Natural-language project state may specialize weight-resident search heuristics

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

A pretrained language model may already contain general heuristics for noticing anomalies, generating alternatives, ranking promising directions, selecting probes, persisting, and backtracking. It does not already contain the current state of a particular project.

The conjecture is that retained natural language can specialize those general heuristics by supplying the project's intent, working theory, prior branches, failures, commitments, and constraints. The language need not encode a complete search procedure. The weights may supply the general competence while the retained natural-language state determines how it applies here.

This note isolates one part of project state, not the whole state available to an agent. Source code, tests, schemas, configuration, execution traces, and other symbolic artifacts also constrain search. The natural-language contribution becomes operative when context assembly places it in a prompt; the resulting semantic operation is instantiated jointly by that prompt and the model's weights. Code complements the weight–prompt pair with independently executed symbolic operations.

The relevant evidence is behavioral. Withholding, replacing, or perturbing the natural-language state while holding the model, symbolic project state, tools, and task fixed should change branch choices and their later consequences. A plausible explanation that leaves search unchanged does not establish the conjecture.

Scope

  • This is an empirical conjecture, not a claim that current models already allocate open-ended search effectively.
  • A retained natural-language artifact is not automatically a prompt. It joins the weight–prompt operation only when supplied as model input.
  • Showing that code or other project information improves search does not by itself establish the distinctive contribution of theory-level organization.
  • Specializing search heuristics does not give the resulting choices acceptance authority.

Relevant Notes: