Behavior-determining organization

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

A system's behavior-determining organization is the retained structure inside its declared boundary that causally shapes how it operates on later inputs. The term distinguishes changes to the producer from changes only to what it delivers. Whether an artifact belongs depends on its later role, not on whether it was originally produced as an output.

The self-improving system definition requires a change to this organization. Such a change can alter later behaviour without improving it; benefit relative to the declared objective needs separate evidence.

Scope

What counts, across representational forms and boundary types:

  • Parameters and weights — a model's weights or a controller's gains.
  • Policies and rules — decision procedures, routing rules, review gates, and acceptance criteria, including the update procedure itself.
  • Memory and retained artifacts — notes, lessons, indexes, and caches that later operation consumes.
  • Workflows and procedures — prescribed sequences, checklists, escalation paths, and, within a declared human-agent boundary, human procedures.
  • Code, tools, and architecture — the executable substrate, available tools, and the structure connecting components.

The common test is whether changing the retained structure can change later operation, with the new inputs and other relevant conditions held fixed. Storage alone is insufficient: there must be a consumption path.

A work product can have a second role here. A software house includes the product whose evolution it remains responsible for. Product code is delivered output, but also retained structure that the house reads, executes, tests, and changes on later requests. A report reused as a decision premise or a generated tool installed for later work can likewise enter the organization. Merely labelling an output as internal does not establish that causal role.

Exclusions

  • Outputs with no retained internal role. A delivered answer or compiled program does not change the producer merely by being good. A compiler that optimizes an output without changing its own later operation has improved that output, not itself.
  • Transient state outside the claimed horizon. Scratch reasoning and intermediate results that cease to affect operation when the episode ends do not support a cross-episode claim. State that governs later decisions within an episode can belong to a system assessed over that shorter horizon.
  • External effects alone. A thermostat changing room temperature changes its environment. That effect is not a change to the thermostat's retained control organization, even though later readings may differ.

Misuse Cases

  • Counting a persistent product patch as learning without testing whether experience changed capacity on later work. Retention and a changed starting state are not by themselves evidence of learning.
  • Excluding every work product because it was first created as output.
  • Excluding evaluators or update procedures when they govern later behaviour.
  • Treating organization as only code or only weights rather than examining the declared boundary and consumption paths.

Relevant Notes:

  • Self-improving system — defined-in: the definition whose organization-change condition this term sharpens
  • Software house — exemplifies: a producer whose maintained product lies inside its functional boundary
  • Operative change — contrasts: whether a change takes effect is separate from whether it targets the organization
  • Behavioral authority — extends: identifies the consumer, channel, and force through which a retained artifact governs operation
  • Representational form — extends: classifies how the organization's parts are encoded and consumed