LLM readers
How text and files should be structured when the reader is a large language model, and how a model that writes for the repository is held to its rules.
Steps: frame · review · state of the art · position · against · baseline · bibliography
Claims
| Claim | Grade | Where |
|---|---|---|
| Longer input lowers accuracy even when the task is unchanged and the added text is irrelevant. | measured | state-of-the-art.md, reading long inputs |
| Irrelevant sentences and passages lower accuracy. | measured | state-of-the-art.md, distraction |
| A fact in the middle of a long input is used less well than at the ends; the size of the effect depends on model and length. | supported with caveat | state-of-the-art.md; against.md |
| Long-context retrieval leans on literal word overlap between question and answer. | measured | state-of-the-art.md |
| Meaning-preserving layout changes move accuracy; the effect is model-specific, smaller in some newer models, and no layout is best for every model. | measured | state-of-the-art.md, format and order |
| Premises ordered as they are used give better reasoning than shuffled ones. | measured | state-of-the-art.md |
| Repeating the question, or the whole prompt, improves accuracy. | supported with caveat | review.md, strand 3 |
| Agents follow instruction files, but the files do not raise task success on average and cost about a fifth more. | supported with caveat | state-of-the-art.md, instruction files |
| Codebase overviews in instruction files do not help agents find the code to change. | supported with caveat | against.md |
| Adherence falls as rules multiply and as a session lengthens. | supported with caveat | state-of-the-art.md |
| llms.txt and AGENTS.md improve model accuracy. | convention, untested | position.md |
| LLM text carries stylistic markers measurable across a corpus. | measured | review.md, strand 5 |
| Detecting LLM text in a single document is unreliable: paraphrase evades detectors, and non-native writing triggers false positives. | measured | state-of-the-art.md, model as writer |
| A model reviewing its own output without external feedback does not reliably correct it; tests and tools do. | measured | state-of-the-art.md; baseline.md |
| Define before use, one name per thing, one place per fact and cutting what does not serve transfer to models. | supported with caveat | position.md |
| Human working-memory limits and visual layout cues do not transfer to models as such. | supported with caveat | against.md |