@3xhaust/oh-my-design 0.16.1

package/core/theory/voice.md

@@ -0,0 +1,595 @@
+# Voice — decision material
+
+Copy reveals whether its language fits the speaker, listener, surface, and task before visual
+polish can rescue it. A forced line break, repeated cadence, or empty hedge can prompt a
+contextual review, but no individual signal establishes authorship. This file records the
+jobs those patterns can obstruct and the repairs that preserve evidence.
+
+---
+
+## Statistical observations are not writing instructions
+
+Generated-text studies have measured aggregate differences in predictability, sentence
+length, and paragraph complexity within particular datasets. Those distributions can
+describe a corpus; they cannot establish who wrote an individual passage or whether it does
+its job. Genre, language, accessibility needs, short UI labels, and technical constraints can
+all produce the same measurements.
+
+Do not manufacture sentence-length variance, surprise, typos, asides, or roughness to imitate
+a distribution. Rhythm follows the speaker's thought and the surface's job. When prose feels
+mechanical, inspect the discourse: repeated propositions, abstract agency, prompt mirroring,
+misplaced caveats, or a register that does not fit the relationship. Repair that cause and
+let sentence shape follow. Detector scores and stylometric thresholds never gate shipping.
+
+---
+
+## How people actually read on the web
+
+The finding that shaped modern web writing came from Jakob Nielsen and John Morkes in 1997.
+In their study of how users approached web pages, 79 percent of test users always scanned
+any new page — only 16 percent read word by word. Twenty-three years later, Nielsen's 2020
+follow-up confirmed the proportion had not moved. Scanning is not a failure mode; it is the
+default mode of a reader whose attention is distributed across millions of competing pages.
+
+The same pair published the implications in "Concise, SCANNABLE, and Objective: How to Write
+for the Web" (Morkes and Nielsen, 1997). Testing five writing styles against a control, they
+found the concise version scored 58% higher in measured usability, the scannable version 47%
+higher, and the objective version (no promotional language) 27% higher. A site that was all
+three simultaneously scored 124% higher. A follow-up rewriting Sun.com pages produced 159%
+improvement. These are not marginal effects; they are the largest usability gains in the
+literature for a single intervention class.
+
+Nielsen's 2006 eye-tracking study of 232 users across thousands of pages documented the
+F-pattern: a full sweep of the first line, a shorter sweep of a second line, then a vertical
+scan down the left edge. The pattern is a failure state — it emerges when content lacks
+visual hierarchy and the reader falls back to a heuristic. The design response is not to
+format text to fit the F; it is to give the reader entry points that break the fallback.
+
+What these studies mean for sentence construction:
+
+**Front-load information.** The first sentence of a paragraph and the first clause of a
+sentence carry disproportionate weight. If the important claim is buried after a preamble,
+the scanner never sees it. The inverted pyramid — conclusion first, support after — is not
+a newspaper convention; it is the writing shape that matches how web users extract value.
+
+**One idea per paragraph.** A paragraph that opens with claim A and closes with claim B is
+two paragraphs. Splitting them is not padding; it is giving each claim the surface area to
+be scanned and absorbed independently.
+
+**Concrete nouns over abstract nominalizations.** "The facilitation of user onboarding" and
+"helping people get started" convey the same claim. The second is scannable; the first
+requires parsing. Abstract nouns — facilitation, optimisation, utilisation — slow processing
+and carry no additional precision. Every nominalization is a candidate for replacement with
+a verb or a specific noun.
+
+---
+
+## Plain language is not dumbing down
+
+Daniel Oppenheimer's 2006 study "Consequences of Erudite Vernacular Utilized Irrespective
+of Necessity" (*Applied Cognitive Psychology*, 20(2), 139–156) ran five experiments with
+Stanford students, manipulating the vocabulary complexity of texts and measuring the authors'
+perceived intelligence. The result reversed the folk assumption: as complexity increased,
+estimated author intelligence decreased. The mechanism is processing fluency — difficult
+text feels effortful to read, and the reader attributes that effort to the writer's
+unclear thinking rather than to the word choice. The effect held regardless of essay quality
+and regardless of readers' prior expectations.
+
+Plain language is the confident register. It is the register of a person who understands
+the subject well enough to explain it without armour. Mailchimp's public content style guide
+(styleguide.mailchimp.com) names this "plainspoken" and defines it as stripping away the
+hyperbolic language, upsells, and over-promises that the reader has learned to distrust —
+not because simplicity is good in itself, but because clarity is what trust is built on.
+
+37signals states the same principle in their communication guide: "If your words can be
+perceived in different ways, they'll be understood in the way which does the most harm."
+Plain writing reduces the number of ways a sentence can be understood. That reduction is a
+quality, not a constraint.
+
+Condition → choice → reason: when a sentence contains a long Latinate noun where a
+shorter verb would serve — "provide an indication of" → "indicate", "make a determination"
+→ "decide", "in the event that" → "if" — replace it. The test is not whether the long form
+is correct but whether the short form loses any precision. In the overwhelming majority of
+cases, it does not.
+
+---
+
+## The voice chart: writing from product principles
+
+Torrey Podmajersky's *Strategic Writing for UX* (O'Reilly, 2019) introduces the voice chart
+as the instrument that connects product principles to copy decisions. The chart holds rows
+for vocabulary, verbosity, grammar, and concepts — filled in for each dimension by asking
+what the product principle demands of the words, not what sounds nice. A financial product
+whose principle is "trustworthy" chooses precise vocabulary over colloquial; a product whose
+principle is "approachable" accepts shorter sentences and native vocabulary over formal
+constructions. Podmajersky's formulation: "These principles define what the experience is
+trying to be to the people who use it. Then, the voice can do its job of conveying those
+product principles with every word."
+
+The voice chart makes copy decisions auditable. Without it, every copy choice is a taste
+decision — and taste decisions are re-litigated on every page, by every person who touches
+the product. With it, the question becomes: does this sentence match the registered
+voice dimensions? That is a question with a testable answer.
+
+Condition → choice → reason: before writing a word of copy, name the product principle
+each section is serving. When no principle can be named, the copy is decoration; decoration
+is the first thing the scanner skips.
+
+---
+
+## The review-mining move
+
+Joanna Wiebe at Copyhackers documented what she calls the voice-of-customer method: instead
+of writing copy on a blank page, mine the language that real users have already written —
+in reviews, support tickets, forum threads, comment sections — and use their phrasing
+directly. The argument is that the customer names the value of the product more precisely
+than the copywriter can, because the customer experienced the problem before the solution
+existed and reaches for the words that match that experience.
+
+The method is a search operation. Identify the exact problem your product solves. Find
+online communities — reviews, subreddits, Help forum threads — where people describe that
+problem in their own words. The phrases that appear repeatedly, with emotional precision,
+are the phrases the copy should use or echo. The user who writes "I used to spend three
+hours a week reconciling this manually" has handed you a sentence. The copy that quotes
+their situation back to them is the copy that converts.
+
+This connects directly to the scout's community captures. A Reddit thread where users argue
+about exactly this product category, a Hacker News comment naming why a redesign failed —
+these are not sentiment data. They are voice data. The scout captures them; the hand uses
+them. A voice study that records how a site writes is the positive model; the community
+captures record the phrases the audience already uses for the problem.
+
+---
+
+## Korean copy: 한국어 카피의 목소리
+
+Korean product copy requires decisions that English-language guidance does not settle.
+Literal transfer can produce translation-shaped prose, so judge the result against its Korean
+speaker, listener, situation, and register rather than inferring its origin.
+
+### Speech-level commitment: choose one and hold it
+
+Two common written registers in Korean product copy are 해요체 (하다 → 해요, 됩니다 → 돼요)
+and 합니다체 (하다 → 합니다, 됩니다 → 됩니다). They can imply different relationships:
+합니다체 often supports institutional precision, while 해요체 often supports conversational
+warmth. Neither is a universal product default; choose from the voice contract, audience,
+situation, and required authority.
+
+Toss committed to 해요체 across every string in its product and published the rule
+explicitly: all copy uses 해요체, without exception, regardless of context or screen.
+Their stated reason is consistency of experience — 합니다체 anywhere breaks the register
+the rest of the product has established. The rule also extends to avoiding over-polite forms
+(~시겠어요?, ~께) that amplify formality past what the relationship calls for.
+
+What the `SLOP-KO-REGISTER-MIX` rule detects — 해요체 and 합니다체 alternating within one
+text node — is a review candidate. It may expose unplanned register drift, or it may be
+required by quoted speech or a deliberate change of speaker. Context decides.
+
+Condition → choice → reason: before writing Korean surface copy, record its speech level in
+the voice contract. Hold it across strings that share one speaker and relationship, including
+applicable error, empty, and action states. When the register changes, identify the changed
+speaker or situation rather than treating inconsistency as personality.
+
+### Vocabulary temperature: 고유어 over 한자어 where precision allows
+
+Korean has two vocabulary strata available for almost any concept: native Korean (고유어)
+and Sino-Korean (한자어). Most product copy defaults to Sino-Korean because it feels
+professional and precise. This instinct produces text that is correct, stiff, and forgettable.
+
+Toss's 8 writing principles include "Easy to speak" — their test is whether the sentence
+sounds wrong when said aloud, not whether it looks wrong in print. They cite 한자어 and
+literary 문어체 constructions as the primary cause of copy that passes a proofreading
+read but sounds like documentation. Examples from their practice: 송금이 완료되었습니다
+becomes 5000원을 보냈어요 — "송금" (Sino-Korean for remittance) replaced by "보내다" (native
+verb for send), "완료되었습니다" (formal Sino-Korean completion) replaced by "보냈어요"
+(native past tense in 해요체). The semantic content is identical; the temperature is
+completely different.
+
+The working vocabulary test: say the sentence aloud. Does it sound like something a person
+would say to a friend explaining what just happened? If not, identify the Sino-Korean
+construction and replace it with the native equivalent. "결제 진행" → "결제 중". "이용
+가능합니다" → "쓸 수 있어요". "확인 후 진행하세요" → "확인하고 다음으로 가세요". Each
+substitution costs no precision and removes the institutional chill that Sino-Korean carries.
+
+### One sentence, one useful job
+
+Toss's principle 4 — Focus on key message — is a useful test for compact product surfaces:
+one short UI string usually carries one primary job that can be read in a breath. Korean's
+agglutinative structure tempts long sentences because each extension costs only a morpheme,
+and the grammatical seams are invisible in a way that English subordination is not. The
+more jobs a compact string carries, the harder it becomes to scan.
+
+Use breath as contextual evidence, not a universal sentence rule. A legal notice, tutorial,
+or technical explanation may need a longer sentence. When a compact surface combines
+unrelated jobs, split at the change; retain a connective when it carries a real logical
+relationship that juxtaposition would lose.
+
+### What Korean products actually sound like
+
+Successful Korean consumer products such as Toss, 당근, and Kakao show that consistency is
+not the same as informality. Their compact action and state copy is precise within its chosen
+register. Heroes often lead with the user's change; documentation, comparisons, and feature
+details may explain mechanism when mechanism is the useful fact.
+
+배민 (Baemin) operates in a distinct creative register — warm, witty, deliberately playful —
+but the underlying commitment is the same: a consistent voice that was designed and held,
+not allowed to drift. Their UX writing interview at bcut.baemin.com describes the first
+UX writer's mandate as "ensuring the text inside the app communicates clearly with users
+and protecting consistency." The register is brand-specific; the structural commitment is
+universal.
+
+### Document-structure narration: write the content, not that content is coming
+
+Korean product copy can lose its job when it narrates the document before presenting it.
+"아래는 그 기록이에요" (below is the record), "다음은 기능 목록입니다"
+(the following is the feature list) — these sentences announce the organisation of what
+follows rather than delivering it. A tutorial or long document may need signposting; a small
+product surface usually benefits from starting with the fact or action itself.
+
+`SLOP-KO-SIGNPOST` fires on the deterministic patterns: `아래는 … 기록/내용/목록/정리` and
+`다음은 … 입니다/이에요`. The fix is deletion: remove the announcing sentence and start with
+the content it was announcing. "아래는 그 기록이에요." → delete it; begin with the first item.
+
+The negative to hold in mind: "아래 버튼을 누르세요" does not fire. 아래 used as a spatial
+adverb (without the topic marker 는) is navigation copy, not structural narration.
+
+### Calibration evidence: what the measured gap between human and pipeline prose reveals
+
+The following lessons come from a direct calibration: a 16,000-character Korean personal tech
+essay by Evan Moon (evan-moon.github.io/2020/12/29/2020-retrospective/, published December 2020,
+fully human-authored) was used as the baseline. A parallel version covering the same facts was
+written by following voice.md and the humanize skill rules faithfully. The two versions were
+measured against the same metrics. The gaps name what our rules actively get wrong.
+
+**Measured diff table (N=26–36 sentences per version, character lengths exclude spaces):**
+
+| Metric | Human essay | Our version |
+|---|---|---|
+| Sentence length min | 8 chars | 8 chars |
+| Sentence length max | 114 chars | 48 chars |
+| Sentence length mean | 50.4 chars | 19.9 chars |
+| Sentence length σ | 26.1 | 9.8 |
+| Ending diversity (unique/total) | 0.86 | 0.96 |
+| Connective rate (per sentence) | 0.42 | 0.00 |
+| C-11 violations (connective+comma) | 1 | 0 |
+| Parenthetical asides | 6 in full article | 0 |
+| Unresolved uncertainty markers | 5 | 0 |
+
+**Lesson 1 — Register context comes before a speech-level default**
+
+The sampled human essay uses the informal narrative register (-다 style) throughout: 해였다,
+것 같다, 생각이었다, 모르겠다. None of these are 해요체 or 합니다체. Korean personal essays,
+회고글, and reflective blog posts are written in this register — it is the written voice of a
+person narrating their own experience, not the addressed voice of a product speaking to a user.
+
+Applying 해요체 to every Korean genre can produce a newsletter register: warm,
+present-tense, and second-person-inflected, which may fit one UI voice but not a formal
+service, first-person retrospective, or developer note.
+
+Condition → choice → reason: identify speaker, listener, situation, genre, and authority,
+then choose 해요체, 합니다체, -다 style, or another evidenced register in the voice contract.
+Product copy has no universal speech-level exception to that decision.
+
+**Lesson 2 — Connective overcorrection: the tell is the comma, not the connective**
+
+The sampled human essay uses connectives at a rate of 0.42 per sentence — 15 connectives
+across 36 sentences, including 하지만, 그래서, 그런데, 물론, 또한, 사실, 결국. Not one
+is followed by a comma (C-11 count: 1, a naturally-occurring case). The human connective
+rate is high because connectives carry logical structure; they are not AI tells by themselves.
+
+The pipeline version, shaped by the SKILL.md rule against "connector abuse," produced zero
+connectives across 26 sentences. That is not more human — it is a different kind of abnormal.
+Abrupt juxtaposition without connectives reads as choppy, clipped, produced by a sentence-by-
+sentence generator rather than a mind that knows where the paragraph is going.
+
+What the C-11 rule identifies is the narrow pattern of a connective immediately followed by
+a comma — 하지만, / 그러나, / 또한, / 따라서, — for contextual review. The rule does not
+say that the author is known or that all connectives should be removed.
+
+Condition → choice → reason: clustering and punctuation can make connective use feel
+mechanical, while one connective may mark a necessary logical turn. Ask whether each opening
+expresses a real relationship. Repair redundant structure, not the words as a class.
+
+**Lesson 3 — Parenthetical asides: a positive model our rules never named**
+
+The sampled essay contains six parenthetical asides across 16,000 characters — roughly one per
+1,000 characters, or once every three to four paragraphs: "(살면서 노래방을 이렇게 안 가본 게
+처음…)", "(결국 이러다가 번아웃이 왔었지만, 습관은 잘 안 고쳐지더라…)", "(그리고 스스로 뭔가
+꼰대 같다는 생각이 들어서 이걸 깨닫고 한동안 우울했었다)". These are self-interruptions —
+second thoughts, humor, and corrections that carry information the main sentence cannot.
+
+The pipeline produces zero. The tells list names what to remove; it has no positive model for
+this form. An essay with no parenthetical asides in 3,000+ characters is measurably flatter
+than the human baseline.
+
+Condition → choice → reason: in essayistic Korean prose, preserve a supplied second thought
+when it changes the meaning or carries honest uncertainty. Do not add an aside merely because
+the passage lacks one; a manufactured interruption is no more natural than a manufactured
+cadence.
+
+**Lesson 4 — σ and unresolved uncertainty: preserve what remains unresolved**
+
+The human essay's sentence-length σ was 26.1; the pipeline's was 9.8. The maximum sentence
+length in the human essay was 114 characters (no spaces); the pipeline capped at 48. The old
+burstiness instruction was the wrong abstraction. The human passage follows thought instead
+of managing a sentence-length pattern; its measurements describe one sample and are not a
+recipe to reproduce.
+
+The sampled human essay closed its mentoring section with: "하지만 아직도 잘 모르겠다." Thirteen
+characters. The pipeline produced: "이게 올해가 남긴 가장 큰 질문이에요." That is a packaged
+takeaway. It names the lesson; it does not live in the uncertainty. The human's 13-character
+sentence is more trustworthy precisely because it refuses to resolve what has not been resolved.
+
+The full article closed with: "2020년에는 결국 이렇게 답을 찾지 못한 질문으로 회고를 마무리하지만,
+내년 2021년 회고에는 이 질문과 고민에 대한 답을 찾아서 회고에 적는 것을 목표로 잡아야겠다." The
+human commits to finding the answer next year rather than having found it this year.
+
+Condition → choice → reason: when an essay section ends, ask whether a tidy conclusion is
+honest or performed. A genuine takeaway — something the writer actually learned and can state
+— earns the closure. A resolution that smoothes over ongoing uncertainty changes the claim.
+If the answer does not exist yet, retain that uncertainty instead of packaging a lesson.
+
+### Uniform sentence endings and not-X-but-Y: pre-handoff checks, not linter rules
+
+Two Korean patterns require contextual pre-handoff review and cannot become safe
+deterministic rules:
+
+**Three or more consecutive identical sentence-final endings** (이에요./이에요./이에요. or
+합니다./합니다./합니다.) may make prose feel mechanically parallel in context. However,
+product benefit lists deliberately use parallel endings: "빠릅니다.
+정확합니다. 쉽습니다." is valid copy that would false-positive on every legitimate feature
+section. A linter rule would fire on too much legitimate writing to be trusted.
+
+Pre-handoff check: if adjacent sentences share an ending, ask whether the ideas are truly
+parallel. Keep the form when it supports comparison or recall. If sameness hides a change in
+communicative job, repair that sentence's job rather than forcing a different ending.
+
+**The not-X-but-Y construction** (`~게 아니라 ~했어요` / `~이 아니라 ~입니다`) can become
+performed contrast when no real distinction supports it. It also appears in legitimate copy:
+a brand that deliberately contrasts itself with a
+category norm ("빠른 게 아니라 정확합니다") uses this construction with intent. No narrow
+regex distinguishes the cases reliably.
+
+Pre-handoff check: when this construction appears in the hero or opening paragraph, ask
+whether the contrast is a designed claim (cite it in decisions.md with the brand reason)
+or a hedge that weakens the useful claim. Delete empty hedges; keep designed contrasts. "AI가 만든
+게 아니라, 사람이 썼어요" as a self-description of the tool is a hedge; "비싼 게 아니라
+오래 씁니다" as a product claim is a designed contrast.
+
+Neither pattern has an `omd check` rule. Both belong on the hand's pre-handoff review pass.
+
+Translated English marketing is the loudest failure mode. The sentence structure of
+English product copy — subject, action verb, object, benefit clause — does not map cleanly
+onto Korean's SOV order or its topic-comment preference. Copy that was conceived in English
+and translated reads as translated: the particle choices are safe rather than native, the
+nominalisations are abundant, and the rhythm is wrong in a way that fluent readers feel
+before they can name it. The correct direction is to write the Korean sentence a Korean
+product would have written first, not to translate the English sentence that would have
+appeared on an English-language version of the same screen.
+
+---
+
+## Register: diagnose fit, not origin
+
+Web, product, and assistant copy can drift into a medium-formal register that sounds like a
+support FAQ edited by a committee. The useful question is not who wrote it, but whether that
+register fits this speaker, listener, situation, and surface.
+
+A 2025 study that fine-tuned Llama, Qwen, and Mistral-Nemo models toward natural
+responses (arxiv 2501.05032v1, "Enhancing Human-Like Responses in Large Language Models")
+identified the primary register failures: self-referential disclaimers, impersonal
+deflection, over-structured formatting, and "formal and impersonal" phrasing that is
+"structured, clear, and precise, but lacks the warmth and spontaneity of natural human
+conversation." The fine-tuned models won blind human ratings at 79–90% preference rates;
+general benchmark accuracy was unchanged. The lesson: formality should be chosen, not
+defaulted into.
+
+A register-aware linguistic evaluation (arxiv 2605.23651, "How Human-Like Are Large
+Language Models? A Register-Aware Linguistic Evaluation Framework") measured the
+statistical gap between LLM output and human writing across five registers, including
+instructive online text — the closest analogue to product and help copy. Consistent
+findings: models over-produce nominalisations, favor longer words, inflate attributive
+adjectives, and stack present-participle clause constructions ("helping teams manage",
+"allowing users to see"). Human instructive copy scores higher on Biber's Overt Expression
+of Persuasion dimension and is more lexically variable. The model's version is less varied
+and less authoritative, not more.
+
+### The self-introduction cadence
+
+In product hero copy, "X is a platform that helps teams Y and Z" can merely restate the
+brief instead of giving the reader a useful change or distinction. The construction is not
+an authorship signal and remains appropriate when a comparison, documentation page, tooltip,
+or feature detail genuinely needs category and mechanism.
+
+Condition → choice → reason: when a hero opens with "[Product] is a [category noun] that
+[verb]s you", ask whether category or mechanism is the useful fact. If not, lead with the
+verified user change or distinction. Keep the mechanism where it answers the reader's real
+question.
+
+The same cadence appears in Korean as "X는 Y를 도와주는 플랫폼이에요" or "X는 Z를 관리할 수
+있는 도구예요". The pattern is identical: the product introduces itself by naming its own
+mechanism. Rewrite from the user's change: "월요일 빌드는 제때 나가요" rather than "X는
+빌드 관리를 도와드리는 플랫폼이에요".
+
+### Reflexive over-politeness
+
+Reflexive politeness such as "Please feel free to", "We'd be happy to help", and "Don't
+hesitate to reach out" can import a helpdesk register into a surface that needs direct
+action. Support contexts may need care and an escalation route; decorative politeness does
+not replace either.
+
+Condition → choice → reason: when the polite clause adds no relationship or safety value,
+write the action directly. Keep specific care or escalation language when the situation
+requires it.
+
+### Uniform formality regardless of product voice
+
+A NAACL 2024 benchmark study (DialogBench, aclanthology.org/2024.naacl-long.341) found
+that assistant-AI instruction tuning can actively weaken emotional naturalness — the
+positioning trains the model toward uniform helpfulness, which erases the register
+variation a specific product voice requires. A fintech product designed to feel like a
+knowledgeable friend and a legal SaaS that wants institutional precision both receive the
+same medium-formal output if the writer does not actively set the dial.
+
+Condition → choice → reason: read the product's voice study before writing a word. Then
+choose a register that is measurably lower or higher than the default formal medium.
+"Lower" is not "chatty" — it means committing to vocabulary the product's actual users
+use, at the warmth level the product's principles demand. The voice chart (Podmajersky)
+is the tool that makes this concrete and auditable.
+
+### Over-nominalization and present-participle clause stacking
+
+The register-aware study (arxiv 2605.23651) identified two structural tells in LLM
+instructive text that appear below the level of word choice: over-nominalization
+(converting verbs to abstract nouns — "the facilitation of onboarding", "the
+optimisation of workflows") and present-participle clause stacking ("a platform for
+managing, tracking, and reviewing your pipeline"). Both are addressed individually in the
+plain-language section of this document. They are named here as register signals too: in
+isolation each is acceptable; clustered in a hero or feature section they can obscure actor,
+action, and consequence. Test whether a direct verb or concrete noun carries the same fact
+more clearly in the chosen register.
+
+### Deterministic rule: declined
+
+The self-introduction pattern is regex-describable:
+`\b\w+ is an? (platform|tool|system|solution|app|service) that (helps?|allows?|enables?|lets?)\b`.
+It is not a safe linter rule. Product comparison tables, developer documentation, feature
+detail sections, and UI tooltip copy all use constructions that match this pattern
+legitimately. Without reliable hero-node detection, the rule would false-positive on too
+much legitimate writing to be trusted. It belongs on the hand's pre-handoff review pass,
+in the same category as the not-X-but-Y construction documented above under Korean copy.
+
+---
+
+## Static copy and live dialogue have different situations
+
+OMD's voice system covers static product writing and conversational replies without
+collapsing them into one register. A page has no shared turn history: its headline, body,
+state, and action must stand on their own. A live reply has a speaker, listener, prior turn,
+and immediate purpose; it should use established context instead of repeating the request or
+narrating the answer before answering. Acknowledgement belongs in dialogue when it repairs a
+relationship, confirms a consequential instruction, or responds to an observed event. It is
+not a generic opening.
+
+The dialogue-system research below concerns timing, overlap, and assistant evaluation. Those
+mechanisms do not transfer to static page copy. The broader requirement that register fit the
+speaker, listener, and situation applies to both surfaces.
+
+**Text Overlap / OverlapBot** (aclanthology.org/2025.sicon-1.10): studies backchannels,
+overlapping responses, and turn-taking timing in real-time text chatbots. The paper
+reports that OverlapBot produces 130% more chatbot turns than a standard system through
+behaviors like backchannel signals and proactive interruption. Backchannels, overlap
+timing, and turn interruption are properties of live dialogue — a landing page does not
+respond, does not take turns, and has no equivalent to "음" or "I see" in a hero section.
+Applying these findings to static copy would be a category error, not humanisation.
+
+**DialogBench** (aclanthology.org/2024.naacl-long.341): a benchmark for evaluating LLMs
+as conversational dialogue systems across 12 tasks. Its findings about emotional
+perception and everyday-life knowledge quality apply to chatbot response evaluation, not
+to page copy. The study's principle — that the assistant-AI positioning degrades register
+naturalness — does transfer as a writing decision (documented above), but the benchmark
+tasks and metrics themselves do not.
+
+**Are Human Conversations Special?** (arxiv 2403.05045): an architectural study of how
+transformer attention mechanisms differ across domains — conversation, web content, code,
+mathematics. It identifies no surface-text patterns and transfers nothing to writing
+decisions. It is cited here to prevent a future reader from importing it: the paper's
+conclusion is about domain specialisation in attention heads, not about copy register.
+
+---
+
+## Sources
+
+- Nielsen, J. (1997). "How Users Read on the Web." Nielsen Norman Group.
+  nngroup.com/articles/how-users-read-on-the-web/ — 79% scan finding; scanning as default
+  web reading behavior, not an aberration
+- Morkes, J. & Nielsen, J. (1997). "Concise, SCANNABLE, and Objective: How to Write for
+  the Web." Nielsen Norman Group. nngroup.com/articles/concise-scannable-and-objective —
+  58% usability gain for concise writing, 47% for scannable, 124% combined
+- Nielsen, J. (2006). "F-Shaped Pattern For Reading Web Content."
+  nngroup.com/articles/f-shaped-pattern-reading-web-content-discovered/ — F-pattern as a
+  fallback behavior, not a design target
+- Oppenheimer, D.M. (2006). "Consequences of Erudite Vernacular Utilized Irrespective of
+  Necessity: Problems with using long words needlessly." *Applied Cognitive Psychology*,
+  20(2), 139–156. doi:10.1002/acp.1178 — plain language increases perceived intelligence
+  via processing fluency; complexity reduces it
+- Feature-based detection of AI-generated text (Springer, *AI and Ethics*, 2025) — human
+  writing has greater sentence-length variance and standard deviation than model-generated
+  text; confirmed across academic, journalistic, and creative domains
+- PMC study on academic science writing (pmc.ncbi.nlm.nih.gov/articles/PMC10328544/) —
+  sentence-level diversity in length and paragraph complexity as primary detection signals;
+  over 99% accuracy using stylometric features alone
+- arXiv 2408.04647, "Distinguishing Chatbot from Human" — GPT sentences longer on average;
+  human writing has wider range and higher variance across sentence lengths
+- Podmajersky, T. (2019). *Strategic Writing for UX.* O'Reilly Media. — voice chart as
+  instrument connecting product principles to copy decisions; voice applied like a spice:
+  too little is flavorless, too much inedible
+- Mailchimp Content Style Guide (styleguide.mailchimp.com) — "plainspoken" as the core
+  voice dimension: strip hyperbolic language, upsells, and over-promises; clarity above
+  entertainment; active voice, positive language, plain English
+- 37signals, "The 37signals Guide to Internal Communication" (37signals.com/how-we-
+  communicate) — "If your words can be perceived in different ways, they'll be understood
+  in the way which does the most harm"; plain writing reduces interpretive surface
+- Wiebe, J. / Copyhackers, "How to do rapid-fire review mining"
+  (copyhackers.com/how-to-do-rapid-fire-review-mining/) — voice-of-customer method:
+  mine real user language from reviews and forums; the customer names the value better
+  than the copywriter
+- Toss Tech. (2022). "토스의 8가지 라이팅 원칙들."
+  toss.tech/article/8-writing-principles-of-toss — eight principles built from A/B test
+  data: Predictable hint, Weed cutting, Remove empty sentences, Focus on key message, Easy
+  to speak (avoid 한자어 and 문어체), Suggest over force, Universal words, Find hidden emotion
+- Toss Developer Center, UX Writing Guide
+  (developers-apps-in-toss.toss.im/design/ux-writing.html) — 해요체 as the product-wide
+  standard; active voice preference; one message per sentence; native Korean over Sino-Korean
+- 배민 UX Writing interview (bcut.baemin.com/6287/) — first UX writer mandate: clear
+  communication and consistency of voice; register as a designed commitment, not a default
+- Qian et al. (2025). "Enhancing Human-Like Responses in Large Language Models."
+  arxiv.org/abs/2501.05032 — DPO fine-tuning removes robotic formal register; self-intro
+  disclaimers, impersonal deflection, and encyclopedic structure identified as primary tells;
+  79–90% preference for natural responses with no benchmark degradation
+- Jiang et al. (2025). "How Human-Like Are Large Language Models? A Register-Aware
+  Linguistic Evaluation Framework." arxiv.org/abs/2605.23651 — LLMs over-produce
+  nominalisations, longer words, attributive adjectives, and present-participle constructions
+  across all registers; human instructive text is more persuasive and more lexically variable
+- Ou et al. (2024). "DialogBench: Evaluating LLMs as Human-Like Dialogue Systems."
+  aclanthology.org/2024.naacl-long.341 — DIALOGUE-ONLY: benchmark for conversational
+  agents; finding that assistant-AI positioning weakens emotional naturalness transfers as
+  a writing principle (choose register deliberately); evaluation tasks do not transfer
+- Liu et al. (2024). "Are Human Conversations Special? A Large Language Model Perspective."
+  arxiv.org/abs/2403.05045 — EXCLUDED: architectural study of attention mechanisms by
+  domain; no surface-text patterns; does not apply to copy writing decisions
+- Lala et al. (2025). "Text Overlap: An LLM with Human-like Conversational Behaviors."
+  aclanthology.org/2025.sicon-1.10 — EXCLUDED: backchannels, overlap, and turn-taking in
+  real-time text chatbots; findings are dialogue-system-specific and do not apply to static
+  page copy
+
+---
+
+## Official evidence map and layer ownership
+
+- **Toss writing principles** — https://toss.tech/article/21022 — writer: one key message,
+  remove repeated/empty lines, make the next screen predictable, use one spoken breath,
+  preserve user choice. Eye: read aloud and test CTA prediction.
+- **Toss marketing writing** — https://toss.tech/article/Marketing_Writing — writer: connect
+  the audience's present concern to a concrete next action; do not turn a click-rate tactic
+  into an unsupported product claim.
+- **Apple Writing HIG** — https://developer.apple.com/design/human-interface-guidelines/writing
+  — writer: clear, concise, useful labels in the platform's voice. Hand: place text where it
+  explains the affected control or state.
+- **Microsoft UI text guide** — https://learn.microsoft.com/en-us/windows-server/manage/windows-admin-center/extend/guides/ui-text-style-guide
+  — writer: consistent terminology, direct action labels, sentence-style capitalization.
+- **W3C Writing for Web Accessibility** — https://www.w3.org/WAI/tips/writing/ — writer:
+  descriptive headings/link text and plain language. Hand: expose the same meaning to
+  assistive technology.
+
+Ownership is deliberate. `omd-writer` establishes facts, audience language, voice, and
+surface copy in `.omd/copy-deck.md`. Copy-editor eye judges fact fidelity, scan, repetition,
+CTA prediction, register, emotion, and recovery in context. `omd copy --check` verifies only
+explicit structure, state applicability, exact unresolved sentinels, and fact-ID links.
+
+Screenshot phrases and common connectives are prompt/eval observations, never global lint:
+legitimate brands and sentences use them. AI detectors, perplexity thresholds, and sentence-
+length distributions are forbidden as shipping gates: domain, language, accessibility copy,
+and short UI labels create false positives, and a statistical signal cannot establish
+authorship or copy quality. Use those ideas only to prompt a contextual human/eye review;
+never reject a deck because a detector score or distribution looks machine-like.