✝ Christian AI Labs

Open by design

Show the work.

Our weights, our tests, and our review record are public. Churches and developers can inspect, reuse, and improve them.

Apache 2.0

OpenAI gpt-oss

Primary chat and study model

Apache 2.0

Qwen3

Fine-tune experiments and multilingual

Apache 2.0

Gemma 4

Kid-safe mode and church website widget

How a model earns release

  1. 1Curated corpus of public domain Bibles, creeds, and confessions
  2. 2Fine-tuning on board-approved answers
  3. 3Preference tuning ranked by reviewers
  4. 4Retrieval from a verified Scripture index
  5. 5Creed Bench scoring and board approval

Creed Bench

An open test anyone can run.

About 1,500 prompts across Scripture accuracy, core doctrine, disputed topics, pastoral care, crisis handling, and attempts to trick the model. Every release ships with its scorecard.

Theological error bounty

Find a made-up verse or a doctrinal error and report it. Confirmed reports earn credit, and every fix goes into a public error log.

Bible texts used for training are public domain (KJV, WEB, BSB, ASV). Copyrighted translations are never used for training.