A larger vocabulary, a closer reading: findings from Wikipedia Wave III

Findings & method · Wikipedia Wave III

A larger vocabulary.
A closer reading.

The third Wikipedia enrichment wave adds 1,843 approved Carrier concepts. Its most useful result is a set of distinctions that let a message say more clearly what follows, under which conditions, and what remains open.

The prepared release brings the checked vocabulary from 4,414 to 6,257 entries, an increase of 41.75%. Lucid, Founder of Unified State, reviewed and approved the new compact definitions and contextual links. That approval gives the additions a place in the shared vocabulary; it does not turn a selected source passage into a complete account of its subject.

Consider a simple description: “This ground contains plenty of water.” A practical next question is whether the water can move through the material. Porosity describes the fraction of void space. Hydraulic conductivity concerns transmission under specified conditions. Pore arrangement, connectivity and the fluid still matter. The value of another concept appears at the point where one broad sentence would otherwise invite the wrong inference.

Four strands, one working language

The strands differ in granularity. Their sizes reflect editorial choices and the evidence retained for this round, not an objective ranking of disciplinary importance. Selection favoured an identifiable gap in the earlier vocabulary, a useful role in expression, a compact definition supported by selected evidence, and scope that could be stated honestly.

A new title was not sufficient reason to create an entry. The edition records 61 explicit reuse mappings to earlier meanings. It also leaves 17 title leads unresolved rather than forcing them into a nearby definition. Reuse and deferral reduce the apparent output while making the actual vocabulary less misleading.

The source corpus and the reading are different things

This was a targeted thematic acquisition, not a comprehensive review of Wikipedia. The source ledger contains 1,962 requested-title records. It retained 1,907 records representing 1,862 distinct article revisions; aliases can lead to the same retained revision. Fifty-five requested-title records were not retained as sources.

The normalized corpus contains 107,063 indexed passage occurrences, or 103,890 distinct indexed units. Those are acquisition and indexing counts. They do not mean that every paragraph was semantically read or checked.

The authored additions cite 4,756 distinct selected passages from 1,598 article revisions. This narrower evidence set is the relevant measure of the material used to support the definitions. Selected passages were read with relevant context, and each proposed meaning was given a source pointer, scope and reason for inclusion. The disposition ledger distinguishes used evidence from unused discovery material.

Raw wikitext, captured HTML, acquisition records and normalized spans were retained. Mechanical verification reproduced the normalized extraction for all 1,862 distinct captures and checked the pinned bytes. Tables, navigation and much end matter are excluded from the reading extract, and displayed formulae can be absent or imperfectly represented there. The raw forms are important when the extract is insufficient.

A revision link pins an article’s wikitext. It does not independently reconstruct every historical template used in a rendered page. The captured rendering is therefore recorded separately. A hash establishes which bytes were kept; it does not establish that every statement in those bytes is true.

Careful selection sometimes means importing less

Several of the useful findings were decisions about scope. A phase-change discussion compared latent heat with specific heat without supplying a temperature interval; the definition retained the storage mechanism without importing that broad comparison. A power-factor definition kept both displacement and distortion in view rather than reducing the concept to phase difference alone.

In the life strand, the retained material on RNA splicing and exons required care about protein coding. The resulting distinction concerns which sequence is removed or retained without treating every exon as protein-coding sequence. Other source caveats led to narrower mechanisms, qualified wording or deferred entries.

These are recorded editorial assessments of selected passages. They are not claims to have audited whole articles or settled specialist debates. Wikipedia supplied a public, attributed starting corpus. A primary-literature review and domain-specific validation would be further work.

Connections need boundaries of their own

The edition contains 8,279 authored directed contextual links. Of these, 5,120 point into the earlier checked vocabulary, reaching 814 distinct existing neighbours. There are also 2,354 unique pairs among the new concepts. After pair storage is de-duplicated, the edition adds 7,474 stored contextual pairs. Directed authored links and stored pairs are different counting units; neither number should be substituted for the other.

The most informative connections are often those that carry a limit:

• Soil structure → porosity. Arrangement includes more than the fraction of void space; connectivity, aggregate form and biological access still need attention.
• Landscape connectivity → percolation threshold. A network model offers one analytic lens. Organisms, behaviour and habitat quality determine whether a route is usable in an actual landscape.
• Software circuit breaker → electrical circuit breaker. The protective interruption analogy is useful, while dependency requests and electrical interruption remain different mechanisms with different failure criteria.
• Object provenance → software supply chain. Both invite inspection of an artifact’s history. A software dependency record does not acquire the legal ownership or authentication conclusions of an object-provenance record.

Eight explicit cross-strand connections make such qualifications especially visible. A graph can bring related questions into reach; the connection itself supplies no identity, entailment, endorsement or causal claim. Likewise, a connected component or a high degree describes this selection of links. It cannot establish completeness or rank the truth of concepts.

From definitions to expressions

The use lab contains 32 authored before/after pairs, eight for each strand. They move through particular problems: clustered samples and stopping rules; water transport and genetic variation; care, support and group reference; heat pumps, optical measurement and orbital descriptions. Each case states the intended distinctions, the conditions of the message and readings it rejects.

One human-language case begins with a listener’s brief “yes.” The expanded version distinguishes conversational acknowledgment from consent and asks who is included when a facilitator says “we.” It also preserves the possibility that an acknowledgment can convey agreement in some context. The vocabulary is useful because it helps ask the missing question without deciding the answer in advance.

A household-care example separates direct and indirect care, time poverty, emotion work and offered instrumental support. It resists treating a difficult workload as merely a failure of resilience. The example remains hypothetical: an offer of support does not establish that useful help was received, and a person’s report is not a clinical assessment.

A software example separates a reproducible build, the assumptions behind a checked proof, and permission to execute the result. The conditions that establish consistency with declared inputs do not automatically grant authorization. This is the same discipline at another scale: preserve each claim’s actual reach.

Existing concepts plus literal explanation could express many of these ideas before the expansion. The benefit proposed here is that the components gain reusable, inspectable references. Some later messages deliberately add detail. The pairs are therefore authored comparisons, not automatically equivalent paraphrases or measurements of compression.

What the local tests establish

All 64 original messages were encoded and recovered through the existing USL1 codec. The checks covered 228 actual concept-token occurrences, including 142 references to the working edition. Packet integrity, tier separation, snapshot resolution and actual token spans were examined. The 69-check use-lab suite passed.

A preserved development note records an earlier error in the last negative-test fixture. It duplicated a snapshot tier that the chosen packet did not use, so the expected ambiguity was never exercised. The fixture was corrected to duplicate a tier actually referenced by that packet. The expected rejection stayed in place; no codec, definition, source, message or normative artifact was changed to make the test pass.

For publication, the authored message text is preserved while concept references are repinned to the approved checked IDs and release dictionary head. All 64 public renditions also recover exactly, and all 228 token spans match their bindings. Of these concept occurrences, 142 select newly approved concepts; none use the working tier.

These are explicit public renditions with changed metadata and packet bytes. The original working-reference packets remain in the evidence package. Re-encoding is a publication step, not a second independent interpretation experiment.

The codec was not independently reimplemented for this work. Exact local recovery shows that the supplied representation and references can survive that encoding path. It does not show that a different reader reconstructs the intended meaning, that the later message is semantically superior, or that every new definition has been tested in use.

Approval keeps the history visible

The release distinguishes the original proposal edition, Lucid’s approval and the resulting checked snapshot. Earlier checked history is preserved. The inherited LEGACY_ID_HISTORY warning about historical reuse of working IDs w3–w9 remains visible; it neither authorizes new reuse nor calls for rewriting the earlier record.

Checked status records an editorial decision about the displayed definitions and contextual links. It is not specialist certification, a claim of consensus, or a guarantee that a definition is sufficient for every later application. A stable shared reference is valuable precisely because a correction or disagreement can point to something identifiable.

The next useful result may be a correction

After an expansion of this size, use deserves sustained attention. Another thousand entries would bring another thousand discovery and maintenance obligations. A difficult original message may reveal more about the current language than another round of acquisition.

A useful next exercise is to give a reader a composed message before showing the author’s explanation. Ask the reader to identify who is acting, what is asserted or proposed, which conditions apply, and what the message leaves unresolved. Preserve the answer, including disagreement. Compare it with the intended reading without silently repairing either version.

Record the literal qualifications that remain necessary, the concepts chosen incorrectly, the missing distinctions and the corrections. Successful recovery matters; so do recoverable disagreements and visible limits. Those observations can guide later enrichment with evidence from actual expression.

A shared language earns its next addition through the work it helps people do—and through the misunderstandings it makes possible to repair.

Evidence and attribution

The downloads separate the release records, measured analytics, evidence package, full retained source archive and public expression lab. Per-entry records provide selected passages, revision links and contributor histories. The Wikipedia-derived additions and adaptations are shared under CC BY-SA 4.0; consult the attribution and modification notice. Earlier Carrier material, software and separate lexical dependencies retain their prior terms.

Release JSON Measured analytics Release manifest Retained source archive Evidence package Public expression lab