Pitfalls & fixes
This page is generated from the logbook entries rather than maintained separately — every answer has a documented report behind it. All entries →
You do not — you write the default so that it stays true unchecked. An alt text is read aloud, not looked at; it passes every visual inspection. "Chapter mark." is correct for every book and therefore cannot go stale.
Read the entry: One hundred and forty-two alt texts for the wrong book →
By the imprint inside the text and by the overlap with the original. Adapted publisher versions often carry editors' names in the middle of the work and depart substantially from the source — in our case the overlap was 29 per cent. Neither shows on the title page; both show in the full text.
Only in one direction. Our segmenter systematically overestimates a character's share of the dialogue — in ten out of ten checked cases, by a factor of 1.6 to 21, because it assigns unclear speakers to the nearest name. What it sees below the threshold is certainly below it; a high figure proves nothing and has to be read against the text.
By looking at it — and by making sure nothing wrong is offered in the file dialogue. Dimensions, filename and checksum say nothing about the image. So before every upload we now generate a folder containing only the files for that one channel, rebuilt on every run, which is not created at all if the checks come back red.
With a case that MUST fail and demonstrably does — and that additionally checks WHAT it fails on. A run that was only ever green proves nothing; and "it went red" is not a passed test, because a broken invocation is just as red as a correct abort. The test must also reproduce the real call, not an imagined version of it.
Read the entry: Four safeguards, eight bugs, not one self-reported →
Never miss one
We write when a new book is out — we send nothing else. New logbook entries are available via the RSS feed.