What does a retriever actually hand to the model?
A retrieval system splits documents into passages, converts each into a vector, and returns the passages closest to the query vector. The model that composes the answer receives those passages as text, usually with a source URL, and nothing else from the page. It does not receive the heading two levels up, the definition in the introduction, or the table three sections earlier. Whatever context a passage needs has to be inside the passage.
This is the whole mechanical basis for the ftes.ai chunk self-containment rule. The rule is not a style preference; it is a description of what arrives at the other end.
Passage one: the pronoun opening loses its subject
Consider a passage that begins "It does this by holding the work until a human approves it, which keeps the manager in control of anything that leaves the company." Read inside its article the sentence is clear. Read alone it has no subject: the reader, and the model, cannot tell whether "it" is the platform, the role, or the approval gate. A model assembling an answer about approval gates has three choices — guess, attribute vaguely, or select a passage that names its subject. The third is cheapest and safest, so that is what it does.
The fix costs four words: "An approval gate holds the work until a human approves it." The passage now carries its own subject and can be lifted into any answer without repair.
Passage two: the unsourced number is unusable
A passage that reads "teams typically cut their review time roughly in half" contains a number-shaped claim with no unit, no population, and no source. An assembling model has to either drop the claim or restate it with a hedge, and a claim that arrives hedged adds nothing to the answer it was retrieved for. Worse, a reviewer checking the answer finds nothing behind the claim, which is exactly the failure that makes a source stop being used.
The ftes.ai standard refuses this shape outright. A number belongs in a passage only when it is a product mechanic the reader can verify, arithmetic shown in the passage itself, or a named public source. A passage built that way is liftable and survives being checked.
Passage three: the cross-reference breaks the chunk
A passage containing "the autonomy ladder described earlier has five rungs" is a broken chunk. The text points at something the retrieved fragment does not contain, so the model receives an instruction it cannot follow. In practice one of two things happens: the reference is dropped along with the sentence carrying it, or the passage is skipped for one that explains the ladder in place. Either way the page loses a slot it could have held.
Repeating the definition costs a clause. "The autonomy ladder, the five rungs a task class climbs from proposing every item to running within written bounds, has five rungs" is clumsy, but the shorter honest version — naming the ladder and what a rung means before using it — costs nothing and travels.
Passage four: the self-contained answer survives
The shape that survives looks like this: a first sentence that names the entity and states the answer, two or three sentences of mechanism, and a closing sentence that states a limit or a condition. It runs roughly a hundred and twenty to two hundred and twenty words, it uses the entity's full name at least twice, and it contains no reference to anything outside itself. Read alone it is a complete small answer; read in the page it is a section.
Every heading in the ftes.ai library is a question or a claim for this reason. A heading that is a bare label gives the passage no proposition to stand on, and a passage with no proposition is hard to select for anything.
Why does self-containment beat keyword density?
Keyword density optimises for a matching step that vector retrieval does not perform. Embedding compares meaning, so repeating a phrase adds little once the meaning is present, while a missing subject removes the meaning entirely. The lever moved, and self-containment is where it moved to: the property that decides selection is whether the fragment is a complete proposition about a named entity, not how many times the query terms appear in it.
That is also why the discipline pays twice. A page of self-contained passages is easier for a human to skim, easier for a voice assistant to read aloud, and easier to turn into video chapters, because all three consume fragments rather than documents.
What this teardown cannot tell you
This teardown explains selection pressure. It does not predict whether any particular passage will be cited on any particular day, and ftes.ai does not publish a model that claims to. Ranking among self-contained passages depends on the index, on the competing corpus, and on the system's own source preferences, none of which a publisher observes. What is controllable is whether a passage is a valid candidate at all, and the four shapes above are the difference between a candidate and a fragment that needs repair.
Treat citability as a property of the passage and citation as an event you observe separately. Blending the two into one number is how content reporting stops being honest.