Position

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

Known

Finding	Grade
Per-word reading time is linear in surprisal	measured
So, for reading time, even spreading of information buys nothing at fixed length and total information	measured (argument in `state-of-the-art.md` §2)
Non-uniform surprisal lowers judged acceptability	supported with caveat
Languages keep dependencies short, between random and optimal	measured
Long dependencies cost reading time in controlled sentences	measured
In a corpus of 7 languages, written center-embedding never exceeded degree 3; in speech it was practically absent	measured (corpus)
Reading time tracks propositions more than words	supported with caveat
A missing antecedent costs a bridging inference	measured
Ideas high in the text hierarchy are remembered better	measured
Readers who use headings summarize best	measured
The best cohesion level depends on reader knowledge and skill	measured
Classic readability formulas are a weak baseline for adults	measured

Each item below is "not seen in this review", which is a claim about the search, not about the world.

A controlled test of definition-to-use distance in reference notes.
A test of heading or list nesting depth on retrieval from notes.
A test of the surprisal linking function on reread or searched text.
The complexity of ordering concepts when weights are in-degree minus out-degree.
A test that RST tree depth predicts reading cost.
Whether lab effects on single reading hold for texts that are skimmed and searched.

One map from five scales to findings, grades and checks, so a rule can be traced to its evidence.
The Jensen argument written out, with the verdict split by measure: reading time or acceptability.
The note-ordering problem restated as scheduling with precedence, with what is and is not known of its complexity.
Checks sorted by what they compute exactly without a language model (baseline.md).
A reader condition on every density rule, because cohesion and support effects reverse with expertise.

Practice runs ahead: sentence targets of 15, 20 and 25 words (EC 2011; Style Manual AU 2024) come with no test in the sources read.
Practice ignores evidence: US state law still fixes a Flesch floor (NAIC 2025) after Redish 2000, who already criticised state Flesch floors in insurance law, and Crossley et al. 2017.
Practice agrees: the legalese advice matches Martinez et al. 2022 on center-embedding (p. 6: unpack clauses into separate sentences).
Critics and the standard meet: Redish 2000 and Schriver 2017 name usability testing as the benchmark.
Consequence for checks: a readability or length number is a flag for a reader to judge, never an error.
Not seen in this review: any government test of its own word cap.

The brief said Futrell et al. 2015 use MinLA as a baseline.
They use the minimum over projective orders; unconstrained MinLA on trees is the baseline of Ferrer-i-Cancho et al. 2022.

Long-context models use information best at the start or end of the input (Liu et al. 2023).
That is a locality effect at document scale; whether human findings transfer to such readers is untested here.