Carried, Word for Word

Safina is the Arabic and Persian word for a ship: a vessel that takes on its cargo in one place and delivers it, unchanged, in another.

The word carries two older echoes for us. When Shaykh Abbas Qummi, who compiled Mafatih al-Jinan, wrote his great guide to Allama Majlisi's Bihar al-Anwar — "the Oceans of Lights" — he called it Safinat al-Bihar: "the Ship of the Oceans", a vessel for crossing a sea of hadith. And the Prophet described his own household in the same image, in a hadith narrated by Shia and Sunni scholars alike:

مَثَلُ أهلِ بَيتي فيكُم كَمثَلِ سَفينَةِ نوحٍ مَن رَكِبها نجَا ومَن تخَلّفَ عَنها غَرِقَ

"The likeness of my household among you is the likeness of Noah's Ark: whoever boards it is saved, and whoever stays behind is drowned."

— as al-Nuʿmānī quotes it in his Kitāb al-Ghayba

safina is also the name of the fourth tool we have built for the people who prepare our material, and it opened this week.

Four tools, one purpose

IlmFlow is our effort to make sourced answers about the Shia tradition — faithful to their sources — available to people who currently get neither.

We have written about sending a student to the hawza: not a person, but a large language model, trained to learn how scholars think. And we have written about the three tools that give it material to learn from, a library to research in, and a place to keep its notes. That library had a gap. It held books. It did not yet hold what the scholars have said.

daftar is where the teaching material is written — by scholars, in their own language, and reviewed before it is used.
fihrist is where the sources in books are found — which book and which page to open, never a quote to copy.
diwan is where our own writing is kept, exactly as it was written.
safina is where the spoken teaching is kept — the lectures, lessons and speeches that the scholars' own offices publish on their websites, exactly as they published them.

Much of what a scholar teaches never reaches a book. It is given in a lesson, a talk to students, a speech on an occasion — and very often that is where a ruling is explained in the scholar's own words: not only what the ruling is, but why. Those talks are published by the scholars' offices, on websites that change, move and are sometimes rebuilt from scratch. safina keeps them.

What safina keeps, and how

The words, exactly as published. safina reads each page with a recipe written for its site, turns it into plain text with its headings, its emphasis and its notes, and then checks it: the text safina keeps must match, character for character, the text the page itself shows. A page that cannot be matched exactly is set aside and marked, never tidied into something close. Today that is about one in thirty of the sessions of Tafsir-e Tasnim. Each will be read again when its recipe learns how — from the copy of the page safina already holds, without asking the site for anything again.
Where and when, with every text. Each text carries the address of the page it came from, when safina read it, and a fingerprint of the text itself. A quotation taken from safina can always be traced back to the page it came from, and checked against it.
Every version. If a site changes a page, safina keeps the new version beside the old one. Nothing is overwritten.
With permission, recorded. safina reads a site only with the permission of the office that publishes it, and that permission is written into safina itself — who gave it, when, and how — beside the pages it covers. A permission that is only remembered does not count.
Politely. safina reads a site the way a considerate visitor would: one page at a time, with a pause of ten seconds or more between requests. It reads each site's robots.txt — the file in which a site tells visitors like it what to leave alone — and follows it, unless the office that publishes the site has given us permission, recorded in safina, to read further. It says who it is on every visit, and the owner of any site it reads can find out who we are, and how to reach us or ask us to stop. Reading all of Tafsir-e Tasnim takes most of a day.
Walled in. The part of safina that reads other people's websites can reach those websites and nothing else of ours. And nothing from another site ever runs here: safina keeps the words, never the site's code.

The first two

safina opens with two collections, each kept with the permission of the office that publishes it:

The speeches of Ayatullah Muhammad Taqi Mesbah Yazdi — nearly eight hundred of them, from almost four decades, 1982 to 2020.
Tafsir-e Tasnim, the commentary on the Qur'an of Ayatullah Abdullah Jawadi Amoli — more than four thousand sessions, from 1983 to 2024.

More will follow, from the offices whose permission we hold.

Searching as a reader reads

Persian and Arabic can write the same word in more than one way: two forms of the letters ya and kaf, with or without vowel marks, with or without the half-space that joins a word's parts, with Persian or Latin digits. safina's search treats all of these as the same word, as a reader does, and a word of three letters or more also finds the longer words it begins.

And when a search finds nothing, safina is careful to say what that means. It tells you which forms of each word it looked for, and how much of the collection it searched. "Not found" is a statement about the search — never a claim that the scholar never said it.

Reading, and working with an AI

Members read in a reading room laid out like diwan's: right to left for Persian and Arabic, a list of contents for a long lecture, footnotes that jump to their notes and back, and every version the site has published.

safina is also built for working with an AI research assistant. As with diwan, you give your assistant its own key, and it can search safina, read any text in full with everything that says where it came from, and list what safina holds. It can change nothing. And it follows the rule we follow ourselves: safina gives your assistant every text in its own language, never a translation. If your assistant translates a passage for you, it must say that the translation is its own.

Before we opened safina, we asked an AI assistant to try it — and told it nothing about it first. It used every tool correctly, and it found two things we had missed. Some pages join the parts of a word with an invisible character that many programs quietly delete. And one site sets the honorific after a name in small, raised type, so that in plain text it runs straight into the name. Both are kept exactly as the sites publish them — that is safina's first rule — and we are now deciding how a quotation should carry them. It was a good first test of the way we mean to work.

One key, still

safina opens with the same sign-in as daftar, fihrist and diwan. daftar says who you are; safina decides, from its own list of members, what you may do there. Like the others, it is by invitation, and there is still no sign-up form anywhere. If you would like to work with us, contact us and tell us about the work you have in mind.

What we got wrong on the way

safina replaces an archive that we opened, and closed, on the same day. It could keep a copy of a web page, but to show you one it ran the original site's own code inside ours, and parts of it came under a software licence we do not use. That was not a risk we would take, with other people's sites or with our own. We closed it that evening and built safina instead. Nothing was lost: the archive had not yet collected what we needed.

And on the morning safina went live, we found that the page its visits point a website's owner to — the page that says who is reading their site, and why — did not yet exist. By the afternoon it did. It is exactly the kind of thing we said we would tell you.

What is next

The recordings. Many of these lectures exist as recordings. Next, safina will keep each recording beside its text, with a player on the page.
Translation. Later, members will be able to read a passage in English, French, Spanish, Italian or German, with honorifics in full and every verse of the Qur'an in Ali Quli Qarai's translation — each marked clearly as a translation, and never kept in place of the original.
More voices, from the offices whose permission we hold.

And further ahead: when the student we are sending to the hawza has finished its training, safina will be one of the libraries it reads. When it explains a ruling, we want it to be able to take you to the lecture where the scholar explained it himself — in his own words, exactly as he said them, with where and when.

Keep what was said. Say where it came from. Change nothing.