Ed — AI Bible Translation Reasoning

Ed (עֵד, witness) is an AI-assisted Bible translation reasoning platform. It uses the Textus Receptus (TSR) Greek text, the Masoretic Hebrew tradition, and the family stack (Mahalath, Tirzah, Milcah, Hoglah) to translate under configurable principles — reconstructing the principles behind the King James Version, not copying KJV wording.

The Bible is a self-defining, self-referencing whole: any term is interpreted in light of every other canonical occurrence.

See docs/vision.md and docs/translation-principle-semantic-layers.md.

ed --help
ed policies
ed policies --format json --out policies.json
ed target-languages
ed target-languages --format json --out target-languages.json
ed manifest --out ed-manifest.json
ed build-bundle fixtures/john-1-1.json --policy evidence-based --out bundle.json
ed build-bundle fixtures/john-1-1.json --policy evidence-based --format markdown --out bundle.md
ed build-debate-prompt bundle.json --target-language en --out prompt.txt
ed build-bundle fixtures/genesis-1-1-cantillation.json --policy masoretic-traditional --out hebrew-bundle.json
ed decide bundle.json --target-language en --mock --out decision.json
ed decide bundle.json --target-language en --mock --prior-decision decision.json --out revised-decision.json
ed decide hebrew-bundle.json --target-language en --mock --out hebrew-decision.json
ed decide bundle.json --target-language es --mock --out decision-es.json
ed explain decision.json
ed explain decision.json --out decision.md
ed evaluate-decision decision.json fixtures/evaluations/john-1-1-en.json
ed evaluate-decision hebrew-decision.json fixtures/evaluations/genesis-1-1-en-cantillation.json
ed run-evaluation-suite fixtures/evaluation-suite.json --mock --out suite-report.json
ed explain-suite suite-report.json --out suite-report.md
ed run-evaluation-suite fixtures/evaluation-suite.json --mock --format markdown --out suite-report.md
ed run-evaluation-suite fixtures/evaluation-suite.json --mock --case-id genesis-1-1-en-cantillation --work-dir suite-artifacts
# optional trace output (overwrites by default; --trace-append to accumulate):
ed decide bundle.json --target-language en --mock --trace-out trace.jsonl
ed explain-trace trace.jsonl --out trace.md
# optional local Ollama adapter:
ollama create ed-dictalm-q8-final -f ops/ollama/ed-dictalm-q8-final.Modelfile
ed decide bundle.json --target-language en --ollama-model ed-dictalm-q8-final --prompt-out prompt.txt --out decision.json
# optional Milcah specialist pressure-test adapter:
ed build-milcah-request bundle.json --target-language en --out milcah-request.json
# ...or seed the request from an existing decision instead of a fresh mock:
ed build-milcah-request bundle.json --base-decision decision.json --out milcah-request.json
ed decide bundle.json --target-language en --milcah --out decision-milcah.json
# optional thinking review packet:
ed decide bundle.json --target-language en --ollama-model hf.co/dicta-il/DictaLM-3.0-24B-Thinking-GGUF:Q8_0 --ollama-think --capture-thinking --out decision.json
ed review-thinking decision.json --out thinking-review.md

Use --capture-thinking with an Ollama thinking model only when the raw reasoning trace should be stored in the decision metadata for later documentation/review. By default Ed records that thinking was present without persisting the trace. Local Ollama decisions request JSON-mode output by default; use --no-ollama-json-mode only for models that fail under constrained formatting. If a thinking model returns private reasoning but no final rendering, Ed fails the debate and records diagnostic stop/length metadata in the trace without storing raw thinking text.

Live Ollama output is only credited with the witnesses the model actually cites (validated against the bundle); if it cites none, Ed records a review note rather than implying grounding. If it cites an id that is not in the bundle, Ed names the invented ids in a review note rather than dropping them quietly — a fabricated citation is evidence the rendering needs more scrutiny, not less. Repeated citations of one witness are counted once. Structured debate JSON can be long, so --ollama-num-predict defaults to 1024; a generation that stops at the length limit is flagged with a review note (raise --ollama-num-predict and re-run).

Rule templates & web UI

The built-in policies are the seeded set; you can also author your own translation-rule templates — a policy plus optional tone, custom_guidance, and a term_glossary (fixed renderings for salient lexemes) — and translate under them. Templates are stored as JSON under ~/.ed/templates and are usable anywhere a policy id is, including --policy.

ed templates save my-rule.json          # validate + store a rule template
ed templates list
ed build-bundle fixtures/john-1-1.json --policy my-rule --out bundle.json
ed build-bundle fixtures/john-1-1.json --policy ./my-rule.json --out bundle.json  # or a file

An interactive UI is available with the web extra:

pip install 'ed[web]'
ed serve                                 # http://127.0.0.1:8799

Author/edit a rule template (fork a built-in or start blank), paste or upload an evidence fixture (source unit + witnesses), pick a target language, and translate — the decision (rendering, reasoning, unresolved questions, cited witnesses) renders live. Save the template to reuse it from the CLI.

Family stack

Sense layer (design)

Design work for a Hebrew/Greek sense layer — so that what a source word means comes from codified, debated, corpus-derived evidence rather than a model's memory — is recorded in ADR-006 – ADR-014 of docs/architecture-decisions.md. Headlines: Mahalath owns the source-language sense inventory and Ed owns rendering; one sense per lexeme by default, with splits as justified exceptions; no English lexicon — English is purely a rendering under policy; reference lexica are witnesses Ed may visibly disagree with; and terms with no English equivalent get a consistent multi-word rendering or a debated composite coinage plus a reader-facing dictionary entry.

The working paper behind those decisions — options not taken, the licensing research, and the build order — is docs/lexicon-options.md. Nothing here is built yet.

Review

A point-in-time functional + code review of 0.2.0 is in docs/review-2026-08-11.md. All eight findings were actioned in 0.2.1 — see the remediation section at the end of that document.

Develop

pip install -e ".[dev]"
pytest

Install .[keturah] or the sibling Keturah checkout when validating the full capability-manifest integration. Ed can run its standalone fixture workflow without Keturah installed.

Install .[milcah] or the sibling Milcah checkout to run ed decide --milcah; the default development/test install keeps that integration optional.

License

Apache License 2.0.