The most underrated sefer I own

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

I bought Derech Tevunos because Ramchal wrote it. I learned Mesillas Yesharim, like everybody. I learned Derech Hashem. This one I opened twice, filed it as the logic one, and put it back on the shelf, where it has been keeping company with a lot of other people's copies.

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 distinctions a subject can be examined along.

Chapters 7 and 8 say what each move does to the claim it hits, and that is the part I would put in front of anybody.

A סתירה shows a claim is false, and the claim is rejected. A דחיה shows only that the proof could be read another way, and the claim drops back to doubt. A פירכא never rejects anything at all; it unsettles, and unsettling is the whole of what it is permitted to do. Doubt, for Ramchal, is where every claim starts and where it returns the moment its support is taken away.

Knocking down a proof does not prove the negation. He says it flat, in chapter 8, and almost nobody's instincts run that way. A sugya is full of moves that feel like victories and are not, and he is the one who tells you which are which.

Or take his handling of אם. A conditional asserts the dependency and nothing else, so a statement whose halves are both true but unrelated is false, and one whose halves are both false is true when the dependency is real. Material implication calls the first of those true. Ramchal calls it false, in 1742, and he is the one describing how a sugya argues.

Why nobody runs it

The procedure is sound, and the procedure is 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, and I want to be precise about which half.

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 nearest neighbour. An answer outside the vocabulary fails validation and the file never draws. That is clerical work against a closed list, which is the one kind of work these things are reliable at.

So I ran it. There are 9 passages on this site, 5 of Ramchal's own examples and 4 opening sugyot. For each one a model proposed the labelling and I checked it by hand, sentence by sentence. It got things wrong. I have no measured error rate for you and I am not going to invent one.

What made the checking survivable is the drawing. A sentence labelled wrong takes the picture with it: a resolution floating with nothing above it to resolve, a rail running to the wrong row, a fold closing over an argument that never settled. You catch it before you have finished reading the row.

The drawing is doing more work than the labels

Here's the thing I would rather not say out loud.

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. Anybody promising otherwise is selling something.

Ramchal opens the sefer by asking that we look at the claims and not at the ones making them, and then spends 11 chapters teaching a person how to do that. He wrote a manual. Manuals are for using.

The ask

The passages are drawn. The sefer is here, Hebrew and English, all 11 chapters, and chapter 9 is the vocabulary the drawing reads. A file you label yourself gets drawn by the same code these were.

If you find a sentence I labelled wrong, I want to hear about it. Every label on this site is a claim about a sugya, and a claim about a sugya is something I am allowed to be wrong about.