the folder the pipeline writes to on the day of the download.