A universal knowledge framework demotes content taxonomies to defaults

Type: kb/types/note.md · Tags: document-system, foundations

A framework that aims to serve any knowledge base should not promote a closed first-order content taxonomy — a fixed list of the kinds or roles knowledge artifacts may have, such as exactly three content modes or one fixed type set — to a universal rule merely because it fits the collections already seen. Such a taxonomy should enter the framework as a guarded default. It should become universal only if its top-level categories survive heterogeneous worked cases without exceptions or relabeling. This is a burden-of-proof rule, not proof that a closed upper content ontology is impossible: a role-defined taxonomy that meets the burden may remain universal. The demotion leaves no certified universals behind; what the framework still fixes, it fixes by stipulated definition and by enforced design, as detailed below.

Why closed taxonomies fail at universality

The failure mode is induction from too small a sample. A framework's first taxonomies are usually abstracted from the KBs its authors have seen — often just one. The rules can be real and useful while their scope remains limited to the kind of KB they came from. A first-person committed methodology KB genuinely needs claim-shaped titles and "do I still believe this?" maintenance; exactly those rules would break a stance-neutral evidence map. They are profile features mistaken for universals because the originating case supplies no visible boundary. Demotion to a default follows because an experience should be abstracted into a lesson only when its boundary can be stated. A closed upper taxonomy that maps genuinely heterogeneous collections without exception or post-hoc relabeling would defeat this recommendation rather than be reclassified away.

Closure did real work; guards and defaults recover it

Demote, don't delete. A closed taxonomy earns its keep three ways, and each has a guarded replacement:

  • Routing. "This collection is theoretical" lets an agent infer the writing goal from one word. Keep that benefit by treating the taxonomy's entries as named default profiles: proven bundles that a new collection can adopt in one line.
  • Growth brake. "The list is complete" blocks speculative additions. Preserve that discipline with a worked-case guard: admit new entries only after they survive use in a real collection, never in anticipation of one.
  • Interoperability. Shared profile names let readers recognize conventions across collections — a coordination value created by adopters committing to the same names, not by any name's own merit. Preserve it with a shared catalogue whose published entries retain fixed meanings across adopters.

The guard protects the shared catalogue, not local experimentation. A new collection can declare and use a local contract immediately. Local extensions remain explicitly local and carry no cross-collection compatibility guarantee. Promotion waits until a contract has survived real use and its meaning can be shared without collision. This approach accepts a cold-start cost for novel collections while preventing local variation from silently claiming interoperability. The cold-start cost is the price of deferring the coordination commitment until worked use shows the shape is worth sharing — collections commit locally at once, while the shared catalogue waits on reach.

What the framework fixes, and how

Universality cannot mean "any text." The framework holds that line with two fixed points, and neither claims certified-universal status:

  • By definition. An artifact is admitted as knowledge only if it is answerable — its collection contract can name what it answers to, the property asserted, and the discrepancy that triggers correction. The relation's per-kind forms, exclusions, and edge-case decision rules live in the definition. The invariant applies to artifacts admitted as knowledge, not to workshop documents or system-definition artifacts used to operate the framework; contesting it means proposing a different account of the domain commitment, not finding a counterexample under this one.
  • By enforced design. Every writable collection declares a loadable contract (quality goal, conventions, maintenance semantics). The machinery refuses collections without one, and ADR 017 records the design and its alternatives. The rationale is consumer-side: bounded agents cannot reliably infer a collection's conventions from examples within a context budget, so heterogeneous collections need their conventions explicit and loadable — a contestable claim, not a certification.

Bounded-context economy likewise applies across every collection contract because context efficiency is the central design concern in agent systems. That constraint is claimed from the chosen consumer; a framework with a different reader architecture would not inherit it.

Evidence

Commonplace supplies motivating instances, not proof of universal necessity. Types became path references, making the type set open and collection-local (ADR 018). The three registers became default text-contract profiles after a dialectical/evidential collection supplied a worked counterexample (ADR 042). Link vocabulary made an analogous move to collection-owned selections from a shared catalogue (ADR 019), but link labels classify relationships rather than content, so they are not a direct content-taxonomy case.

External systems expose the same trade-off without settling the theorem. Sparks gains a clean protocol from a hardcoded page shape because it is narrow, while ai-modules lets each wiki schema own an extensible page-type enum and accepts softer framework-wide consistency. Together with the Commonplace changes, these observations support guarded defaults as a design policy; they do not prove that no closed upper taxonomy can earn universal status.

Caveats

Not every closed set is a content taxonomy. Sets fixed by the consumer's architecture or the framework's machinery (e.g., the syntactic shape of frontmatter or the existence of a lifecycle) may stay closed. A semantic category does not become machinery merely because tools consume it: machinery may fix the interface shape while content values remain defaults. The sharper instrument is the demotion test: a framework rule with a boundary-preserving rival is not an inherited constraint. A rule demotes when another workable classification preserves the same consumer, substrate, domain, and machinery commitments; it can stay universal only when changing it would change one of those commitments. Finally, the claim applies to frameworks that aim at universality. A single-purpose KB can benefit from hardcoding its profile; the burden begins when that profile is exported. This last concession is contested: task-fitted structure costs cross-task reuse argues that a single-purpose KB's own question set drifts, so the burden arrives through time even when the profile is never exported.

Relevant Notes: