Against: controversies, limits and failed assumptions

This file gathers the evidence against the tree model and its reuse: the critique of OHCO, the costs of strict models, fragments read out of context, and the limits of the evidence used here.

Text is not one tree

  • Renear 1996 concludes that "no version appears to be free from counterexample" of the OHCO thesis.
  • Its counterexamples come from literary and textual scholarship: verse against sentence, enjambment, strike-outs, variants, link anchors.
  • Schmidt 2010 (abstract) goes further: embedded markup hides the text and carries obsolescent, subjective data, beyond overlap.
  • Haentjens Dekker 2017 makes hierarchy optional and orders only the text.
  • Against the critique: Renear 1996 itself keeps that perspectives often determine hierarchies, and TEI 2026 still builds on XML trees, with workarounds for what does not nest.
  • Whether ordinary documentation overlaps as literary texts do is not known (position.md).

Strict models cost

  • Dayton 2010 (abstract): half of the users of single sourcing or content management reported significant downsides, and about 1 in 10 respondents had seen an implementation fail.
  • Batova 2018 (abstract): at one company, writers' motivation fell when component content management was framed only as efficiency.
  • Martens 2006 (abstract) relates the little use of XML Schema's extra power to the complexity of its specification.
  • HTML's outline from nested sections stayed unimplemented for over a decade and was replaced by heading levels (WHATWG 2022).
  • XInclude was never shipped by browsers; an answer cites the cost of loop detection among the reasons (SO 2011d).
  • S1000D's schemas are more rigid than DITA's (WSE 2018b), and a do-it-yourself DITA setup gives little support for reuse (WSE 2014a): opinions.
  • The Composite pattern's uniform interface loses type safety (Wikipedia 2026d), and the evidence for design patterns in general is weak (Zhang 2012).

Fragments reused without their context

  • Batova 2014: segmentation strips content of its larger context, so translators cannot judge it and make context errors; best practices do not resolve the conflict between consistency and adaptation.
  • WSE 2018g: readers who dip into documentation need repetition at read time; reuse tools serve this only if each use reads on its own.
  • DITA needs @copy-to so that a topic used twice gets its own links and navigation (OASIS 2015): without it, position-bound links are shared.
  • Notion gave up several parents per block because permissions became ambiguous (Notion 2021, the owner's account).
  • Copies are not always wrong: some code clones help maintenance, as file-structure reviews (Kapser 2008), linked, not repeated.

Limits of the measured evidence

  • The corpus studies of XML are old: Choi 2002, Mignet 2003, Mlýnková 2006, Grijzenhout 2011. Choi and Mignet were read through Mlýnková 2006, Grijzenhout through its abstract.
  • Their documents are web XML of every kind (feeds, data, configuration), not documentation.
  • Juergens 2010 measures copies in requirements specifications; its reading and inspection costs are computed from published rates, not observed.
  • Dayton 2010 surveyed 1,000 STC members in 2008-05 and 276 answered (27.6%), so the respondents selected themselves; read by abstract.
  • Fleckenstein 2023 pools studies of students' essays, not documentation or linters; its heterogeneity is high.
  • Johnson 2020 is a preprint, read by abstract, on one wiki.

Limits of the practice evidence

  • Standards state rules, not outcomes; tool documentation states behaviour, not effect.
  • Notion 2021 is a company describing its own design.
  • Stars measure interest, not use; standards older than GitHub are understated (state-of-the-art.md).
  • Schematron 2026 scores under 10 of 20 and is used only for its own edition claim (sources.md).
  • The Q/A threads are opinion; the GoF reading in SESE 2022c is secondary.

What was not found

  • No reader study comparing reuse by reference with copies.
  • No replication of Juergens 2010 on user documentation; Luciv 2018 and Nosál 2016 count duplicates without consequences.
  • No study of overlap frequency in technical documentation.
  • No evaluation of the DITA conref or key mechanisms on consistency, in this search.