Prediction is not the hard problem.
Authority is.

Models are fallible and regimes move. The durable problem is controlling what a system may do when its inference is wrong. Neurobrick treats prediction as input and authority as an explicit, auditable state.

Inference is advisory

A model, rule or researcher may produce a proposal. Being right does not create permission.

Authority is explicit state

Live-capital access must be represented, bounded, attributable and readable by the gate that consumes it.

Safe direction is asymmetric

Automation may tighten a guardrail or reduce risk inside an envelope. Opening or expanding authority requires a different class of proof.

Evidence is scoped

A passing receipt supports only its system, surface, instrument and observation window.

The self-approval door

A boundary under observation, not an absolute slogan.

The intended boundary is human authority for opening or expanding live-capital access, with bounded machine authority in the safe direction. The dated system review states which observed permissions support that distinction and where actor attribution, surface coverage or effective state remains open.

Scope

This is a governance architecture, not a track record. Returns, balances, strategy parameters and investment performance are outside the public surface. The question here is whether authority is bounded and whether the evidence supports that claim.