Load-bearing vocabulary collisions should be prevented or visibly scoped at write time

Type: kb/types/note.md · Tags: computational-model

Technical terms are hard to use in a knowledge base because natural-language content offers no reliable, mechanically enforced namespace for an unqualified technical sense. In the risk model proposed here, a technical sense is load-bearing when misresolving it changes what an agent does, which maintenance regime applies, what authority a claim carries, or whether a fallback is licensed. Symbolic systems use scope to keep identifiers apart: two maps in different namespaces need not collide, because formal position participates in resolving meaning. Natural language has no equivalent: readings are delimited only by grammar and context — conventional, fallible cues. A lexical technical sense without a visible clause frame or schema position therefore binds globally and weakly:

  • Globally — no marker delimits where the binding applies; every co-loaded occurrence of the word is a potential use site of the technical sense.
  • Weakly — no validator enforces the binding; a reader or an LLM pass can resolve the wrong sense, and nothing necessarily detects the error.

This is the corpus-scale form of a pathology already named at session scale: LLM context is composed without scoping names it collision. At corpus scale the colliding senses live in durable notes written months apart, and the collision can recur whenever the notes are co-loaded.

The failure is at composition, and it is silent

Context assembly is concatenation, not import. A symbolic system resolves name collisions at link time — qualified names, renaming, or a duplicate-symbol error. Concatenated natural-language content has no rename step and no linker error. An agent can read two notes that use one term in different technical senses, blend them, and produce a conclusion that appears to inherit authority from both notes while following from neither. A reader sometimes notices the clash, but that noticing is itself an inference — vulnerable to distance, phrasing, and load — and it emits no error signal when it fails to fire. This is the shape of natural-language content has no reliable dereference: a formal operation replaced by an inference that may work but cannot be relied on as a control.

The deficit is not inherent to natural language as a medium: books are prose, and they do not suffer this collision problem in practice. What protects a book's vocabulary is not formal scope but a boundary object. Opening a book is a manual act of entering an attentional frame that stays active for the whole reading session; a chemistry textbook's reduction never collides with a rhetoric textbook's reduction because no reader is inside both books at once. A knowledge base removes that boundary object: retrieval pulls notes across boundaries that used to separate books, with no signal that a crossing happened. The risk, then, is not that natural language lacks formal scope — ordinary reading has survived without it for as long as books have existed. It is that KB composition strips away the informal single-source boundary that let natural language get away with lacking it, and puts nothing in its place.

Distillation — the retired term that named both entailment-preserving reshaping (methodology → skill) and evidence-to-rule generalization (traces → preference rule) — illustrates the stakes. The two operations carry different maintenance regimes: one is checkable against its retained source, the other must earn authority through testing. The bare word marks no boundary between them, so co-loaded notes could blur which regime applies.

Write-time prevention is the conservative default

Three candidate remedies offer different levels of assurance. Reader vigilance — noticing the clash at read time — is a useful backup but too silent and fallible to be the only control. Symbolic disambiguation — per-note sense declarations ("in this note, X means…") that readers resolve before composing — can work, but it adds work to every read and write, and a missing declaration silently restores the original ambiguity. Write-time prevention or visible scope — reserving each unqualified load-bearing term for one sense across co-loadable artifacts, and placing necessary reuse in a schema slot or linked clause frame — removes the dependence on read-time inference altogether.

Retrieval-time disambiguation looks like a fourth remedy: treat the collision as word-sense disambiguation and let the composer assemble context in which each occurrence resolves correctly. But it decomposes into the first two. Where the retriever resolves senses from declared markers, it is symbolic disambiguation executed mechanically — stronger than leaving declarations to readers, but the declarations must still exist and travel with every excerpt, so writing them is write-time scoping by another name. Where it resolves senses semantically, it is reader vigilance moved into the pipeline: the same silent inference, weakest exactly here. Classic word-sense disambiguation leans on sense inventories a model's priors already know, but a KB's load-bearing senses are local coinages defined in single notes — an excerpt that drops the qualifier leaves only the bare word, and the prior pulls toward the common reading. The assurance a retrieval solution adds is precisely its symbolic, write-time ingredient; the semantic remainder inherits the failure profile of the reader it replaced.

Session-level isolation does not substitute for this, and neither does a retriever that simply keeps colliding notes apart — that only relocates the same move earlier in the pipeline: if a task needs both artifacts, separating them defeats the composition the task requires. The both-notes-needed case rules out separation, not co-loading: annotating both notes with sense-distinguishing markers — provenance metadata, sense qualifiers — does make the co-load safe, but that is not a rival remedy. It is visible scope executed as metadata: the same write-time scoping by another name. Body-composability needs either a collision-free vocabulary or visible, checkable scope for the technical use, and collection-local vocabulary is reliably safe only when a colliding sense cannot co-load — which a cross-linked KB makes hard to guarantee.

Devices rank by binding strength

Schema-defined positions offer the strongest scope available: a link-label token, a tags: value, a type-spec field, or a frontmatter key can be unambiguous because it occurs in a validator-defined slot. Position can disambiguate even when spelling alone cannot. Moving a binding from natural-language content into such a slot is codification applied to vocabulary itself. The natural-language devices rank below it by binding strength:

Device Binding Failure mode
Definition note + vocabulary entry Corpus-wide natural-language convention, not mechanically scoped Drift or conflicting use
"In this note, X means…" Local declaration, not preserved when excerpted A citation can lose the qualifier
Clausal binding + required link (actionable) Grammar-scoped, link-anchored convention Mis-predication or a missing link
Rare compound token (trace-learning) Distinctive exact string, corpus-auditable Accidental reuse or jargon opacity
Link label / field / tag in a defined slot Positional, validator-enforced Schema ambiguity or content misuse

The two strongest rows compose. A rare hyphenated compound occupying a validator-defined slot is both cheap to audit corpus-wide and mechanically checkable in position — derived-from and its siblings, used as link-label relation names, are this composition, with the spaced phrase (derived from) left free for ordinary prose. Neither ingredient alone suffices: a bare rare compound still depends on writers keeping it out of unscoped natural-language content, and a slot alone does not stop a common word placed there from dragging in its ordinary associations. At the other end, a definition note by itself is the weakest device — it states a binding but neither enforces nor delimits it.

Rare compounds make the collision check lexical

The practical default for new internal technical terms is a compound of two or more words. A coinage creates no formal scope; its value is that it changes the kind of check collision control requires. Auditing whether a common word already carries a technical sense demands semantic classification of many occurrences; for a rare exact string, rg lists the few uses directly, turning most of the audit into a cheap lexical screen with semantic reading reserved for the residue. The component words also constrain each other's reading, offering a partial gloss before the definition loads, and an unfamiliar compound imports fewer established associations than a redefined common word — fewer, not none.[^pseudowords] Trace-learning and discovery lifecycle carry their technical senses on the searchable compound rather than the common bare words learning and discovery.

[^pseudowords]: No string is semantically neutral ground: even pseudowords — orthographically legal non-words — elicit form-based meanings that humans and language models converge on (de Varda et al., 2024). A coinage trades a strong established prior for a weak form-driven one rather than opening a blank namespace entry.

Borrowed terms follow the same risk logic: context engineering is a low-risk borrowing because the local sense aligns with established use, while distillation was high-risk because its machine-learning association fails to mark the boundary the KB needed. The cost of coinage is opacity to outsiders — coined terms are banned from outward-facing copy — and each coinage is still a global name every reader carries: the cost side of the minimum viable vocabulary is the naming set that most reduces.

Prototype shorthands are not aliases

A familiar prototype may introduce a broader technical category in outward-facing language, but only as a visibly approximate gloss, not as a second canonical name. “Prompts, code, and weights,” for example, can orient a reader toward natural-language, symbolic, and distributed-parametric artifacts. The mapping is not a synonym table: a natural-language artifact need not be a prompt, a symbolic artifact need not be code, a prompt can contain both natural-language and symbolic operative parts, and weights are only one kind of distributed-parametric state.

The prototypes select common instances or uses, whereas the categories name representational forms. Since artifact analysis separates representational form from consumption path, lineage, and behavioral authority, treating a prototype as its category would collapse axes the vocabulary exists to keep separate. Production method does not determine form: a program-generated note can remain natural-language, while a hand-authored schema can be symbolic where a validator assigns its elements defined consequences.

The approximation must therefore carry visible scope — “in practical terms,” “roughly,” “prototypically,” or an explicit example frame — and the canonical term must return before a load-bearing classification, maintenance rule, schema value, or inference depends on the distinction. “Prompts are the most familiar example of operative natural-language artifacts” preserves the boundary. Letting a prototype alternate unqualified with its broader category creates the same globally weak binding this note warns against.

A predicate can receive limited grammatical scope: the technical sense is licensed only where the word predicates an explicit subject in the same clause, as with actionable in actionable methodology — "a methodology is actionable for an operator when…". Grammar alone is weaker than a distinctive string: the frame depends entirely on writer discipline, no validator enforces it, and an exact-string search cannot confirm it because every candidate clause needs semantic reading. That is why clausal binding is admissible only with a required definition link. The link does not prevent mis-predication, but it makes each intended technical use visible and spot-checkable — under the convention, an unlinked actionable simply has its ordinary reading. A corpus audit on 2026-07-18 found all 36 bare uses of actionable across the library to be ordinary ones: exactly the many-innocent-occurrences profile that makes capturing a bare common word dangerous, and evidence the frame convention has held so far.

The invariant targets load-bearing technical senses

The rule is not "eliminate polysemy" — polysemy is normal in natural language, and a KB that fought all of it would fight its own medium. The target is load-bearing polysemy, and the proposed invariant is:

A load-bearing sense should be carried by a validator-defined schema position, a rare compound, or a clause-bound predicate whose technical use links to its definition. No unqualified term should carry two load-bearing senses across co-loadable artifacts. Writers should prevent or visibly scope collisions rather than delegate resolution to readers.

Today the invariant is held by writer discipline and manual audits. The candidate mechanical controls — a reserved-term registry, a coinage collision screen in the write path, a naming-review gate, slot-escape linting, and a clausal-binding link check — are collected as a design object in write-time vocabulary collision controls.

Prevention fails sometimes, so the invariant owes a recovery path

Write-time prevention is a control, not a guarantee. Writers scope terms under incomplete information: the colliding sense may not exist yet when a term is chosen, or the fact that would disambiguate may be missing at authoring time. Some misresolutions will therefore ship despite the invariant, and since enforcement without structured recovery is incomplete, the invariant owes two contingencies beyond prevention:

  • Detection — a way to attribute a downstream problem to sense misresolution. This is the hard half: by this note's own argument the failure is silent, so detection cannot wait for an error signal. The proven instrument is the corpus audit — sweep a suspect term's occurrences and classify each against its declared sense. A cheaper per-artifact probe — re-deriving a suspect conclusion with each technical term's intended sense spelled out, to see whether the derivation survives — is plausible but untested.
  • Remediation — once a misresolving term is identified, determine what disambiguating information is missing and put it in place: promote the sense to a stronger device in the ranking, split the senses across distinct terms, or retire the word entirely and rescope its former traffic.

The remediation audit should be classification-led and baseline-anchored. Record each technical occurrence against a stable path-and-heading locator before rewriting it; choose among the replacement category, a narrower prototype, a precise local noun, and deliberate preservation; then reconcile the result with a broad residual search. This ordering matters because a blind replacement cannot distinguish the captured technical sense from ordinary editorial use, named machinery, or historical wording, while a zero-hit goal would destroy those legitimate uses. Independent verification of the occurrence ledger supplies the missing check that a self-reported corpus rewrite cannot provide. The same procedure applies when the successor is another category term, not only when a term is retired without replacement.

The distillation retirement (ADR 053) is a worked instance of the full loop: a 464-occurrence audit detected two load-bearing senses with opposite maintenance regimes sharing one word, and remediation retired the term, moved the boundary into lineage link labels, and routed the ampliative traffic to a coined compound. The definition note that was supposed to hold the binding had existed, impotently, throughout the drift.

Open Questions

  • Do definition notes need a health warning as a type? They can bind concepts to explanations safely (what is a term?) but bind words to senses only weakly.
  • Which of the proposed enforcement mechanisms earns adoption first — and does the deterministic subset (registry uniqueness, slot-escape linting) deliver enough of the invariant to defer the semantic gates?
  • Detection of shipped misresolutions currently rides expensive corpus audits triggered by suspicion. Is there a cheaper standing signal — a review criterion, a co-load-time check — that flags sense misresolution before a full audit is warranted?

Relevant Notes: