The most Underrated Sefer I Own
Derech Tevunos is a specification, and the work just got cheap
In short
Derech Tevunos is a procedure: closed vocabularies, a test for each label against its nearest neighbour, and a fixed rule for what every move does to the claim it lands on. Running it by hand costs hours per page, which is why a sound method sat unused. That cost is what changed, and with this site i hope to show it.
4 sentences of Aramaic in Shabbos 5b.
12 steps underneath them, and not one of the 12 is written in the Gemara.
Ramchal counts them out in chapter 10: two claims, four syllogisms, one elucidation, reconstructed line by line until every premise lands on something nobody argues with. Then he says the thing that has bothered me for years. Leave one step out and you have not understood the argument.
He is right, and look at what being right costs. Four sentences took him pages. The longest passage I labelled for this site runs 77 sentences, and that is two dapim of Bava Kamma.
The Sefer I own and did not learn
If anybody was equipped to finish this one it was me. I write software, I like logic systems the way other people like crosswords, and I went through the whole Semantic Web era like it was a parade: RDF, OWL, description logics, triple stores, ontologies, knowledge bases. I have owned Derech Tevunos for 10 years and I have not finished it.
I bought it because Ramchal wrote it. I learned Mesillas Yesharim, like everybody. I learned Derech Hashem. This one I have opened dozens of times, and dozens of times I got a few pages in and put it back on the shelf shortly afterward.
A dense treatise of pure logic with nothing to look at is very hard to process, and there is not one diagram in the sefer. Those 12 steps under the 4 sentences of Shabbos 5b are a picture, and they arrive as prose, in order, each one leaning on something a paragraph back. So you draw it yourself, in your head, on every page, and it goes when you close the book. Every attempt ended the same way: I could tell you the method was sound and not what it had just proved.
What is in it
Derech Tevunos is a specification.
Chapter 9 takes every sentence a sugya can contain and gives it a name: 7 elements, 19 leaves, each one a thing a sentence does to a sentence before it, each with a test that separates it from the label sitting next to it. Chapter 3 says that whatever the surface was, a question, a flourish, an ellipsis you finish in your head, the claim underneath is a subject and a predicate in one of 11 manners, and you put it in that form before you do anything else with it. Chapter 4 says how two statements can stand to each other, including the case where two Tannaim look like they are arguing and are not. Chapter 11 gives 24 הבחנות, the d
Chapter upon chapter of prose builds up a taxonomy that can classify every sugya in existence because the Ramchal can hold the entire Gemara in his head at the same time. Once you can classify a sugya you can build visuals on top of it so logic flow is a thing that can be eyeballed. Choose a classification system that is format enough and the visuals can have mechanics, and mechanics mean demonstrability via playing around. That was the dream!
Why nobody runs it
The procedure is sound, but too expensive.
Bava Metzia 21b to 22b runs 57 sentences. Abaye states a claim in the first one, it comes under attack 10 separate times over the next 33 rows, and the support that finally lands on it sits in row 51. Labelling that by hand, every label naming the row it acts on, is an afternoon. Then you have one page.
There is a second problem sitting under the first one. A Talmid Chacham already does most of this in his head, fast, with no names for any of it. The person equipped to produce the labels has no use for them, and the person who would gain from them cannot pay for them. A method whose only qualified users have already outgrown it does not go anywhere.
What changed
The expensive half got cheap.
The question a model answers here is narrow. What does this sentence do to the sentence before it? The answer has to be one of 19 labels, it has to name the sentence it acts on, and it has to survive Ramchal's own test for telling that label from its siblings. An answer outside the vocabulary fails validation and the file never draws. That is clerical work against a closed list, which is the kind of work LLMs are quote reliable at.
So I ran it. There are 9 passages on this site, 5 of Ramchal's own examples and 4 opening sugyot. It got things wrong, or at least inconsistent which is understandable for LLMs. More importantly, the original translators of Derech Tevunos themselves explain that parsing a sugya can never and will never be absolutely deterministic. We've already accepted all of this decades ago.
What makes it categorically better today is the drawing. A sentence labelled wrong takes the picture with it: a resolution floating with nothing above it to resolve. You catch it before you have finished reading the row.
The drawing is doing more work than the labels
Here's a fact of life: Very few people are interested in the deep analysis of dialectical process. A great many people are interested in a well-made drawing with a good icon set. I am on the third generation of the icon set, and that is where the reach is.
If this sefer gets used, it will be because somebody saw the shape of an argument, found it beautiful, and got curious about what produced it.
What I am not claiming
The machine sorts sentences into categories a person specified. It does not learn, it has no Yiras Shamayim, and it replaces no Rebbi, no Chavrusa and no Beis Medrash. The purpose of all of this is as an aid. Nothing more.
The site
The homepage opens on a passage instead of an index, because the drawing is the thing. It is 4 sentences from Pesachim 16a and 18b, the shortest of the examples Ramchal works through himself: R. Eleazar says liquids cannot become unclean by Torah law, a proof from Yosi ben Yo'ezer's testimony lands under it, and Rav Huna answers. Ramchal labels two of those four himself, in chapter 9, so the drawing of this one can be checked against the sefer that specified it.
The drawing is called Sugyascade, which is the link in the nav: sugya plus cascade, because a cascade descends over a series of ledges, which is the shape, and because each stage of one is set off by the stage above it, which is the argument.
That reads fine at 4 sentences. The opening sugyot run to 36, 57 and 77, and there the sentence a move acts on can be 30 rows above it, which is what the rail and the fold are for.
A rail runs up the gutter rather than through the rows, in a colour no verdict uses, so following one is never confused with the red of a refutation or the amber of a doubt. A fold keeps the result of the stretch it closes over, which is what makes it safe to read past. The bar above the rows holds the state of play, every claim and where it currently stands.
Nothing below the frontier is readable: a row or two of blur, then nothing rendered at all.
The control beside the rows moves that frontier one sentence at a time, which is how a sugya arrives when you learn it, and it is the only way I have found to read a passage I have not seen before without the ending spoiling the middle.
/sugya has all 9, and /docs has the sefer, Hebrew and English, all 11 chapters, with chapter 9 as the vocabulary the drawing reads.
/byo draws a file you labelled yourself. Drop it or paste it and it is parsed and drawn by the code that draws these 9, so what the page accepts and what the site ships are the same format by construction. Nothing is uploaded: the file lives in your tab and the parse runs there. A file it refuses prints every fault at once, each with its path in the JSON.
If something is broken, or a label looks wrong to you, I want to hear about it. I am @deusaquilus on X and aioffe on LinkedIn.