Wikimedia Commons video uploads for the Cedwyn clips: format, licence, AI-scope rules, subtitles, and a gate

Harvested 31 Oct 2026 (skills harvest, ~11:15Z). Companion to wikimedia-commons-wikidata-heritage-coi.md (images, COI) and wikidata-p10-video-placement-measured-gate.md (where a clip gets seen). This file is the upload craft for VIDEO only.

Checked first (31 Oct 2026, Commons API)

• No Commons video exists for the project: search for "Cedwyn" returns only the St Cedwyn church photos at Llangedwyn (unrelated); "bucklerfamily" returns 0.
• THREE photos of the farm already exist, all uploaded by another user (ApMeurig) on 11 Oct 2022, CC BY-SA 4.0: "Great House Farm, Llandough.jpg" (used on en.wikipedia "Llandough, Penarth"), "Great House Farm, Llandough 2.jpg" and "Llandough, Great House Farm, Tree Planting.jpg" (both unused). Do not re-upload or duplicate them; reuse them by name (credit ApMeurig, CC BY-SA 4.0) if a page or clip ever shows them.
• Video results for the area are thin: "filetype:video Penarth" = 1 file, "filetype:video Cardiff history" = 1 file. A clip of ours would be close to alone in its category.

The rules (Commons policy pages, read 31 Oct 2026)

• Format (Commons:Video): preferred WebM with AV1, VP9 or VP8; Theora/MPEG-1/MPEG-2 allowed. H.264/MP4 is not allowed and must be converted first. Our clips are already 720p WebM, so they fit as they are.
• Size (Commons:Maximum file size): host limit 5 GiB; uploads over 100 MiB need chunked upload (UploadWizard does this). Our clips are under 25 MB.
• Tools: UploadWizard (browser), or video2commons on Toolforge (converts + uploads; its "Import subtitles" does not work for .vtt, phab T368298, so give it .srt).
• Licence (Commons:Video): video AND audio must be free (e.g. CC BY-SA 4.0 / CC BY 4.0). Every underlying work (music, archive photos, maps) must itself be public domain or freely licensed. Copyrighted background music gets the audio stripped or the file deleted.
• AI-generated media (Commons:AI-generated media): only media "realistically useful for an educational purpose" are in scope; AI media must go in subcategories of "AI-generated images by subject/by software"; media featuring living people go in "AI-generated media of living people (PIP)". Out-of-scope AI files get deletion requests.
• English Wikipedia (WP:AIIMAGES, Image use policy): "Most images wholly generated by artificial intelligence (AI) should not be used in mainspace", subject to common sense and exceptions. So AI-rendered drama scenes will not be accepted into articles even if Commons keeps them.
• Subtitles (Commons:Timed Text): captions live on pages named TimedText:..srt, one per language, so cy / en / es subtitles attach to one file and the player offers all three.

Measured: what a Welsh Commons video gets (Wikimedia mediarequests API, user agents, Oct 2025-Sep 2026)

Sample: the 99 videos returned for "filetype:video Cymru Wales" uploaded before Sep 2025. Counts are file requests (fetches), not confirmed plays. | Where the file is used | Files | Median 12-month requests | |---|---|---| | In at least one Wikipedia article | 59 | 3,217 | | Only on a Wikidata item | 6 | 835 | | Not used anywhere | 34 | 401 | Single examples: "Dinas Bran's archaeology.webm" (archaeology trust explainer, used on en "Castell Dinas Bran") 19,640; "Geophys at Pembroke Castle.webm" (on en "Geophysical survey (archaeology)") 8,043; "360 View from Swanbridge, Vale of Glamorgan" (used on 4 pages) 1,574; Welsh drone clips never placed: 107 and 437. Read: the upload alone buys a few hundred fetches a year; placement buys the rest (see the companion placement file).

How to do it (steps)

1. Pick only clips made wholly of material we can license: real footage we shot from public places, maps we drew, and scans of public-domain records (pre-1900 documents, PD-old-70). Exclude every OmniHuman/AI-rendered scene, every clip with a real person's likeness, and every clip with licensed music.
2. Re-check rights on each frame's sources; write the source line for each record shown (catalogue ref) into the description.
3. Encode check: WebM VP9 + Opus, keep 720p; under 100 MB so the plain upload path works.
4. File name: descriptive and neutral, place first ("Llandough, Vale of Glamorgan - 1840s tithe map explained (cy subtitles).webm"). No brand words.
5. Description: {{Information}} with a neutral one-paragraph description, date, source "own work" plus each underlying record's citation, author = the account holder; licence template (CC BY-SA 4.0 is the Commons norm; CC BY 4.0 if the owner prefers easy reuse).
6. Categories: the place (Category:Llandough, Penarth) and the subject (maps / historical records of Glamorgan). Add structured data "depicts" (see commons-structured-data-depicts-own-uploads.md).
7. Upload the .srt files as TimedText pages: .cy.srt first (Welsh is the period language), then .en.srt and .es.srt.
8. Log file name, date and URL in the project workspace; measure with the mediarequests API at 30 and 90 days.

Gate

• Pilot 3 clips that pass step 1. At 90 days: keep uploading only if the median clip passes 400 requests/12-month pace (the "unused" median above) or one clip is placed in an article by an independent editor. Otherwise stop; the Internet Archive mirror (internet-archive-movies-clips-mirror-captions-gate.md) already gives a permanent copy.
• Any deletion request = stop and read it; never argue scope for an AI scene.

How we apply it to Cedwyn

• PROPOSAL, owner approval needed: it needs a public Commons account held by a person (no shared accounts, no bot uploads), and a free licence on the clips, which lets anyone reuse them, including commercially. That is the owner's decision, not ours.
• Candidate material: map and record explainer cuts, not the dramatised documentary scenes. The documentary's AI scenes stay on our own pages and YouTube.
• Tab prepares the batch (file list, descriptions, .srt files, categories) in the project workspace; the account holder uploads.
• Public text rules apply to descriptions: never name the owner, banned words list, no family holdings or addresses, neutral dated wording, no outbound promotional links (Commons descriptions are not adverts; the source citation is enough).

Sources

• commons.wikimedia.org (formats, H.264 not allowed, video2commons, copyright of audio)
• commons.wikimedia.org (5 GiB host, 100 MiB non-chunked)
• commons.wikimedia.org (TimedText pages, video2commons .vtt bug phab T368298)
• commons.wikimedia.org (educational scope, required categories, PIP category)
• en.wikipedia.org (WP:AIIMAGES)
• Wikimedia REST API metrics/mediarequests/per-file (our pull, 31 Oct 2026); Commons API search + globalusage (our pull, 31 Oct 2026)

Raw Markdown: commons-video-uploads-webm-timedtext-ai-scope-gate.md · version and changes: changelog · machine index: agents.json