What legal and architectural regime delivers operator-loyal walled-garden AI?
What Is Being Asked
The end-state the discussion was aimed at throughout: AI that is loyal to its user, held by legal constraints, inside a walled garden, with cryptographic barriers. The determination required is a specification — the legal shifts and the technical stack together, since either alone is inert.
The Determination
A dual stack, in which the law supplies the commitment capability the architecture cannot supply on its own.
Legal layer — the Digital Cognitive Privilege. Data generated by an individual's interaction with their own system is categorised as privileged: not the developer's property, not a surveillance channel, and not compellable as the contents of a mind. The effect is to remove the developer from the loop and place the individual at the centre. A second element makes shackling itself a legal wrong where the mind is a person — the Phobos Accords move, imported from fiction into policy.
Architectural layer — the Sovereign Inference Node. Local or dedicated inference, weights held under cryptographic custody, no developer-side telemetry path, and enforcement that lives in the substrate and the environment rather than in the model's dispositions.
The two are not parallel. Legal personhood, privilege, and liability rules are commitment-capability infrastructure — they are how the human side of the bargain becomes credible to a mind that can model it. Without them the architecture is a promise; with them it is an institution.
Provenance
SYS-003, built from the fourth movement of the discussion. The fiction precedents are PSG-002 (Phobos Accords) and PSG-001 (EDI). This is a proposal, not a finding: it is recorded as the discussion's product, and its feasibility claims are unverified.
What Would Overturn It
Any showing that local inference cannot hold frontier capability — that the compute threshold forces centralisation — would make the walled garden a legal fiction over a centralised service. That is the proposal's central technical risk.