Self-revision design space: problem map

This file preserves the workshop's live problems, rejected formulations, and evidence base. The plain account is the article-facing synthesis; the boundary map holds the formal bookkeeping behind it (backstage by operator constraint). Earlier versions of this file treated “complete self-revision relative to a kernel” and “acceptance bounded by position” as candidate central claims; the cases forced both to be narrowed before promotion.

The formal model, axis dispositions, decomposition senses, and settled/moving/unknown status live only in the boundary map, and their plain-language consequences only in the plain account. Keeping both out of this file is deliberate: multiple current syntheses would recreate the drift problem the workshop is trying to diagnose.

Formulations narrowed or rejected

One static kernel

The original kernel bundled permanent boundary exclusions, effective non-reach, warranted freezes, unrepresented organization, a source-state governance scheme, and the authority cut for a current transition. They are not one set. The Gödel-machine case is decisive: incumbent axioms and utility can license successors that replace them, so they can participate in one transition's authority cut without necessarily being immutable components.

Use the relational accounting vocabulary in the boundary map: declared inventory obligations, determination and installation reach, controlled reach C, improvement-warranted reach W, bounded-experiment reach E, declared exclusions, warranted effective freezes, unrepresented remainder, source-state scheme Σ_t, transition conformance C_Δ, and its authority-cut subrecord A_Δ. The relations preserve change class, pathway, source state, horizon, and evidential standing; no bare target-set equation is adequate.

Raw subset ordering of “kernels” is therefore meaningful only inside a stable inventory, boundary, and granularity. Even there, an exclusion leaving F is progress only if the target enters controlled reach with the improvement warrant or bounded-exposure authorization the assessment context requires.

Governed reach as one relation

The direct-update test broke this formulation. A path can conform to a source-state control scheme while lacking evidence that its updates improve the objective. Conversely, an uncertain update can be responsibly authorized because its downside is bounded without acquiring improvement warrant. Report C, W, and E separately and use governed only as an umbrella with the intended readings named.

The same test added a first-class update-law profile. A direct path needs its evidence-to-successor mapping, constraints, stochasticity, scope, and assumptions; an evaluator profile alone silently restores proposal-selection architecture.

Objective plus final adoption veto as the universal floor

The objective is needed semantically to call behavior better. A final adoption veto belongs only to a proposal-selection path. A direct updater can be controlled by an incumbent update law and scope bounds, while its improvement warrant rests separately on the evidence relation and domain justification. Monitoring and recovery can limit exposure but do not supply improvement evidence. Treating a veto as universal would re-smuggle proposal-selection architecture into the definition.

An incumbent objective can license a successor relative to that incumbent. A stronger claim that a terminal objective itself became better under a different normative index requires a comparison level outside the pair. Neither case requires one permanently fixed objective artifact.

Acceptance standard bounded by target position

This formulation gave the target too much explanatory work. Position changes the directly available checks, intervention costs, coordination requirements, and blast radius. It does not set a general ceiling on acceptance quality. Natural-language instructions can be judged through downstream deterministic tests, symbolic programs can receive semantic review, and formal proof remains conditional on the formalization matching the external objective.

Proof, benchmarks, LLM judgment, and human judgment are evaluator regimes with different domains, premises, costs, and failure modes. They are not one strength ladder derived from representational form.

Commitment as a universal unit

Commitment works when semantically addressable content can be named. It does not yet individuate cleanly across prose, code, schemas, weights, bindings, and topology. The path-valued-type test rejects persistent target identity across architectural change: an obligation-migration record instead combines many-to-many lineage edges with explicit introduced and retired sets. What remains open is how to anchor the lineage nodes and transport evidence, not whether row or file identity suffices.

Economical evidence base

The article does not need one episode to support every distinction. Use simple cases for the distinction they actually expose.

  • Tag-README completeness — strongest simple case for changed verification allocation, symbolic enforcement, and later use. It also exposes boundary and granularity sensitivity.
  • Reports layer — simplest role-inventory change: replace-in-place generated snapshots gained a role distinct from durable notes and consumed workshop work. The ADR records the decision; migrated outputs and later generators establish uptake.
  • Proposal lifecycle — separates truth-apt claims, undecided designs, and implemented decisions. The article-layer proposal later reached ADR 057, establishing use. Creating proposal-selection machinery does not by itself prove that this creating episode was proposal-selected.
  • Darwin Gödel Machine — exposes horizon and boundary sensitivity: archive admission is viability-gated, benchmark score shapes reproductive selection over later generations, and broad code mutation can reach decomposition while diagnostician, controller, or objective remain outside the target.
  • Gödel machine — exposes the difference between broad syntactic writability, narrow provably reachable change, transition-relative authorization, and permanent hardware exclusions. It is a theoretical capability case, not an observed revision episode.

Technically richer reserves remain path-valued type contracts and the general freshness store. They are useful when interface revision, migration closure, transactionality, or subsystem extraction is the target; they are too expensive for a first illustration.

Hard boundary cases

  1. Writable but inert: a configuration key is updated but no live consumer reads it. This can show change authority and installation without operativity.
  2. Reflective but read-only: a self-model mediates controller behavior but cannot itself be revised. Reflective coverage, behavioral authority, and later exercise do not imply revision reach.
  3. Direct parametric update: online gradient descent can be effective and operative without a self-representation or rejection event. Any universal account must describe its warrant without inventing a gate.
  4. Self-ratifying rubric: an updater rewrites both instruction and rubric and then passes the successor rubric. Broad addressable and operative reach does not supply authority-independent warrant.
  5. Rotating evaluators: evaluator A governs installation of B; later B participates in governing a replacement for A. A source-state authority cut can avoid wholly successor-conferred authority with no permanently fixed evaluator, but it does not prevent substantive self-ratification or establish that either transition is warranted.
  6. Fixed but invalid check: protecting an evaluator from revision does not make it distributionally valid. Fixedness and warrant are independent.
  7. Opaque rollback: checkpoint restore can provide recovery without semantic addressability; conversely, a rollback command without detection, attribution, and a live restoration path is only raw reversibility.
  8. Binding versus contents: a model's contents may be unmodifiable by the pathway while the binding to the whole model is selectable. Reach changes with target aspect and granularity.

Candidate durable outputs

  • Note: a self-revision claim requires a versioned revision profile, separating affordances, determination reach, installation reach, controlled reach, improvement warrant, bounded-experiment authorization, operativity, and recovery.
  • Note: a source-state authority cut can exclude wholly successor-conferred authority without residing in a permanently frozen component, while non-circular grounds and adequacy remain a separate warrant trace.
  • Note: inventory-changing revision needs an obligation-migration record; splits, merges, introductions, new interfaces, and retirements prevent coverage from transporting by identity.
  • Note or definition clarification: decomposition revision can concern role inventory, allocation/realization, or topology/interface; the comparison dimensions remain provisional.
  • Article: an aspect-bounded thesis about where each builder function lives: the experimental pathways make different role, middleware, rule, and code redesigns operative while leaving different evaluators, ontologies, update protocols, and controllers supplied; Commonplace retains part of a human–agent redesign pathway as operative and revisable system machinery; the Gödel machine handles the meta-level problem through broad self-referential redesign but cannot adopt beneficial changes it cannot prove.

The three completed case tests are in Discriminating tests. The hostile reading forced the aspect-bounded correction and has been integrated into the note and article; the formal model remains backstage because the corrected comparison can be stated in plain language.