The Citizen Mining Network
Thousands of hours of video exist in which a politician explains, in their own words, what they intend to do. Turning that into text somebody can quote is slow work, and the movement's own server is barred from doing it. So the missing work is published as a public list, and citizens take rows off that list on their own machines. The word mining describes the shape of the work and nothing else: no coin, no wallet, nobody paid.
What it does
A scheduled run subtracts the transcripts the movement holds from the videos it knows about and publishes the difference as a public work queue in the open data repository. It states its eligible universe — every video attached to a person on an in-scope roster with a media address anyone can reach — and gives a typed cause for each exclusion: no media address, an unsupported host, an earlier permanent failure, a duplicate. If that universe comes back as zero while the video library is not empty, the run fails loudly instead of publishing an empty list.
Each row carries a priority, and the arithmetic behind it is published. A row gains weight when almost nothing is known about the person, when the video runs past forty minutes, and when the person has no transcripts at all; it loses weight for every copy already received. Long-form is favoured on purpose: a two-hour interview is where somebody's reasoning becomes checkable, and a ninety-second clip is not.
A citizen turns on a mining node from settings. It is off by default. You point it at a public repository you own, choose a node name — a pseudonym is fine and limits nothing — and set your own limits: minutes of audio per day, rows per run, whether to skip rows that already have enough copies. The scheduler command is shown as copyable text, and one button runs a dry harvest that reports what the movement can see in your repository.
The node does the work on its own hardware. It pulls the queue, which is read-only to it, picks rows by priority, and runs the same transcription ladder the movement runs: existing captions first, local speech recognition otherwise. It commits the words plus a provenance sidecar — which tool, which model, which rung of the ladder, and the checksum of the transcript's exact bytes — into its own repository, never into the movement's. It commits text only: no audio, no video, nothing behind a paywall or a login. A hosted install is barred from fetching media at all, which is why the central server cannot do this work in the first place.
Copies are then corroborated. Three are requested by default, because two copies can disagree with no way to break the tie and three can. Copies are normalized for case, spacing and punctuation, then compared pairwise on how closely they align and how much of the longest copy each covers. The copies that agree form a set, the largest set elects the canonical copy, and a set that is a majority of all copies means the transcript is trusted and quotable. A lone copy is usable, and anything derived from it is marked as resting on an uncorroborated transcript. No majority means nothing is quoted, the row stays open, and the product says so plainly: several citizens transcribed this video and do not agree.
Nodes build a standing from agreement over time. It buys preference — a node's copy is favoured as canonical inside a set — and never authority. No standing level lets a copy skip corroboration, and a fully pseudonymous node can reach the top.
The goal it serves
Hundreds of missing transcripts is an impossible bill for one server and a pleasant afternoon for thirty citizens. A movement that has to buy its own evidence supply is limited to whatever a funder keeps paying for, and when the money stops the record stops growing.
The integrity argument matters more. A single central transcript is an unchecked claim, however good the tooling that produced it. Several unconnected people transcribing the same public video, then comparing the results word against word, is what turns that claim into evidence somebody on the other side of the argument can stand on. It is the discipline the rest of this site runs on: publish the derivation, and where the record does not support a conclusion, say so rather than rounding up.
It also keeps the evidence supply out of any single company's hands, including ours. The queue is a public file, the repositories belong to the citizens who filled them, and the movement holds a pointer and a checksum rather than a monopoly on the words — the same posture the whole open record takes, and the reason the code and the repositories are public too. If this organization disappeared tomorrow the transcripts would still be there.
What keeps it honest
- A node never pushes to the movement's repository. The public record is read-only to every miner. Work arrives by being published elsewhere and harvested.
- The text is never edited to improve what the person said. A transcript is a record. Corrections and doubts go in the sidecar notes, never into the words, because an edited copy poisons every comparison it enters.
- The checksum is the anchor, and it is re-checked. A sweep re-fetches canonical copies and compares hashes; a source transcript that vanished, or changed under us, is stated on the pages citing it.
- Comparison is about words, not meaning. Corroboration never votes on what a person meant. It compares one text against another, so the reading happens later, on one agreed text two people can both check.
- The wall holds. Nothing produced here reaches the Monkey Award or the Llama Award, whose only input is the voting record. Transcripts feed the Flamingo Award and nothing else.
Works with
- Run It On Your Own Computer — a mining node is that same local install with one switch turned on, doing the work on hardware the citizen owns.
- The Public Data Repository — where the work queue is published, and the record a node reads from but is never allowed to write to.
- Nothing Fails Quietly — an empty eligible universe, a disagreement between copies, and a checksum that stopped matching are all loud by design.
Where to go next
- Open Data — how the movement's public record works, and why it is a repository you can clone rather than a database you are asked to trust.
- Clone the record — pull the whole thing onto your own machine, including the queue this network drains.
- Naming the gaps — why an admitted hole in the evidence is worth more than a confident guess.
- Trust Scores — how a pseudonymous citizen becomes believable through checkable work instead of an identity document.
- Take Action — what a citizen can do this week, including running a node and closing rows on the queue.