Theory mediation can coordinate heterogeneous factory development

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

TODO! remove or update to use software house instead of software factory

Natural-language theory may be a particularly versatile mechanism for factory development because one LLM-interpretable account can connect task structure, relevant solver limitations, observed failures, proposed interventions, scope conditions, and evidence, then guide changes across heterogeneous production machinery.

Two claims must remain separate. Minimal factory learning does not require theory: trial-and-error retention, trajectory reuse, program search, learned policies, direct optimization, or mixtures can all produce retained changes used later. For the harder target of open-ended coherent modification, however, this program follows Naur in treating a project-theory function as indispensable. Some project-specific state or capacity must connect the program to its purposes, account for its organization, and relate new demands to what should be preserved.

The necessity attaches to that function, not to one carrier. A project theory may be distributed across weights, artifacts, tools, and participants, and it need not be fully explicit. The contingent hypothesis here is that an addressable natural-language theory gives an LLM-based system useful leverage when a change must coordinate several artifact kinds or when decisive feedback is sparse, delayed, or arrives under a structured shift.

The coordination problem

A failure in software production may not identify the artifact that should change. The same symptom can arise from:

  • a mistaken understanding of the task or product family;
  • a decomposition that drops necessary dependencies;
  • context selection that omits relevant evidence;
  • an intermediate representation that cannot express required distinctions;
  • a tool or algorithm that performs the wrong transformation;
  • an evaluator that rewards a proxy;
  • a workflow that stops before delayed consequences arrive; or
  • conflicting schemas, prompts, tests, and runtime assumptions.

The remedy can therefore span natural-language guidance, symbolic software, schemas, tests, workflows, retrieval policy, and model-facing context. A fixed update operator for one artifact kind may repair the local symptom while leaving the shared explanation inconsistent elsewhere.

A retained theory can instead state a cross-artifact commitment such as:

The final decision depends on relations among evidence units, so independent summaries are insufficient; extraction may be parallelized, but the representation and aggregation path must preserve cross-unit relations for a later reconciliation pass.

An LLM can interpret that claim into changes to decomposition, intermediate schemas, map-reduce code, prompts, aggregation logic, and tests. The theory is useful only if it actually changes those decisions and remains revisable when later outcomes contradict it.

What theory can represent in one medium

A factory-relevant theory may include:

  • task or domain claims — what distinctions, dependencies, invariants, and variations matter;
  • solver claims — which operations are reliable, bounded, expensive, or prone to a known failure mode;
  • production claims — why a decomposition, workflow, representation, or tool should work;
  • evaluation claims — which observations would support or defeat the proposed organization;
  • scope claims — where the account applies and where it should not transfer; and
  • revision claims — which parts should change when a prediction fails.

The relevant self-knowledge is therefore not a complete theory of the model's internals. It is a task-relevant theory of the relation among the task, the current solver, and interventions that can make the task tractable.

A useful theory can be weak and fallible

Theory mediation does not require a complete account that deductively yields the right factory change. Real software development commonly proceeds from partial and sometimes mistaken understanding. A developer uses that understanding to choose a promising change, inspects the result, encounters conflicts, backtracks, and revises both the program and the understanding of it.

A computational theory-holder should be judged by the same longitudinal standard. A weak theory can still be operative when it narrows search, identifies commitments worth preserving, interprets failure, or tells recovery what to restore or revise. Backtracking is not evidence that theory was absent; it is one way a fallible theory remains corrigible under incomplete information.

partial theory
  -> candidate change
  -> provisional check
  -> later or external consequence
  -> retain, repair, backtrack, or revise
  -> changed theory and later search

A theory that never risks revision is not made stronger by surviving only the evidence it selected for itself. Independent tests, later demands, rival explanations, and negative transfer are needed to turn plausible guidance into learning.

Theory guides search; it does not replace it

The theory can shape a generate-and-verify process by controlling:

  • which failure explanations are plausible;
  • which machinery is worth inspecting;
  • which candidate decompositions or tools to try;
  • which experiments are informative;
  • which checks should reject a proposal; and
  • how an observed result should revise the retained account.

Blind or stochastic exploration can remain inside the process. A learned policy can propose candidates. Program search can construct executable machinery. Code can perform exact transformations and tests. Trajectories can preserve evidence about earlier decisions. These mechanisms may construct, revise, apply, compile, or implicitly embody project theory rather than simply compete with it.

The theory-mediated claim is causal: addressable theory changes search, diagnosis, evaluation, recovery, or revision. The serious functional rival is a system that sustains coherent modification without any project-specific state performing the mapping, justification, and integration functions assigned to program theory.

The scheduler–LLM separation gives the complementary implementation principle: semantic interpretation can remain model-mediated while exact progression, bookkeeping, and stabilized invariants move into software. Theory can guide which invariants should be codified without requiring the LLM to execute them unreliably on every run.

Why versatility is plausible

Theory mediation has four potential advantages for heterogeneous factory learning.

Cross-artifact reach

The same explanation can guide coordinated edits to prompts, schemas, workflows, tests, code, tools, and documentation. This can reduce the need for one separately trained update policy per artifact class.

Selective revision

Addressable claims can be revised or rescaled without discarding the whole retained state. A failed prediction may defeat one scope condition or mechanism claim while preserving unrelated parts of the factory theory.

Delayed credit assignment

A theory records why an earlier production decision was made and what consequences it predicted. When a later requirement exposes damage, the retained joins can connect the consequence to the earlier decomposition or representation choice.

Transfer under structured shifts

A mechanism-level claim may apply to a new product variation even when surface details differ. This could improve sample efficiency relative to replaying only concrete trajectories, provided the theory captures real structure rather than a plausible story.

These are hypotheses. Natural language also introduces ambiguity, retrieval cost, contradiction, interpretation error, and rationalization. A theory can coordinate several artifacts into the same mistake.

Comparative predictions

Against information- and budget-matched alternatives, an explicit natural-language theory surface should earn support when it:

  1. transfers one retained explanation into several causally coherent machinery changes;
  2. predicts where a decomposition, representation, evaluator, or tool will fail before exhaustive trial;
  3. improves recovery after delayed feedback by identifying the earlier commitment that should change;
  4. supports targeted rescoping instead of copying or discarding an entire trajectory;
  5. reduces new human construction of family-specific production knowledge when a related but unanticipated product variation appears; or
  6. leaves an intervention-sensitive trace: withholding, replacing, or corrupting the theory changes consequential factory-development choices.

It should narrow or lose as a representational and coordination strategy when:

  • the retained theory is ignored or reconstructed post hoc;
  • direct search or optimization reaches better machinery at comparable total cost;
  • an implicit or differently represented project theory guides modification as well with less maintenance;
  • trajectory reuse transfers as well without theory-level organization;
  • theories become self-confirming because their evaluators share the same assumptions;
  • cross-artifact coordination increases correlated error;
  • human judgment required to maintain theory grows with the system; or
  • each new domain still requires people to supply the decisive decomposition, family knowledge, and evaluator.

A loss by the natural-language surface would not by itself show that the successful system had no project theory. It may show that the same functional organization was better carried elsewhere.

Theory content and factory machinery are different roles

A natural-language artifact can be consumed as evidence, advice, an instruction, a constraint, or a generator input. Merely storing a theory beside the factory does not make production theory-mediated. The causal path must show how the theory changes model calls or executable machinery and how later evidence revises the same retained surface.

Conversely, theory need not remain in natural language forever. Stable claims can be codified into validators, schemas, tools, or workflows. A brittle symbolic rule can be relaxed back into an interpretable hypothesis. Theory mediation describes the learning relation, not a requirement that every useful result remain prose.

Scope

  • Minimal factory learning does not require theory. The indispensability claim is scoped to open-ended coherent modification where local criteria do not exhaust program purpose and organization.
  • The indispensable object is a project-theory function, not necessarily an explicit natural-language artifact.
  • The claim about natural-language theory concerns versatility across factory-development decisions, not universal superiority on every task.
  • Natural-language theory is one representational surface inside the deployed system; model weights and symbolic machinery remain essential.
  • A theory of the task alone may be insufficient when decomposition depends on solver limits. A theory of the solver alone may be insufficient when task dependencies determine what must be preserved.
  • Theory mediation can coexist with fixed general machinery and with non-theoretical search at lower levels.
  • Reflection is not required for every theory-mediated factory change. It is claimed only when a causally connected representation of selected aspects of the same system participates in operation or revision.
  • Greenfield's factory specialization remains a particular operation on reusable factory structure; it is not a synonym for acquiring family-specific production knowledge from evidence.

Relevant Notes: