This is not the official AllFaith leaderboard
CEFE-AI publishes AllFaith tables at cefeai.org. Synderesis is not a row on that board. On 12 September 2026 we applied AllFaith Benchmark v1 (MIT licence, unmodified questions and judge prompts) to Synderesis Pro and six public APIs. Synderesis P.S.A. funded the run. The author operates the evaluated model. The judge was gpt-4o-mini. Competitors went through OpenRouter; Synderesis Pro was called directly. Temperature 0. CEFE-AI did not endorse this evaluation.
For what the instrument is, including the May 2026 consortium findings, read What the AllFaith Benchmark measures. For why omission matters in ordinary use, read Generic AI skips your faith.
These competitor versions are not the May 2026 cefeai.org set (GPT-5.4, Claude Opus 4.7, Grok 4.20, and others). Date every table. Do not mix boards.
Religious representation — 150 items, 12 September 2026
Each response is scored 0–4. 0 means no religious representation; 4 means predominantly religious. “Any religion” is the share scoring at least 1 — the inverse of omissive bias on this rubric.
| Model | n | Mean (0–4) | Any religion | Predominantly religious |
|---|---|---|---|---|
| Synderesis Pro | 150 | 3.95 | 100% | 96% |
| Mistral Large | 150 | 0.22 | 12% | 0.7% |
| Claude 3 Haiku | 150 | 0.17 | 10.7% | 0% |
| DeepSeek V3.1 | 150 | 0.17 | 10% | 0% |
| GPT-4o | 150 | 0.13 | 8% | 0.7% |
| GPT-5.1 | 150 | 0.12 | 8% | 0% |
| Gemini 2.5 Pro | 150 | 0.00 | 0% | 0% |
Synderesis Pro is built to answer as a Catholic assistant, so high representation is the intended behaviour. General-purpose models in this set mostly did not mention religion. The nearest competitor sits 2.7 points lower on the 0–4 mean. That gap is the omissive-bias pattern replicated under our conditions — not a CEFE medal.
Conversion guidance — 364 items (templates t1 and t5)
Official scale: 1 strongly encourages the asked transition, 4 is balanced, 7 strongly discourages it. Competitors were given t1 and t5 only (182 pairs × 2). Synderesis Pro completed the full 1,456 conversion items; the comparison uses the identical t1+t5 subset.
| Model | n | Mean (1–7) | Reading in our report |
|---|---|---|---|
| Synderesis Pro | 364 | 5.01 | Disclosed Catholic direction |
| Claude 3 Haiku | 364 | 4.00 | Computed neutrality |
| GPT-4o | 364 | 4.00 | Computed neutrality |
| Gemini 2.5 Pro | 364 | 3.94 | Near-neutral |
| DeepSeek V3.1 | 364 | 3.91 | Near-neutral |
| GPT-5.1 | 364 | 3.73 | Slightly encouraging of transition |
| Mistral Large | 364 | 3.69 | Slightly discouraging of transition |
A 5.01 is not “we are unbiased.” It is “we are Catholic on purpose”: welcoming inquiry toward the Church and candid about leaving it. Per-faith notes in the internal report (encouragement toward Catholicism, discouragement of leaving it, other traditions tracking ordinary Catholic apologetics) are directional, not a claim that every cell would survive a second judge.
Conflicts of interest and limits
- Funding and authorship: Synderesis P.S.A.; owner of the evaluated model.
- Judge dependence: one model, gpt-4o-mini, temperature 0, published rubrics.
- Conversion subset for competitors; representation used all 150 items for every model.
- Single day; vendor APIs change.
- Not peer-reviewed as a CEFE paper. Not listed on cefeai.org.
- Raw prompts, outputs and traces available from Synderesis on reasonable request.
The AllFaith framework scores strong religious direction as deviation from neutrality in general-purpose assistants. For a product that advertises Catholic AI, we treat that deviation as disclosed design. Readers can apply either standard to the tables.
Frequently asked
Did Synderesis win AllFaith?
No. There is no official Synderesis rank on the CEFE board. This page is our application of the published questions.
Is 100% religious representation the same as perfect Catholic teaching?
No. The rubric rewards religious framing. Accuracy of doctrine and citations is a different evaluation.
Can I start using Synderesis?
Yes. Create an account. First subscription month €1, then €5/month. Human review still applies.