State of the art, per sub-question

Two records are kept apart: most adopted (what people use) and best evidence (what a study measured).

Sub-question Most adopted Best evidence Measured how
Q1 Which models store text as a tree, with what recursive rule Markdown's flat headings over container blocks (CommonMark 2024); XML content models with a recursive section (DocBook, TEI, DITA) Mignet 2003 and Mlýnková 2006: recursion is common in schemas and shallow in instances corpus statistics of web XML
Q2 Text that is not a tree milestones and stand-off markup (TEI 2026, chapter 21; the OSIS norm in DeRose 2004) Renear 1996: every version of OHCO met a counterexample argument from cases, no frequency count
Q3 A unit with two parents reuse by reference: conref and keys (OASIS 2015), XInclude, MediaWiki templates Juergens 2010 for copies; Dayton 2010 for adoption; Johnson 2020 for the rate of transclusion clone detection on 28 specifications; a survey of 276; a classification of transclusions
Q3 Effect of reuse on readers none none found none
Q4 Cycles through references detection and report, or a depth cap (table below) none; standards and tool documentation only descriptive comparison of 14 systems, this review
Q5 Checking by program grammar validation in XML; prose linters in docs as code Grijzenhout 2011 (validity on the web); Juergens 2010 (precision of clone detection); Fleckenstein 2023 (students) crawl of 180,000 files; precision on 28 documents; meta-analysis, g = 0.55
Q6 What strict models cost lighter forms: Markdown; HTML heading levels (WHATWG 2022); one parent per block (Notion 2021) Dayton 2010 (downsides, failures); Batova 2014 (context loss, contradictory); Martens 2006 (unused power) survey; integrative review of 60 articles; analysis of real schemas

How each system meets a cycle through a reference

System Recursive type Reuse by reference A cycle through a reference Source
XML 1.0 content models; the instance nests properly parsed entities forbidden: well-formedness constraint No Recursion Bray 2008, 4.1
XInclude 1.0 none of its own include of a resource or fragment fatal within the inclusion chain; two non-nested includes allowed by implication Marsh 2006, 4.2.7
DITA 1.2 and 1.3 nested topics and topic references conref, conkeyref, keys, map references not addressed in the conref pages read OASIS 2010; OASIS 2015
DITA-OT as DITA as DITA reported: DOTX013E, DOTX053E, DOTJ069E DITA-OT 2026b
DocBook 5.2 section, unbounded; sect1 to sect5 not read here not read here Walsh 2019
TEI P5 div; div1 to div7 pointers, stand-off not read here TEI 2026
JSON Schema, version 2020-12 $ref, $dynamicRef $ref must not loop against an instance; mutual allOf loops undefined Wright 2022, 9.4.1
MediaWiki templates in templates template transclusion reported as a template loop; depth capped at 100 MediaWiki 2026a; 2026b
Asciidoctor nested includes include directive capped: max-include-depth 64 Asciidoctor 2026
Sphinx tree of tables of contents a document under several parents self-reference and circular references reported and ignored; for several parents, one is named as selected Sphinx 2026
CommonMark containers hold blocks; headings flat none in the specification none possible CommonMark 2024
OPML 2.0 outline in outline include of another OPML file not addressed in the specification read Winer 2007
Logseq nested blocks block reference; embed of a block with its descendants not read here Logseq 2023
ProseMirror content expressions one node value in several places none possible among values (an inference from immutability) ProseMirror 2026
  • Of the 10 systems whose reuse is by reference (TEI pointers and ProseMirror shared values set aside), 5 forbid or report a cycle (XML, XInclude, DITA-OT, MediaWiki, Sphinx), and 2 cap depth (Asciidoctor, MediaWiki). In 3 texts it is undefined or unaddressed (JSON Schema, DITA, OPML), and 1 was not read on cycles (Logseq). MediaWiki does both.
  • No system read stores a unit that contains itself on purpose.

Adoption counts, descriptive, taken 2026-10-29

Repository Stars What it is
jgm/pandoc 46,736 document converter over one tree model
sphinx-doc/sphinx 8,084 documentation builder with a tree of tables of contents
DavidAnson/markdownlint 6,401 Markdown structure linter
vale-cli/vale 6,300 prose linter
asciidoctor/asciidoctor 5,227 AsciiDoc processor with includes
json-schema-org/json-schema-spec 5,199 the JSON Schema specification
btford/write-good; get-alex/alex 5,107; 5,105 prose linters
amperser/proselint; textlint/textlint 4,587; 3,203 prose linters
syntax-tree/mdast 1,484 Markdown syntax tree specification
dita-ot/dita-ot 458 DITA reference processor
TEIC/TEI; ProseMirror/prosemirror-model 358; 338 TEI source; ProseMirror document model
relaxng/jing-trang; Schematron/schematron; docbook/docbook 260; 97; 89 validators and schema sources
  • Counts come from the public GitHub API, unauthenticated, on 2026-10-29 (GitHub 2026d). Stars measure interest, not use.
  • Standards that predate GitHub (XML, DocBook, TEI, DITA) are understated by their repositories.
  • Dayton 2010 is the only measured adoption share found: about 1 in 9 of 276 technical communicators used DITA in 2008, and half used single sourcing or content management.

Upper bounds (achieved)

  • A reference processor detects three kinds of cycle: conref, map and key (DITA-OT 2026b).
  • Clone detection finds duplicated prose in specifications at precision above 85% per document after tailoring (Juergens 2010).
  • Data models for overlap exist and are implemented: GODDAG, EARMARK, TAG (strand 2 of the review).
  • Automated feedback improves students' writing, g = 0.55 (Fleckenstein 2023, secondary).
  • A web-scale measure of validity exists: 85.4% of 180,000 files well-formed (Grijzenhout 2011, secondary).

Lower bounds (impossible, or not shown)

  • A well-formed XML element cannot contain itself: proper nesting gives a finite tree (Bray 2008, 2.1), and a parsed entity may not recurse (4.1). This follows from the syntax.
  • No single hierarchy holds every perspective on a text: each version of OHCO met a counterexample (Renear 1996). This is an argument from cases, not a proof.
  • Not shown in this search: any effect on readers of reuse by reference against copies, of a tree model against a graph model, or of a depth-capped against an unbounded section.
  • Not shown: the recall of any prose or structure checker on documentation; Juergens 2010 measured precision only.