1. Parts that only count together. An eyebrow is a paragraph only when a headline follows it (ADR-046); an epigraph is a blockquote only with its attribution (SPEC-185 D5). Written as independent optional fields, either part can be read alone.
2. Parts inside a matched block. An attribution written as a blockquote's last line (> — **Ada Lovelace**, Analyst) is inside the node a blockquote field matched, and no field can read inside it.
3. A trailing part after a run. A body of paragraphs followed by an attribution paragraph: a greedy body takes the attribution too.
4. A required part that never came. Nothing reports it.

This spec adds a construct for each, in the vocabulary content models already use, and proves them on the one native rune whose real grammar is hand-written: section.

Why

section declares a content model and does not use it. Its schema declares a sequence (eyebrow?, headline?, blurb?, body*), and its transform's first step is to put those fields back together in source order and re-read them with the hand-written sectionHeaderParts. The comment there says why: the sequence "would read a paragraph with no heading after it as an eyebrow". The grammar compositions are audited against (schemaHeaderRoles) is that same hand-written reader. So the header that every composition over section relies on (hero, cta, testimonial, and card when it is composed) is code, not declaration, because the model cannot say "these together or not at all".

Compositions can only use what is declarative. A native rune can fall back to code; a composition cannot. If the epigraph and the inside-the-quote attribution of SPEC-185 exist only in hand-written readers, compositions can never offer them, and the same Markdown means one thing in a native rune and another in a composed one.

What the resolver does with a required field today, for the record. resolveSequence walks fields in order and children in order:

• A field that matches takes its node (or, greedy, the run of matching nodes).
• An optional field that does not match is skipped; the child waits for the next field.
• A required field that does not match drops the child and reports it as content-unmatched, then resolution continues with the next field. So Prose. followed by **Ada** — Analyst, under a required quote and an optional attribution, reports Prose. and then reads the attribution, with no quote: the report names the wrong problem, and the result is wrong.
• A required field reached after the children have run out is skipped silently.

Decisions

D1 — an optional group matches all of its required members, or nothing

A field may be a group: an ordered set of member fields that matches as one.

epigraph:
  optional: true
  group:
    quote:       { match: blockquote }
    attribution: { match: paragraph, itemModel: { fields: [ { name: name, match: strong } ] } }

1. All or nothing. A group matches when every one of its required members matches, in order, from the current position. Then it takes those nodes. Otherwise it takes nothing, and every node it looked at is left for the field after it. This is SPEC-182 D3's rule ("match only when the required parts are there") applied to sibling nodes instead of the parts of one paragraph.
2. Members may be optional, as fields are. An optional member that does not match is skipped inside the group without failing it: an eyebrow before a required headline.
3. A group may be required or optional, as a field is, and may contain groups.
4. Members resolve to their own names, as if written at the group's level: a composition places {% slot name="quote" /%}, not a path. Member names are unique across the whole model, and a duplicate is rejected at load. The group's own name resolves to whether it matched, for conditions.
5. No backtracking across fields. The group tries its members once, from where it stands. A group that fails does not make an earlier field give anything back.

D2 — a block field reads its node's children with a nested model

A field matching a block node (blockquote, a list item's block content) may declare model, a content model applied to that node's children:

quote:
  match: blockquote
  model:
    type: sequence
    fields:
      body:        { match: paragraph, greedy: true }
      attribution:
        match: paragraph
        optional: true
        position: last
        itemModel:
          fields:
            - { name: dash, match: text, pattern: '^\s*[—–-]\s*$' }
            - { name: name, match: strong }
            - { name: source, match: text, pattern: '^\s*,?\s*(.+)$', optional: true }

1. The field still matches the whole node, as today. The nested model reads its children into members named quote.body and quote.attribution: dotted, because a nested model is the node's own, and its members may repeat names used elsewhere (body). Nested models may nest again, to any depth (a.b.c).
2. A composition places a member by its dotted name ({% slot name="quote.attribution" /%}) or the whole node by the field's name, not both (SPEC-145 D11's placed-once rule extends to members).
3. The nested model is any content model, so a group (D1) or a condition can sit inside it.
4. An unmatched child inside the node is reported as content-unmatched, naming the field.

D3 — a field can claim the container's last node

position: last on a field: before the sequence is walked, the last node of the container being read is set aside for that field if the field matches it. Every field before it, greedy or not, then reads the nodes before it; the anchored field takes the node it reserved.

• body (greedy paragraphs) and attribution (position: last, a paragraph with a dash and a strong): the body takes every paragraph but a final attribution. With no attribution line, the last node is not reserved and the body keeps everything.
• "Last" is the last node of the container being read: the blockquote's children for a nested model (D2), the zone or the rune's children otherwise. A node after the attribution line (an image inside the quote) means it is not last, and it is not read as one, which is the authored convention.
• One node. A trailing run is not anchored; no case needs it, and a mechanism for it can be added when one does (ADR-030 rule 6: vocabulary is cheap, mechanism expensive).
• At most one anchored field per container, and it must be the last field of its model. A model with two, or with fields after it, is rejected at load.
• Stated on the field that takes, which is the field the rule is about: an attribution is the last line. (Decided 2026-11-01 over yields: true on the greedy field that gives, which needed a field of lookahead and an "otherwise matches nothing" clause.)

D4 — a required field that matched nothing is reported

1. content-missing is reported for a required field (or required group, or required member of a matched group) that took no node: whether the children ran out first or the field was skipped. It names the field and the rune, at the rune's line.
2. A child skipped by a required field is no longer dropped. It stays for the next field, as it does after an optional one, and is reported as content-unmatched only if nothing takes it. The report then names the real problem (the missing field), and nothing after it is misread on the strength of a field that is not there.
3. Rollout: content-missing is a warning for one release, then an error. Before it lands, a sweep of site/content, the SEO fixtures and the composed fixtures lists every instance, and each is fixed or its field made optional where the rune genuinely accepts the absence.

D5 — section's header grammar is declared, and the hand-written reader goes

section's content model becomes its grammar:

type: sequence
fields:
  header:
    optional: true
    group:
      eyebrow:  { match: paragraph, optional: true }
      headline: { match: heading }
      blurb:    { match: paragraph, optional: true }
  body: { match: any, optional: true, greedy: true }

• A paragraph with no heading after it fails the header group (its required headline is missing) and falls to body: ADR-046, declared.
• The media zones stay where they are. section splits its media and footer zones on --- before it reads its header (sectionZones). The header model applies to the content zone, which a delimited model already expresses; the spec's work item confirms the zone split reads the same through it.
• The header attribute's modes (headline, none) become conditional models (when), which the content model already supports.
• sectionHeaderParts and the reader registered in schemaHeaderRoles are removed. Rendering, the composition audit and the editor read the header from the one declaration, so they cannot disagree. Every existing section, hero, cta and testimonial test passes unchanged: that is the gate.
• SPEC-185 D5's epigraph is then one more optional group inside header, after the blurb.

D6 — compositions declare the same constructs

A composition definition's content accepts group, model and position: last in the same YAML, with the same load-time checks: unknown keys, a group with no required member, duplicate member names, a nested model on a non-block field, position: last on more than one field of a model or on a field that is not its last. Each is reported at definition load, naming it, and the composed-runes guide's Errors table gains an entry for each new code.

Not in scope

• The attribution and epigraph themselves. SPEC-185 defines them; this spec provides what they are declared with.
• Backtracking or general lookahead. D1 and D3 each look at most one field ahead of where they stand. A model that needs more is a model to restructure.
• Inline-level groups. itemModel already reads a paragraph's inline parts in order.

Open questions

• (Settled 2026-11-01: dotted names for a nested model's members, flat names for a group's, D1.4 and D2.1. A dot means "inside that node"; no dot means "one of this rune's own parts". Nesting may go to any depth, a.b.c.)
• (Settled 2026-11-01: position: last on the field that takes, D3.)

Acceptance Criteria

• A content model accepts an optional or required group that matches all of its required members in order, or takes nothing and leaves its nodes for the next field
• A block field accepts a nested model over its node's children, resolved to dotted member names, with unmatched children reported
• A field with position: last reserves the container's last node when it matches it, the fields before it reading the rest, and reserves nothing otherwise
• A required field, group or member that took no node is reported content-missing, as a warning for one release, and a child it skips is no longer dropped
• The sweep of site/content and the fixtures is done, and each content-missing it found is fixed or its field made optional
• section's header is read from its declared model, and sectionHeaderParts and its header-role reader are removed, with every section, hero, cta and testimonial test passing unchanged
• Composition definitions accept group, model and position: last, with load-time errors for each misuse documented in the composed-runes guide

References

• ADR-046: the header grammar D5 declares.
• SPEC-182: paragraph item models, and D3's match-only-when-present rule D1 generalises.
• SPEC-185: the epigraph (D5) and attribution (D6) that depend on D1–D3.
• SPEC-145: D11, placed-once, extended to members.
• SPEC-169: the content models a composition takes.

Why

The look-names fail as polaroid did. SPEC-159 rejected comic and polaroid because a look name is "a specification wearing the costume of a name": a theme cannot render editorial differently without failing to do it. An editorial theme looks editorial by default; an author on a docs site should not be choosing it per quote.

plain exposed the problem. The composed testimonial needed its quote undecorated. Instead of the quote's role deciding that, the author (the composition) asked a decorated rune to remove its decoration, so the testimonial looks right only because of how Lumina happens to be styled.

> is not always a quotation, so the element is the wrong thing to style. recipe reads > as tips, comparison as callouts, conversation as dialogue, and event, organization and symbol accept it as body content. A theme that styles blockquote as a quotation puts a rule and a glyph on a recipe tip, and each of those runes undoes it.

The variety an editorial theme wants comes from elsewhere. A theme wanting different quotations in different parts of a page is a real need, and every case of it has a home that is not an author's look-name: where the quote sits (the region, the container, the reading register, which the theme already sees), what it is (a role, below), and how loud it is (a scale, prominence, D9). A remaining wish for pure variety is a composite intent (SPEC-159 Part 3), named for what the quote is to the reader and defined by a theme or a pack.

Decisions

D1 — four roles of quotation

The roles are the author's choice. What each looks like is the theme's.

D2 — decoration belongs to the theme; the look-names are withdrawn

• pullquote's variant is deprecated: default, accent and editorial name looks. It is accepted with a deprecation warning for a release, then removed.
  • editorial maps to nothing: on an editorial theme a featured quotation looks editorial already.
  • accent's tinted background is the universal tint, which every rune already takes; its extra size is loudness, which is prominence (D9).
• plain is withdrawn before it is released. It was added in v0.45.0 and is unpublished, so this is a changeset edit (as SPEC-182 D7 withdrew the testimonial codemod).
• align stays. It is placement, not decoration.
• A theme retunes a role; it does not redefine it (SPEC-159's bound). Every theme renders a featured quotation as a set-apart statement with its attribution; how it is set apart is its own.

D3 — a theme styles roles, not the blockquote element

1. Each role carries a marker, data-quote="quotation" | "featured" | "excerpt" | "epigraph", on the quotation's outermost element: the <figure> when it is attributed, the bare <blockquote> when it is not, the <aside> for an excerpt. A theme styles the role through it, with its blockquote and figcaption as descendants, and never guesses which wrapper holds it.
   • Its own attribute, one per vocabulary as data-reading, data-media and data-meta-type are. Not data-section, which names a part of a rune's own structure: a > in running text is part of none. Not a general content-role attribute: how text speaks (dialogue, lyrics, a caption; SPEC-159's voice role) is a different question from what kind of quotation it is, and keeping them apart lets one element carry both.
   • Independent of rune names. An epigraph comes from section's header grammar or the page start, a quotation from the core node, and a later composition may make a featured quotation without pullquote, so the theme contract names roles, not runes.
2. The quotation role comes from the core blockquote node. A plain > that no rune consumes has no rune to mark it, so the node emits data-quote="quotation".
3. A rune that reinterprets > produces its own elements and never carries the marker: tips, callouts and dialogue are not quotations, and no longer need to undo a quotation's style.
4. A composition does not style what it reinterprets; it places the rune that is that thing. A composed recipe's tips are notes, so its template places them in a hint (type="note", or a tip type if recipe's conversion shows one is needed), and the theme's hint treatment applies, as for any hint. The same holds for comparison's callouts. Neither keeps tip or callout CSS of its own once it is composed: that is the rule every composition already follows (SPEC-145 D2), stated here because > is where the temptation lies.
5. Lumina's global blockquote rule becomes its [data-quote="quotation"] rule. A blockquote with no marker (third-party content transformed outside refrakt's nodes) gets a minimal, unopinionated fallback at zero specificity, so it is never unstyled to the point of illegibility.
6. The reading register is context a theme may use. A quotation in prose and one in ui (a card body) may be styled differently: Lumina keeps its rule and glyph in prose and sets a quotation quieter in ui. The role is the same; the context differs (SPEC-108).

D4 — pullquote is the featured quotation; the excerpt is an option of it

One rune, because the two cases look the same and differ in how they are read. It keeps its name and gains quote as an alias, the name for its default case; renaming is not decided (open question 2).

• By default it is a featured quotation, read by assistive technology, as every current use of it (this site's examples quote other people) already is.
• excerpt marks a true pull quote: the words already appear in the text, so it renders as an <aside aria-hidden="true"> and a screen reader does not read the passage twice.
• Its structure is <figure><blockquote>…</blockquote><figcaption>…</figcaption></figure>. The authored blockquote's content is the <blockquote>'s; it is never nested in a second one. An attribution, when present, is the <figcaption>.
• It reads its attribution from a paragraph after the quote, **Name** — Source, with the item-model reading SPEC-182 gave compositions (D6 below).

D5 — an epigraph is recognised by position and attribution

A blockquote is an epigraph when both hold:

1. It stands in the epigraph position: after a section's headline (and its blurb, when there is one) and before the body; or at the very start of a page's content, or directly after its # title (many layouts take the title from frontmatter, so the content may have none).
2. It carries an attribution (D6), in either position.

Without an attribution it stays a quotation. That keeps the ordinary opening quotation from being misread: epigraphs almost always name their source, and requiring one is SPEC-182 D3's rule ("match only when the required parts are there") applied to position. No new rune.

• It is part of section's header grammar (ADR-046), as one more optional group after the blurb (SPEC-186 D1, D5). section reads its header from the content a composition places in it, so every rune composed over section reads epigraphs the same way: hero, cta, testimonial, and card once it is composed. That is the intended reach: reinterpretation means the same Markdown means the same thing in every rune. A card's epigraph need not look like an essay's; the theme styles the role by context (D3.6).
• The quote and its attribution are read as one unit, optional together and required within: a lone attribution after a headline is not an epigraph, and stays the blurb it is today.
• A heading inside a body does not open an epigraph. It is the extension most likely to misread ordinary content (a source quoted under a docs heading), and the cheapest to add later.
• A composition that never wants one sets the placed section's header="headline".

D6 — attribution is one shared part

Every role that carries a source reads it the same way and renders it the same way:

• Read from either of two positions, with one item model (**Name** required, then the source after a dash or comma, as SPEC-182 D6 fixed for testimonials):
  • after the quotation, a **Name** — Source paragraph directly following it. Read wherever a rune owns the sequence: pullquote, a composition's model, section's epigraph position. The quote and the attribution are one group (SPEC-186 D1), so an attribution is never read without a quote.
  • inside it, as the blockquote's last paragraph, — **Name**, Source: the leading dash marks it, so a quote ending in bold text is not misread. Read by a nested model on the blockquote field (SPEC-186 D2, D3), so it works in a composition as it does in a native rune, and by the core blockquote node, which gives a plain quotation an attribution with no rune.
  • The after form can ship first. It needs only SPEC-182's paragraph model and SPEC-186 D1; the inside form waits for SPEC-186 D2–D3.
• Rendered as <figcaption>, the name in <cite>.
• A quotation (role 1) with an attribution becomes a <figure>; without one it stays a bare <blockquote>.

D7 — the composed testimonial places a featured quotation

The staged testimonial places {% pullquote %} for its quote, with no variant: a testimonial's quote is a featured quotation, and its look in a card comes from the theme and the ui register, not from the composition. The testimonial keeps what is its own meaning: the rating, and the Review / Person / Rating schema.

The pullquote owns the attribution. It reads the **Name** — Role paragraph (D6) and renders the <figcaption>, so the testimonial no longer has a byline of its own:

• The testimonial's attribution field and its fills (SPEC-182 D6) go: the paragraph is the pullquote's content now, placed with the quote.
• Its byline meta block goes with them. The <figcaption> is the byline.
• Its schema reads the attribution from the pullquote: author and role resolve to the <figcaption>'s name and role nodes, as reviewBody already resolves to the pullquote's quote across the placed rune. The Review / Person graph is unchanged, which the SEO baseline checks.
• author and role stay declared on the testimonial for the attribute form (SPEC-182 D4), and fill the pullquote's attribution when written, so both forms render one <figcaption>.

This amends SPEC-182 D6: the paragraph reading it introduced is kept, and moves from the testimonial to the pullquote, where every role reads attribution the same way.

D8 — what happens to the interim fixes

• plain (#746, unreleased) is withdrawn under D2.
• The unwrap in #747 (no blockquote nested in a blockquote) is right and is folded into D4, where the root becomes <figure>, so pullquote's markup changes once rather than twice.

D9 — a featured quotation's loudness is its prominence

prominence scales a page-section header, and SPEC-183 made it a register for the whole header. A featured quotation is the centrepiece of its block in the same way, so it takes the same axis rather than a new one:

• The gate widens. The prominence facet applies to a rune with a header role, or one marked as a featured quotation or an excerpt (D3). On a plain quotation (role 1) it stays unavailable.
• The steps keep their meaning. quiet and normal are a quotation's own size, prominent and display a set-apart one, marquee a page's centrepiece. The theme maps each step to the quote's type size, as it maps them to a header's; Lumina ramps from the quote's resting size, as it ramps a title (WORK-538).
• accent's extra size (D2) is prominence="prominent".
• A testimonial's quote is normal in a card; a quote set as its own band is display or marquee.

Not in scope

• Callouts, tips and dialogue. They are not quotations (D3.3); their own runes style them.
• Inline quotations (<q>). A separate question.
• A theme other than Lumina. This spec sets the contract a second theme must meet; it does not write one.

Open questions

1. (Settled 2026-11-01: prominence, D9.)
2. Renaming. D4 adds quote as an alias. Whether quote later becomes the name and pullquote the alias is left until the alias has been in use.
3. (Settled 2026-11-01: section's header grammar, so every rune composed over it, and the page start; not headings inside a body. D5.)
4. (Settled 2026-11-01: the pullquote owns it, D7.)
5. (Settled 2026-11-01: data-quote with quotation | featured | excerpt | epigraph, on the outermost element, D3.1.)

Acceptance Criteria

• The core blockquote node marks a quotation data-quote="quotation"; runes that reinterpret > carry no marker
• Lumina styles the quotation role, not the bare element, with a minimal fallback for an unmarked blockquote, and a quieter quotation in the ui register
• pullquote renders <figure><blockquote>…</blockquote>[<figcaption>…</figcaption>]</figure>, never a nested blockquote, marked data-quote="featured"
• {% pullquote excerpt %} renders as <aside aria-hidden="true">, marked data-quote="excerpt"
• An attribution paragraph is read the same way for every role and rendered as <figcaption> with <cite>
• A blockquote after a page or section title with an attribution is an epigraph, data-quote="epigraph"; without an attribution it stays a quotation
• pullquote's variant warns as deprecated, and plain is withdrawn before release
• pullquote accepts the quote alias, and takes prominence (D9), which ramps the quote's size
• The staged testimonial places {% pullquote %} with no variant, the pullquote renders its attribution, and the testimonial's Review / Person graph is unchanged, read from the pullquote
• A composed rune that reinterprets > places the rune that is that thing (a hint for tips) and ships no style for it
• The theme-authoring docs list the four roles as part of the theme contract, and the pullquote and blockquote docs describe roles, not looks

References

• SPEC-159: intent, not look; "retune, don't redefine"; composite intents.
• SPEC-108: reading registers, the context D3.5 uses.
• SPEC-182: paragraph reading (D4, D6) and match-only-when-present (D5); D6 amended by D7.
• SPEC-145: D2, a composition has no style of its own (D3.4).
• SPEC-183: prominence as a register (D9).
• SPEC-186: the content-model constructs D5 and D6 are declared with: groups, nested models, position: last.
• ADR-046: header grammar by position (D5).
• ADR-030: variants are modifiers; semantics are rune-locked.