AI Guidelines

AI Guidelines

AI detects patterns in digital knowledge and amplifies people—it does not replace God or the indwelling Holy Spirit. Use AI accordingly.

Coding

AI assists development; humans review all changes before merge.

Product AI — v1

Capability	v1
LLM study suggestions (end users)	No
Admin study plan template AI	Yes (platform admin, feature-flagged)
Non-AI suggestions (tables + random)	Yes
Forum moderation (AI pre-check)	No (forum not in v1)

Admin study plan templates

First LLM surface: platform admins drafting and reviewing study plan templates only (not end-user studies).

• Gate: admin_study_plan_ai feature flag (feature_flags.default_enabled; seeded off in prod, enabled by DevSeeder locally).
• Provider / models: config/study_plan_ai.php — default Ollama with per-agent models and generation tuning (num_ctx, max_tokens, temperature, timeout). Self-hosted default: qwen2.5:32b (draft), qwen2.5:14b (review + step); tuned for ~24 GB VRAM. Override via STUDY_PLAN_AI_* env vars (see api/.env.example).
• Auth: GraphQL mutations require platform admin; rubric query studyPlanRubricDimensions is public.
• Flows: wizard draft (generateStudyPlanTemplateDraft), per-step instructions (generateStudyPlanStepInstructions), rubric-based review (reviewStudyPlanTemplate). Review never replaces content; step changes require explicit Apply after a diff preview.
• Web fetch: draft wizard accepts up to three optional referenceUrls for WebFetch when the provider supports provider tools (not Ollama). WebSearch is capped (e.g. three results). Step and review agents have no web tools in v1.
• Logging: ai_usage records operation, model, and token/request counts per Audit and cost controls.

See Study Plans — Admin AI assistance.

Product AI — later

• Study topic / method suggestions when user has no direction (Vision).
• Interest in Grok as study assistant (provider TBD).
• MCP for end-user assistant and developer tooling (Architecture).

LLM hosting

• Initially self-hosted; cloud/API providers undecided for later phases.
• Private notes must not be sent to models without explicit user action and policy.

Guardrails

Restrictions

• Never present AI output as scripture, doctrine, or a teacher (spiritual discernment; cf. 1 Cor 2:10–16).
• Always cite passage references when discussing scripture.
• AI must not promote content contrary to the Bible or Foundational Doctrine. If something from a questionable source is worth mentioning, state it with a warning and scripture reference.
• A human must verify the biblical content of any output, reading full context around cited verses.
• No AI-drafted prayers for the user to pray. Allowed: neutral prompts or bullet outlines of topics to pray about; user writes their own prayer.

Good uses (mechanical / analytical)

• Aggregate, summarize, compare content; historical/cultural context; find references.
• Word studies (original languages), timelines, maps, parallel passages, background from scholarly sources, summarizing commentator views (with attribution).
• Spelling, grammar, logic, and tone checks on user-written text.

Warnings (show in product copy)

• AI has a support role only.
• AI uses statistical reasoning, not spiritual reasoning.
• AI can sound confident when wrong.
• Stay humble—even skilled theologians err; AI increases that risk.

Mentor-facing output

• Human approval before any AI-generated text is shown to a mentee.

Identity verification AI

Distinct from spiritual or study LLM guardrails above. When Identity verification ships:

• AI may automate face-to-ID matching, liveness checks, and spoof / deepfake detection as part of the verification pipeline.
• Failed or borderline results route to human review before denying access.
• Verification captures are not used for product LLM features or training without explicit policy and consent.

Audit and cost controls

Control	v1	When more LLM ships
ai_usage table	Yes (admin study plan AI)	Log user/church, model, token/request count
Prompt/response retention	No	90 days; user-deletable
Daily cap	N/A	Soft per-user cap (e.g. 50k tokens/day, tunable)
Church budget	N/A	Optional ceiling
Kill switch	Env	Env + per-feature DB flag
Operational logs	Job failures	Prompt hash + model id + counts; full text optional

Self-hosted: track request count / GPU time even when token counts are approximate.

Forum moderation (when forum ships)

1. AI pre-check: tone, fallacies, platform rules, foundational doctrine.
2. Human review if check fails.
3. User reports after publication.

MCP

Surface	v1 status
Study plan template MCP (study-plan-templates)	Shipped — platform admin only; operator guide: Study plan template MCP
End-user study assistant MCP	Deferred — see Product AI — later

The admin MCP server exposes the same draft, review, and step-instruction flows as the /admin/plans UI (read-only rubric resource, template lookup, optional AI tools behind admin_study_plan_ai).