Videos, Transcripts And Provenance
Someone asking for your vote for the first time has no roll call behind them. What they do have is hours of themselves talking — interviews, town halls, podcasts, speeches — and it is the most checkable evidence there is. It is worth nothing while it is still a link. This is what turns the link into words a citizen can read, quote, and check against the tape.
What it does
The app never downloads media. It does not fetch audio or call a video host for a file — not under a flag, not in any environment. A build guard reads the source tree for any executable reference to a downloader, string literals included, and fails the build on it.
That refusal is the whole reason the citizen mining network exists. The list of videos still missing a transcript is a plain public file; working from it needs no account and no permission. Volunteers opt in on their own computers, transcribe, and publish the words and how they made them into a repository they own. Five hundred missing transcripts is an impossible bill for one server and an afternoon for thirty volunteers.
Distribution also makes the result believable, because a single central transcript is an unchecked claim. Independent copies of one video get compared: a majority agreeing makes the canonical copy trusted, a lone copy is flagged uncorroborated, and copies that disagree trust nothing — we are not quoting that video until they agree.
What the app holds is the result: the transcript text, unedited, because an edited transcript poisons every comparison downstream of it; the records pointing back at the original video; and provenance, a mark on the transcript file, never on the person, saying how the words were obtained. A human-written caption track, a platform's automatic captions that mishear and separate no speakers, and a machine transcription with word timings and speaker turns are marked as three different things. Where a quote rests on a weak one the caveat renders beside it, never in a tooltip: these words come from automatic captions, and speakers are not separated.
One control renders those words wherever they appear: a timecode gutter, a who-is-speaking key with the page's subject set apart, gaps drawn in position rather than skipped. Click a sentence and a button appears under the player — Move play head to 12:58 — and it seeks without playing.
Each video also gets its own page listing every qualification area, whether that video spoke to it or not, each blank stating its cause. The empty rows are the point: she did not talk about economics here and nobody has read this video yet are different facts, and confusing the two turns our gap into her silence.
The goal it serves
The incumbent's structural advantage is that only the incumbent has a record. Sizing up a challenger otherwise means campaign advertising, endorsements, and a party organization's say-so — the exact terrain money and organized influence already own. A two-hour interview is the one place a person's reasoning becomes something you can weigh instead of an impression, and most of it sits in public, unread.
This is what makes that material usable. It is the reason Good New Leaders can be assessed at all, and the evidence base under every Flamingo Award published about somebody who has never held office. Without it, replacing a captured incumbent means asking a party's base to back a stranger on faith.
The insistence on provenance is not fussiness. Machines mishear, and publishing a claim about a named living person out of a machine transcription without saying so is how a movement earns a lawsuit. Saying how the words were obtained lets a citizen price the uncertainty instead of being asked to trust us — the discipline of sourcing to the official record, carried from votes to speech.
What keeps it honest
- The refusal to fetch is enforced, not promised. The software cannot quietly stop obeying it between releases.
- Weak provenance lowers stated confidence and renders a caveat. It never withholds a finding. A rule vetoing findings that rested on a weak transcript was proposed and rejected: the Flamingo Award's absence has two meanings, so quietly declining to assess somebody publishes unknown about them.
- Speaker separation is not trusted further than it was measured. On noisy long-form recordings the machine over-splits badly: one man talking alone for three hours came back as seventy-two voices. Unresolved voices collapse into one honest line rather than a wall of invented names.
- A guess is allowed; an undeclared guess is not. Where the pipeline works out whose words it is reading, it publishes that it guessed, by what method, and how sure it is. That uncertainty moves how much we claim to know about the person, never the assessment.
- Only the subject's own words count. An interviewer's question or an inserted clip is not evidence about the subject; someone merely mentioned in another person's video is left out of their own assessment; and the transcript is where you find the moment, never a replacement for the recording.
Works with
- The Politician Record Page — where findings from single videos roll up into one person's award case, every source video still reachable from it.
- The Rosters — the rosters decide whose channels are worth following, and a person with no channel on record is a stated gap, not a silent blank.
- Your Seats And The Seat Map — where evidence about a challenger turns into a decision you can act on before a primary.
Where to go next
- Good New Leaders — the people with no voting record, and what gets used to assess them instead.
- What feeds the Flamingo — everything on this side of the wall, and what is deliberately kept off it.
- Qualifications — the published areas a transcript is read against, written down before anybody is assessed.
- Naming the gaps — why a missing transcript is published rather than rounded away.
- The action catalog — including adding a video to a politician's record, which you can do this week.