mindGAINS · Wistia library audit
Every active video in the GAINS Wistia account, with a proposed title, a proposed description, and a proposed set of tags drawn from a controlled vocabulary. Nothing here has been written to Wistia — this is the review copy.
Three separate problems overlap across the library: titles carrying Zoom export debris, a tag vocabulary that was machine-generated three months ago, and descriptions that are almost entirely empty.
What changed on 16 September 2026
The library went from 102 active files to 84. Nineteen dead items were cleared — 17 archived (all 14 board meetings, the truncated Re-Imagining Boundaries copy, an Oct 2024 Zoom test, a 33-second fragment) and 2 deleted (a 49-second fragment and a remix duplicate, recoverable until 16 October). Every one had zero plays and none was in a channel. On this account's legacy plan, billing counts active files rather than storage, so archiving is the full saving and deleting buys nothing extra.
One video was added: a Community Discussion recorded that morning. Four member discussions that had been waiting 28 to 301 days were moved into the archive channel, which now carries 83 of the 84 active files. The remaining one is the website welcome video, embedded directly on the GAINS site by design.
Most of the damage is mechanical and came in through Zoom. The recorder's export filename became the Wistia title and was never cleaned, so viewers see the meeting's start time and runtime appended to the topic.
| Defect | Videos | Example |
|---|
Four pairs of videos previously shared a title, and all 14 board meetings were called simply Board Meeting. Every one of those shorter or duplicate copies has since been archived or deleted, and the board meetings are archived. No two active videos now share a title. The full-length copy survived in each case — the 57-minute Re-Imagining Boundaries, the 63-minute Way of Play, the 88-minute Executive Functioning, and the original Iacoboni recording — and each is published in the archive channel.
The account holds exactly 100 tags in two distinct layers, separated by a clean two-year gap. Thirty-three were made by hand between August 2023 and April 2024. The other sixty-seven were generated in 2026 — on upload from May onward, then backfilled across the catalogue in a single sweep on 26 August. Two-thirds of that second layer lands on exactly one video each, which makes it description text wearing a tag's clothes.
| Layer | Tags | Videos | Character | Recommendation |
|---|---|---|---|---|
| Curated Aug 2023 – Apr 2024 |
33 | 59 | Speakers, years, broad domains — porges, education, dei | Map onto the new vocabulary; drop the year tags |
| Auto-generated 2026 |
68 | 44 | Per-video phrases — mother's journey, brain's sight bias | Retire; 68 of them describe a single video each |
The cull did not shrink this problem. Archived media keep their tags, so retiring 19 files removed only one tagging. A new upload on 16 September added three more, including a freshly minted somatic practice — so the account now holds 101 tags, not 100, and auto-tagging is still running on every ingest.
Five facets, each answering a different question, so tags never compete for the same slot. A tag has to divide the library to earn a place: it must land on at least three videos and not on nearly all of them. That rule disqualifies interpersonal neurobiology, the account's most-used tag at 36 videos — in a library published by the Global Association for Interpersonal Neurobiology Studies, IPNB is the universe, not a facet. Dashed tags below fall under the three-video threshold and are candidates to cut.
All 84 videos, oldest first. Select any row to see its current and proposed metadata side by side.
Do the 2026 auto-tags get cleared, or left in place?
They sit on 44 videos. Leaving them means the new vocabulary coexists with 67 single-use phrases, and the tag list stays unusable as a filter.
Suggested: clear the 2026 layer, keep the 2023–24 one, apply the new vocabulary over the top.
Are public titles renamed, or only the untouched Zoom imports?
29 videos sit in the Shared folder with real traffic — the O'Donohue conversation alone has 912 plays and 1,095 visitors from a referrer that hasn't been identified yet. Renaming won't break embeds or URLs, which key off the hashed ID, but the change is publicly visible.
Suggested: apply to the Zoom imports first, review the high-traffic public titles separately.
Do the 10 existing hand-written descriptions get overwritten?
They read "with Bonnie Badenoch and Sharon Stanley" — short, human, accurate. The proposed descriptions are longer and drawn from transcripts, but they're machine-extracted.
Suggested: keep the human line, append the extracted summary beneath it.
Two names need a human, not a rule.
Dan Siegel (May 2024 session) is vague, and nothing in the metadata says what the talk covered. Dan Siegel Webinar Recording, Oct 2023 was uploaded in 2024, so its title date and upload date disagree.
Four limits on the proposals in this report, stated plainly so they're weighed rather than discovered later.
Transcripts were sampled, not read in full
Added 16 September 2026, correcting an understatement in the original. Before analysis, every transcript was reduced by a single rule: keep the opening 1,400 words, then 12 evenly-spaced 170-word windows across the remainder. Every downsampled file is exactly 3,452 words — a median of 35% of the spoken words, ranging 21% to 56%, with the longest sessions faring worst. 72 of the 83 kept under a quarter.
It samples across the whole session rather than truncating it, so late material is not systematically invisible — on a 1h51m session the sample still reaches 01:43. But the resolution is coarse, and the consequence runs deeper than any single row: a topic discussed only in the unsampled portion never became a theme, so it never became a candidate tag, so it is absent from the 55-tag vocabulary — and nothing on this page would reveal the omission. The eleven analysed videos carrying no topic tag are the visible symptom; there may be an invisible one.
Tags were assigned by keyword rules, not by reading
Each video's proposed tags come from pattern-matching its extracted summary against the vocabulary. Expect the accuracy of a careful skim — right most of the time, wrong in ways that are obvious on any specific row. Eleven analysed videos still carry no topic tag at all.
Speaker tags came only from titles and existing descriptions
Never from transcripts. The speech-recognition name list contains Louis Cozolino beside Lou, Tina Payne truncated from Bryson, and two non-people (Living Journal, The Practitioner). The cost of that caution is coverage: 15 speakers tagged, and guests named only in the audio are missed.
Descriptions come from unreviewed transcripts
83 of the 84 active videos have a proposed description built from the transcript extraction. The 19 previously excluded — 14 board meetings, 4 fragments, 1 remix duplicate — have since been archived or deleted. The single remaining gap is the 16 September upload, which has a transcript but has not been through the extraction pipeline; its tags come from keyword rules over the raw transcript and its description is a placeholder. Proper nouns throughout are unreliable; one existing description already misspells Marco Iacoboni as "Iacaboni".