You know the move. Someone produces a pillar — a long, well-positioned piece that was expensive in attention, brand risk, and craft. A YouTube essay. A field-guide ebook. A board-ready argument that took months to earn. The strategy that follows is almost automatic: cut it up. Text posts. Image cards. Blog extracts. Clips. You tell yourself you are maximising return on the investment. Quietly, you also tell yourself a second story: the cuts are descendants. The pillar is the parent. Everything else is lesser.

That second story is the lineage fallacy. It feels respectful — it protects the dignity of the hard work — and it is the wrong metaphor for what the cuts can do.

What changes the picture is not a marketing tip. It is a grain problem. When you ask how a fifty-idea pillar relates to everything you already know, you are forced toward a general answer. When you pull the interesting, meaning-complete pieces into separate units, each piece can be judged on a different axis: what it contradicts, what it extends, who it is for, why it matters now. The pieces do not invent new source material. They acquire addresses the undifferentiated whole was too coarse to hold.

I treated the cuts as lesser descendants of the pillar. That was the wrong metaphor.

Hierarchy was a comforting map

The old map looks like this:

It matches how budgets are approved and how brand teams talk. The pillar carries narrative, evidence, and qualification. The derivatives carry reach. In that framing, anything short is a compromise — a promotional scrap of the real thing.

The emerging map is a graph, not a family tree:

Those artefacts are not automatically smaller versions of the original. They can be new acts of cognition performed over the original — if the unit is closed enough to stand as a claim, and honest enough to keep its receipt.

Worked miniature: same pillar, two cuts

What this book is actually claiming

The important proposition is not a slogan that fragments are always better than wholes. That claim is false and usually sold by people who never built a proof surface. The claim is causal and narrower:

The set of individually resolved pieces can express more usable meaning than the undifferentiated whole, because each piece can form relationships the whole was too coarse to hold.

Usable meaning, here, is not vibes. It is the precision and number of legitimate joins you can form: contradict, extend, evidence, apply, re-project, compile for a moment. Averages destroy those edges. Addresses create them.

Industry language often calls the cutting step “content atomization” and optimises for channel coverage.¹ Coverage can be logistics. It is not the same doctrine. This book is not about getting more assets from one post. It is about semantic addressability — why closed units increase the precision of relationships a canon can form.

The objection that arrives immediately

“So you want us to shred long-form and live on quotes?”

No. The whole remains strongest as territory and proof: narrative coherence, evidence, qualification, the place a sceptical reader goes to falsify you. The pieces are strongest as interfaces and joins. The valuable system is parent source plus addressable fragments plus provenance plus corpus relationships — never floating slogans with good typography.

“Isn’t this MetaWriter with different branding?”

MetaWriter already showed that multi-idea conversations forced into one container weaken every idea, and that completeness can belong to a set of articles rather than one fat artefact. This book continues the grain ladder downward and names the relational mechanism: why claim-sized closure increases legitimate joins. Conversation Is the REPL already argued that a document is the wrong grain for a dialogue turn and that a claim with edges is turn-sized. We will use that bridge, not re-derive it.

What you will be able to do

By the end of this book you should be able to:

• name relational resolution and relational grain without sliding back into lineage language;
• distinguish summary from extraction, interpretation, relational enrichment, and media compilation;
• refuse fragmentation dressed as strategy using a closure-and-provenance rubric;
• walk a pillar into at least ten closed units and compare broad document relations with precise claim relations;
• re-project one claim across honest organisational lenses without rewriting the bronze fact;
• operate a pillar-to-constellation loop that writes new edges back into the canon.

Chapter 2 stops gesturing at the mechanism and defines it. Once you can say why averages destroy edges, the rest of the architecture stops sounding like taste and starts sounding like design.

Ask a serious question of a serious ebook: how does this relate to what we already know? Watch what your answer does. If the book develops several frameworks, tells multiple stories, and carries a stack of qualifications, the honest answer becomes a blur. It develops these three ideas. It resembles those two prior works. Its overall significance is roughly X. You are not stupid. You are averaging.

That average is not a failure of intelligence. It is a property of grain. The document is too coarse to host the joins you actually need. This chapter names the property so the rest of the book can stop arguing from metaphor.

Relational resolution

Relational resolution is the increase in precision with which a meaning-complete fragment can be positioned, interpreted, and joined against a wider corpus.

Three verbs matter. Positioned — where it sits relative to existing claims. Interpreted — what significance it carries under a worldview. Joined — which typed relationships it can form: contradicts, extends, evidences, applies, rephrases, implements. Resolution rises when those operations become specific. It falls when they collapse into “related to AI” or “about marketing.”

A fifty-idea ebook related as one node might support a handful of broad relationships. The same ebook represented as fifty coherent claims can expose many legitimate relationships across prior frameworks, external evidence, client problems, audiences, current events, and contradictory claims — not because the system invented material, but because each idea finally has an address.

Relational resolution is bounded by relational grain.

What “fifty claims” actually unlocks

When the whole is one object, you get answers like:

• it develops three frameworks;
• it resembles two books;
• its overall significance is X.

When it becomes fifty meaning-complete pieces, each piece can be independently resolved:

• this claim contradicts an earlier assumption;
• this one extends a known recognition loop;
• this one is evidence for an attention-sovereignty argument;
• this one connects prompting attention with human attention;
• this one is the ideal answer to a current news event;
• this one belongs in a client proposal;
• this one is interesting to a marketer but not an engineer.

The long document may hold all fifty meanings latently. Latency is not addressability. Size prevents those meanings from being individually targeted. Decomposition does not merely reduce the object. It increases semantic addressability.

Relational grain as a third grain

You already need more than one grain concept if you work with knowledge systems. Storage grain is how files and folders happen to be saved. Semantic closure is the smallest bundle whose meaning closes — the right unit to hand a model without inventing missing premises. Relational grain is different again: it is the unit size that maximises the precision of joins.

Those three do not automatically coincide. A closed claim is often the right handoff to a model. A chapter may be the right storage package. A whole ebook may be the right narrative experience for a human reader who wants proof. The failure mode is forcing one grain to do every job. The document is comfortable for storytelling. It is a poor interface for edge formation.

Worked case: one question, two grains

Why this matches how serious knowledge systems already think

You do not need to rebuild computer science to see the rhyme. Knowledge graphs treat entities and relationships as first-class rather than only storing documents as blobs.² RDF models information as statements with a subject, predicate, and object — the atomic unit is already a relation-bearing claim, not a chapter file.³ Linked data practices emphasise stable identifiers so items can be addressed across contexts.⁴

Those systems are not your publishing stack. They are an existence proof that edge-centric grain is how people build joinable knowledge when they are serious. If your canon only ever exposes pillars, you will keep asking models and humans to invent the missing addresses at query time — expensive, inconsistent, and hard to write back.

Objection: “Isn’t this just smaller files?”

No. A random 200-word clip is small and still relationally dead if it does not close. A function-sized claim with a stable identity and typed edges is small and alive. Conversation Is the REPL already put the dialogue version of this on the table: a SharePoint document is too big to be a turn; a claim with edges is turn-sized. Relational grain is that insight applied to corpuses and publishing systems, not only chat turns.

“Won’t claim-level everything explode into unmaintainable spaghetti?”

Uncontrolled atomisation can. That is why the next chapters insist on closure, provenance, and combination permission. The answer to spaghetti is not to return to undifferentiated pillars. It is to refuse weak units and weak edges with the same seriousness you refuse weak code interfaces. Complexity from bad grain is worse: you pay the cost of a large object and still cannot form a precise join.

Averages destroy edges.

Hold that sentence. Chapter 3 shows how people smuggle five different operations under the single word “repurposing,” and why only some of them increase relational resolution.

Once you accept that relational resolution depends on grain, a second confusion shows up immediately. Teams say they are “repurposing the pillar.” That phrase is doing too much work. Under it sit at least five different operations, only one of which is necessarily lossy, and only some of which increase usable meaning.

If you cannot tell them apart, you will measure success as “we shipped more surfaces” while your graph of understanding stays as coarse as the day the pillar was published.

The five operations

A summary is usually lossy by design. It trades addressability for brevity. That can be the right trade for a table of contents or an abstract. It is the wrong trade when what you needed was a precise edge.

Semantic refraction lives primarily in the middle three operations, then uses the fifth as output. Isolate a claim. Understand its role. Compare it with the larger canon. Find the consequential angle. Compile that into an artefact sized for the moment. The original passage supplied the claim. The new artefact may add:

The additional value comes from the join, not from rewriting the sentence.

Coverage logistics are not relational work

Content marketing has a mature vocabulary for breaking pillars into channel-sized pieces — often under names like content atomization — and the usual success metric is distribution coverage.¹ That practice can be competent logistics. It answers: where else can this investment appear?

Relational enrichment answers a different question: what becomes true, false, stronger, or newly actionable when this closed unit meets the rest of what we know? You can ship twelve channel variants and perform zero enrichment. You can perform one enrichment and ship a single quiet briefing that changes a decision. Confusing the two is how teams get busy without getting sharper.

Chunking helps — it is not the whole thesis

Cognitive and UX research on chunking shows why undifferentiated walls of text fail people: a chunk is an organisational unit in memory, and presenting content in meaningful, distinct units helps scanning, comprehension, and recall.⁵ That is necessary background, not a complete doctrine.

A visually chunked page can still be relationally dead if the units do not close, do not keep provenance, and never get joined. Progressive disclosure sequences complexity along one path for one user;⁶ refraction creates multiple independently addressable joins from one territory. Do not collapse those ideas. Layout is not grain.

Worked path: one passage, two pipelines

Notice what did not happen. Nobody needed to invent a spicier sentence. The upgrade was cognitive work over a faithful unit: significance and relation, not paraphrase inflation.

Objection: “Isn’t enrichment just spin?”

It becomes spin the moment significance detaches from a pointer. Cache the Significance already draws the hard line: brilliance with a citation is archive; brilliance without one is marketing. Keep the claim and the significance claim in different layers. The claim should still be checkable against the source. The significance layer should say what changed relative to the corpus and remain revisable when the corpus changes.

Interestingness-as-diff is the same discipline in another dialect: interesting is a relation between an item and what you already know, not a perfume you spray on text. When you “diff each unit” in later chapters, you are running that discipline at claim grain.

Mark this chapter’s table as definitive placement. Later chapters will say “enrichment, not summary” without rebuilding the taxonomy. Chapter 4 installs the guardrails that stop the middle operations from collapsing into fragmentation and floating slogans.

Once people hear that closed pieces can hold more usable meaning than an undifferentiated pillar, a predictable mistake appears. They cut harder. Shorter clips. More cards. Thinner slices. They call it atomisation and wait for the graph to get smarter.

It does not. Arbitrary clips are fragments. Meaning-complete units interpreted against a worldview are refracted meanings. The metaphor is a prism, not a shredder: a rich source passes through different contextual lenses and separates into several distinct, valid directions of meaning — rather than merely being destroyed into scrap.

Fragmentation breaks meaning. Refraction reveals its different directions.

Whole and pieces optimise different jobs

Before the guardrails, fix the scoreboard. The whole is not competing with the pieces for a single trophy called “better.” Each dominates on different dimensions:

The whole is strongest as territory and proof. The pieces are strongest as interfaces and joins. The valuable system is never fragments without a parent. It is parent source, addressable fragments, provenance, corpus relationships, and context-specific interpretations together.

Five guardrails that separate refraction from scrap

1. Semantic closure

The unit’s meaning closes. A reader or model can use it without inventing the missing premise that made the sentence true in context. Open fragments force the audience to smuggle context — or to hallucinate it.

2. Tractability

Small enough for one cognitive move; rich enough to connect. A whole ebook is rarely one move. A pronoun without its antecedent is not a move either. Claims work when they are inspectable and joinable in a single turn of thought.

3. Provenance

Every significance claim keeps a pointer to exact source territory. Cache the Significance already stated the honesty rule: brilliance with a citation is archive; brilliance without one is marketing. The moment quotes float free of the pillar, you have built a slogan farm with good typography. The pillar’s job is re-specified, not demoted: it is the receipt that keeps confident atoms falsifiable.

4. Combination permission

Pure atomisation can destroy meaning. Setup plus punchline, heading plus paragraph, two mutually reinforcing boxes, an escalation across bullets — sometimes the closed unit is a small combination, not a single sentence. Deterministic boundaries are candidates, not a religion. Refraction allows recombination when closure requires it.

5. No audience-only duplication

A new surface without a new closed claim and without a new honest join is coverage, not refraction. Changing the greeting line for “CFOs” while leaving the same open-ended blurb does not multiply meaning. Re-projection (Chapter 6) is different: same claim, lens-qualified implications, still one bronze truth.

Worked gallery: three failures, one valid unit

Why software and data people already know this feeling

Microservices rhetoric promised that smaller services would always be freer and faster. In practice, smallness helps only when boundaries are well chosen; bad boundaries create distributed mud instead of a clean monolith.⁷ Separation of concerns exists for the same reason: isolate reasons-for-change so components can be reasoned about and recomposed.⁸

Database normalization is the data rhyme. The point is not aesthetic neatness. Right grain lets facts attach cleanly and reduces update anomalies when the world changes.⁹ Coarse rows that mix concerns force average answers the same way coarse pillars do. Zettelkasten-style personal knowledge methods made a parallel bet: atomic notes plus explicit links compound better than undifferentiated long dumps — when the notes actually close and link.¹⁰

These analogies are not identity claims. Your publishing system is not a microservice mesh. They are guardrail rhymes: smallness without good boundaries is not a win.

Part I ends here with a vocabulary and a refusal list. Part II stops arguing in the abstract. Chapter 5 takes one pillar, cuts it into at least ten closed units, and shows the before-and-after of broad document relations versus precise claim relations — the minimum proof burden this doctrine has to meet in public.

Part I gave you names. This chapter spends them. We take one realistic pillar — not a toy sentence — treat it first as a single node in a corpus, then decompose it into at least ten meaning-complete units, and compare what joins become possible. No invented engagement percentages. Only structural before-and-after.

The pillar as one node

Call the work When Drafting Is Free — a mid-length field guide for operators who already publish frameworks, run editorial calendars, and feel the quality signal slipping as generative volume rises. Its spine, stated honestly:

• drafting cost collapsed; human attention did not;
• shipping volume stopped being a quality proxy;
• refusal and receipts became the scarce craft;
• feeds behave like dialogues, so dialogue-sized moves matter;
• long-form remains the falsifiable territory behind short public units.

Ingested as one document node, its relations to a corporate or personal canon tend to look like this:

Those links are not false. They are averages. Almost none of them can be contradicted cleanly, evidenced by a single receipt, or applied as a decision rule without re-reading the whole book. Relational resolution is low because relational grain is wrong.

Ten closed units from the same territory

Now decompose. Each line below is written as a meaning-complete claim a model or colleague could hand to a turn of thought. Combinations are allowed where setup and consequence must travel together. Provenance would point to chapter spans in the real system; here we show the units and the kinds of edges they unlock.

That list is illustrative, not sacred. A real pipeline would mint stable IDs, retain source scopes, and refuse units that fail closure. The point is the shape: ten independently judgeable joins where one document node offered a handful of broad tags.

After — claim-level relations (shape)

Decomposition did not add words. It added addresses.

Edge density without fairy numbers

It is tempting to invent a percentage: “edge density rose 400%.” We will not. What we can say as a shape-statement is enough for operators:

• document grain supports a small set of broad, hard-to-falsify relations;
• claim grain supports many specific, typed relations;
• useful density rises when units are closed and provenance-tethered;
• weak density rises when units are fragments or floating slogans — which is why Chapter 4’s guardrails are load-bearing, not etiquette.

Knowledge-graph practice already treats relationships as first-class citizens rather than afterthoughts on documents.² Your publishing and wiki systems should not pretend otherwise while still filing everything as pillars-only.

Objection: “You cherry-picked the ten best lines”

Good. Cherry-picking is half of editorial work. The other half is refusal. A refraction pipeline that cannot say no will mint weak units until the graph is a junk drawer. The discipline is:

1. prefer units that close;
2. prefer units that can earn at least one honest typed edge;
3. retain combination when a single sentence fails closure;
4. leave the rest in the pillar as territory — silent, available, not forced into public life.

Attention-native publishing already argues that most material should not interrupt anyone. Semantic refraction agrees from the relational side: most sentences are not addressable joins. Ten strong units from a pillar beat fifty scraps.

Chapter 6 keeps the same bronze honesty and multiplies renderings without multiplying truths: one event, several legitimate lenses.

Chapter 5 multiplied addresses along the spine of a single argument. This chapter multiplies something else: renderings per address. A whole document roughly supports one fair summary. A closed unit is cheap to re-project. The underlying claim stays fixed. The implication field changes with role, moment, and intent.

That is not the audience-only clone Chapter 4 rejected. Cloning changes the greeting and leaves the same open mush. Re-projection keeps one bronze truth and attaches different, owned consequences.

Canonical event: Project Orion

As a single document — a post-mortem PDF in a shared drive — the organisational question “what does Orion mean?” again collapses to an average: “it was a messy cancellation; be careful with productisation.” Useful edges die in the blur. As closed, lens-qualified implications, the same bronze supports several independently judgeable joins:

One organisational memory, many disclosed lenses — not rival realities.

What each implication must carry

Re-projection stays honest only when the implication is structured like a claim, not a vibe:

• lens — sales, delivery, people, finance, risk, …
• owner — who stands behind this reading;
• date / as-at — when it was true enough to assert;
• source pointer — back to the bronze event or passage;
• confidence or status — candidate, accepted, superseded;
• conditions — under what circumstances it applies.

Without that metadata, “what this means to sales” becomes tribal storytelling. With it, the implication is a first-class unit you can contradict later when staffing evidence changes. The bronze event does not fork into four facts. The implication graph fans out.

Why the whole document resists multi-lens use

Ask a post-mortem PDF to be sales-ready, delivery-ready, people-ready, and finance-ready in one pass and you get a compromise document that serves none of them at relational precision. That is the same averaging problem as Chapter 2, now across organisational dialects instead of corpus joins.

A closed claim — or a closed implication unit — can be re-lensed at near-zero marginal cost once the bronze is fixed. You are not rewriting the territory. You are compiling significance for a role. That is Chapter 3’s enrichment and media compilation sequence applied inside the enterprise, not only on a social feed.

Temporary maps for a particular intent

The same pattern scales beyond standing departmental lenses. Knowledge work often needs a temporary map: this strategy review, this proposal critique, this week’s risk committee. You still do not want rival bronze. You want an intent-conditioned projection over shared material — a task-shaped view that selects, joins, and interprets without rewriting the estate.

That runtime object is a sibling argument in its own right — call it an intent-conditioned task world when you meet it later in the series. Here we only need the publishing and knowledge-grain consequence: shared bronze material can be interpreted for sales, production, or a temporary task without changing the underlying territory. Separate physical wikis or separate claim fields are governance choices. The cognitive design is one fabric, many disclosed readings.

Objection: “Won’t lenses invent convenient facts?”

They will if you let implications overwrite bronze. The fix is architectural, not motivational:

1. bronze events and source passages are append-only territory;
2. implications are claims with lens and pointer;
3. conflicts between lenses are edges to resolve, not a licence to edit history;
4. when finance and people disagree about Orion, you do not merge their paragraphs into a blurrier PDF — you keep both implications and examine the disagreement.

Personal knowledge methods that bet on atomic notes already assume multiple links out of one note; organisational refraction assumes multiple legitimate consequence fields out of one event.¹⁰ RDF-style statement structure is again a rhyme: the relation is part of the meaning unit, not an afterthought.³

Chapter 7 turns the guardrails into a tool you can run in review: a rubric that passes closed, provenance-tethered units and fails fragments, floating slogans, and audience-only duplicates.

Guardrails that only live in essays get ignored under deadline pressure. This chapter turns Chapter 4 into a review tool. Use it when a human or an agent proposes a unit for the constellation: a claim, a quote-scope, a lens implication, a social card candidate. The question is not “is it catchy?” The question is “is it addressable without lying?”

The refraction rubric

Score each candidate pass / weak / fail on five gates. A unit needs no fails to ship into the claim graph. Weak is allowed only with an explicit repair task. Public compilation (cards, posts, briefings) should prefer clean passes.

Pass / fail gallery

How this sits on prior spines

MetaWriter already enforced hard not-in-scope boundaries so each article-sized unit could be fully legible; completeness lived in the set. This rubric continues that discipline below the article. Conversation Is the REPL already defined turn-sized grain as claim-with-edges. Gate 2 is that idea as a review question: can this participate in one move?

Nielsen Norman’s chunking guidance is a weaker external rhyme: chunks must be meaningful units, not an undifferentiated mess of atomic scraps.⁵ We add provenance and relational promise because publishing into a canon is stricter than layout for a single page view.

Objection: “This is bureaucracy”

Bureaucracy is process that does not change outcomes. This rubric changes outcomes in two places people already bleed time:

• Graph rot — weak units force future readers to re-derive meaning and re-litigate edges.
• Trust rot — floating slogans teach audiences that your short form is uncoupled from proof.

Five gates on a card draft is cheaper than a year of polite confusion. If a team cannot afford the rubric, it cannot afford claim-level publishing. Stay at pillars and honest summaries until the review muscle exists.

Completeness belongs to the set — but only if each member of the set could survive alone long enough to join.

Part III leaves the review desk and installs the operating architecture: pillar-to-constellation publishing, then the weekly loop that writes edges back so the canon improves when pieces travel.

By now the mechanism is not the bottleneck. Vocabulary without architecture still collapses under production pressure. People revert to the hierarchy they can draw on a whiteboard: big thing in the middle, little things orbiting as promotion. This chapter draws a different board.

Three names, one contribution

• Relational resolution — the mechanism (Chapter 2).
• Semantic refraction — the metaphor that refuses the shredder (Chapter 4).
• Pillar-to-constellation publishing — the operator architecture for living with both territory and addresses.

In a constellation, brightness is not the same as hierarchy. Some stars are public interrupts. Some are internal claims. Some are evidence nodes. Some are lens implications that never leave the firewall. All of them keep vectors back to source mass.

The architecture in layers

Notice what is missing: a layer called “lesser descendants.” Compilations are stages with jobs, not children with lower status. Attention-native publishing already split the attention event from the proof surface; constellation publishing generalises that split across the whole set of artefacts your canon can emit.

Claims as interfaces

A useful software rhyme: a claim is less like a paragraph and more like a function. It has a bounded purpose, a stable identity, and relationships to other units. You do not reason about a repository only as one giant main(). You reason about interfaces. Conversation Is the REPL made the dialogue version explicit: documents are too large for a turn; claims with edges are turn-sized.

The social surface is catching up. A quote compiled for an interrupt is becoming the feed equivalent of a claim-with-edges: small enough to move, strong enough to join, honest only while the receipt remains. That does not make the ebook optional. It makes the ebook the library the interfaces compile from.

The pillar contains the ideas. Decomposition gives them addresses. The graph gives them meaning.

No privileged level called “the real content”

MetaWriter already broke the fantasy that one container must hold every idea from a multi-idea conversation. The grain ladder continues:

Each level is useful for a different cognitive job. Articles carry public thesis shape. Claims carry join precision. Quotes carry source-exact exhibits. Interpretations carry significance. Attention artefacts spend scarce human attention. Confusing the levels is how you either publish un-checkable slogans or bury every idea in unread monuments.

What the constellation optimises

Industry atomization still optimises the left column when it only counts channels filled.¹ Linked knowledge practices already assume addressable items and relations on the right.⁴ Your brand, if you have a real canon, is closer to the right column whether your org chart knows it or not.

Objection: “Won’t a constellation confuse the brand?”

Confusion comes from inconsistent claims, not from multiple surfaces. A hierarchy of lesser descendants often hides inconsistency behind tone. A constellation with provenance makes inconsistency visible — which is a feature. When two compilations disagree, you can walk edges to the parent and repair the unit. When a slogan floats free, you cannot.

Brand, under this architecture, is the public face of a trustworthy graph: the units you are willing to interrupt people with, the proofs you keep behind them, and the silence you keep when nothing earns an address. That is stricter than a style guide. It is also how serious canons already behave when they are not performing marketing cosplay.

Chapter 9 turns the architecture into a loop you can run: create, decompose, diff, compile significance, translate, retain provenance, write edges back. That is the operating model that makes constellation publishing a practice rather than a poster.

You do not need a twelve-week transformation programme to start. You need a repeatable sequence that respects closure, refuses vanity volume, and improves the canon when pieces travel. This is the operating model behind pillar-to-constellation publishing.

Seven steps

The source produces the pieces, but analysis of the pieces also teaches you what the source contained.

A Monday-sized start

That is enough to feel the difference between coverage theatre and relational work. If the ten candidates mostly fail the rubric, you learned something load-bearing about the pillar’s grain or your editorial taste. If they mostly pass and still cannot name edges, your corpus map is the bottleneck, not your writing.

What success looks like (without fake metrics)

• You can answer “how does this relate?” at claim precision instead of document blur.
• Short public units remain falsifiable in one hop.
• Re-projections carry lens metadata instead of rival facts.
• The canon gains edges after a publishing cycle, not only new files.
• Most candidate units still go dark — silence remains a valid outcome, consistent with attention-native gates.

You will not get a universal percentage for edge density. You will get a practice you can audit: units, pointers, edges, refusals.

What this book deliberately left next door

• Semantic Decompilation — engineering recovery of claims from source structure (HTML handles, AST grain, deterministic disassembly). Named, not taught here.
• Semantic Experiment Graph — measurement and concept-level heat. Named, not taught here.
• Intent-Conditioned Task World — temporary runtime projections for knowledge work. Cameo only in Chapter 6.
• Full interrupt-budget doctrine — already live in Attention-Native Publishing.

Stepping on those toes would make this book longer and this doctrine blurrier. Relational grain is enough to own.

Closing formulations

Hold any one of these when the lineage fallacy tries to return:

• Averages destroy edges.
• Fragmentation breaks meaning. Refraction reveals its different directions.
• The whole is territory and proof. The pieces are interfaces and joins.
• The set of individually resolved pieces can express more usable meaning than the undifferentiated whole, because each piece can form relationships the whole was too coarse to hold.

Or the short stack:

The pillar contains the ideas. Decomposition gives them addresses. The graph gives them meaning.

If you publish frameworks, operate a canon, or maintain a knowledge graph over organisational exhaust, change the question at decomposition time. Not “how do we get more coverage from this pillar?” but “what closed units does this territory contain, and which precise joins does each one unlock?” Then write the answers back so next week’s question starts richer than this week’s.