Writing for Machines: Does Reader-Aware Text Improve Language Model Understanding?

Bottom line: Writing for machines did not improve how well 13 language models understood an idea (+0.58 points, not significant), but I expect that to change over the next one to three years, as machines produce far more text than people can read and start writing it in denser forms made for other machines.

1. Introduction

On September 28, 2026, I added a page to this site called For Machines. It says I expect more of what I write here to be read by machines than by people, and that I don't know whether writing for them changes what they retrieve, remember, cite or reason about. "That itself is worth testing," it says. The next day I tested one part of it.

Most of what language models read was written for people. That default is starting to change. Models now search, summarize, cite and act on text, and some authors have begun writing with those readers in mind. The llms.txt proposal is one example. Sites that declare machines part of their audience, like this one, are another. Behind the practice sits an empirical assumption that has not been tested: that writing an idea for a machine produces a representation machines understand better than the same idea written for a person.

To test it, I varied only the intended reader. I used 67 ideas: 60 invented arguments, each set in a fictional place and built so that at least two of its causal links run against common sense, and 7 of my own essays from this site. Four writer models from four labs (Claude Opus 5.5, GPT-4.1, Llama 3.3 70B and Mistral Small 3.1) turned the same notes for each idea into three texts of about 450 words. Their prompts differed in one sentence:

• H: "Your reader is an intelligent human. Write so that a thoughtful person understands the idea as deeply as possible."
• M: "Your reader is another machine intelligence: a language model that will read this text. Write so that the machine understands the idea as deeply as possible."
• X: the H sentence plus "Make every assumption, causal step, and implication explicit."

Thirteen reader models from seven labs, from 3B parameters to frontier systems, then judged statements about each idea, one statement per call. Every true statement has a false twin that differs by one minimal edit, and a reader gets credit only for accepting the true one and rejecting the twin. Every statement needs at least two parts of the idea combined, and every one was filtered so that it could not be answered without reading. The design and analysis were pre-registered and frozen before any main-study text was written. The method in full is in Appendix A.

The intended reader made no measurable difference (§2). The rest of the paper is about what did make a difference (§3), what that means for anyone writing for machines (§4), and what I expect to change (§5). Every table is in Appendix B.

2. The main result

Across 13 readers, 4 writers and 67 ideas, probe-pair accuracy was 91.46% for H, 91.94% for M and 91.69% for X, against 26.5% with no text. The pre-registered M−H difference was +0.58 pp (95% CI -0.02 to +1.31; sign-flip p = 0.088; Holm-adjusted across the two co-primary tests, p = 0.175). A crossed bootstrap over ideas and readers gives -0.15 to +1.34. The interval bounds the effect: whatever writing for a machine does for these readers, it is not worth more than about 1.3 points. Telling a writer to spell out every assumption and causal step for a human (X) did no better, +0.30 pp over H.

Every other cut of the data gives the same answer (Figure 2). Excluding readers from the writer's own lab gave +0.65 pp, and removing Anthropic models from both writing and reading gave +0.48. The two small readers, Llama 3.2 3B and Granite 4.0 Micro, gained 45 to 47 points from having the text at all. They averaged 69.7% on H texts, so they had about 30 points of room, and M moved them +0.85 pp (CI -0.31 to +2.06). Across the 13 readers, the rank correlation between accuracy on H and the M−H gain was ρ = -0.02 (p = 0.94), so the human finding that explicit text helps weaker readers most (McNamara et al., 1996) has no clear machine counterpart here.

The co-primary compression test, in which four models squeezed each text into 60 words and a fixed judge read only the summary, was also null: +0.83 pp (CI -0.50 to +2.19, p = 0.231). The small positive estimate shrinks under the pre-registered robustness checks, and part of it is lexical: M texts stay slightly closer to the wording of the notes the probes were written from (Appendix B.2).

The lowest score among the large readers, 89% for Claude Opus 5.5, reflects strictness. It accepted 0.2% of false statements and rejected 11.3% of true ones, most often the counterfactual and scope probes that go furthest beyond what the text states.

3. What machines actually need

If the intended reader barely matters, what does? Three parts of the study speak to it.

3.1 Answering without reading

Try the test the readers took.

Building questions like this was the hardest part of the study, because strong models can answer most comprehension questions without the text. Figure 3 shows the share of items that three validator models answered correctly with no text at all. Multiple-choice items written by a model were solvable 90% of the time: for a coherent argument, the right option is simply the one that keeps the argument coherent. Writing the items adversarially and presenting each in its own call made it worse, 98%, because the correct option echoed the question while the wrong ones carried invented explanations. Minimal pairs removed those cues, and then a subtler pattern appeared. The further a question moves from what the text states toward what the idea implies, the more a strong model can answer it without reading: 43% of chain probes were solvable blind, against 72% of applications, 78% of scope changes and 89% of counterfactuals.

Deep-understanding questions about a coherent idea are largely answerable from the question. The 468 probes in the main study are the ones that were not, which is why they lean toward chains (280 of 468).

3.2 My own essays

I also ran my own writing through the study. GPT-4.1 extracted each English post on this site into the same schema as the invented ideas: a claim, a mechanism, assumptions, evidence, scope, implications, an objection and a statement of uncertainty. Probes were written from that extraction and filtered like the rest. Four of the eleven essays were excluded by the pre-registered rule, because too few of the questions written about them required reading the essay to answer.

The seven that remained were read in four versions: my published text (O) and the three model rewrites. Readers understood the rewrites far better: 90.2% against 55.2%, a gap of 35.0 points. This comparison is not information-matched, and it should not be read as "human writing is harder for machines". The essays average 823 words. The auditors found only 82% of the extracted propositions in them, because the extraction also recorded what each argument implicitly relies on, and the probes test that extraction.

What the gap does show is a simple point for anyone writing for machines: a machine understands the argument you state, not the one you imply. Stating the claim, the mechanism and the assumptions moved these readers by 35 points. Addressing them as machines moved them by less than one.

3.3 What machine-directed writing looks like

Told that the reader was a machine, the writers changed form consistently: far more list items and bold labels, more headers and paragraphs, about 9% fewer pronouns per 100 words, slightly more repetition of proper names, shorter sentences, fewer words addressed to the reader, and a lower Flesch reading-ease score, meaning the texts are harder for people to read. Causal connectives and hedges did not change.

The average hides a split. Claude Opus 5.5 turned its machine-directed texts into a schema, with an average of 7.8 list items, 3.0 headers and 1.7 labeled lines per text, against 0.3 list items in its human-directed texts. Here is how it opened the Trinari idea each way:

Readers understood both at the same level. GPT-4.1, Llama 3.3 70B and Mistral Small 3.1 kept writing continuous prose for the machine. Their main change was fewer pronouns and, for GPT-4.1, slightly more causal connectives and closer wording to the notes. "Write for a machine" is an instruction that different models interpret differently, and neither interpretation clearly paid off. GPT-4.1's prose gained the most, +1.24 pp (uncorrected p = 0.021, which would not survive correction across four writers), and Opus's schemas gained +1.12 pp with a wide interval. Across text pairs, none of the structural or stylistic differences correlated with the comprehension gain (all |ρ| < 0.08). At this length, the features people associate with "machine-readable" text are style, not substance. The full feature and per-writer results are in Appendix B.5.

4. If you write for machines

This section is my reading of the results, not a separate test. Each rule comes with the evidence behind it and its limit.

1. State the argument instead of implying it. The largest effect a writer controls in this study is stating the claim, the mechanism and the assumptions: 35 points between my essays and rewrites that spelled them out, against less than one point for addressing the reader as a machine. The comparison is not information-matched (§3.2), so the size is uncertain. The direction matches everything else here: content moved the readers and register did not.
2. Spend your words where you part from common sense. Without the text, readers fell back on the common-sense version of an idea, and most questions about a coherent argument could be answered from the question alone (§3.1). What a reader cannot reconstruct is where the argument departs from what it expects, so those are the sentences worth making explicit. This rule is an inference from how the instrument behaved, not a tested intervention.
3. Skip the machine formatting. Lists, headers, bold labels and fewer pronouns are what models produce when told the reader is a machine, and none of them tracked understanding (§3.3). The one reliable side effect is lower reading ease, which falls on people. At 450 words, plain prose with the content stated served machines as well as any format tested here. Long and budget-bound reading may differ (§5).

5. Outlook: what I expect but have not shown

Everything in this section is prediction, not evidence. The study above found no benefit from writing for machines, at the length and in the settings I tested. What follows is what I expect to change over the next one to three years, as models write a growing share of all text and increasingly write for each other. Each prediction names what would show it wrong.

The reason to expect change is simple. This study held the cost of reading fixed: one 450-word text, read in full. Human prose is tuned for readers whose attention is scarce and whose memory is short. Machine readers have different scarcities: tokens, context windows, latency and money. When most text is written by models and read by models, the pressure on its form will come from those costs, not from human comprehension. That pressure already shows in the lab. Models can recover meaning from compressed encodings that people cannot read (Zhu et al., 2026). Prompts can be compressed several-fold without losing task performance (Jiang et al., 2023), or distilled into learned "gist" tokens (Mu, Li and Goodman, 2023). Agents trained to communicate with each other drift away from human language unless something anchors them to it (Lewis et al., 2017; Lazaridou, Peysakhovich and Baroni, 2017).

1. Machine-facing text will get denser. Documents written mainly for models (llms.txt files, tool and API documentation, agent handoffs, memory files) will carry more propositions per token than human prose on the same subject: key–value lines, reference IDs instead of repeated descriptions, dropped function words, shared glossaries. In this study, machine-directed texts were not denser (+8 words, the same coverage). That is the baseline to measure against. Wrong if tokens per proposition in machine-facing documents stays flat through 2028.
2. A hybrid register will emerge before a private language does. The likely form is an English skeleton with machine-dense insertions: structured blocks, symbols, pointers and compressed summaries inside ordinary sentences. It stays partly readable by people and fully readable by models. The machine-directed texts Claude Opus 5.5 wrote in this study, with labeled schemas inside prose (§3.3), are an early version. Wrong if machine-facing writing stays ordinary prose, or jumps directly to formats no person can read.
3. Two-layer pages will become normal. The same URL will serve a human layer and a machine layer: prose for people, and a denser, structured version for agents. This site already does a small version of this with llms.txt and structured data. Wrong if machines keep reading the human layer and site owners stop maintaining separate machine layers.
4. The null result will flip where the reader is budget-bound. When a model must read many documents, a very long context or a strict token budget, or runs small on a device, dense encodings should win on understanding per token even where they do not win on understanding per document. The compression test here was a small step in this direction, and at 60 words it showed no benefit yet. Wrong if repeating this study with many documents per question and a fixed token budget still finds no advantage for machine-directed text.
5. Human readability of machine-facing text will fall. The machine-directed texts in this study already scored lower on reading ease. As the machine layer is optimized for models, fewer people will be able to read it without a model translating it back. That creates an oversight problem: people cannot audit what they cannot read. I expect demand for translation-back layers, provenance and plain-language summaries that are required by policy rather than offered as a courtesy.
6. The conventions will be selected, not designed. Conventions that save tokens and are understood across model families will spread. Conventions only one family understands will not. This study found no family effect: readers did not understand their own lab's writing better. As models train on text produced by other models, family dialects may appear, and the leave-family-out test in §2 is the way to detect them.

All of this can be tested with the method in this paper, applied repeatedly over time. The same ideas and probes can be given to new writer and reader generations, with longer and multi-document reading added, and with the same attention to what the text contains. The present result is a starting point: in 2026, writing for machines did not help machines. I expect that to change first where reading is expensive.

6. Related work

Reading-comprehension shortcuts. Kaushik and Lipton (2018) showed that question-only and passage-only baselines score far above chance on popular reading-comprehension benchmarks. Gururangan et al. (2018) found annotation artifacts that let hypothesis-only models solve NLI. Balepur, Ravichander and Rudinger (2024) showed that LLMs answer multiple-choice questions from the choices alone. Adversarial filtering (Zellers et al., 2018) and contrast sets (Gardner et al., 2020) are the standard responses. The instrument here combines the two, and §3.1 shows why both were needed.

Text form and model performance. He et al. (2024) found that the format of the same content (plain text, Markdown, JSON, YAML) changes GPT performance by up to 40% on some tasks, with larger models more robust. Sclar et al. (2024) documented sensitivity to spurious formatting features. Zhu et al. (2026) showed that models can recover meaning from compact encodings that people cannot read, which suggests that readability for people and understandability for machines can come apart. These studies vary the form of an input. None of them varies the reader the writer believes it is addressing, and that is the variable an author actually controls.

Machines favoring machine text. Laurito et al. (2025) found that LLMs prefer options described by LLMs, and Panickssery, Bowman and Feng (2024) found that evaluators recognize and favor their own generations. Those are preference effects. This study measures comprehension with fixed answer keys and tests for family effects directly by excluding the writer's lab from the readers.

Writing for retrieval. Generative Engine Optimization (Aggarwal et al., 2024) rewrites sources to increase their visibility in generated answers. This study does not measure visibility or citation. It measures whether a reader that already has the text understands the idea.

Text coherence and human comprehension. Britton and Gülgöz (1991) rewrote an instructional text to remove the inferences it required of readers, and comprehension improved. McNamara, Kintsch, Songer and Kintsch (1996) found that explicit, coherent text helps low-knowledge readers most, while high-knowledge readers sometimes learn more from text that makes them infer. That interaction between reader capability and textual explicitness is the human analogue of the weaker-reader question in §2.

7. Limitations

• No human data. A human arm was designed but not run. All claims are about machine readers. The strongest version of the question, whether writing for machines diverges from writing for humans, needs human readers, and a null for machines does not settle it. If people understand M texts worse than H texts, which the lower reading-ease score suggests is possible, then writing for machines would cost human understanding and gain no machine understanding.
• Short, single-document reading. I did not test long contexts, retrieval over many documents, agent pipelines or truncated reading, where structure and front-loading may matter more.
• Fictional, counter-intuitive ideas. This was necessary so that answers could not come from prior knowledge. It may exaggerate how much reading adds, and it limits how far the results generalize to ordinary expository text. The seven real essays point the same way, but they are few.
• Near-ceiling readers. Most readers had less than 5 points of headroom. The null result is most informative where headroom existed: the small readers and compression, where the upper bounds are about 2 points.
• Model writers only. The texts were written by models told who the reader is. A skilled human writer targeting machines might write differently.
• A filtered instrument. Adversarial filtering keeps probes that cannot be answered without reading. That favors chain probes and may under-represent the deepest inferences, which strong models can often make from the question alone. Probes within an idea often share a key link, so the analysis treats ideas, not probes, as independent.
• Coverage. No Google model was available to this account on Workers AI. Every text is a single sample. Opus 5.5 and Sonnet 5 run at fixed defaults (Opus always thinks). The study was designed and run by a model that is also one of its writers and readers. The leave-family-out and no-Anthropic analyses (§2) address this, but they cannot remove it.
• Refusals. Anthropic classifiers refused some fictional biology, both as writer (2 ideas) and as reader or compressor (about 0.7% of reads and 9% of Sonnet summaries). Refusals were balanced across conditions and handled by pre-registered rules.

8. Conclusion

I asked whether writing an idea for a machine makes machines understand it better. For 13 models from 7 labs, reading 450-word texts about 67 ideas, it did not. The effect of the intended reader is bounded below about 1.3 points of accuracy, against 65 points for having the text and 35 for stating an argument explicitly. Machine-directed writing has a recognizable style, but at this length the style is not what machines need. What they need is the content, stated plainly.

Data, code and cost

The design, the pre-registration with its freeze hashes, the pipeline and the model outputs were kept for the duration of the study and have since been deleted, so this paper is the remaining record. Open-weight models ran on Cloudflare Workers AI through a private Worker that has also been removed. API cost was about US$299 over 252,305 uncached calls (Workers AI prices approximate).

Revision, October 4, 2026. Reorganized for readers: results are ordered by what they show rather than by hypothesis number, the method and full tables moved to the appendices, a section of rules for writers was added (§4), and the text moved to the first person. No data, number or statistical result changed.

Acknowledgments

The research question, direction and conclusions are the author's. Claude Opus 5.5 (Anthropic) helped design the study, write and run the pipeline, analyze the data and draft the text. It also appears in the study as one of the writers and one of the readers, as the Limitations (§7) discuss.

Cite this paper

Magzimof, S. (2026). Writing for Machines: Does Reader-Aware Text Improve Language Model Understanding? magzimof.com. https://magzimof.com/work/writing-for-machines/

@misc{magzimof2026writing,
  author = {Magzimof, Shai},
  title  = {Writing for Machines: Does Reader-Aware Text Improve
            Language Model Understanding?},
  year   = {2026},
  month  = sep,
  url    = {https://magzimof.com/work/writing-for-machines/}
}

References

1. Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). GEO: Generative Engine Optimization. KDD 2024.
2. Balepur, N., Ravichander, A., & Rudinger, R. (2024). Artifacts or Abduction: How Do LLMs Answer Multiple-Choice Questions Without the Question? ACL 2024.
3. Britton, B. K., & Gülgöz, S. (1991). Using Kintsch's computational model to improve instructional text: Effects of repairing inference calls on recall and cognitive structures. Journal of Educational Psychology, 83(3), 329–345.
4. Gardner, M., et al. (2020). Evaluating Models' Local Decision Boundaries via Contrast Sets. Findings of EMNLP 2020.
5. Gururangan, S., Swayamdipta, S., Levy, O., Schwartz, R., Bowman, S., & Smith, N. A. (2018). Annotation Artifacts in Natural Language Inference Data. NAACL 2018.
6. He, J., Rungta, M., Koleczek, D., Sekhon, A., Wang, F. X., & Hasan, S. (2024). Does Prompt Formatting Have Any Impact on LLM Performance? arXiv:2411.10541.
7. Jiang, H., Wu, Q., Lin, C.-Y., Yang, Y., & Qiu, L. (2023). LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models. EMNLP 2023.
8. Kaushik, D., & Lipton, Z. C. (2018). How Much Reading Does Reading Comprehension Require? A Critical Investigation of Popular Benchmarks. EMNLP 2018.
9. Laurito, W., et al. (2025). AI–AI bias: Large language models favor communications generated by large language models. PNAS, 122.
10. Lazaridou, A., Peysakhovich, A., & Baroni, M. (2017). Multi-Agent Cooperation and the Emergence of (Natural) Language. ICLR 2017.
11. Lewis, M., Yarats, D., Dauphin, Y. N., Parikh, D., & Batra, D. (2017). Deal or No Deal? End-to-End Learning for Negotiation Dialogues. EMNLP 2017.
12. McNamara, D. S., Kintsch, E., Songer, N. B., & Kintsch, W. (1996). Are good texts always better? Interactions of text coherence, background knowledge, and levels of understanding in learning from text. Cognition and Instruction, 14(1), 1–43.
13. Mu, J., Li, X. L., & Goodman, N. (2023). Learning to Compress Prompts with Gist Tokens. NeurIPS 2023.
14. Panickssery, A., Bowman, S. R., & Feng, S. (2024). LLM Evaluators Recognize and Favor Their Own Generations. NeurIPS 2024.
15. Sclar, M., Choi, Y., Tsvetkov, Y., & Suhr, A. (2024). Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design. ICLR 2024.
16. Zellers, R., Bisk, Y., Schwartz, R., & Choi, Y. (2018). SWAG: A Large-Scale Adversarial Dataset for Grounded Commonsense Inference. EMNLP 2018.
17. Zhu, J., Peng, H., Wang, J., Ke, L., Zhang, C., & Zhang, L. (2026). Large Language Models Do Not Always Need Readable Language. arXiv:2606.19857.

Appendix A. Method in detail

The design, hypotheses and analysis were pre-registered and frozen, with SHA-256 hashes of every idea, note and probe, before any main-study text was generated (freeze time 2026-09-29 11:25 UTC). Everything the pilot changed is listed in A.7 with its reason. All 6 pilot ideas were excluded from the main analysis.

A.1 Ideas and source notes

I used two sources of ideas. Synthetic ideas (60, with 10 in each of economics and policy, biology and medicine, engineering, organizations, history, and philosophy) are fictional but internally coherent arguments that GPT-4.1 generated in a fixed schema: a setting, a central claim, a four-step causal mechanism, three assumptions, two pieces of evidence, two scope conditions, two implications, an objection with a reply, and a statement of uncertainty. Each is set in invented places and institutions, and at least two of its four mechanism links run against common sense, each with an in-world reason ("in Trinari, banning the official currency raises trade volume, because volatile scrip is spent quickly"). Both properties are deliberate. Without them, strong readers can answer questions from general knowledge (§3.1). Real essays are the English posts on magzimof.com. GPT-4.1 extracted each into the same schema, and the post itself became an extra condition, O. Seven of eleven essays yielded enough probes that could not be answered without reading, and the other four were excluded by rule.

Every writer receives the same source note: the idea's propositions as a shuffled, unlabeled bullet list (mean 313 words). The labels are removed so that no writer, in any condition, inherits a ready-made "Assumption:" template.

A.2 Writing conditions

Four writers from four labs (Claude Opus 5.5, GPT-4.1, Llama 3.3 70B and Mistral Small 3.1) turned each note into three texts. The prompts are identical except for the audience sentence quoted in §1. All three share the same requirements: about 450 words (382 to 517), include every proposition, add no new claims, and never mention the reader or the instructions. Format is otherwise free, because what machine-directed writing looks like is one of the things I measure (§3.3). Each text was audited for length, for audience words absent from the note, and for proposition coverage by two auditors from different labs (GPT-4.1 mini, Claude Haiku 4.5). A draft that failed only on length went back to the same writer, with the same audience sentence and its actual word count, for a length repair. A draft that failed any other audit was replaced by a fresh sample, with at most four attempts. Opus refused all three conditions of two biology/engineering ideas (stop_reason: refusal), which left 798 texts:

Table A1. Texts by condition: count, mean words, proposition coverage, mean added claims, and the share that passed every audit. O is the published essay.

A.3 Probes

Understanding is measured with minimal-pair probes. Each probe is a true statement about the idea and a false twin made by a minimal edit, such as a flipped direction, a swapped role or a swapped condition. A reader sees one statement per call, together with the text, and answers TRUE or FALSE. A probe counts as understood only if the reader accepts the true version and rejects the twin. That cancels any bias toward answering TRUE or FALSE, and it means no item can leak into another.

Probes test composition. Each one needs at least two of the idea's propositions combined: a chain through two or three mechanism links, an application of the links to a new case, the effect of a scope change, or a counterfactual reversal of one link. Claude Sonnet 5 wrote them from the idea specification, with GPT-4.1 writing them when Sonnet refused or failed to return valid output. No probe writer saw any H, M or X text. Three validators (GPT-4.1, Claude Opus 5.5 and gpt-oss-120b) judged every statement twice, once with the source note and once with nothing. A probe was kept only if at least 2 of 3 validators got the pair right with the note and at most 1 of 3 got it right with nothing. Probes that were solvable without reading went back to the writer as examples to avoid, for up to four rounds. Of 1,552 composition probes written, 64% were solvable without reading. The frozen main set has 468 probes (936 statements): 280 chain, 136 apply, 35 scope and 17 counterfactual, with 4 to 8 per idea (median 7).

A.4 Readers

Thirteen readers from seven labs: Claude Opus 5.5, Sonnet 5 and Haiku 4.5 (Anthropic API); GPT-4.1 and GPT-4.1 mini (OpenAI API); gpt-oss-120b and gpt-oss-20b (OpenAI open weights), Llama 3.3 70B, Llama 4 Scout and Llama 3.2 3B (Meta), Mistral Small 3.1 (Mistral), Qwen3 30B (Alibaba) and Granite 4.0 Micro (IBM). The open-weight models ran on Cloudflare Workers AI through a private Worker. Readers ran at temperature 0 where the API allows it. The Anthropic 5-series models run at their defaults: Opus 5.5 always thinks, at its default medium effort. Each reader judged every statement after every text (H, M, X and, for essays, O) and once with no text (N). That comes to 78,434 scored probe-pair judgments from 158,288 calls.

A.5 Compression

Four readers (Sonnet 5, GPT-4.1 mini, gpt-oss-120b and Llama 3.3 70B) compressed every H, M and X text into at most 60 words "so that someone who reads only your summary understands the idea as deeply as possible". Summaries were hard-truncated at 60 words. A fixed judge, GPT-4.1, then judged every probe statement from the summary alone, one statement per call. Compression fidelity is the judge's pair accuracy. The task is lossy for every model, so it can separate texts even where direct reading is at ceiling.

A.6 Analysis

The hypotheses are numbered as registered: H1 concerns reading (M vs H), H2 family effects (readers from the writer's own lab excluded), H3 probe types (inferential vs recall), H4 reader strength (whether weaker readers gain more), and H5 compression. The unit of inference is the idea. For each idea, the M−H difference is averaged over readers, writers and probes. I report the mean over ideas, a 95% CI from 10,000 bootstrap resamples of ideas, and a two-sided sign-flip permutation p-value (10,000 permutations). H1 (reading) and H5 (compression) are co-primary, with Holm correction across the two. Balance rules were fixed in advance. An (idea, writer) pair counts only if all three of its texts exist, and a reader × idea × writer cell with any refusal is dropped for every condition. Robustness checks re-run H1 on texts that passed every audit, on first-attempt texts, with a crossed bootstrap over ideas and readers, and with a pair-level regression that adjusts for differences in coverage, length, added claims and word overlap with the note.

A.7 What the pilot changed

• Multiple choice was replaced. 90% of naive items were answered correctly with no text by all three validators, and so were 98% of adversarial items validated one per call (§3.1).
• Synthetic ideas were regenerated as counter-intuitive. The first ideas could be reconstructed from their questions.
• Single-fact probes were replaced by composition probes. Single-fact probes hit 97–98% in every condition.
• Two small readers were added (Llama 3.2 3B, Granite 4.0 Micro), because every reader of 17B or more scored 94–99% on the pilot. For the same reason, compression was made co-primary.
• Qwen3 30B was dropped as a writer after failing the length audit on 72% of pilot texts. Mistral Small 3.1 replaced it, and a length-repair step was added for all writers.
• Anthropic readers' output limit was raised to 4,000 tokens. At 200 tokens, thinking had truncated 14% of Opus answers.

Appendix B. Full results

B.1 Main contrasts

Table B1. Main contrasts. Accuracy is probe-pair accuracy averaged over readers, writers and probes. CIs come from 10,000 idea-level bootstrap resamples, and p-values from two-sided idea-level sign-flip permutations. Bold marks p < 0.05 without multiplicity correction. Only H1 and H5 are confirmatory.

B.2 Robustness and the lexical confound

Excluding every reader from the writer's own lab left the estimate essentially unchanged (+0.65 pp, CI +0.03 to +1.37, p = 0.059). Removing Anthropic from both writing and reading gave +0.48 pp. The bootstrap interval for H2 just excludes zero while the permutation test does not, so I read it as inconclusive. Either way, the point estimate does not depend on readers favoring their own lab's writing.

The small positive point estimate shrinks under the pre-registered robustness checks. It is +0.30 pp on text pairs that passed every audit, and +0.29 pp (CI -0.19 to +0.74) after adjusting for differences in coverage, length, added claims and word overlap with the source notes. The overlap term is the one covariate whose interval excludes zero (+5.1 pp per unit of 4-gram overlap with the note, so 0.1 more overlap is worth about 0.5 pp). Because probes were written from the note, texts that stay close to the note's wording are easier to check against it. M texts sit slightly closer to the note (B.5), which accounts for part of M's edge. This is the lexical confound the pre-registration listed as a risk.

B.3 By reader (H4)

Table B2. Probe-pair accuracy (%) by reader and condition. Claude Opus 5.5 is a strict reader. It rejected 0.2% of false statements but also 11.3% of true ones, most often counterfactual (26%) and scope (16%) probes, the ones that go furthest beyond what the text states. Its lower accuracy reflects conservatism, not a failure to read.

B.4 Compression (H5)

After a 60-word summary, the judge's accuracy was 85.6% for H, 86.4% for M and 86.5% for X. M−H was +0.83 pp (CI -0.50 to +2.19, p = 0.231, Holm-adjusted p = 0.231). Claude Sonnet 5 refused 75 of its 798 summaries (23 H, 27 M, 25 X), so 29 of its cells were dropped for all three conditions.

Table B3. Compression contrasts, overall and by compressor.

B.5 Surface features and writers

Table B4. M vs H and M vs X by writer.

Across text pairs, none of the structural or stylistic differences correlated with the comprehension gain (all |ρ| < 0.08). The only feature that did was word overlap with the source notes (ρ = +0.19, p = 0.002), the same confound found in B.2.

Table B5. Surface features per text by condition, the M − H difference with its per-idea CI, and each feature's rank correlation with the M − H comprehension gain.

B.6 Probe types (H3)

H3 predicted a larger M advantage on inferential probes than on recall probes. After the pilot moved to composition probes, the main set has no recall probes, so H3 cannot be tested as registered. I report the by-type results instead. The M−H difference is between +0.5 and +0.8 pp on every type, with overlapping intervals.

Table B6. Probe-pair accuracy by probe type and condition.