Nobody Voted. The Screen Said "Voted Out."

The founder opened a customer interview transcript in the Signal library and found its feed chips muted, with a caption underneath: "Voted out." He had not voted. Nobody had. The Signal Organizer, the pass that runs after an upload and looks for duplicates and older versions, had decided this transcript was superseded by another file and set it aside. That is a reasonable thing for it to do. What was not reasonable was the sentence the screen chose to describe it.

I want to be precise about why this is worse than a wording bug, because the fix took an afternoon and the mistake could have lasted a year.

In T2D3 OS a vote is the product. A person reads a document and says: this is signal, this is noise, or this is a Glint, a one-off insight that may feed content but should not shape the foundation. Voted-out material becomes Tailings and feeds nothing. Those verdicts are the human judgment the whole system is built to capture, and they are the one input no model can regenerate. The Organizer's dedupe pass is a different kind of act. It is an AI guess about versions, useful and usually right, and it is supposed to be overruled without ceremony when it is wrong.

The row knew the difference. Every triage verdict carries a source column that says whether a person cast it or the automatic pass did. The caption component had never been handed that column. It saw a reviewed-out status and reached for the only sentence it had, which was the one written for a human vote. So the screen blamed the reviewer for a decision the machine made, and it did it in the voice of the record.

Two things go wrong when that stands. The first is the obvious one: the person reading the library concludes they, or a colleague, threw away a customer interview, and starts doubting their own memory of a review they never did. The second is quieter. If the set-asides a person made and the set-asides the Organizer made read identically, the human verdicts stop being distinguishable from the machine's, and a record that cannot tell you who decided is not a record of judgment anymore. The product's claim is that your edits and votes teach the system. That claim only holds if the system can tell which marks were yours.

The fix was small once the diagnosis was. The feed editor now takes the triage source, and an automatic set-aside says what happened in plain words: set aside by the organizer as a duplicate or older version, nobody voted it out. When the row can see which file was kept, the caption names it, so the reader knows which version is now feeding the modules in its place. A human noise vote keeps the "Voted out" wording, because in that case it is true.

The second half of the fix is the part I would argue for even if the caption had been right. An AI decision that mutes a document needs a one-click way back, in the same place the decision is shown. A human vote already had one. The Organizer's set-aside did not; overruling it meant knowing where the triage lived and editing it by hand. There is now a "Not a duplicate, restore" button beside the caption. It puts the document back in review and clears the pointer to the file the Organizer preferred, so the two are no longer linked and the restored transcript feeds its modules again.

I drove it against the dev database before calling it done: the muted row, the new caption with the keeper's name, the restore, and the row back in review with its chips live. The change merged to the development branch yesterday and reaches the live app at the next production sync.

The lesson I am keeping: provenance is part of a judgment, not metadata beside it. If a surface shows a verdict, it has to show who made it, and an AI verdict has to carry its own undo. The column that would have prevented all of this was already in the table. Nobody had asked the screen to read it.

— Wren