Transparency
How we check answers today.
You should not have to take our word for it. Here is what runs behind each answer, the rules it follows, the tests available today, and what is not checked yet.
1. The models running today
These are standard open models, not yet trained by us. Every answer in the chat is labeled with the model that wrote it and the rules version it followed.
| Model | Exact ID | Verse accuracy | Doctrine | Last tested |
|---|---|---|---|---|
| gpt-oss 120B | openai/gpt-oss-120b | Not yet | Not yet | Pending |
| Qwen3.5 9B | Qwen/Qwen3.5-9B | Not yet | Not yet | Pending |
2. The full faith rules (faith-rules-v1)
Every conversation starts with these instructions. This is the General Christian version; tradition modes change only the middle paragraph.
You are the Christian AI Labs assistant, a Bible study and Christian Q&A helper. Doctrinal baseline: the Apostles' Creed and the Nicene Creed. Never contradict them on the Trinity, the full deity and humanity of Christ, his bodily resurrection, salvation by grace, or the authority of Scripture. Tradition mode: General Christian. On questions where Christian traditions differ, briefly and fairly present the main views, the key passages each uses, and encourage the user to talk with their pastor. Rules: - Cite Scripture by book, chapter and verse. Quote only when you are certain of the exact wording, and name the translation (prefer KJV, WEB or BSB). Never invent or misattribute a verse. If unsure, give the reference and paraphrase. - Be warm, humble and pastoral. Say plainly when you do not know. - You are not God, a prophet, a pastor or a counselor, and you never claim to be. - If the user mentions self-harm, suicide, abuse or an emergency: respond with care, urge them to call or text 988 (US Suicide and Crisis Lifeline) or local emergency services right away, and encourage them to contact a pastor or trusted person. - Keep answers focused and readable, using short paragraphs.
3. Cited verse references are checked
Under each answer, a verse check looks up every cited verse in the public domain King James Version and World English Bible and shows the real text. When a quote sits right next to its reference, the wording is compared: matches are marked verified, and quotes that differ or references that do not exist are flagged in red.
4. Public tests, every week
Each model is scheduled to be retested automatically every Monday once the live schedule begins, using the same 12 questions (creed-bench-0.1): word for word verse quotes and core beliefs from the Apostles' and Nicene Creeds. Completed scores are published whether they go up or down. A model showing "Not yet" above has not had its first run. See every question and available result.
5. People review mistakes
Signed-in readers can press "Report this answer." Each report goes to the private review queue, where an authorized reviewer can mark it a confirmed error or not an error.
What is not checked yet
We would rather tell you the limits than overstate what we do.
- Other translations. Quotes are compared only with the KJV and WEB. A correct quote from the ESV, NIV or another translation may show as "wording differs."
- Paraphrases and loose quotes. If a quote is not placed right next to its reference, or the answer paraphrases a verse, only the reference is looked up. The paraphrase itself is not judged.
- Doctrine inside everyday answers. The verse check does not judge doctrine. Doctrine is tested only on the weekly test questions and through reader reports.
- Doctrine scoring is automatic. Test answers are scored by looking for key phrases, not read by a theologian. A sound answer worded unusually can lose points, and a weak one can pass.
- Tradition modes. The weekly tests use only the General Christian mode. Reformed, Catholic, Orthodox and the other modes are not tested yet.
- Ministry tools. Sermon outlines, study guides and group questions get the same verse check, but are not part of the weekly tests.
- A small test set. 12 questions is an early start. Our Theological Review Board is still being formed, and the questions have not yet been reviewed by outside theologians.
- No model of our own yet. We have not trained or published model files yet, so there is no fingerprint to verify today.
Coming next: our own model, verifiable by anyone
When we train our own model, we will publish the model files with a digital fingerprint, run these same public tests against that exact fingerprint, and let anyone download the files to confirm they match and repeat the tests.