TorahML applies machine learning to the Hebrew Bible to surface deeper patterns in the Torah — and tests every one against chance. Free alpha by SHEMIM.
TorahML = Torah + Machine Learning. Machine learning for deeper Torah study.
TorahML (Torah + Machine Learning) uses machine learning, information theory and statistics to search the Hebrew text of the Tanakh for structure a reader can miss: shared roots across books, word-play, recurring key words, numeric correspondences, parallel passages and unusual letter sequences. Its purpose is to reveal deeper layers of Torah knowledge — honestly. Every candidate pattern is compared with baselines built from shuffled or comparable text, and only patterns that beat chance are presented as findings. As AI grows more capable, TorahML is built to put stronger models to work under the same rule: models may propose; only evidence decides.
Machine learning surfaces, measures and tests candidate patterns. Humans interpret their meaning. Untested candidates remain exploratory, and failed null tests stay on record. Our working rule: Read the Tanakh. Test the patterns. Keep the nulls.
Why machine learning?
AI Torah study can help readers compare more text and test more precise questions. TorahML combines machine learning with statistics and reference tools for Tanakh search and Hebrew root search. Not every tool is a trained model: dictionary lookup, gematria arithmetic and several statistical tests are deterministic. Some methods require optional models or data.
- A confidence-tiered resonance graph links Hebrew roots and cognates. Optional semantic embeddings add similarity links; each path is only as strong as its weakest link.
- Context models disambiguate ambiguous Hebrew word forms, with performance scored on held-out text.
- Compression and information-theory measures assess how surprising or repetitive a passage is, with length controls.
- Sequence models of cantillation (te'amim) measure predictability and test for anomalies against shuffled accent sequences. The reported result is no anomalies.
- Leitwort analysis tests the clustering of recurring key words, with false-discovery control.
- Parallel-passage alignment measures textual differences against composition-matched controls.
- Paronomasia scoring surfaces Bible word-play candidates through sound similarity, validated against independently marked naming etymologies.
- The experimental Torah codes / ELS statistical test compares pre-specified equidistant-letter-sequence claims with shuffled text and an independent control corpus.
- The gematria calculator offers 26 methods, including explicitly experimental rules. The experimental Cartography tool adds base-rate checks for numerical coincidences across effective ciphers.
- Experimental Discovery searches separate exploration from frozen hold-out challenges. Narrative and motif challenges use permutation tests; concept and phrase retrieval reports coverage and transfer evidence, not statistical significance.
Built for 4444
TorahML was built primarily as the research tool behind 4444.com, an educational publishing project by SHEMIM: seventy-two interlocking frameworks and one campaign, each a proposed model published to be tested, not believed. TorahML supplies the Hebrew text, roots, cognates, gematria and statistical checks those frameworks draw on. Learn more about the project at 4444.com/about.
Alpha: please read
TorahML is in alpha and has not been tested. Being reachable on a public domain does not mean it is production-grade, complete, accurate, secure or safe to rely on. Features, results and data can change or break without notice, and some reference text comes from unreviewed OCR or automated extraction. You use TorahML entirely at your own risk; SHEMIM cannot take any responsibility for how it or its results are used (see Terms). Check anything that matters against primary sources.
How to read the results
- Patterns are compared with shuffled or permuted baselines before they are shown as findings; a striking pattern that fails its baseline is reported as such.
- Exploratory lab tools are off by default and labeled speculative.
- TorahML is a research instrument, not a religious or halakhic authority.
Sources and license
Original TorahML content is © 2025 SHEMIM, licensed CC BY-NC 4.0; the engine code is MIT. Texts, dictionaries and other third-party data keep their own licenses and credits: see Sources & licenses. “4444” is a registered trademark.
Contact
Operator. SHEMIM, New York, USA. Contact: [email protected].
Frequently asked questions
What is TorahML?
TorahML applies machine learning to the Hebrew Bible to surface deeper patterns in the Torah — and tests every one against chance. Free alpha by SHEMIM.
What does ML stand for in TorahML?
ML stands for machine learning. TorahML = Torah + Machine Learning. It combines machine learning, information theory and statistics for deeper Torah study.
Can AI reveal hidden messages in the Torah?
AI can surface candidate patterns and measure whether they beat chance. That does not establish a hidden message or authorial intent. Meaning is a human judgment. TorahML reports failed null tests too, and labels untested or experimental results as research leads.
Who built TorahML and why?
SHEMIM built TorahML primarily as the Hebrew Bible research tool behind the educational publishing project 4444.com.
Is TorahML free?
Yes. TorahML is a free non-commercial research tool. Public readers do not need an account. Original content is CC BY-NC 4.0, engine code is MIT, and third-party data keeps its source licenses.
Is TorahML production-ready or safe to rely on?
No. TorahML is alpha software and has not been tested for production use. Use it entirely at your own risk and check anything that matters against primary sources.
Is TorahML a religious authority?
No. TorahML is a research instrument, not a religious or halakhic authority. Numerical matches and statistical scores do not establish religious meaning.
Where does the data come from?
Reference data includes the Westminster Leningrad Codex / OpenScriptures morphhb, JPS and other Sefaria translation editions, separately attributed dictionaries and scholarly research data. The Sources & licenses page lists credits and source-specific terms.