Against: null results, critiques and failed fits

This file gathers the evidence against typed units: null results, a failed replication, critiques of type grammars, and the costs practice reports.

Null results for the typed format itself

Horn 1969: the same content with or without the typed format gave adjusted posttest means of 19.56 and 17.65, not a significant difference (t = 1.17, 38 df).
- The authors give two readings: the visible features matter less than the content analysis, or the classroom test was too coarse. They favour the second.
- Among those who finished (15 typed, 10 prose), the typed group reported more study time (5.21 against 3.87 hours), not significant; the prose finishers had started higher (pretest 14.00 against 8.73).
Kools 2008: neither the number of headings nor the Information Mapping order changed search speed; only having headings did (abstract).
Both tests are small and old; neither was replicated in this search.

Zwaan 1994 found that framing one text as literature or as news changed reading time and memory (secondary).
Triantafyllopoulos 2021 ran three online experiments and found no difference in comprehension by genre expectation.
Its authors conclude that fiction and nonfiction do not exert universal effects, especially in digital reading.
So a type label read by the reader may not change reading at all.

Text structure instruction lost its advantage at the delayed posttest (Bogaerds-Hazenberg 2020).
Far transfer was small and inconsistent (Hebert 2016).
Pyle 2017 found larger effects on researcher-made measures and with researcher implementers (abstract); such measures often inflate effects (an inference, not tested there).

Black 1979: of the story grammars, one was formally adequate; they missed many stories and accepted a class of non-stories.
The same paper argues the grammars add nothing to models of story content.
Trabasso 1985 moved the field to causal networks of events, a content account (title and venue only).
Lesson for the topic: a type can be checked by a program, but passing the check does not make a unit good of its kind.

Horn 1969 lists eight map types; a 2018 answer by the author of Baker 2013 lists six, "originally seven" (WSE 2018a). Information Mapping 2026 (example only) lists none on its method page.
DITA's own specification says many DITA types are not closely connected with Information Mapping (OASIS 2015, 2.2.1.4).
The canonical set has changed over time, so "the" information types are a convention, not a finding.

DITA forbids nested sections and untitled topics; a type that needs them is no longer DITA (OASIS 2015, B.11).
S1000D allows only added constraints on a fixed schema, where DITA allows derived types (WSE 2018b).
Deriving a type needs a developer, says one asker; the answer calls a simple derivation trivially easy (SO 2013a).
Tool and set-up costs fall on small teams (WSE 2014a).
Writers' motivation fell when component content management was framed only as efficiency (Batova 2018).
One topic per task fragments a help system into many tiny pages; the critic proposes grouping related tasks (Johnson 2013).
A story template forced on content harms the story, says a practitioner (WSE 2020a, example only).

No vendor claim on typed content and model accuracy was read in full; one page timed out and is not cited.
Mortezaagha 2026's trade-off between structure-aware and similarity retrieval is reviewed in progressive-disclosure.

Most formal sources were read by abstract only; the CCM reviews by title only (sources.md).
Information Mapping 2026 (5 of 20) and Baker 2018b (9.5 of 20) are below the threshold and are examples only.
Company and committee pages describe their own method; DITAWriter 2026 is an informal list by one practitioner.
Q/A threads are opinion, with no data.
Minimalism (Carroll 1987; Lazonder 1993), the lineage of the task type, was not read; its results are not used.