Borrowing can operate through retained artifacts or weight activation

Type: kb/types/note.md · Tags: foundations, methodology, context-engineering

Commonplace has treated borrowing from programming, philosophy, law, cognitive science, and other fields mainly as artifact borrowing: interpret an external idea, justify its transfer, adapt it to the target system, and retain the resulting claim or methodology in the KB. Pretrained agents add a second route. If an external methodology is already coherently represented in model weights, a prompt can activate it directly without first restating it in Commonplace.

The two routes are:

retained borrowing
external source → adaptation → retained artifact → context → behavior

weight-mediated borrowing
external tradition → pretraining → model weights → compact cue → behavior

The distinction changes what it means to borrow from a mature discipline. A field can contribute not only explicit concepts that Commonplace adopts into its writable theory, but also compact names and frames that select substantial pretrained methodological structure at execution time. Established traditions therefore form part of the usable intellectual substrate of an LLM agent even when their contents are not copied into the KB.

The routes have different properties:

Property Retained artifact Weight activation
Context cost requires loading some explicit content can be very low when a compact cue suffices
Inspectability high low
Selective revisability high low
Versionability high model-dependent
Project-specific adaptation direct indirect unless supplemented in context
Fidelity to a canonical interpretation can be specified depends on the model's learned reconstruction

This means the source-adoption policy and weight-mediated activation answer different questions. The adoption policy asks whether an idea deserves to become part of Commonplace's retained theory or methodology. Weight activation asks whether an already learned external methodology can be used as part of the current computation. A concept can be useful through the second route without passing the bar for explicit adoption, because using model competence does not make that competence a retained Commonplace commitment.

The routes can also compose. When the model already contains most of a useful methodology but Commonplace needs a project-specific variation, the context can supply only the explicit delta:

weight-resident baseline methodology
        +
retained project-specific modification
        +
current mission and constraints
        ↓
operative methodology for this task

For example, a task could supply the gloss “preserve intent and constraints while delegating execution-time choice of means,” then add Commonplace-specific constraints on irreversible changes, evidence, or review. That hybrid uses the bounded mechanism in intent-framed delegation is a control regime, not a short prompt without assuming that the bare cue Auftragstaktik reconstructs the same package in every model. The weight-resident methodology provides context-efficient background structure; retained artifacts specify the parts that must be precise, writable, auditable, or different from the conventional doctrine. The retained delta is not only departures. It also holds selections among equally conventional alternatives, which the weights cannot supply because they hold every alternative and not the project's choice (a specific intent may out-yield local rationales, but contingent facts stay separate names these arbitrary conventions), and restatements of practices the model already endorses where they must bind rather than advise (system-definition artifacts are crystallized reasoning under context scarcity separates that authority-bearing class from recoverable heuristics).

This hybrid explains why the existence of broad model knowledge does not remove the need for an external KB. The model weights provide a large but opaque and effectively read-only methodological library. The KB provides the writable complement: claims and methods whose exact content the system wants to inspect, revise, version, and selectively activate. The design question is therefore not simply whether knowledge exists in weights or in artifacts, but which parts deserve explicit retention and which can safely remain a parametric dependency.

The distinction also creates a new transfer risk. Artifact borrowing can be reviewed against a source and the target-side mechanism. Weight-mediated borrowing additionally depends on whether a particular model actually reconstructs the intended methodology from the cue. A familiar label may activate a distorted, underspecified, or internally mixed representation. Where that difference would change consequential choices, the route therefore needs either a disambiguating gloss or target-side evidence from representative cases. Familiarity with the label is not evidence of activation fidelity. Economical activation trades context cost against methodological fidelity.

Relevant Notes: