Self-revision design space

Goal

Explain where architectural redesign occurs in the systems compared by the reflective self-improvement article. The experimental pathways expose different slices of organizational redesign while leaving different evaluators, update protocols, ontologies, or controllers supplied. Commonplace has made parts of a human–agent builder loop explicit, retained, and operative, with an affordance for turning some current organizing machinery into a later revision target. The theoretical Gödel machine supplies the other comparison point: extraordinarily broad self-referential redesign, but only when its benefit is provable from the incumbent formalization. The plain account is the integrated synthesis; the hostile reading records the counterexamples that corrected it.

Along the way, develop only as much supporting theory as the explanation needs: what makes a change to the parts list harder to control than a change within a part, how procedure, warrant, and permission to experiment differ, and how coverage claims survive a reorganization.

Three operator constraints (2026-08-03) bind this workshop:

  1. Use-first. Extract into the article as soon as a result becomes load-bearing. Do not wait for a complete catalogue or model.
  2. Open design space. Completeness of any axis catalogue is doubted; preserve the selection test rather than implying the inventory is exhaustive.
  3. Plain language first. The central explanation must stand without special notation. Formal records (Σ_t, C_Δ, coverage relations, migration records) are backstage bookkeeping: keep them only where they earn their place, promote them only if a real consumer needs them, and never let an article-facing claim depend on them.

Where the work stands

The current central claim is not that Commonplace's maintainers can redesign it while other research teams cannot. It is that Commonplace retains parts of the human–agent redesign pathway itself—not only its outputs—in artifacts and authority paths that later operation consumes. The reports, proposal-lifecycle, tag-README, and type-contract episodes show repeated revision of organizing machinery; the proposal-to-article sequence supplies one clear case in which newly installed design machinery was reused in a later architectural change. This supports an affordance for moving up another meta-level, not a claim that Commonplace occupies an achieved infinite level, has complete architectural coverage, or delivers autonomy, compounding, or superior outcomes.

The hostile reading rejected a whole-system inside/outside classification. Continual Harness demonstrates repeated sub-agent-role redesign; Autogenesis versions and reuses agent implementations; Self-Harness can propose structural mechanisms and retained middleware; DGM descendants can reorganize agent code. The surviving comparison is aspect-bounded: name the redesign class a path reaches, then identify the still-supplied machinery. An operative builder-level change is not yet an internal builder loop unless the path remains available for another challenge.

The first thesis — complete self-revision relative to a declared kernel — was narrowed by the case tests: "kernel" bundled several different boundaries (permanent exclusions, currently unreachable targets, warranted freezes, unrepresented organization, and the per-change authority arrangement), and the Gödel machine shows authority can rotate — incumbent machinery can license its own successors. The surviving plain form is temporal rather than spatial: nothing must stay outside forever, but incumbent conditions govern each transition, and passing that structural test still is not evidence the change was good.

A second claim — acceptance standard bounded by representational form — was also narrowed: form changes which checks are directly available and what coordination costs, not a general ceiling on achievable assurance.

Both narrowings, and the three distinctions the tests forced (procedure vs warrant vs bounded experiment; coverage re-asked after reorganization; demonstrated vs writable), are stated without notation in the plain account.

Why a separate workshop

The KB fixes the local distinctions — self-improving system, reflective system, reflective coverage across forms, effective update space, behavioral authority, the Gödel-machine corner, omitted vs frozen — but not their composition into one comparison the article can use. The adjoiner holds the proof → benchmark → semantic-judgment material in the article's register.

Evaluation boundary

  • Compare systems on demonstrated changes, stating procedural control, evidence of benefit, and permission to experiment separately; report raw write access only as the outer envelope. Note each claim's evidence standing: declared, theoretically possible, demonstrated, or routinely exercised.
  • A revision claim is relative to a declared boundary, objective, target, pathway, and horizon; changing one changes the claim. State the ones that matter in prose — do not require the full coordinate apparatus to make an article-facing point.
  • Keep three questions separate: what a target affords, how a change through it is controlled and warranted, and whether the declared inventory is covered.
  • "Decomposition change" has three provisional senses — role inventory, duty allocation, topology/interface — and one revision can span several; classify at the claimed granularity.
  • An axis or distinction earns its place only by discriminating something a real case needs; artifact kinds and correlated bundles do not qualify.
  • Absolute completeness is not auditable: an omitted piece of organization is discovered only by counterexample.

Initial questions

  1. Resolved by the hostile reading: several papers make narrower organizational redesign operative; the claim survives only per named redesign class, not as a binary placement of whole systems.
  2. What is the minimum the article must say about "who had the last word and why was their word worth anything" for each Commonplace episode it cites?
  3. When a reorganization splits or merges roles, what is the cheapest honest way to say which prior assurances still hold?
  4. When is a change properly called a decomposition change rather than a large edit — and at what granularity?
  5. What evidence could ever support "nothing consequential is unrepresented," beyond failing to find another omission?
  6. Which parts of Commonplace's current improvement process can become targets through that same continuing process, and which still form an external meta-level?

Working artifacts

  • Plain account — primary: the article-facing explanation, notation-free by constraint.
  • Problem map — live problems, narrowed formulations, evidence base, hard cases, promotion candidates.
  • Revision profiles and moving boundaries — backstage formal synthesis; parked unless a consumer needs a record the plain account cannot carry.
  • Discriminating tests — the three case tests that forced the narrowings; backstage.
  • Hostile reading — strongest counterexamples, the aspect-bounded correction, and article-integration consequences.

What closes this workshop

Practical closure, not completeness:

  1. the plain outer-builder-loop explanation is promoted into the article (and its load-bearing distinctions into kb/notes/), or rejected with the failure recorded;
  2. the "something outside each change" successor of the kernel thesis is promoted, narrowed further, or dropped;
  3. every promoted claim passes the plain-language constraint — stated without the workshop's notation;
  4. the formal apparatus is either consumed by a real need or explicitly retired with the parked records left in place;
  5. remaining threads are routed to notes, peer workshops, or explicit open problems.

Complete file listing (generated at build time)