Theory building has distinct epistemic, structural, and implementation precedents

Type: types/note.md · Tags: learning-theory, self-improving-systems, theory-builder

A system can criticize a theory of an external subject without representing itself. It can use a causally connected self-representation without criticizing it. It can also edit persistent files without formulating criticism of what a theory says. These possibilities separate three questions when comparing systems: how errors are criticized, what the representation is about, and how changes persist and affect operation. A precedent for one does not establish the others.

A theory builder names the system that proposes theories, acts on them, criticizes what they say, and builds on the result. Reflection adds a relation between a system and a representation of itself. Retained text and code around a fixed model supply one implementation. Keeping these questions separate makes both attribution and experiment design more precise.

Popper supplies the epistemic basis

Popper's schema, P1 → TT → EE → P2, describes problems, tentative theories, attempted error elimination, and the further problems that result. Criticism can replace a theory or expose a new problem; the schema does not prescribe local edits or a repository layout (A Realist View of Logic, Physics, and History). His treatment of linguistic formulation and criticism includes informal arguments before symbolic formalization (Epistemology Without a Knowing Subject).

Commonplace adds the conditions for attributing this method to a particular system: its theories are stated in localized units, guide its decisions through what they say, are criticized for what they say, and are revised through iteration, where the result of criticism shapes the next round. How far that result persists is graded. These conditions contain no success condition. Whether a builder improves its capacity for future action is a separate learning claim, tested rather than assumed. Artifact availability alone does not establish the conditions, and a changed decision does not establish improvement. The epistemic basis therefore leaves work for a concrete account of consumption, consequences, and persistence.

EITHER and FORTE serve a narrower role as precedents for connecting failed consequences to candidate repair locations. Their proof-guided repair shows what addressability can make possible under a supplied representation. It does not define every tentative theory, require small revisions, or establish the same capability for interpreted prose. One invocation over a supplied theory and training set is a theory builder at a low grade of persistence: the repaired theory is a new conjecture, re-tested on the examples, and nothing persists beyond the run (theory builder, boundary cases). Their precedent value lies in assigning blame, not in persistence.

Reflection supplies the structural relation

Computational reflection supplies the causally connected self-representation relation inherited by the reflective-system definition. Runtime models and requirements representations give concrete ways to expose selected aspects of a system to its own operation.

In Rainbow, probes and gauges update an architectural model; a constraint evaluator triggers adaptation through supplied strategies and effectors. Its specified types, properties, operators, and strategies determine the available adaptations. That is a working model-to-system path, not evidence that the system learns those choices (Rainbow (snapshot required), architecture-layer infrastructure and architectural style). Requirements Reflection proposes runtime requirements objects synchronized with an architectural representation. The synchronization is a design challenge in that paper, not an evaluated learning mechanism (Requirements Reflection (snapshot required), challenges 1–2).

These comparisons matter when the theory concerns the modifying system's organization. A runtime self-model can be causally operative while its assumptions are closed to criticism. Conversely, a theory builder whose theories concern an external subject need not be reflective. Improving an adaptation outcome does not by itself establish learning of the model or its revision machinery.

Persistent workspaces supply an implementation comparison

Workspace Optimization keeps model weights fixed while revising typed code and role context. Prediction failures are routed to responsible surfaces, and replay supplies feedback on earlier transitions after edits. Its DreamTeam implementation uses a supplied role decomposition and evaluation machinery (Workspace Optimization (snapshot required), §§2–4 and Appendix B).

This is a concrete comparison for keeping learned changes in artifacts that later calls consume. It does not establish learning of a continuing system's own improvement procedure. The retained source reports adaptation within game runs. Changes carried across the levels of one run meet the iteration condition when criticism aims at what the revised code and context say; their persistence reaches the run, not later sessions. The benefit of the complete arrangement also does not isolate the effect of criticism, addressability, or retaining an assembled theory.

Fixed weights constrain where learned changes persist. They do not hold model processing fixed across different inputs, establish content use, or show that explicit artifacts outperform reconstruction. Those are separate comparisons of the arrangements actually run.

Research positioning and comparison

Commonplace's research arrangement pursues recursive self-improvement within Schmidhuber's broad program, using interpreted methodology and selective codification. This states the research objective and chosen realization, a theory builder whose learning is under test. It does not replace the Popperian account of criticism or the structural condition for reflection, and it supplies no result showing that the combined arrangement works.

A comparison should therefore name which question it asks. A runtime-model comparison concerns what criticism aimed at what a theory says adds to learning beyond a causally operative self-model. An editable-workspace comparison concerns the supplied theory and its use beyond persistence and repair. A persistence comparison concerns which work is kept rather than reconstructed, and for how long. None can inherit the conclusion of another merely because the systems share some machinery.

Scope

These are bounded conceptual and implementation precedents, not a priority claim or an exhaustive ranking of earlier systems. A full claim that a reflective theory builder learns needs connected evidence of theory use, consequences, criticism, and later influence, plus evidence of improved capacity. Partial findings remain reportable at their own strength. Neither their combination nor resemblance to a predecessor establishes superiority over simpler arrangements.

The precedents already supply recurrent use of revised theories, incremental revision, localized repair, a bounded sample-efficiency result, and human-engineered seeds. This program must not claim those features as new. Fixed supplied machinery is not inherently defective when it is warranted general machinery over the declared reach.

Relevant Notes: