What we refuse, and why it is most of what we do
Of 202 questions this pipeline has written, 19 have been published and 167 could not be settled. The refusals are the part worth showing.
Most fact-checking publishes its hits. This one publishes its misses too, and there are far more of them.
As of 27 August 2026 this pipeline had read 96 public statements, written 202 questions out of them, and published 19 answers. One hundred and sixty-seven questions could not be settled at all. Two published answers have since been corrected, and both corrections are still on the site beside the answers they replaced.
That ratio is not a defect being disclosed reluctantly. It is the product. A question this system cannot answer against a primary record is a question it will not answer at all, and the alternative to refusing is guessing in a house style that looks like certainty.
What a refusal actually is
Every refused question names the record that would have settled it. Not a category, not an apology: an institution and a document.
Department of Defense · Iran War casualty reporting records
New Hampshire Superior Court · Verdict record in the first Youth Development Center trial
U.S. House Clerk · House roll call on the TPS extension for Haiti
You can go and look at any of those. That is the point of printing them. A reader who disagrees with a refusal now has something specific to disagree with, which is more than a confidence score would have given them.
Where the refusals concentrate
They are not spread evenly, and the shape of the pile says what to build next.
Show the numbers
| Category | refused questions |
|---|---|
| An agency report | 68 |
| A bill status record | 40 |
| A federal spending record | 15 |
| A Federal Register document | 13 |
| A transcript | 10 |
| A statistical series | 8 |
| A court filing | 7 |
| A recorded vote | 4 |
polylink pipeline, read 27 August 2026. Each refusal names one record; these are the eight commonest kinds.
Agency reports are the largest single gap and the hardest one. There is no single index of what a federal agency has published; there are hundreds of agencies, each with its own newsroom, its own document numbering, and no obligation to make either machine-readable. Forty of the bill-status refusals are a different problem: this pipeline can read who filed a measure and what it authorises, and cannot yet read what a measure's other provisions actually say.
Neither of those is solved by being cleverer about the claim. They are solved by adding a source, and until the source exists the honest answer is the one on the page.
Why the count keeps moving
Every refusal here has been examined more than once. When a new source is added to the registry, every question that failed for want of it is re-opened and tried again, which is why a claim on the pipeline page carries the number of times it has been looked at rather than a single date. Some carry twenty-seven.
So the refused number is not a scoreboard. It goes up when we read more statements and down when we build a source, and both of those are the system working.
What this does not mean
A refusal is not a finding about the speaker. It says nothing about whether the statement was true, careless, or exact. It says only that no record this pipeline can reach settles it, on the day it was tried, and that we would rather print that than a number we invented.
If you want the current state rather than this snapshot, the pipeline is live, and every figure in this piece came from it.