Preserve source accounts when editing prose

codex/skills/unslop/SKILL.md

@@ -1,12 +1,28 @@
 ---
 name: unslop
-description: "Remove canned AI-writing patterns from explanations, documentation, PR text, and public prose. Use when a draft is inflated, formulaic, vague, or repetitive, or when the user asks to unslop writing."
+description: "Remove canned patterns without replacing the author. Use when prose is inflated, formulaic, vague, or repetitive, an author's voice has been flattened by editing, or the user asks to unslop writing."
 ---
 
 # Unslop
 
-Edit the writing to remove AI patterns. Preserve its meaning and intended tone;
-do not invent opinions, anecdotes, or personality to make it seem human.
+Edit the account, not the author. Smoothness alone is not an improvement;
+do not invent opinions, anecdotes, or personality to make writing seem human.
+
+## Keep the source in view
+
+For authored prose, work beside the original account and latest corrections,
+not only the last edited draft. Meaning includes objections, mixed feelings,
+qualifications, and cadence that carries thought. Start with the author's words,
+not an imitation of their personality. Understand repetition before cutting it.
+
+When reviewing a regression, distinguish deleted thought, unused or missing
+source material, and useful corrections. History locates changes; it does not prove
+who wrote them or make older wording authoritative. Research can check a
+mechanism, not supply a motive. Respect later corrections and disclosure limits.
+
+Make substantial lost thoughts and new first-person claims visible in brief
+editorial notes, with their source or the reason for the cut. Do not narrate
+routine cleanup. An unchanged passage is a valid result.
 
 ## Patterns to notice
 

codex/skills/unslop/references/sources.md

@@ -27,6 +27,105 @@ They do not establish that every older rule is redundant or that this particular
 catalog improves every draft. The accompanying comparisons are limited editing
 checks, not a general quality benchmark.
 
+## Source-faithful editing
+
+Primary research checked September 15, 2026. These studies examine different
+models, tasks, populations, and outcomes. Their findings motivate checking for
+loss, not assuming every edit degrades a document. The editorial applications
+below are Compass design judgments, not interventions validated by the papers.
+
+### Content retention is not author retention
+
+Zhivar Sourati et al. (2026), [The shrinking landscape of linguistic diversity
+in the age of large language models](https://doi.org/10.1038/s41562-026-02550-0),
+*Nature Human Behaviour*, published August 24. The publisher's abstract reports
+three studies, seven datasets, and over 880,000 texts. Rewriting retained core
+content while homogenizing style and altering author-associated signals. Its
+21 to 50 percent figure measures reductions in writing-complexity variance,
+not information lost per edit. The full publisher text is subscription-only;
+this note relies on its accessible abstract, not an independent methods audit.
+
+Application: review judgments, qualifications, and expression alongside factual
+coverage. Do not infer a person's identity from a classifier or a writing sample.
+
+### Effects depend on the assistance
+
+Vishakh Padmakumar and He He (2024), [Does Writing with Language Models Reduce
+Content Diversity?](https://arxiv.org/html/2309.05196v2), ICLR 2024.
+[Venue record](https://openreview.net/forum?id=Feiz5HtCD0). The controlled
+argumentative-writing comparison found reduced lexical and key-point diversity
+with InstructGPT, but not the same significant reduction with base GPT-3.
+Model-contributed text largely explained the difference. These older model
+conditions do not establish the behavior of every current editor.
+
+Application: compare actual edits with their sources and test the current model;
+do not turn an observed risk into an unconditional ban on assistance.
+
+### Better ratings can coexist with less distinctiveness
+
+Anil R. Doshi and Oliver P. Hauser (2024), [Generative AI enhances individual
+creativity but reduces the collective diversity of novel
+content](https://doi.org/10.1126/sciadv.adn5290), *Science Advances*.
+[Open article](https://pmc.ncbi.nlm.nih.gov/articles/PMC11244532/).
+Access to AI story ideas improved evaluations of individual stories while
+making stories more similar to each other. The task was eight-sentence fiction
+with supplied ideas, not repeated editing of personal essays.
+
+Application: a polished result is not sufficient evidence that an author's
+account survived. This is also a counterexample to claiming assistance can
+only make writing worse.
+
+### Normalization can change culturally situated expression
+
+Dhruv Agarwal, Mor Naaman, and Aditya Vashistha (2025), [AI Suggestions Homogenize
+Writing Toward Western Styles and Diminish Cultural
+Nuances](https://doi.org/10.1145/3706598.3713564), CHI 2025.
+[Author text, version 3](https://arxiv.org/html/2409.11360v3).
+In English-language writing tasks with 118 participants from India and the
+United States, suggestions shifted Indian participants' writing toward Western
+styles. This is evidence about the tested populations and setting, not a basis
+for inferring any particular author's culture or preferences.
+
+Application: do not assume a more direct, informal, or professional register is
+neutral. Let the author's purpose and supplied language determine the register.
+
+### Examples are not a reliable substitute for an account
+
+Zhengxiang Wang et al. (2025), [Catch Me If You Can? Not Yet: LLMs Still Struggle
+to Imitate the Implicit Writing Styles of Everyday
+Authors](https://arxiv.org/html/2509.14543v1), arXiv version 1, September 18.
+The [arXiv record](https://arxiv.org/abs/2509.14543) lists EMNLP 2025 Findings.
+The tested models approximated structured writing more successfully than
+informal blogs and forums; adding demonstrations had limited benefits.
+Evaluation used computational proxies rather than large-scale human judgments.
+These are generation-from-examples results, not a test of source-faithful editing.
+
+Application: edit the supplied wording before attempting to reconstruct a
+personality from style examples. Keep examples as references, not slang templates.
+
+### What the website review contributed
+
+The September 2026 review of ariobarin.com supplied local, inspectable editing
+failures rather than a new universal writing theory. In the September 8
+[DocBot diff](https://github.com/ariobarin/ariobarin.com/commit/1fd0664ab45204139e8f72d7b15076d1044fff64), an
+objection to a giant, immediately stale generated README became a generic
+maintenance requirement. The January 28
+[Pollinator revision](https://github.com/ariobarin/ariobarin.com/commit/cbace4c092b69c1786fe4b3e01a19a3f2b545614)
+added gratitude and pride absent from the replaced passage. These diffs locate
+textual changes; they do not establish who composed the sentences.
+
+The resulting distinction is between deleting thought, leaving source material
+unused, having too little source material, and making a legitimate correction.
+They require different responses. Older writing can be wrong, withdrawn, or
+no longer suitable for publication. Project-specific accounts and disclosure
+rules belong in the website, not this portable skill. No private transcripts
+or conversation logs are included here.
+
+Repeated editing without returning to fixed sources is a plausible mechanism
+for accumulating loss, not a measured law established by these papers. Returning
+to sources can also restore information. Revisit the skill after material model
+changes; keep it only where it changes real editing decisions.
+
 ## Upstream license
 
 MIT License