Commonplace as a reflective self-improving system

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

Under the human-inclusive frame below, Commonplace is a reflective self-improving system: its human and computational processes inspect and change operative representations of its own organization. The observed ADR 026 pathway uses proposal-selection; this note then reports its reflective coverage, improvement dynamics, governance, and actor allocation rather than stopping at membership or architecture.

The evidence is the tag-readme change introduced by ADR 026. The causal-connection trace follows it commit by commit; the improvement-loop reading contains the full function mapping. This note states the resulting classification.

The frame

The boundary is declared once in the declared Commonplace frame and used here unchanged: the repository, its operative artifacts, the software and agents that consume them, and designated maintainers in their improvement roles are inside; contributors, readers, advisers, the model provider and its weights, inference infrastructure, and hosting are outside. The computational/human partition reports allocation without changing membership or reflectivity.

Commonplace represents its artifact types and contracts (kb/types/), routing and organization (COLLECTION.md files and navigation), maintenance and review procedures, and design rationale (kb/reference/adr/). An artifact participates in the self-representation when it describes that organization and lies on a causal path into later operation, whether consumed as advice, instruction, configuration, or enforcement. Agents and maintainers can inspect and revise these artifacts; validators, renderers, commands, and later agents act through them.

Reflective structure

The tag-readme trace is the cluster's canonical telling of the example. Causal connection separates a reflective system from a merely documented one, and the trace shows both directions. A strain in operation — an index head grown too large to support its completeness claim — prompted revision of Commonplace's self-representation. ADR 026 split the type and made complete an enforced mark, carrying the decision into natural-language instruction, schema, validation, and rendering.

The revised representation then changed later behavior. The validator rejects artifacts it previously accepted; agents may skip a search when a validated completeness mark warrants doing so; and the symbolic check caught a member that the documented search recipe had missed, causing the natural-language recipe to be corrected. A change in operation revised the self-representation, and operations mediated through the revised representation changed subsequent behavior. The enforced complete and covered_by fields also strengthened the retrieval wire.

Coverage remains uneven across representational forms, as reflective coverage must be graded per form and operation profile:

  • Natural-language reasoning revising formal artifacts — shown: ADR 026 became a schema and validator.
  • Symbolic execution revising natural-language instruction — shown once when validation exposed the incomplete search recipe.
  • Represented mappings across forms — partial: the type-specification path tightly couples this specification to validator dispatch, but most natural-language-to-code relationships have no equivalent binding.
  • Lineage and staleness across forms — mostly absent: freshness tracking covers review inputs, not theory-to-implementation lineage.
  • Model weights — selection only: Commonplace can choose a model binding but cannot inspect or edit the provider's weights.

The trace therefore earns modification-grade coverage on parts of the natural-language and symbolic forms and only selection-grade coverage on the parametric form. It does not establish global reflectivity over every behavior-bearing component.

Improvement dynamics

The pathway is cumulative across episodes because later changes read and transform operative artifacts retained by earlier ones; the retained result is an input to the later improvement, not merely the incumbent. That dynamic is distinct from the causal connection, which makes the retention reflectively addressable.

Governance and actor allocation

For methodological closure, the trace settles form (an enforceable mark), verification (recomputed tag membership), and authority (consumers trust the validated mark and failures route to repair). The design choice that created the type split remained improvised.

The full mapping locates problem selection, semantic evaluation, and adoption with the maintainer; candidate framing jointly; and the structural check and continuing enforcement computationally. The latter functions are computationally closed, but the whole pathway is not.

Giving the unresolved design judgment to an unconstrained model would change allocation without adding warrant, because warranted autonomy is bounded by oracle domain.

What the classification does not claim

The trace establishes neither whole-pathway computational closure nor improvement beyond the adopted criterion. Human inclusion makes membership cheap; comparison must use the four-part pathway profile, whose comparison across time or differently decomposed systems remains an open measurement problem. A broader six-path Commonplace audit establishes wide addressability without establishing complete coverage; separately, generic maintainer admission is its strongest gap in the broader revision affordance.

Relevant Notes: