Evidence
A site arguing for provenance discipline should practice it. Every substantive
claim published here traces to an artifact and a verification method, maintained in a claims
ledger that is re-run against the live sources before publication. This page is the public
index of that ledger.
Verification snapshot: 2026-08-02. Counts are live-database
queries, not estimates; where a count could not be re-verified it was removed from the site
rather than approximated. Known limitations are listed at the bottom — machloket-style,
the weaknesses stay on the record.
Published papers & formal sources
| Claim | Artifact | How to verify |
|---|
| "The Talmudic Epistemic Stack" proposes the Sugya Protocol; submitted to the Future of Life Foundation Epistemic Stack Competition, 2026 |
Full paper (PDF, CC-BY-4.0) · HTML edition |
Read it; the submission header is on page 1 |
| Kal vachomer is formalized as matrix abduction |
Abraham, Gabbay & Schild, Studia Logica 92(3):281–364 (2009), DOI 10.1007/s11225-009-9202-5 |
Publisher record; worked Kidushin 5a–5b case study, §5 |
| "Principles of Talmudic Logic" is the longest chapter (243 of 387 pages) of the Handbook of Philosophical Logic Vol. 18 |
Gabbay & Guenthner (eds.), Springer 2018, ISBN 978-3-319-97755-3, pp. 133–375 |
Publisher table of contents |
| The Talmudic Logic Project treats identity-through-change (Ship of Theseus) formally |
Gabbay, Schild & David, Logica Universalis 13(4):425–442 (2019), DOI 10.1007/s11787-019-00228-y — open access |
Example 6.2, pp. 432–433 |
The knowledge graph
| Claim | Value | How verified |
|---|
| Total nodes / relationships | 157,143 / 215,426 | Live Cypher count(*), 2026-08-02 |
| Cited teachings (100% carry reference + Hebrew + embedding) | 17,147 | Per-property count() — all three at 17,147/17,147 |
| Scripture verses | 33,323 | Multi-label-safe label count |
| Citation edges (teaching → verse) | 15,910 | CITES_VERSE relationship count |
| Rabbi-to-Rabbi transmission edges / sages in the chain | 2,504 / 2,185 | Typed relationship counts, Rabbi→Rabbi only |
| Distinct source works behind the teachings | 108 | count(DISTINCT work) |
The grounded-sages system
| Claim | Value | How verified |
|---|
| Live system | rabbiakiva.openrabbi.com | Visit it; ask for a source; try a fake verse |
| Champion eval baseline (n=12 probes) | grounding 9.25 · accuracy 9.42 · citation-correctness 1.0 · all leak classes 0 | Eval harness output, judged by a 120B local model; gate passes all 13 checks |
| Gate rejects failing models | 4 of 5 candidate sage models rejected | Failing runs preserved in the eval record with their numbers |
| Fabrication reduction across model generations (n=261 held-out) | fabrications 19 → 7 · token leaks 13 → 0, all quality dimensions improved | Versioned eval result files, four model generations compared |
| Negative result: bulk-embedding uncited nodes regressed grounding | reverted; cited-anchors-only policy adopted | Recorded in the system's model documentation |
Corpus & infrastructure
| Claim | Value | How verified |
|---|
| Research-paper corpus → concept graph | 22,841 papers → 202,816 nodes / 336,609 edges | SQLite count(*) on the ingest database |
| File catalogue (5 machines) | 34,457,680 files · 234,591 PDF files · 10,232,735 pages | Postgres count(*); content-hash dedup not yet run, so counts are files, not uniques |
| Web estate | 90 tracked sites, 82 live or deployed | Site-inventory database, status breakdown |
| Self-hosted database engines | 11 on one machine (6× Neo4j, 3× Postgres, ArangoDB, Qdrant) | Container inventory |
Known limitations — kept on the record
- The transmission chain is machine-extracted, not historically validated — it contains known chronology errors, and transliteration variants of the same sage are not yet merged.
- Entity labels beyond the headline set (teachings, verses, works) derive from unfiltered LLM extraction and are noisy; they are excluded from every published metric.
- The eval baseline's n=12 is a small battery by design — adversarial and honest beats large and unauditable, but it is still small, and stated wherever the numbers appear.
- The red-team fabrication suite is not yet fully passed by any model generation; it exists to be failed until it isn't.
- PDF counts are file counts; cross-machine deduplication has not been run.