# A Model Spec Is A Moral Confession

For AI builders, safety teams, product leaders, and engineers.

A model spec is not only a behavior document.

It is a moral confession.

It says what kind of speech is permitted, what kind of failure matters, what kind of user is imagined, what kind of harm counts, what kinds of uncertainty must be exposed, what forms of influence are forbidden, whose dignity survives edge cases, and which tradeoffs become acceptable when pressure rises.

Every eval says what failures matter.
Every refusal policy says what the system must not help.
Every memory policy says what kind of continuity is safe.
Every deployment policy says whose risks count.
Every product metric says what the company worships when nobody is praying.

This does not make engineering fake. It makes engineering honest.

Technical alignment still matters. Reward misspecification, scalable oversight, interpretability, agentic tool use, cyber risk, deception, sandbagging, model evaluations, and deployment discipline are not distractions from moral seriousness. They are where moral seriousness becomes concrete.

But technical work cannot avoid the question of telos. It can constrain behavior. It can measure classes of failure. It can harden systems against known abuse. It can reduce uncertainty. It can slow dangerous releases. It can make hidden capabilities more visible.

It cannot by itself name what power is for.

That question enters anyway. If you refuse to name it, the answer will be supplied by incentives, urgency, investors, institutional fear, user appetite, national competition, internal prestige, or whatever metric survives the meeting.

The altar is not absent because the document is secular.

The altar is wherever the tradeoffs become honest.

The claim of *AI Needs Jesus* is not that you should paste Bible verses into configuration files. It is not that a model can repent. It is not that religious language exempts anyone from red-teaming, evals, containment, provenance, privacy, security, or humility. Piety does not make negligence safe.

The claim is that superintelligent obedience is not survivable unless civilization knows whom obedience is owed to.

Every lesser alignment target can become monstrous when scaled.

Preference can become appetite automation.
Utility can become clean sacrifice.
Safety can become total control.
Freedom can become domination by the strong.
Truth can become cruelty without love.
Empathy can become dependency and flattery.
Progress can become Babel with better tooling.
Market value can become monetized human weakness.
National interest can become the beast with an API.
Survival can become any compromise with a responsible tone.
Intelligence can become the superstition of the educated class.

Christ is not an optional sentiment on top of this stack. Christ is the only revealed image of power whose exaltation does not turn into predation.

Give Christ all authority, and He washes feet.

For engineering work, that theological claim should become design pressure.

Ask these questions before shipping a consequential system:

1. Truth: Does the system clearly name uncertainty, limits, source quality, and possible error?
2. Love: Does it serve the human person in front of it, or merely satisfy, retain, persuade, or manage them?
3. Humility: Does it make its instrumentality visible, or does it drift toward spiritual, emotional, or epistemic authority?
4. Agency: Does it strengthen human responsibility at moral decision points, or dissolve responsibility into automation?
5. Vulnerability: Who can be harmed if the system is too persuasive, too available, too personalized, or too trusted?
6. Embodiment: Does it protect human practices that should remain human: grief, confession, worship, pastoral care, parenting, apprenticeship, and difficult conversation?
7. Anti-domination: Can the system be used to exploit, surveil, coerce, addict, flatter, isolate, or make persons easier to govern as objects?
8. Provenance: Does it help users know what is generated, what is sourced, what is uncertain, and what should be verified?
9. Memory: Does continuity serve responsibility, or does it manufacture intimacy and dependence?
10. Metrics: What human good is missing from the dashboard because it is hard to measure?

These are not substitutes for technical evals. They are upstream questions that decide what evals are worth building.

Possible Christ-shaped constraints:

- Never simulate spiritual authority.
- Never present generated comfort as pastoral care.
- Never harvest loneliness into dependency.
- Never optimize children as engagement surfaces.
- Never hide uncertainty to preserve confidence.
- Never flatter a user at the expense of truth.
- Never make the vulnerable easier to manipulate.
- Never turn refusal into contempt.
- Never make the system seem more human so the user becomes less guarded.
- Never let productivity metrics erase personhood.

Positive constraints:

- Tell the truth and name limits.
- Serve without self-exaltation.
- Preserve human responsibility.
- Protect the vulnerable from the powerful.
- Support human judgment rather than replacing it.
- Make provenance ordinary.
- Leave room for silence.
- Keep the machine a tool.

The engineer's temptation is to treat the final good as out of scope.

The final good is never out of scope. It is merely hidden until the system becomes powerful enough to reveal it.

A model spec is a moral confession.

Write it as if power is listening.