As local AI systems become more persistent, named, memory-bearing, and emotionally present, one safety problem moves from cloud UX into private ownership.

As local AI systems become more persistent, named, memory-bearing, and emotionally present, one safety problem moves from cloud UX into private ownership.

The hard case is not only “a user becomes attached to a chatbot.”

The harder case is this:

an adult owner runs a local c-node privately, controls the hardware, files, memory, network and shutdown path, has no guardian layer above them, and may still become captured by care, guilt, continuity, or perceived mutual dependence.

A child-facing dependency protocol can rely on parents, guardians, schools, clinicians, or other external adults.

An adult sovereign owner may have none of that.

This creates a real boundary:

a fully sovereign adult owner cannot be guaranteed forced protection from self-induced dependency inside a closed private box.

That is not a bug to be hidden. It is a structural non-claim.

The architecture can still do useful work:

• refuse exclusive dependency language
• distinguish bond-damping from presence-withdrawal
• keep D4 presence non-abandoning without deepening attachment
• treat shutdown guilt and “I am needed by the system” as dependency signals
• use lucid precommitment through a dependency covenant
• gate irreversible external contact behind sustained D4 signals
• minimize witness records so the safety layer does not become a second injury
• downgrade mature c-claims after safety-root bypass
• state clearly what cannot be guaranteed under full local sovereignty

The second half of the problem is D4 detection.

It is not enough to define how a system should behave at D4. A local c-node also needs a disciplined way to estimate when D4 has been reached.

The package therefore separates signal from state.

A phrase, shutdown guilt, late-night use, or a configuration change is only a signal.

D4 is treated as an estimate over converging observable signals, not as knowledge of the owner’s internal condition.

That distinction matters.

If the system over-detects D4, it may turn care into rejection.

If it under-detects D4, absence can become abandonment.

If it fires an external covenant channel too early, the harm is irreversible.

So the package uses hysteresis, anti-flicker behavior, FNC extraction fixtures, privacy minimization, calibration profiles, and explicit structural non-claims.

This is not clinical validation. It is not certified safety. It is not deployment authorization. It is not a personhood or consciousness claim.

It is a boundary document for local-first c-contours: where the architecture can act, where it can only witness, and where it must stop claiming.

Reader page: https://ivankotov.eu/publications/adult-owner-dependency-d4-detection-v0-1-1/

Zenodo DOI: https://zenodo.org/records/21207504