Against

Terms are defined in review.md; works are listed in the bibliography.

1. Linear or super-linear surprisal cost

For linear: Smith & Levy 2013; Wilcox et al. 2023; Shain et al. 2024.
For super-linear: Meister et al. 2021 (weak for reading time); Hoover et al. 2023; Xu et al. 2023 (some languages only).
For sub-linear in surprisal: Brothers & Kuperberg 2021, whose cost is linear in probability.
Failed replication: the same items under eye-tracking give a logarithmic effect (Buggy et al. 2026, preprint).
Source of disagreement 1: the language model that estimates surprisal (Xu et al. 2023; Oh & Schuler 2023).
Source of disagreement 2: the measure, self-paced reading or eye-tracking.
Source of disagreement 3: the range sampled; cloze norms cannot resolve small probabilities (Buggy et al. 2026).
Production is not comprehension: UID evidence from speakers (Jaeger 2010) says nothing direct about reader cost.
Entropy rate constancy did not hold with neural models (Verma et al. 2023).
Information rate rises and falls with discourse structure, so uniformity is not the only pressure (Tsipidi et al. 2024).

Shorter dependencies in corpora (Futrell et al. 2015) do not by themselves show a cost to readers.
In naturalistic reading, integration cost is weak outside nouns (Demberg & Keller 2008).
The random baseline is disputed: non-crossing may itself result from minimization, so a projective baseline builds in part of the effect (Ferrer-i-Cancho & Gómez-Rodríguez 2015).
The "optimal" of Futrell et al. 2015 is the projective minimum; the unconstrained minimum is lower or equal.
A rival mechanism: retrieval interference, not distance, may carry the cost (Lewis & Vasishth 2005).
Yngve depth: a corpus shows a bias against left-branching but no sharp limit (Sampson 1997).

Karlsson 2007 counts what writers produced; it does not measure what readers can parse.
Karlsson 2007 ties multiple center-embedding to a written stylistic tradition, so its rarity is partly cultural.
Processing limits depend on the grammar: the missing-verb illusion holds in English, not in German (Vasishth et al. 2010).
The illusion appears at degree 2 (Gibson & Thomas 1999), below the corpus maximum of 3.

The reverse cohesion effect depends on skill: skilled high-knowledge readers gained from high cohesion (O'Reilly & McNamara 2007).
It depends on the measure: situation-model tasks, not recall (McNamara et al. 1996b).
One of its results did not replicate in a second experiment (McNamara & Kintsch 1996a).
Expertise reversal: a rule tuned for novices is not neutral for more experienced learners (Kalyuga et al. 2003).
Consequence: no density rule is reader-free; each needs a stated reader.

Lab items: sentences or short passages, read once, often word by word.
Notes: reread, skimmed, searched by heading or keyword.
Rereading changes eye movements: shorter and fewer fixations, longer saccades (Raney & Rayner 1995).
Skimming under time pressure favors early parts of paragraphs and pages, and depends on navigation links (Duggan & Payne 2009).
Competent adults read the same text with distinct strategies; heading users summarize best (Hyönä et al. 2002).
Grade for each transfer from lab to notes: convention, untested.

Formulas count sentence length and word familiarity; they do not define what makes a text hard (Davison & Kantor 1982).
A linguistic analysis of four adapted texts argues that rewriting to a formula can make text harder (Davison & Kantor 1982; no reader testing).
For adults, NLP-feature models predict comprehension better than classic formulas (Crossley et al. 2017).
Coh-Metrix indices are correlates of difficulty; a rule that moves an index is not shown to move comprehension.

Gopen & Swan 1990 report no controlled test; parts of it agree with given-new evidence (Haviland & Clark 1974) and locality (Grodner & Gibson 2005).
Centering's uniqueness constraint holds or fails depending on parameter settings (Poesio et al. 2004).
RST describes structure; this review found no test that RST depth itself predicts reading cost.

Practice: rewrite until a grade tool passes (Vale 2026; Hemingway 2026; textstat 2026).
Principle: Even information density is not supported as a way to read faster.
Evidence: an analysis, not a measurement, argues that adaptations written to a formula can read worse (Davison & Kantor 1982); formulas ignore readers (Redish 2000; Schriver 2017).
Practice: a Flesch floor fixed by statute (NAIC 2025), against the weak adult baseline in section 6.
Practice: splitting to a word cap; the EC warns to keep link words so coherence survives (EC 2011).
Tool caveat: textstat's own README limits SMOG to texts of at least 30 sentences.