Jew.tech  /  Sugya Protocol  /  Paper

The Talmudic Epistemic Stack

Structured Disagreement as a Methodology for Trustworthy, Updateable, Adversarially-Robust AI Knowledge Bases

Joshua Lazoff · Submission to the Future of Life Foundation Epistemic Stack Competition (“Lab Leaks, Black Holes, and Eggs”) · v1.0 · 2026 · CC-BY-4.0

Worked examples, full proofs, schemas and the 987-paper corpus index are appendix material in the accompanying repository (epistemic_stack/); curated entry points in its README.

1. The problem and the thesis

FLF seeks “workflows and methodologies for using AI to produce reliable, trustworthy knowledge bases” that compound: “if structured analyses become reusable, refineable artifacts, every serious investigation enables future work … by the same or different people.” The dominant AI paradigm fails this almost by definition. Retrieval-augmented generation yields a fluent narrative whose provenance is a list of links and whose reasoning is not a manipulable object. A single Bayesian probability is worse in a way the competition deliberately stresses: as we note in §7, six competent Bayesian analyses of the same COVID-origins evidence span twenty-three orders of magnitude. A system that emits one number has not represented the state of knowledge — it has hidden it.

Our thesis: the problem FLF poses was substantially solved, for legal and intellectual disputes, by the editorial method of the Talmud, and that method is now mechanizable. The Talmud is a format and a process, not merely an old text. Its defining move is to preserve the structure of disagreement — question (kushya), objection, answer (terutz), distinction (chiluk), and machloket: a disagreement recorded with both sides, their proponents, and their support, never flattened to a winner even when a practical ruling issues. It records honest non-resolution (teiku) as a first-class output. It is layered (Mishnah → Gemara → Rishonim → Acharonim); each layer cites and argues with the previous without overwriting it — an append-only, provenance-preserving, adversarially-edited knowledge commons maintained across fifteen centuries and a globally distributed contributor base. The bridge from “good format” to “buildable system” is Dov Gabbay's formal program.

2. What FLF is buying (the criteria)

FLF splits the task into three layers — ingestion (attribute claims to sources with provenance), structure (inference + discourse structure; “similar but not identical” claims; evolution over time), and assessment (rhetorical-vs-evidential weight; correlated evidence treated as independent; cruxes; what's missing; confidence accounting for out-of-model error; distinguish what the debate settled from what it merely performed settling). From these we use eleven dimensions throughout: provenance, structure, assessment, adversarial robustness, scalability, interoperability, updateability, faithful uncertainty, crux discovery, reusability, and settled-vs-performed. The last is the Talmudic competence par excellence: the genre exists to separate a binding ruling (halacha) from the reasoning theater around it, and machloket (live dispute) from teiku (unresolved) from psak (ruling).

3. The Talmud as a knowledge-commons format

Viewed as an information system, four structural properties stand out — each a FLF desideratum:

1. Preserved disagreement (machloket)

Minority and rejected opinions are recorded by name with their reasoning, beside the accepted view (m. Eduyot 1:5–6, explicitly so a later court may rely on them). This is non-destructive editing of a contested base — exactly “distinguish what was settled from what merely performed settling,” and exactly what argument-mapping tools that drive to one winner lack.

2. Layered provenance

Each stratum cites and argues with its predecessors; a claim's authority is a function of who said it, when, 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) is a working interoperable, multi-perspective format that compounded across centuries.

4. Honest non-resolution (teiku)

Unresolvable disputes are recorded and retained as valuable, citable outputs — faithful uncertainty as an editorial norm.

Why a fifteen-century-old format? Longevity under adversarial pressure is evidence of robustness: the method survived dispersion, censorship, and continuous internal red-teaming (its whole genre is objection and counter-objection) while remaining coherent and extensible. This is the defensible core of the “tradition that survived adverse selection may help us survive powerful AI” intuition — not metaphysics, but the observation that a specific epistemic technology has demonstrated extreme durability and is worth porting.

Honest scope. A dedicated adversarial check (§10) found that the specific “Judaism-survived-adverse-selection → useful-for-RSI-AI” thesis is articulated in the literature by others (e.g., Zibulevsky 2024; Kalman/FLI 2023), not as a documented institutional position of any single lab; AE Studio’s public alignment work is the “Neglected Approaches” agenda and its AI-antisemitism research. We therefore treat the survival argument as a motivating hypothesis with independent support and rest the case on the formal logic and the case studies, which stand on their own.

4. Making it computable: Gabbay's apparatus

The Talmudic Logic Project (Gabbay, Schild & David, Logica Universalis 13(4):425–442, 2019) has formalized specific Talmudic rules we reuse: kal vachomer (a fortiori) via matrix abduction (Abraham, Gabbay & Schild, Studia Logica 92(3):281–364, 2009); Talmudic deontic logic (AI & Law 19, 2011); temporal/contrary-to-time conditionals (AI & Law 20, 2012); Talmudic public-announcement logic that deletes accessibility links rather than worlds, i.e. reversible update (J. Applied Logic 11, 2013); the Klal uPrat set rules (HPL 32, 2011); and uncertainty rules — rov (majority), chazakah (presumption) — formalizing safek (Gabbay & Koppel, HPL 32(1), 2011).

Gabbay's general program supplies the rest. Labelled Deductive Systems attach a label (assumptions, sources, reasoning path) to every formula — provenance as a carried object (Labelled Deduction, Kluwer 1996). Abstract argumentation (Dung, Artif. Intell. 77(2):321–357, 1995) models a debate as an attack graph whose extensions are the defensible sets; Gabbay's extensions type each attack by its logical relation, allow attacks-on-attacks (meta-disagreement), add graded support, and require a dialectical tier in which a claim must answer objections before it stands — the kushya→terutz cycle as a soundness condition. Fibring combines logics without reducing one to another — the formal license for “never collapse the dimensions into one score” — and even fibres argumentation frames with Bayesian and neural networks. Non-monotonic/defeasible consequence makes conclusions retractable (the essence of updateability) and AGM belief revision (Frontiers in Belief Revision, Kluwer 2001) gives the rationality postulates for principled updating. Reactive Kripke semantics lets traversing one edge switch others, giving context-aware, self-modifying assessment with completeness theorems.

A real substrate. This provenance algebra runs over a large corpus of Jewish texts the author maintains in Neo4j and ArangoDB — millions of OCR'd English volumes plus thousands of hand-scanned curated works spanning every period (Biblical → Tannaitic → Amoraim → Geonim → Rishonim → Acharonim → modern), with mathematics and psychology holdings that inform the assessment layer. The repository's torah.yaml already encodes the authority tiers (Biblical = 10, Talmudic = 9, Rishonim = 8 …) and Hebrew/Aramaic/Yiddish/English fibring; the same machinery that ranks a Rishon against an Acharon ranks a peer-reviewed paper against a preprint against an anonymous web claim. (The live indices were not reachable from the build sandbox; artifacts are built to load into them — §10.)

5. The Sugya Protocol

The primary object is a Sugya: an append-only argument graph storing the moves, not a bare answer. Node types and their roles:

Invariant: nothing is deleted. Updates set valid_to and add a REVISES edge to a new version; every node carries an LDS label L=(truth, completeness, sources, priority) and a talmudic_status ∈ {accepted, disputed, minority, majority, practical_ruling, unresolved, suspended}.

Seven operations, each a graph mutation backed by a logic: kushya (typed attack) · terutz (defense/reinstatement) · chiluk (matrix-completion distinction) · machloket (paraconsistent preservation) · nafka mina (defeasible consequence) · safek (typed uncertainty: factual/model/source/measurement/adversarial) · teiku (logged non-resolution).

Agents, single-responsibility and deterministically orchestrated (so judges can re-run on a new case): Mishnah (extract claims+provenance, no interpretation) → Gemara (kushya/terutz/chiluk) → Tosafot (cross-link, find repeated assumptions) → Eidim (source authority; mark correlated evidence) → Stam (missing evidence; new questions) → Machloket (reduce to ranked cruxes) → Daat (incentives, meta-level only). Crux priority =

leverage × uncertainty / (tractability+1)

— the assessment payload telling the next investigator where effort buys the most belief-movement.

6. Why the protocol satisfies the criteria (formal core)

Each is a rigorous link to a published result, not a novel theorem (full statements in the repository paper §6).

P1 — Preserved disagreement is safe

Two mutually-attacking arguments yield a Dung grounded extension that accepts neither (both undecided) — natively representing teiku/safek; and paraconsistent restricted-access inference fires classical rules only over consistent subsets, so a contradiction is localized, not globally explosive (ex falso blocked). A practical ruling is a separate node, leaving the minority standing — the Mishnaic norm.

P2 — Updateability = non-monotonic consequence + AGM, as append-only versioning

Adding evidence can retract a conclusion; valid_to + REVISES implement contraction-then-expansion while keeping the prior state for provenance, satisfying Success/Consistency; independence/factorization makes minimal change local, so revision in one subgraph does not perturb independent parts — the property that makes updateability scale.

P3 — Chiluk dissolves false contradictions by completion, not fiat

In the case×attribute matrix (matrix abduction), an apparent “A ∧ ¬A” is resolved by exhibiting a distinguishing column, re-indexing to A|c, ¬A|¬c — both consistent; “same unless distinguished” is indiscernibility by default, the same rule the repo's entity-resolution uses.

P4 — Faithful uncertainty is a fibred vector, and interoperability is provable

Fibring keeps classical/defeasible/temporal/authority/probabilistic verdicts as distinct components (probabilistic via PrAF; calculus may be probability, possibility, or Dempster–Shafer, distinguishing ignorance from risk). And Dung's complete extensions coincide exactly with three-valued stable models of logic programs (Caminada & Gabbay) — so the same accepted/rejected/undecided verdict is recoverable whether a future investigator runs a debate-graph tool or a rule engine.

P5 — Eidim de-correlation prevents double-counting

Independence/factorization plus substructural (resource-counting) proof underwrite a CORRELATED_WITH {ρ} discount, so two dependent sources do not contribute as two independent ones — FLF's explicit ask.

P6 — Reactive accessibility gives context-aware assessment with a sound/complete basis

Reactive Kripke models (edges that switch other edges), with completeness theorems and a proven correspondence to argumentation networks, back a graph whose accepted set is recomputed as context and time change.

7. Worked examples (full Sugyot in the appendix repository)

COVID-19 origins (open, adversarial)

The 2023 Rootclaim debate ended 2–0 for zoonosis, yet six Bayesian analyses of the same evidence span ~23 OOM — proof that one number hides the state of knowledge. Our Sugya keeps both positions as a Machloket, records the judges' verdict as a separate practical_ruling node (settled-vs-performed), and reduces the dispute to ranked cruxes (market clustering vs. ascertainment bias; the A/B two-lineage argument; the furin cleavage site; DEFUSE). The Eidim discount stops the many shared-data market papers from counting as independent confirmations; typed Safek shows the 23-OOM spread is model uncertainty (priors), not data. (Citations: Pekar et al., Science 2022; Worobey et al. 2022; Weissman 2024; reply arXiv:2405.08040; Alexander, ACX 2024.)

LHC micro black holes (closed, defense-in-depth)

The Sugya encodes three independent backstops — Hawking evaporation → cosmic-ray survival → white-dwarf/neutron-star survival (Giddings & Mangano, Phys. Rev. D 78:035009, 2008) — each a Terutz catching the previous layer's failure, and makes the weakest links explicit as Safek: Hawking radiation is unobserved (so the case is built not to need it), and the neutral/slow-accretion corner is least-closed. (LSAG, J. Phys. G 35:115004, 2008; Hawking 1975; Dimopoulos & Landsberg, PRL 87:161602, arXiv:hep-ph/0106295.)

Eggs / dietary cholesterol (messy, individual)

Central work is chiluk: five distinctions dissolve the apparent contradiction between harm cohorts (Zhong, JAMA 2019) and null cohorts (Drouin-Chartier, BMJ 2020; PURE, AJCN 2020) — dietary cholesterol ≠ serum LDL; average ≠ individual (hyper-responders/ApoE4); surrogate ≠ hard endpoint; association ≠ causation; LDL-mediated ≠ TMAO-independent. The grounded extension leaves both headline claims undecided (correct: preserved Machloket + Teiku), with a logged, conditional bottom line rather than a false universal verdict.

8. Self-audit against the criteria

9. Relation to prior art

Four things none of the above combines, which this approach adds: (1) preserved disagreement as standing, citable authority; (2) non-monotonic updateability with AGM guarantees; (3) fibred multi-logic verdicts (no forced scalar); (4) provenance-as-authority (the LDS label is the trust computation, ranked on the same scale as a fifteen-century corpus).

10. Limitations and adversarial self-review (kushya on ourselves)

Overclaim

“Judaism solves all the issues” is unfalsifiable; we do not defend it. We defend the narrower claim (abstract) and the title's “solves” is shorthand for “supplies a battle-tested pattern that meets.”

Is the framing load-bearing or ornament?

§6 shows each operation maps to a distinct mechanism a generic argument map lacks (chiluk = matrix completion; machloket = paraconsistent non-explosion + Dung undecided; teiku = typed output; authority = LDS label on a real corpus). Residual safek: one could perhaps reach the same constructs without the vocabulary — we agree the constructs are what matter; the vocabulary is a guide to which to include.

Attribution

The Judaism-survival thesis is attributed to the independent literature, not a lab (§3); verified.

Corpus not queried live

The Synology Neo4j/ArangoDB indices were unreachable from the build sandbox; schemas/tiers/graphs are built to load into them, and the 987-paper logic corpus was mined directly. Logged, not hidden.

Case-study facts

Adversarial verification caught and fixed real errors pre-publication (market–WIV distance >10 mi not 6; Dimopoulos–Landsberg arXiv id; Keys-equation form; Gabbay–Woods imprints); remaining hedged items are marked Safek.

Teiku (open questions, recorded not papered over)

(i) Does preserving disagreement serve a decision-maker who must act, or abdicate? We give an Executive-Summary interface with a bottom line and the structure, but the trade-off is real. (ii) At what graph size does Sugya readability degrade, and what summarization avoids re-flattening? (iii) Can the fibred verdict be made legible to non-logicians without collapsing it?

11. Conclusion

FLF wants knowledge bases that are trustworthy, updateable, adversarially robust, and compounding. The Talmud is the longest-running working example of such a system; its discipline — preserve the structured disagreement, never flatten it, revise by adding not overwriting, record honest non-resolution — is what makes knowledge compound across contributors and centuries; and Gabbay's program makes that discipline computable. We instantiated it as the Sugya Protocol over a dual-database substrate, connected each criterion to a published result, and demonstrated it on an open dispute, a closed case, and a messy empirical question. 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.

Key references

Gabbay, Schild & David (2019), Logica Universalis 13(4):425–442 · Abraham, Gabbay & Schild (2009), Studia Logica 92(3):281–364 (matrix abduction / kal vachomer) · Abraham, Gabbay & Schild (2011), AI & Law 19; (2012) 20 · Abraham, Belfer, Gabbay & Schild (2013), J. Applied Logic 11(1):63–90 · Gabbay & Koppel (2011), Hist. & Phil. of Logic 32(1):63–69 · Dung (1995), Artif. Intell. 77(2):321–357 · Caminada & Gabbay (2009), Studia Logica 92(3) (complete extensions ≡ 3-valued stable models) · Gabbay (1996) Labelled Deduction; (1999) Fibring Logics; (2008) Analysis and Synthesis of Logics · Gabbay & Rodrigues, eds. (2001) Frontiers in Belief Revision · Gabbay (2013) Reactive Kripke Semantics.

COVID: Pekar et al., Science 2022 (abp8337); Worobey et al. 2022; Weissman 2024; reply arXiv:2405.08040; Alexander, ACX 2024.

LHC: LSAG, J. Phys. G 35:115004 (2008); Giddings & Mangano, Phys. Rev. D 78:035009 (2008); Hawking, Commun. Math. Phys. 43:199 (1975); Dimopoulos & Landsberg, PRL 87:161602 (2001).

Eggs: Zhong et al., JAMA 2019;321:1081; Drouin-Chartier et al., BMJ 2020;368:m513; Dehghan et al., AJCN 2020;111:795; Carson et al., Circulation 2020;141:e39; Wang et al., Nature 2011;472:57.

Context/prior art: Sourbut & Goldhaber (2025); Berg, Carauleanu & Rosenblatt (2023); Toulmin (1958); Kunz & Rittel (1970); Walton, Reed & Macagno (2008); Irving, Christiano & Amodei (2018, arXiv:1805.00899); Groth et al. (2010); W3C PROV-O (2013).

Full proofs, worked Sugya graphs (Neo4j + ArangoDB loaders, Crux/Uncertainty maps), schemas, and the complete 987-paper corpus index with 154 curated Gabbay-program references are in the accompanying repository under epistemic_stack/.