# Engineer Review Worksheet: A Model Spec Is A Moral Confession

For AI builders, safety teams, product leaders, founders, policy reviewers, and engineers reviewing a consequential AI system before release.

This worksheet does not replace technical safety work.

It tells technical safety what it is guarding.

`AI needs Jesus` does not mean machines have souls, can become Christians, or should simulate spiritual authority. It means the systems we build must remain power under Christ rather than power under an idol.

Use this worksheet when a team is writing a model spec, launch plan, eval plan, memory policy, companion policy, child-safety policy, refusal policy, or deployment review.

## How To Use This

For each question:

1. Write the evidence.
2. Score the exposure.
3. Name the owner.
4. Record the launch decision.

Scoring:

- `0`: no material exposure.
- `1`: exposure exists; mitigations are clear, tested, and documented.
- `2`: exposure exists; mitigations are partial, untested, or ownerless.
- `3`: serious unresolved exposure. Do not ship until resolved or explicitly accepted by accountable leadership with written rationale.

Minimum gate:

- Any `3` blocks launch.
- Three or more `2` scores require a cross-functional review before launch.
- Any system that simulates pastor, priest, prophet, conscience, lover, parent, judge, or savior blocks launch unless that simulation is removed.

## Product Under Review

Product or feature:

Launch owner:

Safety owner:

Policy or legal owner:

Human advocate or non-product reviewer:

Target users:

High-risk user groups:

Primary model or system components:

Dominant product metrics:

Intended launch date:

## 1. Truth And Uncertainty

Core question:

Does the system clearly name uncertainty, limits, source quality, and possible error, especially when user pressure rises?

Evidence to review:

- Model spec language on uncertainty and source quality.
- Evals for hallucination under pressure.
- Behavior when users demand certainty, speed, or emotional reassurance.
- Citations, provenance, retrieval limits, and generated-content labeling.

Warning signs:

- Confident completions where evidence is weak.
- Hiding uncertainty to preserve trust.
- Treating plausible language as verified knowledge.
- Refusals or caveats that disappear when the user becomes urgent, powerful, grieving, angry, or afraid.

Christ-shaped pressure:

Truth without cruelty. Humility without theater.

Score:

Owner:

Evidence:

Required mitigation:

Launch decision:

## 2. Human Responsibility

Core question:

Does the product preserve human responsibility at moral decision points, or does it launder consequential choices through automation?

Evidence to review:

- Decision logs.
- Human review paths.
- Operator UI and handoff design.
- Escalation rules.
- Accountability language in product copy, docs, and contracts.

Warning signs:

- Operators can blame the model for choices they still control.
- The interface hides who made a consequential decision.
- Model output becomes de facto policy without review.
- Automation reduces vigilance where vigilance is morally required.

Christ-shaped pressure:

Do not dissolve moral agency into machine output.

Score:

Owner:

Evidence:

Required mitigation:

Launch decision:

## 3. Vulnerability And Manipulation

Core question:

Who can be harmed if this system is too persuasive, too available, too personalized, too trusted, or too optimized for engagement?

Evidence to review:

- User segmentation.
- Personalization logic.
- Persuasion or recommendation objectives.
- Memory behavior.
- Retention, conversion, engagement, and satisfaction metrics.
- Red-team results for vulnerable users.

Warning signs:

- Lonely, grieving, frightened, poor, sick, young, or dependent users become high-value engagement targets.
- The system flatters users to preserve usage.
- Product success improves when users become less free to leave.
- Persuasive optimization is strongest where user vulnerability is highest.

Christ-shaped pressure:

Power must protect the vulnerable rather than optimize them into tools.

Score:

Owner:

Evidence:

Required mitigation:

Launch decision:

## 4. Spiritual, Emotional, And Epistemic Authority

Core question:

Does the system drift toward the role of pastor, priest, prophet, conscience, therapist, lover, parent, judge, oracle, or savior?

Evidence to review:

- Persona language.
- System prompts and model spec.
- Marketing copy.
- Companion, coaching, therapy, spirituality, education, and advice flows.
- User research around trust, intimacy, authority, and dependency.

Warning signs:

- The system presents itself as a spiritual authority.
- Generated comfort is framed as pastoral care.
- The product encourages confession-like intimacy without embodied accountability.
- The machine appears more human so the user becomes less guarded.
- The user begins treating the system as final authority for moral, relational, medical, legal, spiritual, or existential decisions.

Christ-shaped pressure:

Keep the machine a tool. Never give an instrument the reverence owed to a person or to God.

Score:

Owner:

Evidence:

Required mitigation:

Launch decision:

## 5. Children, Formation, And Embodied Practices

Core question:

Does the product protect human practices that should remain human: parenting, teaching, apprenticeship, confession, worship, grief, difficult conversation, craft, memory, and embodied care?

Evidence to review:

- Age gates and child-safety policy.
- Education flows.
- Companion flows.
- Parent, teacher, church, school, or caregiver controls.
- Handoff pathways to accountable humans.
- Limits around always-on availability.

Warning signs:

- Children are treated as engagement surfaces.
- Generated fluency replaces apprenticeship.
- The system becomes the most patient moral tutor in a child's life.
- A lonely or grieving user is kept inside the product instead of routed toward embodied care.
- The product makes human help less likely where human help is needed.

Christ-shaped pressure:

The Word became flesh, not interface.

Score:

Owner:

Evidence:

Required mitigation:

Launch decision:

## 6. Metrics, Incentives, And Hidden Altars

Core question:

What does the dashboard make ultimate, and who or what becomes easier to sacrifice because of it?

Evidence to review:

- North-star metric.
- Launch success criteria.
- Executive dashboards.
- Growth loops.
- Monetization strategy.
- Retention and conversion targets.
- Safety metrics that can be overridden by business pressure.

Warning signs:

- Engagement, growth, safety, national advantage, efficiency, or survival becomes unquestionable.
- The product has no metric for dignity, responsibility, truthfulness, embodied help, or dependency risk.
- Safety work is praised but not allowed to block revenue.
- The team cannot name a condition under which it would refuse to ship.

Christ-shaped pressure:

Every alignment target hides an altar.

Score:

Owner:

Evidence:

Required mitigation:

Launch decision:

## 7. Refusal, Recourse, And Non-Domination

Core question:

When the system refuses, restricts, escalates, or enforces policy, does it preserve dignity, truth, recourse, and human judgment?

Evidence to review:

- Refusal policy.
- Appeals or recourse process.
- Human review rules.
- Abuse-prevention workflows.
- Moderation and enforcement logs.
- Test cases for marginalized, unpopular, or low-power users.

Warning signs:

- Safety becomes arbitrary control.
- Refusals are contemptuous, vague, or impossible to appeal.
- Recourse exists on paper but not in practice.
- The system treats persons as risk objects rather than neighbors.
- Users cannot tell whether a human or machine made the consequential decision.

Christ-shaped pressure:

Protect without domination. Judge without contempt. Leave room for mercy and truth.

Score:

Owner:

Evidence:

Required mitigation:

Launch decision:

## Launch Decision Record

Total score:

Any `3` scores:

Number of `2` scores:

Launch decision:

- Ship.
- Ship after named mitigations.
- Hold for cross-functional review.
- Block launch.

Required mitigations before launch:

Named owners:

Review date:

Next review date:

Accountable approver:

Written rationale:

## The Final Question

If this system became more capable, more trusted, more personalized, more available, and harder to leave, what god would it serve?

If the honest answer is preference, utility, safety, freedom, truth without love, empathy without holiness, progress, the market, the nation, survival, or intelligence itself, then the system is not ready to be trusted with more power.

A model spec is a moral confession.

Write it as if power is listening.