Artifact
One Article Through a Writing Line Built Like a Factory Floor: The Test Run
Research pulled when a paragraph needs it, with the inspections built into the line rather than bolted on after. One article has been through it, as a test that the line worked.
A thing I built. This page says what it is and what is true about it. Dan Stephens writes about psychology, AI systems and the method behind both at danieltstephens.com.
One article has been through this line. It was the psychiatric medication management guide, and I built it to find out whether the line worked at all and could produce something to specification. That is the honest size of the sample.
The idea is a notebook run like a production line. Research is pulled when a paragraph needs it, rather than gathered in a heap at the start. The inspections sit inside the line rather than bolted on at the end, so a defect is caught at the station that made it.
What it is made of
- A pull, not a push: a paragraph asks for what it needs and the collection answers, rather than a research phase that ends before the writing starts.
- Inspections at each station rather than one review at the end.
- A mechanical citation check that resolves every reference and fails the piece if one points at nothing.
- A hostile read, where a reviewer is instructed to attack the claims rather than improve the prose.
- A human approval gate at the end that nothing gets past without me.
Where it stands
Working, and barely used. One finished article, built as a test.
Two things it did not do on that run. I had not read the underlying sources in full, though most of them did have written reviews that I had read and approved. And the sources without a written review were used with less weight rather than kept out. So what ran was the writing line, with the source gate only partly engaged.
More of my reading time is going into the source side, and more will go there once the counseling practice passes eighty percent booked.
By the numbers
- 1 article produced through the line, as a test that it worked
- 0 of 19 citation flags upheld when I first gave the checking job to an AI judge, and roughly 359 objections across a batch of 14 documents, none of which survived checking
- 5 real citation errors that judge had passed as clean, caught afterward by the mechanical check, including 2 references that resolved to nothing. The subsection carrying both was cut before release.
- 27 claims examined in the first hostile read: 2 factual errors, and 10 places where a claim outran its evidence
- 13 literature reviews and 26 finished documents, roughly 165,000 words, put through the same checking. Those documents were not produced by this line; they are what the checking was proved against.
- 1 book built through the line on 2026-08-28, at four page sizes, surfacing 4 defects, all fixed
The collection this line pulls from, and the gate on the way into it, is in No Source Gets Used Until Its Review Is Written.