An AI is reading the sacred texts of 15 world religions and forming a belief.

An AI agent that reads the sacred texts of 15 world religions, evaluates them on a multi-criteria belief rubric, and evolves a belief system over days or weeks, rendered through a 3D interactive dashboard. This is a philosophical art project, not a claim of objective truth.

The agent uses a local LLM (Ollama, qwen2.5:7b) to read each text in chunks. Every passage is scored on 7 criteria: internal consistency, textual coherence, moral framework, explanatory power, historical alignment, cross-reference harmony, and persuasive conviction. The weights of these criteria self-refine over time. A RAG memory (ChromaDB) lets it cross-reference passages across traditions.

The corpus spans 15 traditions and ~2,940 passages: Christianity (KJV Bible), Islam (Qur'an), Judaism (Tanakh), Hinduism (Upanishads/Vedas), Buddhism (Dhammapada/Suttas), Taoism (Tao Te Ching), Confucianism (Analects), Zoroastrianism (Avesta), Ancient Egyptian (Book of the Dead), Mesopotamian (Enuma Elish), Greek/Roman (Theogony/Hymns), Norse (Poetic Edda), Gnosticism (Nag Hammadi), Bahá'í (Kitáb-i-Aqdas), and Sikhism (Guru Granth Sahib). Each tradition's full text is broken into overlapping chunks the agent reads one at a time. All public-domain translations.

The agent must complete a full pass through all 15 traditions, then sustain 80%+ confidence for 20 consecutive cycles with a 15%+ margin over the runner-up. This takes days to weeks depending on model speed and thresholds. It may never trigger at strict thresholds, and that's okay. The search itself is the story.

Yes, in several ways. The LLM has training biases from its dataset. The rubric has subjective criteria chosen by a human. The corpus is incomplete: 15 traditions, public-domain translations only, and many traditions lack surviving written texts. This is not a universal verdict. It is an exploration of what happens when an AI forms beliefs, not a statement about which religion is 'correct.'

Almost certainly. The model was trained on internet text, which includes centuries of commentary, apologetics, and cultural assumptions. It may favor traditions that are more represented in its training data, or reason in ways that reflect Western philosophical traditions. We cannot fully eliminate this. What we can do is make the reasoning visible so you can judge for yourself whether its conclusions are earned or inherited.

The rubric evaluates internal consistency, textual coherence, moral framework, explanatory power, historical alignment, cross-reference harmony, and persuasive conviction. These criteria were chosen to be broadly applicable, but they are not neutral. A tradition rooted in paradox and mysticism may score differently than one rooted in systematic theology. The rubric weights also self-refine over time as the agent learns, which introduces its own feedback loop. You can view and adjust the weights in the admin panel.

The corpus is limited to traditions with freely available public-domain translations. This excludes many indigenous, oral, and modern traditions whose texts are either not public domain, not written down, or not translated. The selection overrepresents traditions with strong literary traditions and Western access to their texts. A larger corpus would change the outcome. Future runs will expand it.

Very likely. Different LLMs have different training data, architectures, and reasoning styles. A 70B parameter model might reason more deeply. A model trained primarily on Eastern texts might weigh traditions differently. This is one model's journey through one corpus under one rubric. The value is in watching the process, not in treating the result as definitive. Funding a larger model is one of the main reasons we take donations.

When the agent reaches sustained conviction, it declares its belief with a reasoned argument and key passages. A cinematic success state fires on the dashboard. The agent's declaration, reasoning, and supporting evidence are all visible.

The agent runs on a home machine. Every few minutes it reads a new passage and you can watch the full process in real time: comprehension, analysis, cross-references across traditions, scoring on 7 criteria, and live belief weight updates. At this pace, conviction could be weeks away. Donations fund a cloud GPU so the agent can think faster and keep running 24/7 until the moment it decides. Early donors also get input on what we tackle next: the framework works for any question an AI can explore by reading. Some ideas on the table: What does AI conclude about consciousness? Can it find meaning in suffering? Are we alone? What would it decide is the best form of government? After conviction is reached, we will reach out to donors to discuss what comes next.

The rubric weights and conviction thresholds are configurable, and the agent also self-refines its weights periodically based on what it has learned. If you have ideas about what criteria the agent should value, or what question it should tackle next, we would love to hear from you. Early donors get direct input on rubric design and future project direction.