Quote satellite · Evidence type · One-way authority

Give Quotes Access Without Canonical Authority

Selected source-exact quotes are more valuable than chunks and more dangerous than claims. File them in a typed satellite: immutable text, advisory interpretation, reverse lookup without graph citizenship, and promotion of normalised meaning — never prose-as-truth.

By Scott Farrell · LeverageAI · Article-only field note

Here is a design problem that shows up the moment your publishing system and your knowledge graph start talking to each other. You already have strong excerpts. They are source-exact, structurally bounded, and often more useful as access points than a whole chapter summary. The temptation is immediate: promote the best quotes into the main wiki as claims. The danger is just as immediate: you have just put rhetoric into the navigational map.

That is not a marketing problem. It is an epistemic typing problem. If you isolate the excerpts completely, they become unusable from the concept page you are standing on mid-proposal. If you grant them ordinary graph citizenship, you pollute the map with lexical style, force the janitor to maintain verbatim text, and create a path where popularity or a careless merge can look like truth.

Better quotes may be useful claim-grain access points — but putting prose in the canon is dangerous. So give them their own store, with one-way edges into the IP map.

This article delivers that store as a concrete architecture: schema shape, immutable versus derived fields, reverse index, promotion rules, invalidation, and a worked demonstration that the main wiki can retrieve related quotes without treating those quotes as support for its own truth claims.

What nearby pieces already own

This is a standalone design note, not an extender of a parent ebook. A few siblings already carry load-bearing pieces of the surrounding system; I will cite them rather than re-derive them.

• Attention-Native Publishing owns the quote-card as an attention event: the interrupt, the silence, and the long-form as proof surface.¹
• Semantic Refraction owns relational grain — why meaning-complete pieces become better interfaces than document-sized blobs.²
• Semantic Decompilation owns the method: deterministic structure supplies handles; AI supplies judgment without rewriting source.³
• The Prompt Is Source owns stage-relative source in the compiler cascade — useful orientation for why intermediate artefacts keep different authority rules.¹⁶
• Why Richer RAG Metadata Still Cannot Hold Relational Meaning owns the arity argument: relational meaning is not a unary annotation you paint onto one record.⁴

What this piece owns is narrower: the epistemic type of a selected quote, and the traversal-versus-authority rule that lets it stay discoverable without becoming a citizen of the canon.

Two objects that look interchangeable and are not

A claim is a normalised proposition:

A quote is an exhibit:

Evidence, in the ordinary sense, is material that supports or tends to support a proposition; it is not the same object as the proposition itself.⁵ A citation points a reader to a source so the claim can be checked; the citation is not itself the claim.⁶

So when a quote-card pipeline selects a line because it creates useful tension in a feed, that selection is a judgment about human attention — not a judgment that the sentence is the canonical navigational form of the idea.¹ Compress for lean-in and compress for navigation are different targets. Mixing the types has one genuinely dangerous failure mode, which we will run as a case later: the claim-oriented janitor.

Why “just put good quotes in the wiki” fails

Graph citizenship is not storage. In a knowledge graph, the distinction that matters is which nodes participate in navigation — appear on the map, cost the walker attention, and demand janitor maintenance. Recurring meaning earns residency; merely existing in a source does not. That citizenship discipline is developed at length in the BI for Soft Data work; the short version here is enough: every citizen is a maintenance liability.

If every high-scoring excerpt becomes a claim, you risk:

• duplicating ebook prose throughout the graph;
• making lexical style look like semantic structure;
• widening the map with material that adds no new navigation;
• forcing the janitor to maintain verbatim text under rules designed for propositions;
• allowing the same thought to appear as several superficially different “claims.”

If, on the other hand, you keep selection only as rendered cards and never file the intermediate record, you throw away a silver layer you already paid for. A publishing pass that already did deterministic disassembly, isolated judgment, combination, and scoring has produced addressable semantic units. Filing that output is cheaper than pretending semantic decompilation must start from zero later.³

The missing move is a quote satellite: a typed store that behaves like a wiki when explicitly queried, but does not receive ordinary IP-graph citizenship.

The satellite architecture

Think in layers of authority, not in product brands:

Bronze and structural silver already have homes in the wider stack — keep raw territory so the map can be rebuilt; keep stable IDs so selection is reproducible.⁷⁸ What this article specifies is the satellite and its contract with gold.

Schema shape (composite quote record)

Here is a concrete record shape. Field names can change; the separations cannot.

Immutable versus derived fields

Data integrity is the property of remaining accurate and consistent over a lifecycle; immutability and controlled mutation policies are standard ways to protect it.⁹ Write-once, read-many storage is the extreme form of that discipline: once written, the payload is not modified in place.¹⁰ The quote satellite applies WORM thinking to text and scopes even if the rest of the record is mutable.

Read/write contract

Access without citizenship: the reverse index

One-way outbound edges protect the IP wiki:

That keeps the primary map clean. But real work runs the other direction: you are on a concept page mid-newsjack or mid-proposal and you want the best source exhibit. One-way graphical edges alone would make that query impossible without a full scan.

So you add a deterministic secondary index, not a second graph of equal authority:

A database index is a secondary structure that speeds lookup without changing the semantics of the primary records.¹¹ A materialized view is a stored query result that can be refreshed when underlying data changes; it is not the system of record.¹² Treat quotes_by_target the same way: rebuildable access structure, not a citizen edge list.

That is access without citizenship.

Promotion: use earns meaning, not residency of prose

The satellite is an excellent candidate generator. It should not decide canon membership at ingestion.

Promotion triggers (illustrative, not a scoring formula):

• selected in several independent publishing runs;
• surfaced in unrelated wiki investigations via get_excerpts_for;
• associated with several existing IP concepts;
• repeatedly used to explain the same mechanism;
• exposes a relationship absent from the main graph;
• becomes the basis of a new article or framework.

When promotion fires, promote the normalised meaning, not the excerpt and not its heat:

This matches the spirit of derived pages that remain subordinate to their supports and are first to be invalidated when evidence changes.¹³ It also matches the evidence-package discipline: a claim may travel with a verbatim exhibit and a resolvable pointer, but the exhibit does not become the claim.¹⁴

Popularity is not a promotion input for truth. Engagement, if you later attach it as telemetry, is a signal about human lean-in — useful for publishing experiment design (the territory of a later semantic-experiment-graph design), useless as janitor authority. Resonance is not canonicity.

Invalidation

Invalidation must follow the type boundaries:

If a gold claim was promoted from a quote that later goes stale, the claim does not automatically die — but its evidence package must show a broken or superseded exhibit, and a lint pass should surface the gap. Preserve disagreement and broken links as visible state; do not quietly consolidate them into a smoother story.¹⁵

Failure case: the claim janitor meets verbatim text

Consider an illustrative pair of source-exact quotes filed in the satellite (wording simplified for the page; treat as illustrative, not a historical incident):

Both are useful. Both relate to attention-native publishing. A claim-oriented janitor with a loose “merge similar ideas” objective sees adjacency and age, and emits:

That failure is the specialised form of hallucinated consolidation: fine for some claim cleanup, catastrophic for verbatim exhibits.¹⁵ The structural fix is not a longer prompt begging the janitor to be careful. It is type-aware rules:

• Janitors that merge, fade, or paraphrase operate only on gold propositions.
• Quote records are re-edged or dropped, never rewritten.
• Composite quotes store component scopes; they do not invent a third prose string that was never in bronze.

If you need a single navigational node, promote a normalised claim and attach both quotes as exhibits. The map gets one citizen. The evidence layer keeps two immutable receipts.

Worked demonstration: retrieve without elevating

Illustrative walk-through. You are on concept.tight-intent-loose-method preparing a short response. You need language that carries, not a new synthesis of your own doctrine as if the doctrine proved itself.

The main wiki retrieved related quotes. It did not treat those quotes as support for its own truth claims. That is the proof burden of this design: discoverability without circular self-evidence, and without map pollution.

For human publishing, the same reverse index feeds a quote-card pipeline without forcing every selected line into gold. The card remains an attention artefact; the chapter remains the proof surface; the satellite is the filing system between them.¹

Objection: “Why not one table with a type column?”

Objection. Store quotes and claims in one graph database. Mark type: quote versus type: claim. Skip the “satellite” metaphor.

Answer. The storage engine is not the point. The contract is. You can implement the satellite as a namespace, a separate collection, or a typed subset of one store. What you cannot skip:

• immutable text fields with a hard ban on paraphrase janitors;
• advisory interpretation that never auto-promotes;
• no reciprocal citizen edges for exhibits;
• a reverse index or equivalent query path;
• promotion that creates a new normalised claim rather than elevating prose.

If your “one table” enforces those rules, you have built the satellite under another name. If it only adds a type label while the same janitor merges neighbours, you have built a more convenient path to words nobody wrote.

Objection 2. One-way edges make reverse discovery slow or impossible.

Answer. That confuses graph edges with indexes. Secondary indexes and materialized projections exist precisely so primary records keep clean semantics while access patterns stay fast.¹¹¹²

Objection 3. Engagement proves which quotes should become claims.

Answer. Engagement proves which exhibits lean humans in under a particular channel and moment. It does not prove which normalised propositions should organise your worldview. File heat as advisory telemetry if you wish; never wire it into truth or janitor decisions. Designing that telemetry graph is a later problem — a semantic experiment graph territory — not a reason to collapse types today.

Near-term implementation without boiling the ocean

You do not need a full semantic decompiler to start. If you already run a quote-card or selection pipeline, the short path is:

1. Keep chapter-level summaries as the coarse map (if you have them).
2. Export high-rated selections: exact scope IDs, text, combinations, score, reason.
3. Store them outside the primary IP map (namespace or collection).
4. Optional cheap pass: one possible_claim, one semantic_role, zero to three candidate IP links with reasons.
5. Treat all interpretation and edges as advisory.
6. Build quotes_by_target from validated or candidate links_out.
7. Promote only through use, review, or repeated convergence — and promote meaning, not prose.

That avoids both extremes: one claim per chapter (too coarse) and every paragraph as wiki prose (too fine). Quotes become semantic working drawings — more meaningful than arbitrary chunks, more source-faithful than summaries, not yet canonical graph citizens.

The structural selection substrate — deterministic handles, stable IDs, combinations without rewriting source — is already described in the decompilation and text-as-home-turf work.³⁸ Your job here is to file that output under the right epistemic rules.

What this deliberately does not build

Out of scope, on purpose:

• a full marketing heat system or ranking by engagement as authority;
• making every paragraph a silver record;
• re-teaching bronze/silver/gold warehouse theory beyond the labels needed for orientation;
• re-deriving attention-native card design, relational grain, decompilation method, or metadata arity.

Those pieces either already live in published siblings or belong to later designs (publishing as an active sensor, experiment graphs, orientation capital). This article only has to land the missing type and its authority rule.

The doctrine to hold

Source-exact quotes are more valuable than chunks and more dangerous than claims. Keep them in a typed satellite:

• Immutable text and scopes — re-edge or drop, never rewrite.
• Derived interpretation as advisory — candidates until validated.
• Two-way traversal via reverse index — access without citizenship.
• One-way authority — IP claims do not rest on self-quoting exhibits as proof.
• Promotion of normalised meaning through use — not of prose, not of popularity.

Ingest broadly, decompose structurally, retain excerpts cheaply, and grant graph citizenship only to meaning that proves it can travel.

Or shorter:

Give quotes access. Do not give them canonical authority.