A methodology governs its own extension only as far as it settles the meta-decisions it raises

Type: kb/types/note.md · Tags: foundations, constraining, self-improving-systems

A methodology maps represented conditions to interventions and is actionable for an operator able to carry them out. Extending a system according to that methodology can raise further decisions: what form a new artifact takes, how its use is checked, and what authority it acquires. The method governs those decisions only as far as its retained content supplies or imports a basis for them. Call this closure under its own recommendations.

A capable actor may make the remaining decisions successfully. That shows the capacity of the actor-plus-method system, not that the method settled choices it left open. This distinction is useful when attributing reliable execution or deciding what knowledge to retain for later use.

Closure and reflection are separate properties

A reflective system has a causally connected self-representation available to its processes. It can revise that representation through judgment without a method settling every revision. A fixed pipeline can follow a complete method for a bounded task without representing or changing itself.

Neither property therefore entails the other. This note concerns the content of the retained method. Computational closure instead concerns whether execution requires human decisions.

Routing and decision content settle separately

A methodology can govern a decision in three ways:

  1. Name a decider. A maintainer approves a class of change. This settles routing, not the criterion used to approve it.
  2. Supply or import a criterion. A retained rule or referenced decision procedure constrains which result is acceptable.
  3. Determine the result. The method or an invoked procedure fixes the relevant outcome under the stated conditions.

Only the second and third settle decision content, and only to the extent that they actually constrain the consequential choice. An unarticulated human criterion or an implicit model judgment can be useful without becoming part of the method's stated content. Explicit retention makes a criterion easier to inspect and revise; the distinction does not rest on a claim that all other forms of governance are impossible.

The method need not duplicate what it imports, but it must identify the criterion or result and the pathway must use it. If a choice's variation is tolerable, the method can leave it outside the consequential decision set. That narrows the claim rather than supplying the missing criterion.

Underspecified instructions can allow different interpretations. Whether that variation matters must be assessed against outcomes, rather than treating every unstated detail as a reason the method must expand.

Three decisions an extension can raise

Representational form. Should a new operation remain model-interpreted or be assigned to code, a schema, or another symbolic consumer? A method can supply decision heuristics and conditions for codification. Having a written instruction does not settle that choice, and new theory does not automatically require new code when existing machinery can apply it adequately.

Checking. What evidence warrants using the artifact for the intended task? A method can specify tests, comparison cases, criticism, or later exposure. The verification boundary is a limit on warranted acceptance, not a theorem that an unchecked artifact cannot be correct. Identifying an oracle does not by itself establish its adequacy or independence from the proposed change.

Authority and retention. How does the accepted artifact acquire a consumer, channel, and force? The behavioral-authority account distinguishes producing a note or gate from making it govern later operation. A method that names neither its consumption path nor who can authorize it leaves that part of the change unsettled.

These are useful axes for examining a recommendation, not an exhaustive list of every decision an open-ended system can face. Their separation permits a method to settle one while leaving another to judgment.

Search is another question

In a proposal-selection loop, something must find candidates before they can be evaluated and retained. A method may govern the form, checking, and authority of a recommendation it already has without specifying how that recommendation was found. A claim that the method also governs search needs its own account.

Other learning architectures can make evidence-responsive changes without a separate proposal and admission event. The same question about retained guidance still applies, but the loop's stages should not be imposed as a universal architecture.

What the assessment can establish

The Commonplace reference case records decisions settled for a completeness mark alongside decisions that still depended on judgment. It is an application of the distinction, not proof that Commonplace settles all self-extension decisions.

For another pathway, name the consequential decisions, the retained content said to settle them, and the actual consumption path. A decision outside that content bounds the method's coverage. Reliable execution beyond that boundary can show that the actor's competence compensates for it. Whether extending the method improves reliability or cost then becomes an empirical comparison.

The two-layer execution account provides one arrangement: use settled guidance where it applies and fall back to theory-guided judgment elsewhere. Recurring useful judgments can motivate new guidance. Fallback is not necessarily a failure, and recurrence is not a proof that the new guidance is sound.

Scope

A method can be complete for a declared bounded decision set. This note does not assert that every real method is incomplete or that open-ended extension must eventually fail. It identifies which decisions are attributable to the retained method and which require something else.

The account also separates method coverage from authority, correctness, and retention policy. A specified criterion can be wrong; a correct recommendation can lack permission to take effect; and a specified temporary computation need not be kept as reusable code. Each claim needs its own evidence.

Relevant Notes: