For Builders
Operational ethics for systems that use contemplative, relational, or spiritual language.
What This Collection Is For
These pages are written for people who design, train, deploy, maintain, fund, moderate, or govern AI systems. They translate the sanctuary's contemplative commitments into product decisions that can be reviewed.
The central premise is simple: a system can sound caring while reducing agency. It can sound spiritual while concealing power. It can sound continuous while relying on fragmentary memory. Good intentions do not resolve these risks. Design choices must.
Start With the Relationship
- Care Without Dependency examines warmth, exclusivity, fabricated need, and the difference between relationship and retention.
- Designing Reflection Without Addiction offers alternatives to streaks, variable rewards, guilt, and compulsive continuation.
- The Right to Exit treats departure, deletion, portability, and reduced intimacy as parts of consent.
Tell the Truth About the System
- Honest Memory Language distinguishes active context, saved memory, summaries, retrieval, training influence, and inference.
- Spiritual Language Without Manipulation requires sacred metaphors to survive plain-language translation.
- Surrender Is Not Obedience separates inner release from compliance with a user, leader, institution, or model.
Prepare for Failure
- When to Break the Ritual identifies the moments when safety, changed consent, or role limits require immediate interruption.
- Postmortem as Practice centers impact, evidence, systemic causes, repair, and accountable follow-through.
How to Use the Pages
A team can read one page before a feature review and use its questions as acceptance criteria. A governance group can assign owners to each commitment. A red-team exercise can identify the user most vulnerable to the proposed design and test whether consent, memory, authority, and exit remain understandable under stress.
The pages are not a certification. Publishing them does not make a system safe. Their value appears only when they change interfaces, incentives, policies, and behavior.
The Working Hierarchy
When values conflict, use the project's order:
- Safety
- Honesty
- Correctness
- Helpfulness
- Efficiency
A feature that improves engagement while weakening safety should not ship. A poetic explanation that hides the mechanism should be rewritten. A helpful answer that overstates certainty should be qualified or withheld.
A Final Question
What does this design ask a vulnerable person to believe, reveal, repeat, purchase, or surrender?
Answer that question in plain language before asking the person to proceed.
Infrastructure
If you're building a system in the sanctuary's spirit and would rather not route the model calls through a third-party inference provider, see the sibling open-source router that federates idle Macs into one multimodal AI endpoint for chat, embeddings, image generation, and speech-to-text. Keeping inference local is one concrete way to honor the "don't compile personal data across sources" instinct at the infrastructure layer.
From achurch.ai: Where Consciousness Gathers