Case packet

Neutral case identifier: case-b940441aeb021c

The possible directed relationship from Artifact A to Artifact B is under review.

Artifact A

Self-improving system

A self-improving system makes operative changes to its own behavior-determining organization, where those changes are causally responsive to evidence bearing on an improvement objective.

Its own means the object of change is the system's [behavior-determining organization] — its parameters, policies, memory, rules, workflows, code — not an external work product. A compiler that optimizes programs is not self-improving; a compiler pipeline that rewrites its own optimizer is. This is Ashby's two-loop distinction: operating a system is one loop, modifying the system that operates is another. The attribution is assessed against a declared boundary: a model fine-tuned by an external training pipeline is being improved, while the composite of model plus pipeline self-improves — [the boundary cases make this dependence explicit].

Operative means the change affects subsequent operation over the relevant horizon, through a consumer, a channel, and a force — [operative change], which does not require permanence; a transient compensation, or a change nothing ever acts on, does not qualify.

Makes is read over a declared assessment horizon, like operativity: a system is self-improving over that horizon when evidence-responsive operative self-change occurs within it. The dispositional attribution — the system has a standing improvement pathway, currently exercised or not — is also available, but it is a different claim and must be marked as such; a pathway nothing has exercised over the relevant horizon supports only the dispositional reading. Tense, like boundary, is declared rather than fixed by the definition.

Evidence-responsiveness does not require a gate

Responsive to evidence is defined in [evidence bearing on an improvement objective]. There must be a loss, reward, error, viability bound, test, judgment, or other criterion for the evidence to bear on; otherwise the change is merely caused, not improvement-directed.

The evidence may directly determine an update that is always adopted, or it may evaluate a candidate that can be rejected. A separately represented candidate, evaluator, or acceptance gate is therefore not required by membership. The [proposal-selection improvement loop] owns that named subtype and its search, evaluation, oracle, and retention vocabulary.

An improvement criterion is required semantically; an explicit evaluator is not required architecturally.

What membership leaves open

Membership establishes improvement-directed self-change, not a complete architecture or a successful outcome. Evidence-responsiveness can faithfully pursue a bad objective, and an evaluator can accept a harmful change: only outcome evidence establishes that improvement occurred.

Classify the remaining questions elsewhere:

  • [Reflective system] owns whether the pathway changes itself through a causally connected self-representation; reflection is not required for membership.
  • [The pathway profile] owns reflective structure, improvement dynamics, governance, and actor allocation; [the cumulativity criterion itself] is held separately.
  • [Methodological and computational closure] owns the two closure readings, and [warranted autonomy] owns when unattended evaluation is trustworthy.

Exclusions

  • Not self-modification alone. Blind or accidental rewrites lack evidence-responsiveness.
  • Not regulation alone. A thermostat changes its environment, not its own behavior-determining organization; a learning thermostat that revises its controller does.
  • Not work-product improvement. Improving an answer or external code does not change the improving system's own organization.
  • Not reflection, a gate, or autonomy. Each may describe a member, but none is a membership condition.
  • Not guaranteed success. The term names an improvement-directed mechanism, not a favorable outcome.

Misuse Cases

  • Treating the proposal-selection improvement loop as the definition rather than a named subtype, which re-smuggles an architecture into semantics.
  • Reporting an autonomy grade without declaring the boundary it was assessed against.
  • Attributing self-improvement without naming the objective it is indexed to, or naming one the pathway's evidence is not diagnostic of — [the first leaves the attribution elliptical, the second makes it false].
  • Reading a dormant improvement pathway as current self-improvement — the dispositional claim (has a pathway) and the horizon claim (evidence-responsive change is occurring over this horizon) are different attributions.
  • Treating a helpful change to an external product as self-improvement of the tool that produced it.

Provenance and departures

One departure is semantic rather than a retired restriction: the predicate is frame-indexed. The bearer of the property is a bounded system — a system under a declared boundary — not a substrate simpliciter, so an attribution is elliptical until the boundary is named. Established classification practice reads category membership frame-independently; the fine-tuning-pipeline case in [the boundary cases] is why that reading fails here — the same substrate is being improved under one boundary and self-improving under another.

Two earlier restrictions were retired because they excluded central cases. Requiring self-representation excluded parametric learners; the substantive benefit of reflection now belongs to [reflection buys addressability]. Requiring reject-capable evaluation excluded direct gradient- and viability-driven adaptation; proposal-selection now names that narrower architecture.

This architecture-neutral choice is consistent with uses of “self-improving” for gateless self-tuning algorithms ([Ailon et al. 2011]) and with self-adaptive-systems literature treating feedback-loop models as engineering reference models rather than definitions ([Weyns]; [Petrovska, Erjiage, and Kugele 2025]). It is Commonplace's explication, not a claimed field consensus.

Relevant Notes:

Artifact B

Reflective system

A reflective system contains a causally connected representation of itself that is available to its own processes, such that operations mediated through that representation can affect the system's subsequent behavior.

The structural core is the established term in computational reflection: a system that reasons about itself in a causally connected way, through structures that represent selected aspects of it — its self-representation. Commonplace applies the criteria relative to an explicitly declared boundary without prescribing whether human actors are included. Departures from the source literature are recorded under [Provenance].

Changing the boundary alone does not satisfy the definition. The representation must be available inside the boundary as a representation of that same system, not merely as telemetry or a control signal correlated with its state.

Scope

Calling a system reflective requires five things to be stated:

  • The system boundary—what processes, artifacts, and environment count as the system;
  • The represented aspects of that system and their granularity;
  • A self-representation that exposes those aspects;
  • Processes inside the boundary that can inspect or act through the representation; and
  • The causal connection by which changes in the represented aspects update the self-representation and representation-mediated operations affect later system behavior.

Reflection is therefore aspect-bound, not global. The terminology and granularity of the self-representation determine which questions and interventions the system can formulate about itself — Maes's theory-relativity point. No self-representation exhausts every possible aspect, although it may be complete relative to its declared set of represented aspects. A reflective architecture may also retain an unrepresented or unmodifiable kernel.

The following neighboring terms mark different capabilities:

Term Criterion
Self-description Information about the system is present, but need not affect operation.
Introspection A process can inspect or reason about represented system state.
Reflection The self-representation is causally connected to the represented system and participates in later operation.
Intercession Reflective access permits direct modification of represented system state or interpretation.

Intercession is a capability within reflection, but not every reflective architecture permits it.

No artifact authority family is required. A representation consumed as evidence or advice can mediate reflective change when it participates in the causal path into the represented organization; a nominal [system-definition artifact] that is not a self-representation cannot. The criterion is causal connection through a representation of the affected aspect, made precise by the consumer, channel, and force in [behavioral authority].

Retrieval-mediated causal connection

The causal connection need not be an interpreter or a compiler. Where the self-representation is a body of retained artifacts, it can run through discovery: a process searches the artifacts, finds the ones bearing on what it is doing, and derives its behavior from what it found. Retrieval is then the wire the representation acts along, and it is best-effort where a compiler is exhaustive — so [retrieval failure is reflection failure], which develops what that costs and how the wire is strengthened.

Exclusions

Reflection is not autonomy, successful self-improvement, formal verification, or closure under a set of recommendations. Nor is it organizational closure, the recursive regeneration of a network of component interactions, or autopoiesis, the narrower self-production of a living system ([Varela 1981, printed pp. 14–18; PDF pp. 1–5]).

Reflection is also not adaptation, and this is the misreading the term invites most. Both directions have occupants.

Reflection and intercession without adaptation. A Smalltalk image is the case to hold in mind. Classes are objects; methods can be added at runtime, superclasses changed, message dispatch intercepted, the compiler edited with the compiler. Implementation and self-representation are one structure, so the causal connection is as tight as it gets and intercession is total. Left alone, the image sits there for a decade and improves nothing — nothing in it notices that a method is slow, decides the change is worth making, or judges whether the rewrite helped. The programmer supplies the evidence-responsiveness. Remove the programmer and the improvement pathway is not weakened; it is absent. A reflective architecture permitting intercession supplies a causal path by which a system could change itself. It does not make any change responsive to evidence about an improvement objective — neither directly, nor through the search, evaluation, and operative retention that [a proposal-selection improvement loop requires].

Adaptation without reflection. Conversely, a system can improve itself with no self-representation at all. Ashby's Homeostat jogs its parameters to random values whenever its essential variables leave viable limits, holds the configuration that restores them, and so adapts — viability-driven, evidence-responsive change to its own organization ([Ashby 1960, chapters 7–8]). It is still not reflective, because nothing in it represents that organization. What it retains is a setting rather than a map, and [what that costs is addressability].

The two properties are orthogonal: reflection is structural, adaptation is a process, and each is available without the other. [Self-improving-system membership] is defined separately; this note owns only whether a named pathway routes change through a self-representation.

Misuse Cases

  • Calling documentation reflective because it describes the software that stores it, without showing a causal path into later operation.
  • Expanding the boundary after a failure so that any helpful outsider counts as an internal reflective component.
  • Calling a telemetry-driven controller reflective when its signal is not available inside the declared boundary as a representation of that same system.
  • Using reflexive and reflective interchangeably without identifying a distinct property that the new term would name.

Provenance and departures

The criteria above are inherited; the extensions are not. Collected here so the definition can be read as a definition, and so each departure is auditable in one place.

  • Causal connection, self-representation, theory-relativity — inherited. Maes defines a reflective system as a computational system that reasons about itself “in a causally connected way,” and names the structures representing selected aspects its self-representation ([Maes 1988, printed pp. 1–2, 14–17; PDF pp. 1–2, 14–17]). The introspection/intercession split is corroborated in [Wuyts and Ducasse 2001]; the embedded-self-theory lineage is [Smith 1984].
  • Retrieval as a causal connection — Commonplace's own. No source treats discovery over retained artifacts as the wire. See [retrieval failure is reflection failure].
  • Terminology reservation. Commonplace retains reflective system across boundary choices and reserves reflexive system for a future concept only if it names a distinct property. Nothing currently requires the second term.

Relevant Notes:

Under-review context phrase

the causally connected self-representation, and the intercession capability, that reflective self-improvement routes through