{"data":{"kind":"file","path":"README.md","version_id":"ewe7crcem8xgbel3hs69p9vs","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":2136,"modified_at":"2026-04-21T12:21:40.734000","content_hash":"5d5307ea32004bd2586cb91e93c1d0cbed004c8cea11ddfb4b2b8feb3f8c6ac0"},"entries":[],"content":"# luetin/tanakh-gematria-resonance\n\nA **verifiers** environment that rewards a model's answer in proportion to how well the Hebrew translation of its response resonates — numerically — with the Tanakh.\n\n## The reward, briefly\n\nGiven a model's English answer, the environment deterministically:\n\n1. Translates English → Hebrew via a curated perspective-aware lookup + phonetic transliteration fallback.\n2. Computes **mispar gadol gematria** per Hebrew word and for the full passage.\n3. Scores three signals:\n\n| Reward | Weight | Range | What it measures |\n|---|---|---|---|\n| `tanakh_word_coverage` | 0.5 | 0–1 | Fraction of per-word gematria values attested ≥5× in the Tanakh lexicon. |\n| `verse_sum_proximity` | 0.3 | 0–1 | `1 − min(\\|Δ\\|, 100)/100` where Δ = distance to the nearest Tanakh verse's total gematria. |\n| `length_sanity` | 0.2 | 0–1 | Reward 40–250 words; deters gaming via single sacred words. |\n\nCombined via a `verifiers.Rubric` weighted sum. No LLM-judge. Entirely symbolic.\n\n## Dataset\n\nA seeded set of 35 (perspective, question) pairs across 8 domains — cosmology, consciousness, medicine, agriculture, materials, governance, and two \"gap\" taxonomies (knowledge_gaps, engineering_gaps). Derived from the luetin taxonomy.\n\n## Format\n\nModel must place its answer in `<answer>…</answer>` tags. Anything outside is ignored.\n\n## Using it\n\n```python\nfrom verifiers import load_environment\nenv = load_environment(\"tanakh-gematria-resonance\")\n```\n\nOr for RL training on Prime Intellect:\n\n```toml\n[[env]]\nid = \"luetin/tanakh-gematria-resonance\"\n```\n\n## Why it's interesting\n\nMost RL rewards for LLMs are verifiers (math/code), preference models, or LLM-as-judge rubrics. This one is **symbolically grounded in a 2,500-year-old text whose numeric structure predates the modern era.** It provides a reproducible, non-learned, objective reward signal rooted in a fixed corpus.\n\nA model trained against this environment learns to answer in language whose word-level numerics actually appear in scripture and whose passage totals approach canonical verse sums — without ever being told what those are.\n","encoding":"utf-8","truncated":false,"total_bytes":2136},"status":null}