Nine hundred and sixty-six pages in seven minutes and thirty seconds

The test document is a nine hundred and sixty-six page ophthalmology textbook: dense two-column type, plates with captions, tables, and small print inside the figures. It is the hardest thing we read regularly, which is why we keep reading it.

End to end, from the upload to the assembled text, the run takes seven minutes and thirty seconds.

Where the time goes

Almost none of it is the model. The book is split into page images, and the pages are dealt out to whichever cards are free. A page takes as long as a page takes; what matters is how many pages are being read at the same instant. With the fleet warm, that number is in the hundreds.

The parts that do cost time are the ones nobody advertises:

• The split. Turning a large PDF into page images is a single-machine job and it is on the critical path of every run.
• The stragglers. A run is finished when the slowest page is finished. One page on a card that has gone quiet will hold up an otherwise complete book, so a page that takes too long is re-issued to a second card and the first answer to arrive wins.
• The assembly. Nine hundred and sixty-six answers arrive out of order and have to be put back in order, with the page numbers intact.

What we watch for

A reader that invents text is worse than a reader that misses it. The failure we chase hardest is repetition: a model that starts a line, loses its place and writes the same clause forty times. We hold a hand-checked transcription of one difficult page as the reference, and any change to the reading path is measured against that page before it reaches the fleet. The current path takes that page from forty-four missing lines to six.

The economics

The book costs under three dollars of power and machine time. The same book through a commercial interface is about twelve. That gap is the entire business case for owning the plant.