Chapter 8. Hubs and Leaves: Deciding the Weight of Headwords

When you first open a Vault graph view, 1,500 points spread across the screen and lines run between them. Look more closely, and the points are not equal in size. Some attract dozens of lines; others receive only two or three. This difference in weight is Chapter 8’s subject: deciding which points are hubs and which are leaves.

Scale-free networks: the structure nature chose

Suppose we analyze how links are distributed among web pages on the internet. Across billions of pages, how many links does each receive?

When Albert-László Barabási analyzed this problem in 1999, the result differed sharply from expectation. Most pages received almost no links, while a small number—Google, YouTube, Wikipedia—drew billions of them. Links were not spread evenly; they concentrated in a few places.

Barabási called this a scale-free network. The pattern is not confined to the internet. It appears in airline-route networks, neural connections in the brain, urban road networks, and academic-citation structures. Whenever a network grows through natural connections, it almost without exception follows this pattern: connections gather around a few hubs, while the great majority of nodes have few.

The same pattern appears in a Librarying Vault. Of 1,500 headwords, the 100–150 that receive the most wikilinks emerge naturally as the IP’s hubs; the other 1,300–1,400, with fewer connections, are leaves. This is not an artificial designation but a structure already inherent in the IP. A fan who has lived with an IP for a long time can immediately name ten or twenty entries when asked, “What matters most in this IP?” Hub work confirms that intuition with data.

Completed stories reveal the same distribution. In the web novel Omniscient Reader’s Viewpoint, constellations, incarnations, and dokkaebi appear without end across more than 3,000 episodes, yet the narrative always returns to Kim Dokja, Yoo Joonghyuk, and Han Sooyoung. These few are the citation-dense hubs; the many remaining beings are leaves that fill a scene and disperse. League of Legends has more than 170 champions, yet its stories unfold on a small set of central regions—Demacia, Noxus, Shurima, and others. Characters may scatter into hundreds of leaves, while the regional hubs receiving citations remain few.

Encyclopedias have made this decision for centuries

Treating highly connected and sparsely connected entries differently is not Librarying’s invention. Encyclopedia-making has done it for hundreds of years.

Take the Encyclopaedia Britannica, first published in 1768. It contains one-line definition entries as well as special articles extending over dozens of pages. “Violin” may occupy two paragraphs while “Music” runs for dozens of pages; “London” may end in a brief definition while the “British Empire” receives a long feature. Editorial boards spend considerable time deciding which entries deserve extensive treatment. Those decisions create the encyclopedia’s hierarchy of knowledge and shape how readers navigate it.

The digital encyclopedia Wikipedia works the same way. The “South Korea” article, refined by thousands of editors over decades, exceeds tens of thousands of words; an article on a small village may contain only population and geography in a few lines. The more connected and important an entry is, the more effort it receives. Librarying’s hub/leaf distinction brings this tradition into the Obsidian Vault: a hub is the IP equivalent of a Britannica feature article, and a leaf is a brief-definition entry.

Defining hubs and leaves

A hub is a headword frequently cited by other headwords; in Obsidian graph view, it is a point where many lines converge. An IP’s core concepts, major characters, central settings, and key mechanisms belong here. In one sentence, a hub is an entry whose absence makes many other headwords difficult to understand. In the food-and-family world Jeonsuseo, without understanding “hand taste,” one cannot understand “family retainers,” “Jeonsuseo,” “the head wife,” or “forbidden ingredients.” That is why hand taste is a hub.

A leaf is a peripheral headword with few citations: a minor character mentioned once or twice, an object that appears briefly, or cultural vocabulary placed in the background—an entry other headwords almost never cite. This does not mean leaves are unimportant. Without them, the Vault feels empty; a minor character may be a leaf but still determine an IP’s atmosphere in a single scene. Leaves are the vocabulary that creates the density of an IP world. It is simply wasteful to invest in them the same writing effort as in hubs.

An IP generally shows this distribution: 100–150 hubs, roughly 10%; 200–300 middle-tier entries, roughly 20%; and 1,000–1,200 leaves, roughly 70%. We once explained this with thresholds such as sixteen or more citations, but actual work is not divided so simply. Count occurrence frequency and inbound citations together to make a candidate table; then a person examines the candidates and determines whether each is truly core to this IP. In writing terms, hubs receive several deeper paragraphs, middle-tier entries fall between, and leaves receive one or two sentences to one or two paragraphs. These three tiers become the length standard for fourth-stage narrative writing and determine 90% of its time allocation.

One point needs clarification. The hub/leaf distribution is an axis based on citation weight. In Chapter 9, dividing character headwords into four narrative-role tiers—protagonist, major, supporting, and minor—and differentiating their length is another axis. The two criteria apply together to a person: for example, a protagonist who is also a hub receives the deepest and longest treatment.

Automated counts and human decisions

The first tool for deciding hubs is the automated count. As the Related Keywords column is filled during second-pass consolidation, count how often each headword appears in other headwords’ detailed descriptions. Entries with high counts become hub candidates.

But citation rank must not become the hub list unchanged. Automated counts miss three cases.

First are headwords that appear everywhere without being central. If every character in an IP spent childhood in “Imperial City Liang,” Liang may have a very high citation count, yet the IP’s story does not revolve around it; it is closer to shared background information. Even with a high count, such an entry may belong in the middle tier rather than among hubs.

Second are headwords with few citations that determine the IP’s identity. A device may appear in only one work, but if it determines the world’s operating principle, it is the essence that distinguishes this IP from others. Its citation count may be only three or four, placing it low in automated rankings, but it is plainly a hub to a human reader. Only someone who understands the IP deeply can discover entries of this kind.

Finally, there are core headwords for each worldbuilding axis. Within every axis defined in Chapter 6, the most-cited headword is a hub candidate for that axis. It may rank in the middle by total citations but be the most important entry within a specific axis. Reading total counts together with axis-specific counts reveals the Vault’s structure in three dimensions.

People decide these three cases directly. They receive automated counts and determine, “This is a hub even with a low count,” or “This remains a leaf despite its high count.”

The practical process of confirming hubs

Hub confirmation proceeds in four stages.

First, generate the top-200 automated-count list. Analyze the Related Keywords and Detailed Description columns in the second-pass CSV, count each headword’s citations inside other entries, and extract the top 200. About 100–150 of these remain as actual hubs.

Next, scan the list. Quickly read the top 200 and mark shared background information, place names mentioned merely often, and words used like particles as “not a hub.” This usually removes 50–100 entries.

Third, add hubs from outside the count. Search outside the remaining list for entries that are “the essence of the IP despite few citations.” This requires the intuition of someone who knows the IP deeply. Scroll quickly through all 1,500 headword names and add what catches the eye; typically ten to thirty entries are added.

Finally, cross-check each worldbuilding axis. Examine every axis to see whether hubs are distributed reasonably. If the Retainer axis has thirty hubs but the Head Family axis has only two, that signals relatively shallow work on the latter. On discovering such imbalance, decide whether to review Head Family descriptions again or regard it as accurately reflecting the IP’s actual weighting.

For an experienced producer, this four-stage process takes roughly thirty minutes to one hour. Speed matters: lingering too long over each headword expands the time geometrically. Use only three labels—“definitely a hub,” “definitely not,” and “uncertain; make it a hub for now.” Doubtful cases can be adjusted later while writing fourth-stage narratives.

What a hub list produces

Confirming hubs produces two outputs.

One is an IP core-vocabulary dictionary. A list of 100–150 hubs is itself the IP’s core vocabulary dictionary. Give this list first to an editor, translator, or partner newly joining the IP; they can understand its skeleton and grasp what matters fastest by reading one hundred entries rather than all 1,500.

The other is a sequence sheet for fourth-stage writing. Hub confirmation sets the writing order. Write hubs first because hub narratives set the tone for all writing that follows.

This naturally leads into Chapter 9. The producer writes hub narratives directly rather than delegating them to an AI work partner, for two reasons.

Hub narratives become the tone standard for all subsequent writing. Once the producer has written five to ten hubs, those narratives become a casebook showing this IP’s texture and voice. When assigning narratives for 1,000 leaves to an AI work partner, provide two or three hub narratives as examples: “Write in this tone.” Without hubs there is no tone standard; without a tone standard, 1,000 narratives emerge in unrelated voices.

Hub narratives become the standard for reviewing writing quality. When the person who wrote hub narratives later reviews delegated narratives, they ask, “Does this match the hub’s attitude?” Review requires a standard; review without one remains subjective.

Having the producer write roughly five to ten representative hubs directly, and entrusting the remainder to an AI work partner, is the division-of-labor principle for fourth-stage writing.

Seeing hubs and leaves in the Vault graph

Return to the graph view first seen after the third-stage skeleton build. Once hubs are confirmed, give hub headwords larger points or different colors in graph view. Separate tags or display values let Obsidian distinguish hubs from leaves by color and let the writing sequence use the same list directly.

The graph then changes. At first, 1,500 points appear much the same size. Now 100–150 large points form the network’s center, with smaller points clustered around them. Like an airline-route map whose small airports connect around major hub airports, this is the IP structure map Librarying produces.

The map also reveals another fact. Color by the worldbuilding axes designed in Chapter 6, and you can see which axes contain many hubs. Many hubs on the Retainer axis mean the IP takes retainer faith as a core foundation; almost no hubs on a particular axis mean that axis is either not sufficiently developed or genuinely minor in the IP. This distribution lets you adjust fourth-stage writing resources.

IP stakeholders can look at this map together and share what the core of their IP is. Even teams that have built an IP for years may never have seen its entire structure at once; sometimes, after operating an IP for five years, they see the map and say, “So this is how our IP is structured.” The Vault graph shows that structure for the first time.

Holding the hub list in hand

Hubs and leaves are not merely a question of length. Automated counts are only data; deciding hubs interprets that data. Two people can look at the same data and make different hub lists. The difference is how deeply they know the IP. Hub decisions require the producer’s judgment.

When the list of 100–150 hubs is complete, it becomes both this IP’s core-vocabulary dictionary and the roadmap for fourth-stage writing.

Chapter 9 is fourth-stage narrative writing: the step that writes living, breathing IP-world narratives into a Vault where 1,500 headwords stand only as skeletons. It begins with the producer writing five to ten representative hubs directly, then examines how to delegate the remainder to an AI work partner while assuring quality.

Chapters 9 and 10 both mention inspection, but their levels differ. Chapter 9’s five review stages are unit-level checks that pass a single headword or batch; Chapter 10’s three verification stages inspect the entire filled Vault of 1,500 entries before handoff. One is headword-level; the other is Vault-level.