How to use the Alexandria cleaning review site. Read this once, then keep it open for the first week. Nothing you do here changes a book on its own.
Alexandria is a library of about 2,700 books. Many were scanned or converted years ago and small faults came along with them: words stuck together, page numbers sitting inside sentences, forewords by people who are not the author.
Software finds those faults and proposes a fix for each one. You decide which proposals are right. The team then applies only what you approved. You are the judgment in the loop, and the software is the tireless part.
Open the site, type the shared password and your name, and press Sign in. Ask Philip for the password. Everyone on the review uses the same one, so there is nothing to set up and nothing to wait for. You stay signed in on this browser, so you only do this once on each machine.
The name is not optional. It is the only thing that says who you are, so every decision, report and note you record is signed with it. Put your first name in. Sign out is at the top of the panel.
Reading copies hold only what the author wrote. These come out whole, not cleaned up: footnotes and endnotes, editor's and publisher's notes, introductions, forewords and prefaces by anyone other than the author, and page numbers or running headers that leaked into the text.
These stay: the author's own preface, foreword, introduction or dedication; chapter, part and section titles; anything that is part of the work. And the same fix can be right in one book and wrong in another: verse numbers are content in the Bible, line breaks are content in poetry, speaker names start each speech in a play.
Proposed A tool found the problem and worked out a fix. Nothing has been changed in the library yet.
Implemented The fix is now in the library. Readers see it.
There is nothing in between. A book and problem pair is one or the other.
Every decision has a note box. The note is optional, but a sentence saying why is worth more than the click on its own, especially for Reject, Unsure and Reverse.
A decision does not change a book by itself. It records what you think, and the team acts on it in a later run.
A card shows one change in its natural setting so you can judge it by reading, not by guessing.
In the changed sentence, text coming out is
struck in red and text going in is
green. Red with no green means something is being
deleted. For a glued line, the green mark is only the whitespace that was
lost: a single space in a book that reads as prose, a line break in a book
that reads as verse.
An amber card with a badge is one the tool could not decide on its own. Those need your eyes most.
The card also carries the problem name, the paragraph number, and the note line for that problem if one has been written.
Every sentence in every book has one short id, and it never changes:
the chapter number, a dot, then the sentence's place counting from the
first sentence of the whole book. So 3.4831 is sentence
4,831 of the book, which falls in chapter 3.
The running number is what makes the id unique, so the number on its own
is enough: type 4831 in the find box and you get the same
sentence. The chapter is in front only because it tells you at a glance
where in the book you are.
Ids are on the cards, on the change history and on the reports already filed. Click one and it is on your clipboard.
Quote the id whenever you say where something is: in a note on a
decision, in a problem report, in a message to Philip. "The comma in
3.4831" needs no explaining; "the comma near the top of
chapter 3" does. On a book page, Find a sentence
turns an id back into the sentence, with its chapter and paragraph, and
Report this spot on a card carries the id
straight into the report form.
Some problems have hundreds of thousands of instances. Nobody can look at them one at a time, and they do not need to: within one book the same fault is nearly always the same fault.
So the site shows you about ten examples from a book, and your approval becomes a rule for that book and that problem. The counter tells you where you stand, for example "you have looked at 10 of 7,389".
If ten is not enough to be confident, press Show 10 more and keep reading until you are. There is no prize for going fast, and the rest of the book rides on your answer.
One problem is footnote removal. In most books, a small number glued to a word is a footnote marker, left behind when the note itself was stripped out, and taking it away makes the sentence read correctly again.
So on the Bible, those cards get Reject, with a note like "these numbers are verse numbers, they are content in this book".
Your note is the valuable part. It becomes a standing exception on that problem, shown on every card of that problem from then on, so the next person does not have to work it out again and the tools can be taught the same guard.
The general shape of this is worth keeping in mind: ask whether the thing being removed carries meaning in this book. Page numbers never do. Verse numbers, sutra numbers, and numbered sections do.
Cleaning by problem finds only the problems we already know about. The random audit is how we find the ones we do not.
The Random audit panel hands you a book to skim. Claim it, open it in the reader, and read a few pages the way an ordinary reader would. Then pick one of three:
Describe what you saw, not what should be done about it. "The page numbers are inside the sentences, like this one" is exactly right. Working out the pattern and the fix is our job.
Every audiobook has a narrator and a handful of characters, each with a voice and a short performance direction telling the voice how to speak.
A book takes a couple of minutes. Listen to the sample line. If the performance is not right, adjust the performance direction, press Regenerate, and listen again. When it sounds right, Approve.
Regeneration is not instant. The request is queued and a worker picks it up, so come back to that book later.
Write directions the way you would tell a person: "older, dry, unhurried", "young and nervous, speaks quickly". Do not try to spell out the sound phonetically.
Scenes are pictures made from a description of a moment in the book. Your question is simple: does the picture match the scene as the book describes it?
If it does not, adjust the prompt, press Regenerate, and look again. Then Approve if it is right, or Reject if the scene should not be used at all.
Watch for the wrong century, the wrong place, the wrong number of people, and anything the author did not put there.
Choose Unsure and write one sentence about what is bothering you. That is the whole procedure.
Unsure is not a failure and it is not a delay. The rule we work by is that a reading copy must never lose real text, so leaving something alone costs nothing and a wrong removal costs a restore. If you are hesitating, that hesitation is information.
Sign your uncertainty with a reason: "this looks like a stage direction, not a footnote" is enough for the next person to pick it up.
The Activity tab on the Books page shows everything everyone did: decisions, approvals, reports, audits, audiobook work, with who did it and when. You can filter it by person, problem and action, and it refreshes on its own.
Use it to see what has been picked up since you last looked, to avoid working over the same book as somebody else, and to find your own work from yesterday.
The Alexandria reader is at app.alexandria.wiki/read/<slug>, and the small read icon next to a book in the list opens it. Anywhere a card leaves you wondering how the passage really reads, open the book and look.
The reader is what the customer sees. If something looks wrong there, it is worth reporting, whether or not any tool has flagged it.
Nothing on this site edits a book directly. Decisions are recorded, changes are applied later in a run that keeps a backup of every file it touches, and anything already applied can be reversed.