CASE-062: Simplified texts for restricted and for full readers: a gain and a loss documented on different texts, documented nulls and a reversal, and a design that tests both on the same renderings

Found on 2026-10-10 by a search the owner asked for: a rendering made from the source alone that measurably helps a named restricted receiver and measurably loses information for a full-access one, for CLAIM-199. No single study was found that measures both on the same renderings. This case is therefore stipulated: a constructed design, whose two halves and whose nulls rest on documented studies, each cited below.

The case

What is documented. Text simplification, by hand or by algorithm, makes a rendering from the source text alone, without the situation the text is about. In CLAIM-199’s terms it is a garbling.

• A gain for named restricted readers. A. Cerga-Pashoja, J. Gaete, A. Shishkova and V. Jordanova, Frontiers in Psychiatry 10:546, 2019, doi:10.3389/fpsyt.2019.00546: 243 adolescents and adults with high-functioning autism, in a cross-over design, read texts simplified with the OpenBook tool and the originals; “both adults and adolescents with ASD gave more correct answers for the simplified” texts (mean 11.2 against 10.0, p < .001; the Bulgarian adolescents’ gain, 7.4 against 6.8, was not significant). There was no neurotypical arm. L. Javourey-Drevet, S. Dufau, T. François, N. Gala, J. Ginestié and J. C. Ziegler, Applied Psycholinguistics 43(2):485–512, 2022, doi:10.1017/S014271642100062X: 149 French second-graders read 20 texts in normal and hand-simplified versions; comprehension was “significantly better for simplified than for normal texts”, and gains “were greater for participants with weaker reading performance” (scientific texts only; p = .051 for the interaction). The authors concede that “some semantic nuances might be lost” but did not measure it: the same questions served both versions. G. Leroy et al., Journal of Medical Internet Research 15(7):e144, 2013, doi:10.2196/jmir.2569: 99 lay readers answered correctly on 63% of items for an algorithmically simplified health text against 52% for the original (p < .001).
• A loss for full readers. S. Agrawal and M. Carpuat, “Do Text Simplification Systems Preserve Meaning? A Human Evaluation via Reading Comprehension”, TACL 2024, doi:10.1162/tacl_a_00653, arXiv:2312.10126, Table 2: 112 native English readers answered reading-comprehension questions at 78.33% on the originals, 77.22% on a human “Elementary” simplification, and from 76.11% down to 67.78% on system simplifications (20.50% for one system); “over-deletion is the key culprit that makes questions unanswerable”.
• Nulls for restricted readers. A. Szarkowska, I. Krejtz, O. Pilipczuk, Ł. Dutka and J.-L. Kruger, Across Languages and Cultures 17(2):183–204, 2016, doi:10.1556/084.2016.17.2.3: with 44 deaf, 33 hard of hearing and 60 hearing viewers, “no benefit of editing down the text of subtitles, particularly in the case of intralingual subtitling and deaf viewers. Verbatim subtitles displayed with the higher presentation rate yielded slightly better comprehension results”.
• A reversal. D. Burnham et al., Journal of Deaf Studies and Deaf Education 13(3):391–404, 2008, doi:10.1093/deafed/enn003, abstract: “more proficient readers show better comprehension than poor readers and are better able to benefit from caption rate and, to some extent, text reduction modifications”. The gain from the garbling went to the less restricted readers.

What is stipulated. The design that would put both halves on one set of renderings:

1. Take matched passages, each with an original and a simplified version. OneStopEnglish, which Agrawal and Carpuat use, already gives each article at three levels, so the materials exist.
2. Fix probe items in advance, by annotators who see both versions, in two kinds: P-kept, answerable from both versions, and P-lost, answerable from the original only.
3. Draw readers from a named restricted population (for instance L2 readers at CEFR A2, or readers with autism) and from fluent native readers. Each reads one version of each passage.
4. Add probes on the interpersonal relation (stance, politeness) as well as content, so the study reaches CLAIM-087.

CLAIM-199 predicts that restricted readers gain on P-kept items, that full readers lose on P-lost items, and that no restricted reader of the simplified version beats full readers of the original on any item.

What it can show

On its documented halves, both parts of CLAIM-199’s picture, apart. A garbling raised usable information for restricted receivers whose rules could not extract it as easily from the source (the autism and grade-2 studies, Leroy). A garbling lost information that full readers could have used (Agrawal and Carpuat, for system output). Neither contradicts Blackwell’s order (THEORY-156): a gain for a bounded receiver is the claim’s gap between what a rendering makes possible and what a receiver can use (THEORY-176’s move for restricted rules).

The nulls show the defeat condition being tested in the field, and sometimes coming out on the null side. That fits the claim’s own hedge: garblings “help a receiver only when they do work the receiver cannot”.

As stipulated, the design is built to test the claim’s ceiling directly: whether any restricted reader does better with the rendering than a full reader does with the source, on the same items.

What it cannot show

• The central condition is met only across studies. No documented study gives one rendering a gain for a restricted reader and a loss for a full reader. The gain studies re-use questions answerable from both versions, so they cannot see a loss; the loss study has no restricted arm. The ceiling is never checked within one study, and the design that would check it has not been run.
• Content, not relation. Every documented probe is a content question. None touches the interpersonal relation that CLAIM-087 is about, so the case supports CLAIM-087’s improvement only by analogy.
• The strongest reply. The full readers’ loss may be noise where it matters most: 78.33% against 77.22% for the human simplification, a point. And the restricted readers’ gains may be gains in speed or effort, not in information their rules could not extract. A defender of reading the access gap as effort, not rule class, would say the documented studies cannot tell those apart, and that is right: none controls reading time.
• The reversal. Burnham’s result runs the other way from the pattern the claim’s gloss suggests: within a restricted population, the better readers gained more from the garbling. It does not defeat the claim, which says a garbling can help, not that it helps the most restricted most. It does show that the gain is not automatic, and it is known here only from the abstract.
• What would close the gap. The stipulated design above. Its P-kept and P-lost items, with the two populations, could serve as items in the discriminating study (CASE-055), whose named populations and stated tasks it shares. The thresholds, sample sizes and rates are for its preregistration and are not fixed here.

Sources

None of these works is held in the record as a LIT. They are cited here by identifier. Agrawal and Carpuat is a machine-learning work; a grep of a fresh clone of the anthology on 2026-10-10 found no entry for it.

• Cerga-Pashoja, A., J. Gaete, A. Shishkova and V. Jordanova (2019), Frontiers in Psychiatry 10:546, doi:10.3389/fpsyt.2019.00546, https://pmc.ncbi.nlm.nih.gov/articles/PMC6692437/. Read by the search online in the PMC full text; no saved copy, not re-checked here.
• Javourey-Drevet, L., S. Dufau, T. François, N. Gala, J. Ginestié and J. C. Ziegler (2022), Applied Psycholinguistics 43(2):485–512, doi:10.1017/S014271642100062X. Read by the search online; no saved copy. Whether its quotations come from the full text or the abstract is not recorded; unverified here.
• Leroy, G. et al. (2013), Journal of Medical Internet Research 15(7):e144, doi:10.2196/jmir.2569. Read by the search online; no saved copy, not re-checked here.
• Agrawal, S. and M. Carpuat (2024), Transactions of the ACL, doi:10.1162/tacl_a_00653, arXiv:2312.10126. Read by the search in the HTML version, Table 2; no saved copy, not re-checked here.
• Szarkowska, A., I. Krejtz, O. Pilipczuk, Ł. Dutka and J.-L. Kruger (2016), Across Languages and Cultures 17(2):183–204, doi:10.1556/084.2016.17.2.3. Read directly (the abstract quoted was checked against a saved copy).
• Burnham, D. et al. (2008), Journal of Deaf Studies and Deaf Education 13(3):391–404, doi:10.1093/deafed/enn003. Abstract only.
• Hurtado, B., L. Jones and F. Burniston (2014), Journal of Intellectual Disability Research 58:822–829, doi:10.1111/jir.12097. Seen by the search and set aside: it compares Easy Read leaflets with baseline knowledge, not with the original text.