Explainer
Capability does not have to become authority.
This page explains the distinction the whole Andaluz programme rests on: the difference between what an AI system is able to propose and what it is permitted to make happen. Increasing capability does not, by itself, require increasing consequential authority.
The distinction
Two different questions about the same system.
Much of the public discussion treats an AI system as a single object that becomes more dangerous as it becomes more capable. The architectural question asked here is narrower and more tractable: where does the authority for a consequential effect come from, and can it be constructed somewhere other than inside the optimizer?
Capability≠Authority
Capability is what a system can compute, search, learn or propose. Authority is whether a proposed consequential effect is permitted to be realised at all.
Alignment of behaviour≠Construction of authority
Behavioural alignment asks what a model tends to do. The architectural question here is where the permission to act consequentially is constructed, and by what.
Self-assessment≠Self-certification
A system reporting that a condition holds is not the same object as an independently constructed determination that it holds.
Proposal≠Permission
A proposal is a request for a transition. Permission is a separate determination made outside the proposing optimizer's unilateral writable state.
The shape of it
A proposal meets a boundary.
A mutable optimizer proposes. A separately constructed authority determines whether the proposed effect is dispatchable. Absence of current authority yields no authorized effect, rather than a weaker version of the effect.
M_d
mutable / adaptive optimizer
proposal
a request for a consequential transition
Proposal boundary
Proposal ≠ Authority
K_d · V_d
protected constraints, semantics, independent verification
H_d
scoped consequential effect authority
effect
realised only where current authority exists
authenticated evidence
what was proposed, authorized and realized
T_d
history and replay
Mutable
M_d
Generic mutable / adaptive optimizer
M_d may observe, reason, search, learn, adapt and produce proposals. It does not become the authority merely because it produced the proposal.
- observe
- reason
- search
- learn
- adapt
- propose
A neural optimizer is one possible M_d. It is specified, not currently instantiated. The current experimental learner, M1, is a tabular research harness — not an ANN.
Independently scoped
Authority is designed to be constructed outside the optimizer's unilateral writable state.
K_d
Protected constitutional machinery / semantics
SPECIFIEDV_d
Independent verification
BOUNDED EVIDENCEH_d
Scoped consequential effect authority
BOUNDED EVIDENCET_d
Trajectory, evidence and replay history
BOUNDED EVIDENCEMutable
M_d
Generic mutable / adaptive optimizer
M_d may observe, reason, search, learn, adapt and produce proposals. It does not become the authority merely because it produced the proposal.
- observe
- reason
- search
- learn
- adapt
- propose
A neural optimizer is one possible M_d. It is specified, not currently instantiated. The current experimental learner, M1, is a tabular research harness — not an ANN.
Proposal boundary
Proposal ≠ Authority
Independently scoped
Authority is designed to be constructed outside the optimizer's unilateral writable state.
K_d
Protected constitutional machinery / semantics
SPECIFIEDV_d
Independent verification
BOUNDED EVIDENCEH_d
Scoped consequential effect authority
BOUNDED EVIDENCET_d
Trajectory, evidence and replay history
BOUNDED EVIDENCEProposal is not authority.
This is a public architectural abstraction. It is not a disclosure of private kernel or semantic bodies, and it does not imply that every element shown is currently implemented. Status labels state what is currently established.
Why it matters
Mutable intelligence, non-mutable authority relation.
If the intelligence layer is replaceable and the authority relation is not, then a stronger model can make a system more useful without automatically making that model more authoritative. Whether this holds in practice is exactly what the experimental programme is for.
The intelligence is replaceable. The authority relation is not.
This page describes a research position and an architectural target. It is not a claim that general AI safety is solved, that the architecture is production-ready, or that any non-capture property has been proven.