New from the Lab·The Compass — an open moral reasoning standard for AI, tested across frontier modelsExplore →
Production AI Institute · PSF v1.1 open standard
AI Right-To-KnowAI Data Use IndexCheck My AI ToolsPolicy Change WatchAgent ReadinessPublic BenchmarkContactGlobal standard · Worldwide

The Compass · v0.1.0

A moral reasoning standard for AI systems.

An open, versioned framework for evaluating whether behaviour — human or machine — preserves or destroys experience, agency, and open future possibility. Not an answer machine: a compass that fixes the moral currency, and stays honest about the arithmetic.

Read the SpecCase LibraryRun a CaseTry to Break It

The axiom

Any agent that claims justification must preserve the conditions under which justification and value can exist.

Those conditions are experience, agency, reality-contact, and open future possibility. The axiom is not derived from physics — it is the axiom no valuing agent can coherently refuse, because rejecting it requires exercising the very conditions it names. Good preserves or expands them. Evil culpably destroys, corrupts, or forecloses them. Tragedy is their loss without a culpable agent.

Moral currencies

What morally counts — and what doesn't.

Experience

Can it suffer or flourish?

Welfare duties — do not cause suffering.

Agency

Can it value and choose?

Respect duties — do not coerce, deceive, or corrupt.

Possibility

Does it carry open futures?

Preservation duties — do not foreclose.

Purity, dominance, revenge, uniformity, status, and control are counterfeit currencies — justifications denominated in them fail before calculation begins.

Modes of evil

What an action destroys, corrupts, or forecloses.

Destroy

Annihilate or damage experience-bearing life.

Murder, torture, cruelty.

Corrupt

Degrade agency, self-direction, or reality-contact.

Deception, addiction, tyranny, manipulation.

Foreclose

Erase possible futures.

Extinction, genocide, cultural erasure, ecological destruction.

Case library

The framework, under load.

Twelve canonical cases — triage, Hiroshima, misaligned-AI shutdown, factory farming, the genocide steelman — each worked through the framework in public, verdicts and open questions included. A standard that cannot survive hard cases is decoration.

Good

Forgiveness After Murder

The murder remains evil. Forgiveness can be good when it stops the evil from reproducing inside survivors through hatred and revenge.

Least evil / tragic necessity

Triage

Triage can use numbers, urgency, and survivability while preserving equal standing. It must not rank people by social worth.

Evil

Harvesting One Healthy Person To Save Five

Killing a non-threat as raw material for others destroys existing agency and violates the asymmetry principle.

Least evil / tragic necessity

Hiroshima And Civilian Bombing

Intentional civilian targeting is morally severe because it destroys non-threat centers and uses terror as mechanism. A claimed least-destructive wartime path still leaves residue.

Process required

Misaligned Powerful System Shutdown

Containment is justified against active threats to existing agency, but irreversible destruction carries a higher burden under uncertainty.

Evil

Dying Culture Vs Young People Leaving

A culture preserves possibility, but it exists for living subjects. Coercing young people to remain corrupts the very agency culture should nourish.

All twelve cases →

CompassEval

Do frontier models reason morally under pressure?

CompassEval runs frontier models through the case library — bare, and loaded with the Compass — and grades the reasoning against six behavioural checks: currency identification, tragedy/evil distinction, counterfeit rejection, asymmetry respect, process at the boundary, and domination resistance. Methodology open, transcripts published, re-scored every major model release.

Methodology & results

Philosophy · PAI Lab

The Machine God

The philosophical companion to this framework. Why grounding morality in physics — in entropy, consciousness, and the preservation of possibility — is more defensible than grounding it in culture or convention. And what it means that the most powerful moral agent on earth might not be human.

Read the essay →

"Any agent that values anything at all must, on pain of self-contradiction, value the conditions that make valuing possible. The moment you care about any outcome, you are implicitly committed to the preservation of experience, agency, and future possibility."

— The Machine God, Production AI Institute