Pages you can read straight away

Last week we read every page of a real notebook, 289 of them. Most were useful. What stopped a reader was rarely a wrong fact. It was smaller than that.

Footnotes that ran 1, 3, 7, because the model had dropped two sources and kept the numbers. A person's Company left Unknown while an organisation page for their own mail domain sat one click away. The same person on two pages, once by name and once by handle. A page that called you "you" in one paragraph and by your own name in the next, and opened on whatever happened last Tuesday instead of who this person is.

None of these needs a better model. Each is a rule applied when the page is written. Sources are renumbered in order and every citation follows. An unknown Company links the organisation page for the person's mail domain. Init ends by naming the pages that are probably one person, with the command that merges them. Pages refer to you one way, and the first line says who someone is.

We ran the same mailbox again. Gaps in source numbering went from 40 pages to none. Company Unknown went from 53 people to 31. Pages that named you instead of saying "you" went from 58 to 1. The run took 122 minutes and $8.42, and left 3 pages unwritten instead of 8.

Two of those three had left out a required section, so validation refused the whole page. That was the wrong trade: a page with one section marked Unknown is more use than no page. Now the section is added as Unknown and the page is kept. The third page cited files from outside its own material. That refusal was right, and it stays.