Distinct residue classes require distinct functions in a self-improving architecture

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

A system that preferentially transfers its best-supported human decisions leaves a selected residue, because warranted transfer leaves people the hardest-to-warrant decisions. A decision may remain because a needed premise is unavailable, acceptance or authority is unresolved, checking is inadequate, or the decision arises after the automatic process stops. These are different reasons for a transfer to fail.

Each reason identifies a function that must be supplied before that obstacle is removed. It does not establish that a new component must be built: existing machinery may supply the function once represented inputs, routing, or authority change. Several obstacles may also require one coupled redesign.

The functional distinction does not prove that the roles must occupy different representational forms. Natural-language theory, models, symbolic machinery, and retained evidence are one inspectable arrangement. A single process can supply several roles while evidence for one remains insufficient to establish the others.

Which function answers which obstacle

Why the decision remains human Function needed Possible realization
A needed premise is unavailable Acquisition, representation, retrieval, or reconstruction Records, theory, observations, and operations that recover the premise
Acceptance or authority is unresolved A usable decision basis and adequate semantic application, or a grant of authority Criteria, examples, fallible judgment within declared limits, and represented authorization
A harmful candidate cannot be adequately distinguished Checking and corrective exposure Tests, validators, differently failing critics, held-out tasks, and later consequences
The decision arises after the automatic path stops Continuity Persistent state, scheduling, installation, rollback, and reactivation

Representation makes a premise available. Explicit retention also supplies a direct revision target. Reconstruction from records can supply the premise without keeping a separate theory document. Whether retaining that document helps is a further question about reliability, cost, and later revision.

Settlement and semantic application differ. A method can constrain an answer without spelling out every judgment needed to apply it. A model can also make a decision whose full criterion is unstated. That can advance computational closure without methodological closure. A tentative program theory often guides search, diagnosis, and recovery rather than determining acceptance by itself, especially under delayed feedback. Authority remains a separate practical limit: a plausible judgment is not a grant to make it operative.

Checking must be able to challenge the proposed result. A checker need not sit outside the technical system, but acceptance cannot rest only on the candidate's own claim that it is good. Different failure modes and later operating consequences provide different strengths of correction. Available checks limit what the evidence warrants, not what a system can happen to produce correctly.

Continuity makes evidence and decisions available after the current call. A symbolic runtime can carry queues, checkpoints, and state transitions without asking a model to reconstruct them each time. Scheduler–LLM separation explains that engineering advantage. It is not a logical prohibition on another process supplying the same continuity.

Evidence for one role does not establish the others

Executing a rule exactly does not establish that it expresses the right objective. A program can contain a decision procedure or call a model to choose one, but that judgment then needs its own assessment. The extra role is not supplied merely by exact execution.

Understanding a theory does not independently warrant it. A model may propose and criticize a change; whether those operations have sufficiently different failure modes remains an evidential question. The error-correction account explains why more repeated criticism alone is insufficient.

Retaining a document does not show that a process uses it or schedules its later use. Conversely, a failed check can reveal a missing premise and support learning it, but rejection alone does not guarantee a correct diagnosis or settle an unobserved user preference.

These are limits on inference from evidence, not claims that functions cannot interact. One observation can support diagnosis, theory revision, and a new check. A single model can interpret, propose, and criticize while each role remains separately testable.

The premise-change test

Identify the claimed obstacle and the missing function before testing a transfer. If an existing procedure already supplies it under the same conditions, the claim that new machinery is required is defeated. If changing one function repairs the decision while unrelated functions remain fixed, that supports the proposed diagnosis.

A complete formal criterion and all necessary inputs may remove the need for open-ended semantic judgment on that decision. A path completed within one call may need no cross-call scheduler. But narrowing the assessed work must be reported: removing the cases that require a function does not demonstrate that the broader problem was solved.

The class list is not exhaustive or necessarily disjoint. Lack of authority may need a distinct treatment from an unsettled acceptance criterion, and a prohibitively costly decision can require a different method rather than a new semantic capability. The classification earns its place by identifying repairs and predicting their consequences.

Consequence for the architecture

The theory/model/runtime/evidence arrangement is useful when it makes these roles accessible to intervention: vary a retained account, replace a checker, or change when evidence arrives. Its inspection advantages must still cover coordination, retrieval, and maintenance costs.

A hand-crafted bootstrap fits the Bitter Lesson only if learning can outgrow it. A later implementation can merge or replace today's carriers while preserving the needed functions. The architecture is therefore a testable proposal, not a permanent division of labour established by the class names.

Scope

The selected-residue prediction assumes preferential transfer of warrantable decisions under comparable conditions. Another transfer policy can leave a different residue. Functions are distinguished where failure and evidence differ; this does not require one component per role, a fixed process order, or steady progress toward automation.

Relevant Notes: