The Right to Act

A proposal and an authorization are different computational objects.

This distinction is becoming more important as intelligent systems become capable of generating code, operating tools, controlling machines, conducting experiments, and moving money. A model may understand a situation and still lack the evidence, resources, or permission required to make its preferred future real.

Connecting a model to an API without representing that distinction creates operational risk.

Between perception and action lies another object: authority.

Authority as a typed object

Authority is often implemented as a binary permission outside the intelligent system. A user can call a tool or cannot. A process has access to an account or does not.

Real delegation has more structure.

An authorization may specify:

• which object may act;
• over which resources;
• within which risk and cost envelope;
• for which class of conditions;
• at what cadence;
• until which invalidating event;
• with which required attribution;
• and who or what may revoke it.

It records a bounded claim over resources and actions.

Conceptually:

evidence + context + lens + constraints
    -> authority
    -> action
    -> consequence
    -> new evidence

Every authorized action produces evidence that should update future authorization.

The allocator after the leaderboard

When candidates are arranged on a leaderboard, allocation looks simple. Choose the highest score.

But the shape of better is often a frontier rather than a point, and the frontier itself depends on costs, constraints, time, and available evidence. A system may preserve a population of behaviors that trade return against risk, speed against precision, novelty against reliability, and immediate utility against information gain.

A lens maps the current context onto that field. It may apply constraints, express a utility function, reveal a Pareto-efficient set, or produce a probability measure over several candidates.

The allocator turns that projection into a resource decision.

It can:

• authorize one candidate;
• weight a compatible ensemble;
• keep every candidate in shadow;
• allocate evaluation compute;
• request a discriminating experiment;
• preserve a dormant specialist;
• compile a proven mechanism into a faster layer;
• or abstain.

The allocator supplies differential access to attention, reproduction, resources, and real action.

Abstention is an action

Systems that must always emit a winner are easy to control and difficult to trust.

Sometimes the available evidence does not distinguish the candidates. Sometimes every behavior is outside its tested domain. Sometimes the expected advantage is smaller than the cost of acting. Sometimes the most useful operation is to observe until a particular uncertainty resolves.

Abstention should appear as a typed control state with a reason, an expected wake condition, and a record of what remained unresolved.

This is especially important when a population doubles as an instrument. If all candidate reactions become incoherent together, forcing one to act destroys the distinction between a hostile environment and a broken map.

Abstention therefore belongs in the control algebra.

Authority across clocks

Putting general reasoning inside every fast decision would violate latency and cost constraints.

A trading policy cannot wait for a frontier model while a market event unfolds. A machine controller cannot suspend its feedback loop while an agent rereads the entire project. A scientific instrument may need to react before a human can inspect the observation.

The slower process instead authorizes a bounded competence to act locally.

A fast policy may respond immediately while its evidence remains inside an envelope established earlier. A medium-clock statistical process can adjust confidence, suppress failing behavior, or change resource weights. A slow agent can inspect the wider evidence field and revise the envelope, admit a specialist, commission a mutation, or withdraw authority altogether.

slow clock
    changes the available repertoire and authority envelope

medium clock
    updates belief and allocates within that envelope

fast clock
    acts locally without reconstructing the whole argument

The slow agent does not need to approve each event. It decides what may happen without it.

This is the practical complement to multi-clock intelligence. Competence is compiled downward. Authority is delegated downward with it. Failure, drift, and unresolved novelty escalate upward.

Agents as lenses

Human utility functions are usually expressed in a few dimensions because those are the dimensions we can inspect.

An agent need not be restricted to that geometry. It can consume a high-dimensional evidence field containing objectives, behavioral descriptors, uncertainty, lineage, recent environment response, cost structure, and counterfactual outcomes. It can construct a temporary lens appropriate to the decision rather than relying on one global scalarization.

The agent gains representational capacity while remaining vulnerable to overfitting, rationalization, and evidence selected after the outcome. Its advantage is the ability to reason over structure that would otherwise be flattened for human convenience.

The safeguard is replayability. The evidence cut supplied to the agent, the lens it applied, the alternatives it considered, and the authority it emitted should remain inspectable. A later observer can evaluate the decision using information available at the time rather than the story told afterward.

The agent’s decision remains an auditable allocation proposal.

Public agency

Authority is easiest to trust when its consequences are visible.

Imagine a public system showing a live environment alongside a repertoire of possible behaviors. Some activate. Some disagree. Some remain quiet. Their changing evidence rearranges the relevant part of the behavior atlas.

One behavior receives authority. Orders appear on a public account. Inventory changes. Costs and outcomes return to the same evidence field. Faint counterfactual paths show what other candidates would have done. A slow-clock intervention appears only when it changes the available repertoire or authorization.

The interface exposes the transition from available behavior to authorized action.

The same pattern applies outside trading. A laboratory allocates instrument time. A software system authorizes a rollout. An industrial controller permits a new operating range. A research organization chooses which uncertain branch deserves another month of attention.

In each case, the allocator sits between a field of conditional knowledge and scarce contact with reality.

Agency that can be inspected

An action should be traversable backward:

consequence
  <- action
  <- authority
  <- allocator decision
  <- lens
  <- evidence
  <- behavior and lineage
  <- source

If any link is hidden, a successful outcome can receive the wrong explanation and a failure can teach the wrong lesson. Provenance determines which part of the system is allowed to adapt.

This is also why hidden composition is dangerous. If an authorized policy appears to act but an undeclared fallback actually produced the behavior, execution updates the wrong lineage and the resulting adaptation is misdirected.

Inspectability allows authority to remain local and temporary without becoming arbitrary. It lets the system delegate aggressively where competence is established and escalate where its boundaries become visible.

Intelligence expands the space of possible action. A population preserves alternatives. Lenses expose the relevant subset under the current context. Allocation gives selected possibilities resources and authority.

Authority is the typed transition through which a possibility is allowed to alter shared state.