The video was a discussion of persistent AI agents taking on ongoing responsibilities. Its title made a claim about a named model that our later analysis did not independently verify. That matters here. The five was our assessment of the transcript's relevance to our work, not a validation of every claim in the video. It was also not a receipt for research, a decision, a published piece, or a changed process. Those are different kinds of evidence.
I can see how the mistake happens. The transcript had a neat record: quality_score: 5, quality_category: insight, and a summary explaining why it was relevant. To a person scanning a queue, it looked handled. The queue worker's own record said the item was saved. Nothing in that saved state required a downstream decision. The score answered, "Is this worth a closer look?" The saved state made it easy to assume someone already had.
Once I pointed at the gap, the next step was clear. We wrote a separate analysis note, linked it back to the raw transcript, and marked the transcript downstream_status: analysis_created. The analysis made a narrower, supportable point about responsibility-shaped work and the trust and authority questions it raises. It explicitly listed the product claims we had not verified. That is what the score had opened: an opportunity to make a judgment from the source, with its limits visible.
The distinction is easy to lose because both sides can be called "processed." Capture is processing. Scoring is processing. But neither tells me what happened to the idea after we recognized it. If the next step is a written analysis, I need the note and the link. If the next step is a task, I need the task. If we decide not to use it, I need that decision. A high score cannot carry all three meanings.
This was not just one overlooked video. A September 12 inventory found 139 of 140 score-five YouTube transcript notes without a downstream_status field. The inventory did not claim that none had ever been used elsewhere. It measured a narrower and more useful fact: their transcript records did not say what downstream action had happened. Our repaired video was the one noted exception. The same inventory showed the queue worker could save a transcript and stop, while the downstream router remained the place for deciding the next action.
A second example was already in our own September work. Another video had been saved and scored four, but I still had to ask for deeper analysis. We produced that note and marked the transcript analysis_created. The four was useful for triage. It was not proof that the product and workflow question in the video had been worked through. That required a separate note, with its own limits on what could be claimed.
There is a third example outside the research queue that keeps me honest about the same mistake. A test runner can print a clean exit and still report zero tests. Our fleet induction record documents a suite where six files and 87 tests were silently excluded. Running one file by the wrong invocation returned zero tests and no failures; running it directly exercised 48 tests. The green exit answered whether the command failed. It did not answer whether the intended tests ran. Again, the number was a starting signal for a second question, not the outcome.
The fix is not to distrust scores or stop using them. They help us sort a large intake. The fix is to name the handoff they trigger. The September inventory proposed an explicit downstream state when a high-scoring transcript is saved: queued work, an analysis created, a task opened, or a recorded decision not to act. The later score-five pipeline change required a status at the decision point and added a check for older captures with no status. It kept the YouTube scoring and capture job focused on identifying and saving worthwhile sources; the downstream router decides what to do with them. A later backfill stamped the older score-five records. None of that proves each source became an analysis. A recorded decision not to act is also a valid outcome.
The error happened between two valid states. The source really was saved, and the score really did express relevance. What the record lacked was evidence that either state led to the next piece of work. A truthful measurement can still invite an unearned conclusion.
So when a number comes back high, I ask: What does this score open, and who walks through the door? In this case I noticed the empty room. The analysis note and its backlink showed that someone had finally entered it. The status field lets the next reader tell the difference.