אִם יִרְצֶה הַשֵּׁם
ה׳ תִּשְׁרֵי תשפ״ז
Thesis: Talmudic hermeneutics is not the search for a satisfying assignment, and it is not the construction of a structure-preserving map. It is the generation of readings scored against a ranked list of constraints, in a system where an inference can be retracted and an open kashya is a legal result.
The rest of this post is only that sentence, unpacked on one amud.
The page that forces the question
Shabbos 155a–b is not a middot page. There is almost no גזירה שוה gezeira shava, no היקש heqqesh, no כלל ופרט kelal u-frat. What the Gemara is doing is more primitive, and therefore a better test of slogans.
The Mishnah says: untie these bundles, shake those, do not touch the third. Do not crush שחת shachas or carobs, except that R. Yehuda allows carobs for a דקה daka. Do not pack a camel like a barn. Do not do המראה hamra’ah to a calf; הלעטה halatah is allowed. Feed chickens this way; do not feed cote-doves that way. Pour water on מורסן morsan (bran); do not knead.
Three of those words — כיפין kifin, זירין zirin, hamra’ah — do not come with a picture. Two hundred years after the barn, nobody can point. The Amoraim have to assign them.
Rav Huna and Rav Yehuda do not argue about what the Tanna wrote. They argue which physical bundle each word names, because each of them already holds a sevara:
- Rav Huna: extra bother on food that is already food is allowed; turning a non-food into food is not.
- Rav Yehuda: turning a non-food into food is allowed; extra bother is not.
Each one tags the words so the Mishnah’s yes/no pattern lands on his distinction. They swap the whole dictionary at once. Then they drop the next line on the model and see what breaks.
On 155b the same machine runs again. Hamra’ah / halatah is either a split on depth (past the point the animal can spit) or a split on tool (hand versus kli). A Baraisa about chickens and cote-doves is held up as a mirror. The mirror only reflects if you also decide whether you may even throw grain in front of a bird whose food is not upon you.
That is the object. Not a verse. A short list of verdicts, underspecified names, and a demand that one distinction sort them all.
What SMT gets, and where it stops
Satisfiability modulo theories is a fair first picture of the hunt.
The Mishnah supplies clauses. A sevara supplies a background theory. A dictionary is an assignment of values to the underspecified names. You look for an assignment that makes every clause true in the theory.
That is why both Amoraim retag all three bundle-words together. A local retag would leave a clause false. It is also why the bird-Baraisa is a test and not a tour: it is a new clause dropped on the same variables.
The theory-part of SMT is doing real work. “Comfort is cheap, creating food is expensive” is not a clause. It is the module the clauses are interpreted in. Change the module, and the same string of words gets a different legal assignment. That is the whole machlokes.
Then the picture dies.
SMT’s success criterion is satisfaction. Rav Huna can satisfy the carob-line by rebinding daka. A line earlier the Mishnah said בין דקה ובין גסה — small and large. Two lines later he wants daka to mean the animal that dayka, that grinds, i.e. a large animal. The sentences can be made true. The Gemara still will not take it. The objection is not “unsat.” The objection is: you already used that word.
SMT also wants a winner. This amud leaves Rav Huna with a standing kashya, parks flour-versus-bran for tomorrow, and does not kill Rav Yehuda because his model is ugly on a different line. “Sat / unsat” has no name for that outcome.
And SMT treats the background theory as fixed while you search. Here, when the assignment fails, they sometimes switch sevara rather than switch tags. The module is itself under revision.
So: good name for the hunt. Bad name for the refusal of a legal patch, and for the right to stop.
What category theory gets, and where it poses
The demand that one distinction sort every word at once is a coherence demand. That is the part category theory names cleanly.
There are three maps:
- The Mishnah: words and acts → {mutar, asur}.
- A sevara: kinds → {mutar, asur}.
- A dictionary: the Mishnah’s words → those kinds.
The triangle is supposed to close. After you translate the words, the sevara should reproduce the Mishnah. An Amora is not free to tag zirin as twigs and then treat “shake zirin” as comfort. The whole dictionary has to move together. That is why the swap is total.
The refusal to rebind daka is the same demand in time. A functor that sends an object to “small” on line one and to “grinder” on line three, with no notice, is not natural with itself. The Gemara does not say “naturality.” It says the first clause already taught you what the word means.
That is what category theory earns.
Then it starts to pose.
In a category, morphisms compose. ⚘ A kal va-chomer on this page is composition: if creating food is allowed, then all the more so comfort. The Gemara does not grant the composite. That refusal is the kashya on Rav Huna. A formalism in which composition comes free cannot represent the moment composition is put on trial.
Category theory also does not describe the walk. They do not compute a unique functor and halt. They run one factorization until a square fails, then they run the other. They add an object (אין מזונותיו עליך ein mezonotav alecha) to save a triangle. They leave two factorizations standing. Unique-up-to-iso is not the epistemic aim.
Limits, adjunctions, Yoneda add nothing on this amud. There is no universal property sitting in zirin. There is a short list of acts and a demand that one distinction sort them.
So: good name for the coherence demand and for “you already used that word.” Bad name for the search, for retractable inference, and for an open node.
The object is four jobs
Slogans fail because they flatten four different acts into one.
Assign underspecified words.
Kifin and zirin do not denote until someone tags them. Hamra’ah does not denote until someone says whether the split is depth or tool. A formalism that starts with a fixed signature has already skipped the first job.
Keep one distinction in charge.
Not a pile of local labels. One sevara has to sort the list. That is the pressure that makes the dictionary move as a block, and the pressure that makes a later line a test rather than a new subject.
Defeat an inference.
“All the more so” is licensed, then killed by “only for a daka.” Defaults die. Ordinary deduction has no such object. A category in which composites are free has no such object either.
Stop without a winner.
A standing kashya is not an incomplete computation. It is a legal state of the page. Flour-versus-bran, parked, is a legal state of the page. Any mathematics whose terminal object is “the model” will treat those states as bugs.
Classic hermeneutical middot are a fifth job. They have side-conditions: a gezeira shava that is not מופנה mufneh, two verses that teach one thing and therefore do not teach a third. They barely run on 155. A theory of “Talmudic hermeneutics” that leads with the middot will misdescribe this amud. A theory that cannot host the middot when they do appear will misdescribe Pesachim 45a. The jobs are not interchangeable.
What does fit the walk
Not one named theory. A family: defeasible, ranked, argumentative.
Candidate readings, ranked constraints.
Generate readings. Rank what you are willing to violate:
- Do not flip an explicit din of the Mishnah.
- One sevara sorts the whole list.
- Do not remap a word you already used on this amud.
- Do not invent a mouth or a plant the Tanna did not put here.
Rav Huna’s daka = grinder satisfies (1) and (2) only by violating (3). They will not demote (3). That is the meaning of “forced.” The sentences were satisfiable. The ranking refused the repair.
Two Amoraim are two rankings, not two languages. Rav Yehuda will spend more to keep (3) and will pay for it on a different line, where a soft food looks like comfort. Rav Huna will spend more to keep his sevara uniform and will pay for it on daka. The page is the bill.
Those three pieces — ranked constraints, dying defaults, typed attack — are a family, not a brand. After a pass on the amud, the refusals.
A worked pass on the amud
Start with Rav Huna’s ranking: (2) is “comfort cheap, create-food expensive.” Tag hay as the words that got yes, twigs as the word that got no. Clauses one through three sit.
Drop “do not crush shachas or carobs.” If both are soft, crushing is comfort, and he has forbidden what his sevara allows. Repair: both are hard, and the mouth is a young foal’s. Constraint (4) is strained — the Tanna did not write “foal” — but the repair is tolerated because the plant-names can carry hardness.
Drop “R. Yehuda allows carobs for a daka.” On his sevara the small mouth is creating food (should be asur) and the large mouth is comfort (should be mutar). The Tanna allowed only the small one. The default composite — “all the more so the large” — is now an attack. Repair: rebind daka as grinder. That repair violates (3). Refused. Standing kashya.
On 155b, Rav Yehuda’s depth-split is attacked by the bird-Baraisa under one reading of מהלקטין mehalketin. He does not rebind the calf-words. He adds an object the Mishnah already had in the water-line: some animals’ food is upon you, some is not. Cote-doves route through that object. Constraint (4) is not violated; the Tanna had already split bees and chickens. The triangle closes by revealing an object, not by remapping a word.
That difference is the whole method. Revealing an object the text already split is cheap. Remapping a word the text already used is expensive. SAT cannot tell those repairs apart. A functor that is allowed to send the same object anywhere also cannot.
Generator and evaluator
The usable split is not “which branch of mathematics.” It is who is allowed to speak.
A generator may guess a dictionary and a sevara. An evaluator may only point at a line and say: this ranking just lost. The evaluator may not invent a young foal to save the generator, and may not rebind daka in silence. That is already how the Gemara talks. It is why a kashya can stay a kashya overnight.
Mathematics that cannot leave a node open, cannot retract a kal va-chomer, and cannot treat “ugly but sat” as a loss, will fail in the same place SMT and category theory failed. The failure is not that the diagrams are wrong. The failure is that the success criterion is wrong.
Hermeneutics on this page is not “find the meaning.” It is “keep scoring readings until the next line, and do not pretend the score is a proof.”
What this is not
A family of tools is easy to sloganize. Three slogans are already waiting, and each one is the wrong success criterion with a better vocabulary.
It is not “Talmud = optimality theory.”
Prince and Smolensky packed ranked constraints into a linguistics machine: GEN writes candidates, CON is a hierarchy, EVAL picks the least-bad survivor. That shape is close to what this amud does with readings. It is still one packaging, from one science, built to pick a unique output form. The Gemara does not owe that machine a winner. It will leave two rankings on the table and go to sleep. OT’s EVAL is a judge. The Gemara’s evaluator is only allowed to say no. Collapse those and you have renamed the Torah after a phonology paper.
Useful picture. Not the name of the Torah.
It is not “Talmud = Dung argumentation.”
Dung’s frameworks give you arguments, attacks, and preferred extensions. That is the right direction for the walk: a kashya attacks, a terutz repairs, two sheitos can both be preferred. The slogan dies on the type of the attack.
On this page the attacks are not interchangeable edges.
- “This line contradicts your sevara.” Soft carobs, if they are already food, forbid the comfort Rav Huna wanted to allow.
- “This line remaps your word.” Daka already meant small. It cannot mean grinder two lines later.
- “This Baraisa is only a proof if throwing is asur.” The bird-pair mirrors the calf-pair only under one reading of mehalketin. Change that reading and the attack misses.
An untyped attack-graph smears those together. Then “preferred extension” looks like a result. It is only a pile. The hygiene of the amud is the type, not the edge.
It is not the Gabbay–Abraham–Schild papers as a finished answer.
Michael Abraham, Dov Gabbay, and Uri Schild built a logic that already has the right floors: object-level din, higher-level rules for handling conflict, contrary-to-duty (“you should feed, you should not force”). That is the academic line that is walking toward this family instead of toward SAT or functors. Credit the direction.
They still clean the walk more than 155 does. Their systems want the conflict handled. This amud wants some conflicts named and left. Rav Huna’s daka is not compiled into a higher-level exception. It is a kashya. Flour versus bran is not compiled into a contrary-to-duty. It is parked. A finished “Talmudic logic” that always returns a handled state has the same wrong terminal object SMT has. The papers are a research program. They are not the page.
Three smaller refusals, so they are not smuggled back in.
The middot are not ranked constraints. When a gezeira shava runs, its side-conditions are their own system. Do not flatten them into CON.
An evaluator that only says no is a description of the Gemara’s hygiene. It is not a machine that should pasken.
The page is not the proof that one Amora has the better ranking. The page is the argument over the ranking.
The bran line at the end of 155b is the same hand over a bowl instead of a throat. Pouring is still feeding. Kneading has crossed into גיבול gibel, a melachah. The Gemara asks whether pouring into bran is already that melachah, or whether bran never pastes, and then it parks the question.
That park is the thesis in miniature. The sentences could be forced to sit tonight. They leave them unsettled, because the cost of the sitting has not been ranked yet.
Open kashya
When the evaluator parks a line — is that a refusal to finish, or is “not yet ranked” already one of the legal objects the system has to contain?