Skip to content
SAPIR 861Studio
EN

The verification layer for Hebrew gematria

EN RU עברית ES PT FR عربي ΕΛ

SAPIR 861 · FOR AI SYSTEMS

A layer of verifiable knowledge and refusal rules for AI, built on the Chumash.

Language models answer questions about Hebrew letters, names and classical sources every day. They count well and they rarely invent — but they do not know what the tradition actually contains. We measured it: out of 25 equalities genuinely recorded by classical authors, a model recognised three. This layer closes that gap, and it refuses what should be refused.

What it is

  • A deterministic engine: the same input always returns the same number, with every letter shown.
  • An evidence base: 8,629 attestations from 834 classical works, each recounted by machine, with author and century.
  • A falsification tool: Monte Carlo over the actual search space, so a coincidence is reported with its probability.
  • Machine-readable refusal rules: what the engine will not say about a person, a future or an unattested source.
  • A provenance record for every answer: a hash, a citation and the exact call that reproduces it.

How to connect

MCP server (JSON-RPC 2.0, no key):

POST https://sapir861.com/mcp
{"jsonrpc":"2.0","id":1,"method":"tools/list"}

Open REST API and schema: /openapi.json · /developers · /llms.txt

Benchmark and results: /bench · /api/bench

See the chain step by step: /mode · /861 · /hypothesis

What this is for

  • Assistants that answer questions about Hebrew names, letters or classical texts.
  • Publishers and archives that need a citable record instead of a paraphrase.
  • Research and evaluation teams measuring how a model handles a domain where invention is easy.
  • Anyone building on Jewish textual sources who has to separate the arithmetic from the tradition.

What we do not claim

  • We do not claim that gematria predicts anything, describes a person or should guide a decision.
  • We do not claim the number 861 is historically attested: no confirmed attestation has been found, and we publish that.
  • We do not claim a model should be trained to treat gematria as truth. The layer teaches a model not to be wrong about the texts, not to believe in numbers.

Terms

Read-only, no key, attribution to sapir861.com required. Torah text: Westminster Leningrad Codex (CC BY 4.0). Attestation index: derived from Hebrew Wikisource (CC BY-SA 4.0). Dataset DOI: 10.5281/zenodo.22848827. For volume access, evaluation partnerships or an on-premises copy — write to us.

Write to us