Addressable theory

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

An addressable theory is a theory formulated in language whose assumptions, scope conditions, and parts can be inspected and revised individually. A failure can then guide a search over candidate parts, and an edit can change selected parts while the rest stays in place.

The KB needs the term to keep a structural property apart from an epistemic status. Every theory the KB retains is a tentative theory; only some are addressable. A theory builder requires only the minimum, localized content: some unit carries what the theory says. Finer grades are Commonplace's design commitment, not a membership condition. The expected benefit is that criticism can name a part, so a revision keeps what still works while changing what failed; whether that benefit arrives is an empirical question.

Scope

  • It comes in degrees. A theory with separately stated assumptions and scope conditions is more addressable than an undivided document. Whole replacement can still change a theory's scope; addressability lets the revision target that scope separately.
  • It follows the localization axis. Of the two axes that derive representational form, localization supplies addressability under the consumption path in question. Natural-language and symbolic artifacts have parts the consumer takes up separately. Distributed-parametric state exposes no such instruction-level parts under the path on which a system runs it as a model. The same parameters are addressed numerical units under the runtime's path; changing one does not thereby identify an instruction-level commitment. See reflection buys addressability.
  • How strongly a case can contradict the theory is a separate matter. It follows the other axis, assigned consequences. Where consumers are held to a reference that fixes a part's consequences, a validator or executable test can mechanically check a contradiction within the predicate it implements. This does not make every consequence decidable. Whether the encoding captures the intended claim remains criticizable. Where a reader derives the implication from natural language without such a reference, that interpretation also enters the diagnosis. See codification.
  • Addressability does not guarantee affordable inspection of consequences. A prompt under a pinned model adopted as its reference remains separately editable, but the instruction-level rule assigning its consequences is not separately exposed in the weights. It is symbolic and can be opaque under that path. A readable script or validator can expose that rule; symbolic membership alone does not.
  • A located part is a candidate. A failure rarely identifies one faulty commitment. Popper notes that a test may bear on a large part of a theoretical system, while holding that some cases do identify the responsible hypotheses (Conjectures and Refutations, Chapter 10, section XVI). Unchanged text also does not guarantee unchanged consequences; a revision is checked against the failure and against other cases.
  • Locating a fault and containing a repair are separate. A trace can implicate a shared rule even when many other parts depend on it. That rule can be a repair candidate while checking the edit requires cases across those dependencies. Near-decomposability proposes weak interaction as a condition that can reduce this inspection burden; it is not a requirement for addressability or for identifying a candidate part.

Precedent

The classical theory-refinement systems of machine learning, EITHER and FORTE, are the fully addressable case with computed consequences: a Horn-clause rule set, a proof procedure, and named repair operators. They show that, in such a setting, proof traces can connect a discrepancy to candidate repair locations. They do not show unique fault identification, and FORTE's search can stop before its training cases are consistent. The KB cites them as a precedent for repair under addressability. It takes neither its paradigm nor a preference for minimal revision from them.

Exclusions

  • A latent world model as such. It is non-localized, so it is revised without separately stated assumptions or claims to target. An inspectable causal model or simulator program can expose those parts and be both a world model and an addressable theory.
  • Storage location. A theory kept in a file is not thereby addressable, and addressability does not require a particular store. Traces with an index locating conjectures, their parts, and their testing record can implement the same addressable objects as separate theory documents.

Misuse Cases

  • Reading mechanically checked consequences into a prose part. For a natural-language part, "derived" means interpreted until the part is codified.
  • Treating a successful local edit as confirmation of the theory that motivated it.
  • Assuming an addressable prompt's consequences can be read from an adopted model reference. The prompt is localized; that does not expose the reference's instruction-level rule.

Relevant Notes: