Jew.tech  /  Sugya Protocol

The Sugya Protocol

Modern AI knowledge tools treat knowledge as answers: they retrieve sources, emit a narrative or a single probability, and discard the reasoning. The Sugya Protocol treats knowledge as structured disagreement — an append-only argument graph in which nothing is ever deleted, every claim carries its provenance as part of the object, both sides of a dispute keep their standing, and an honest unresolved question is a first-class output rather than a failure.

The protocol is the methodology set out in The Talmudic Epistemic Stack, a submission to the Future of Life Foundation Epistemic Stack Competition. Its argument is narrow and testable: the problem of building knowledge bases that are trustworthy, updateable, adversarially robust, and compounding was substantially solved — for legal and intellectual disputes — by the editorial method of the Talmud, and that method is now mechanizable using Dov Gabbay's formal logic programme.

What a sugya is

A sugya (סוגיא) is the Talmud's unit of argument: a self-contained passage that poses a question, raises the strongest objection against the proposed answer, answers the objection, draws distinctions where two sources appear to contradict, and arrives at either a ruling or an acknowledged impasse — with every one of those moves left visible on the page. It is not a summary of a debate. It is the debate, stored.

In the protocol, a Sugya is the primary object: an append-only argument graph that stores the moves, not a bare answer. Where a retrieval-augmented system produces fluent prose whose provenance is a list of links and whose reasoning is not a manipulable object, a Sugya produces a graph that a later investigator can attack, extend, re-run, and revise without starting over.

The deliverable is the graph and the method, not an answer; the answer, in the Talmudic sense, is the preserved structure within which future investigators continue the argument rather than restart it. The Talmudic Epistemic Stack, §11

Four properties of a knowledge commons

Read as an information system rather than as scripture, the Talmud has four structural properties — each of which is a requirement that modern knowledge-base design keeps rediscovering.

1. Preserved disagreement — machloket מחלוקת

Minority and rejected opinions are recorded by name, with their reasoning, beside the accepted view. The Mishnah says explicitly why (m. Eduyot 1:5–6): so that a later court may rely on them. This is non-destructive editing of a contested base — and it is precisely what argument-mapping tools that drive toward a single winning position cannot do.

2. Layered provenance

Each stratum cites and argues with its predecessors. A claim's authority is a function of who said it, when, and in what chain of transmission. Provenance is the primary key, not bolted-on metadata.

3. Append-only revision

New rulings are added as commentary literally surrounding the original; nothing is deleted. The Talmudic page (daf, דף) is a working interoperable, multi-perspective format that compounded across centuries.

4. Honest non-resolution — teiku תיקו

Unresolvable disputes are recorded and retained as valuable, citable outputs. Faithful uncertainty is an editorial norm, not an admission of failure.

Why treat a fifteen-century-old format as engineering evidence? Because longevity under adversarial pressure is itself the argument. The method survived dispersion, censorship, and continuous internal red-teaming — its entire genre is objection and counter-objection — while remaining coherent and extensible. That is an observation about a specific epistemic technology's demonstrated durability, not a metaphysical claim.

The node types

A Sugya graph is typed. Each node kind corresponds to a move the Talmud already had a word for, and each carries a distinct role in the assessment.

Seven operations mutate that graph, and each one is backed by a published logic rather than by convention:

The invariant: nothing is deleted

One design decision carries the rest of the protocol.

Revision is supersession, not overwriting. The superseded node stays in the graph with a closed validity interval, and a REVISES edge connects the new version to the old one. The prior state remains available for provenance; the current state is whatever is valid now. Formally this is AGM belief revision — the pair implements contraction-then-expansion while retaining the earlier state — and the independence and factorization postulates keep the change local, so revising one subgraph does not perturb independent parts. That locality is what lets the scheme scale: an append-only store whose every update forced a global recomputation would not survive contact with a real corpus.

Every node carries its label

Provenance is not metadata sitting beside the assertion. It travels with it, as part of the object the inference manipulates — Gabbay's Labelled Deductive Systems applied directly. Each node carries:

L = (truth, completeness, sources, priority)

together with a talmudic_status drawn from {accepted, disputed, minority, majority, practical_ruling, unresolved, suspended}.

Because the label carries both priority and source, the same machinery that ranks a Rishon against an Acharon also ranks a peer-reviewed paper against a preprint against an anonymous web claim. The label is the trust computation.

Where to look next: the crux ranking

A Sugya does not only record what is known. It scores what is worth attacking, so the output tells the next investigator where effort buys the most movement in belief:

crux priority = leverage × uncertainty / (tractability + 1)

Three case studies

COVID-19 origins — an open, adversarial dispute

The headline number is the argument: six competent Bayesian analyses of the same evidence span roughly twenty-three orders of magnitude. A system that emits one probability has not represented the state of knowledge — it has hidden it. The Sugya keeps both positions as a Machloket, records the 2023 Rootclaim debate's 2–0 verdict for zoonosis as a separate practical_ruling node — distinguishing what was settled from what merely performed settling — and reduces the dispute to ranked cruxes: market clustering versus ascertainment bias, the two-lineage A/B argument, the furin cleavage site, DEFUSE. The witness-assessment layer stops the many papers built on the same market data from counting as independent confirmations, and typed safek shows that the 23-order spread is model uncertainty in the priors, not disagreement about the data.

LHC micro black holes — a closed case, defence in depth

Here the Sugya encodes three independent backstops — Hawking evaporation, then cosmic-ray survival, then white-dwarf and neutron-star survival — as nested Terutz nodes, each catching the failure of the layer before it. What makes the encoding honest is that it marks its own weak points as Safek: Hawking radiation has never been observed, so the safety case is deliberately constructed not to depend on it, and the neutral slow-accretion corner is flagged as the least closed. A structure that records where it is weakest is more trustworthy than one that reports only its conclusion.

Dietary cholesterol and eggs — a messy, individual question

The central operation is chiluk. Five distinctions dissolve the apparent contradiction between cohort studies finding harm and cohort studies finding none: dietary cholesterol ≠ serum LDL; average ≠ individual (hyper-responders, ApoE4); surrogate ≠ hard endpoint; association ≠ causation; LDL-mediated ≠ TMAO-independent. Once the distinctions are drawn, the grounded extension correctly leaves both headline claims undecided — a preserved Machloket plus Teiku — and the output is a logged, conditional bottom line rather than a false universal verdict.

Teiku, applied to itself

The most unusual section of the paper is §10, in which the objection operation is run against the paper's own thesis and whatever survives is logged.

The strong version of the claim is refused outright: “Judaism solves all the issues” is unfalsifiable; we do not defend it. The provenance of the survival argument is itself checked, and found to be articulated in the independent literature rather than being the documented position of any single lab — so it is demoted to “a motivating hypothesis with independent support,” and the case rests instead on the formal logic and the case studies, which stand on their own. The corpus indices were unreachable from the build sandbox, and the paper says so rather than implying live results: “Logged, not hidden.” Adversarial verification caught real errors before publication — a mistaken distance figure, a wrong arXiv identifier, a misstated equation form, incorrect imprints — and the remaining hedged items are marked Safek.

Then three questions are recorded as teiku — open, not papered over:

This section matters more than the results. A methodology that claims to preserve disagreement, register uncertainty, and record honest non-resolution is only credible if it does those things to itself. Section 10 is the protocol applied reflexively: it raises the strongest objections to its own thesis, narrows the claim until it is defensible, discloses its own gaps, and leaves three questions open rather than resolving them by assertion. That is the demonstration — not the case studies.

Against the prior art

Argument mapping, defeasible argumentation, and provenance standards each solve part of this. None of them combines the parts.

Four things none of the above combines: preserved disagreement as standing, citable authority; non-monotonic updateability with AGM guarantees; fibred multi-logic verdicts with no forced scalar; and provenance-as-authority, where the label is the trust computation.

Read the paper