# Domain — PR replies, issue replies, review comments

Covers replies on pull requests and issues, code-review comments, and discussion-thread responses. Run with `professional-pass.md` (short-answer weighting: factuality, specificity, templatedness).

## Human baseline

Direct, specific, proportional. Maintainers answer the point in the first sentence, quote the exact code or error, disagree plainly, and say "I don't know" or "won't fix" when that's the truth. Terseness is the norm, not rudeness. **Read the thread and the maintainer's other replies first — match that register**, not a universal politeness standard. With no thread to sample, this baseline applies.

## AI tells in this domain

| Tell | Fix |
|---|---|
| Praise/thanks opener by default: "Great catch!", "Thanks for the detailed report!" | Open with the answer. Thank people when it's genuinely warranted (first contribution, unusual effort) — not as a reflex |
| Restating the question/issue back before answering | Delete; the thread already contains it |
| A wall of bullets for a one-sentence answer | One sentence. Length proportional to stakes |
| Hedged non-answers: "There could be several factors…", both-sides mush | Commit: the most likely cause, the check to confirm it, or an honest "I don't know" |
| Boilerplate empathy on bug reports: "I understand how frustrating this must be" | Ask for the repro detail you actually need |
| Sign-offs: "Hope this helps!", "Let me know if you have any questions!" | End at the content |
| Promising work in the reply instead of doing it ("I'll look into this") when the fix is at hand | Do it, then link the commit |
| Perfectly uniform comment structure across a review (every comment: praise + issue + suggestion) | Vary; some comments are one word ("nit: typo"), some are paragraphs |

## Rules

1. **Answer first.** Verdict/answer in sentence one; reasoning after, only as needed.
2. Cite artifacts: `file.py:214`, the commit SHA, the error text verbatim, the doc link. A claim about code points at the code.
3. Disagree plainly with a reason ("This breaks the retry path — see #388") — no apology wrapper, no praise sandwich.
4. State uncertainty honestly and cheaply: "not sure — does it reproduce on 2.4?" beats three hedged paragraphs.
5. Wontfix/out-of-scope: say so, one reason, link the policy or issue where it was decided. Don't soften it into ambiguity the reporter must decode.
6. In review comments, distinguish severity explicitly (blocking vs nit) the way the repo already does.

# Domain — long-form journalism

Covers features, investigative and data stories, explanatory news, interviews, and a reporter's first-person account of reported events. Reported narrative belongs here even when it opens on a scene, and whether or not its sourcing is complete: a feature with missing attribution is a journalism draft with a check 5 finding, not a personal essay. Personal and literary essays, fiction, press releases, wire copy and opinion columns do not belong here. Run with `professional-pass.md` (article-like weighting) plus the outline and QUD checks in `discourse-pass.md` §1–3. Evidence base: one human-side corpus of Traditional Chinese long-form journalism (`languages/zh.md` §1b, ledger `ZH-NEWS-CORPUS-2026`) and its 169-article close reading; the machine side (`zh.md` §1c) is a register-to-register contrast group of 119 synthetic pieces from three models under one base prompt (the Grok run carried an added no-tools line), not a measurement of this venue's machine output; English journalism is not measured at all. Rows citing M say so, and no row is a cutoff.

## Human baseline

Lead and body are two registers: the standfirst gives the result, the first paragraph of the body puts a person, a place or a number in front of the reader before any definition. Subheads carry the transitions, so paragraphs rarely open on a connective. Quotations keep the speaker's spoken texture and are often attributed by context or by the language's post-posed attribution (in the measured Traditional Chinese corpus, 「說」 after the quotation); in that corpus most quotation marks enclose terms, not speech. Quotation marks and attribution syntax follow the target language and venue. The reporter appears in a few fixed roles: stating method, steering a source back, recording a silence, and, in a first-person account, narrating what the reporter went to see and observed; not delivering a verdict the reporting did not establish. Pieces end on a quotation, a fact, a return to the opening person, or an open question; a summary or a moral is rare (7 of 169 close-read articles). Non-corrective follow-ups arrive as a dated update block; a correction of a false statement, or a rolling update at a venue that revises the body, edits the body with a dated note (rule 6).

## AI tells in this domain

| Tell | Fix | Evidence |
|---|---|---|
| Preview opening: 「本文將探討…」, "This article examines…" | Open on the person, the place or the number | close reading 1/169 |
| Definition-first lead that explains the topic before anything happens | Concrete before abstract; the definition comes after the reader has a reason to want it | close reading: scene, person or number lead in most of 169 |
| `某某表示：「書面句」` as the only quotation pattern: name, colon, a complete written sentence | Vary the frame around the quotation only: attribute by context or a post-posed 「說」, move the name, split the lead-in. The quoted words are never changed on an existing draft; if a transcript is available, restore the speaker's own wording from it | T human side: colon lead-in is a minority pattern (zh.md §1b); M: the `：「` lead-in share is 0–24% across three models (zh.md §1c), a proxy for the lead-in only; the full name + 「表示」 + complete-sentence pattern is unmeasured on either side |
| No spoken texture in any quotation that comes from speech (interviews, recordings): every such quotation is complete, tidy, particle-free. Quotations from written records (filings, emails, prepared statements) are faithful when tidy and are not this tell | On write, keep the speaker's repetition, particles, code-switching, self-correction as recorded. On review, report the pattern. On refactor or recreate, change a quotation only from transcript material supplied by the user; otherwise leave it and note it (SKILL.md: quoted material is load-bearing) | close reading counts spoken texture as the norm; unmeasured on the machine side (M's quotation rows in zh.md §1c measure length and sentence count, a proxy that says nothing about particles, repetition or self-correction) |
| A manner adverb on every speech verb (「緩緩地說」「無奈地表示」) | Delete the adverb; the verb, or a gesture after the quotation, carries it | T human side: near zero (zh.md §1b, §2 row) |
| Summary or moral ending: 「總而言之」, "凸顯了…的重要性", "提醒我們…" | End on the last fact, the last quotation, or the person | T human side: 0 per 100k (zh.md §1b); close reading 7/169 summary endings |
| Every paragraph closing on the reporter's verdict sentence | Let the paragraph end on the source's words or the fact; one collecting sentence per section at most | close reading: paragraph-end verdicts are the exception |
| Sections of equal length with the same inner order (context, quotation, verdict) | Depth follows what the reporting found; a section may be one paragraph | professional-pass check 9; close reading: mixed shapes in 98/169 |
| Connectives opening consecutive paragraphs (「此外」「另一方面」「然而」) | Let the subhead or the juxtaposition do the switching | T human side: single connectives are register-normal, paragraph-initial chains are not counted (zh.md §1b, §5); Sepia inference |
| Numbers without a comparison or a local referent | Each figure carries its baseline, its change, or a referent the reader knows; method stated in the body when the reporter built the dataset | close reading: data pieces pair every figure with a comparison |
| Question asked and answered by the reporter in the same paragraph | The paragraph-end question is answered by the next speaker's data or words | close reading: question-as-transition is the norm; unmeasured on both sides (T and M count paragraph-ending questions, a different pattern from a question answered by the reporter within the paragraph) |

## Rules

1. **Two registers.** The standfirst may give the result; the first body paragraph must put a person, a place or a number in front of the reader. No scene is written without scene facts from the reporting (SKILL.md "Never invent specifics"): a missing detail is a TODO or a question, never prose.
2. **Subheads switch, paragraphs do not announce.** Transitions live in the subhead. A single paragraph-initial 「此外」 or 「然而」 is register-normal (zh.md §1b); what fails is the chain: consecutive paragraphs opening on connectives, or a numbered "first, second, third" walk.
3. **Quotations.** Keep spoken texture; attribute by context or by the language's post-posed attribution (Traditional Chinese: 「說」 after the quotation; English: "she said" after it); terminology in quotation marks is not speech and is not counted as a quotation; the speaker's words are never rewritten (SKILL.md: quoted material is load-bearing). The quotation marks themselves follow the target language and venue: 「」 with 『』 nested is measured for Traditional Chinese (T) and is not imposed on English or Simplified Chinese text.
4. **Numbers.** A figure carries a comparison (baseline, prior period, or a referent the reader knows) and, when the reporter built the dataset, the method stated in the body. Round, sourceless numbers fail check 5.
5. **The reporter's first person** appears to state method, to steer a source, to record a silence or a refusal, and, in a first-person account of reported events, to narrate what the reporter went to see and observed; it does not appear to deliver a verdict the reporting did not establish.
6. **Endings.** No summary, no moral, no generic or speculative outlook (check 7 as written). A concrete, sourced next event (a scheduled ruling, a pending vote, a dated data release) is a fact and may close the piece. Non-corrective follow-ups (a later development, a response received after publication) go in a dated update block after the body, in plain informational register, and the body is not rewritten for them. A correction of a false statement, or a rolling update at a venue that revises the body, edits the body and says so in a dated note; the venue's correction policy wins over this default.
7. **Stance (check 4), read for this venue.** The judgment a report commits to is what the reporting established and where the parties disagree, named as such. A piece that asserts nothing it verified, or blurs a documented disagreement into "both sides", fails check 4 exactly as written; committing to the verified facts is the stance.
8. **Density and relevance (checks 2 and 3), read for this venue.** The reader's task includes being placed in the scene, so a concrete detail that builds the picture is information and passes; a generic statement true in any context still fails. This narrows nothing in `professional-pass.md`; it says what counts as information for this reader.

Weighting: article-like (relevance, density, stance), then rules 7 and 8 as the venue's reading of those checks.

# Domain — incident postmortems

Covers incident reports, outage retrospectives, RCA documents. Run with `professional-pass.md` (article-like weighting: relevance, density, stance) plus the outline test in `discourse-pass.md`.

## Human baseline

Blameless toward people, merciless toward mechanisms. The real document has absolute timestamps, exact failure mechanics (the query, the config line, the race window), honest dead ends ("we spent 40 minutes on the wrong hypothesis"), and action items someone actually owns. A team's incident template is a fine container — the tell is filler inside it.

## AI tells in this domain

| Tell | Fix |
|---|---|
| Agentless fog: "mistakes were made", "the change was deployed" with no actor anywhere | Blameless ≠ agentless. Name systems and roles: "the deploy pipeline promoted the config before validation ran" |
| Generic lessons: "we will improve monitoring and communication" | An action item is a change with an owner and a date: "add alert on queue depth > 10k (owner: infra, due 09-15)" |
| Every template section filled to similar length for completeness | Sections earn their length; "What went well: N/A-grade prose" → one honest line or delete |
| Self-praise adverbs: "the team swiftly identified…" | Timestamps carry the speed judgment; let them |
| Hedged root cause: "a combination of factors may have contributed" | Commit to the causal chain you believe, and mark the genuinely unknown part as unknown |
| Moralizing conclusion about reliability culture | End at the action items |
| Round, sourceless numbers | Real duration, real blast radius, real user/request counts — from the actual incident data, never estimated to sound complete |

## Rules

1. **Timeline with absolute times and timezone**, including the wrong turns — the 40 minutes on the bad hypothesis is the most instructive part; models systematically omit failure-within-the-failure.
2. The failure mechanism at code/config level: the exact query, flag, limit, or race. If you (the writer) don't know it, that's a question for the team, not a blank to prose over.
3. Counterfactuals stated honestly: what would have caught it, and why it didn't exist. No "the system worked as designed" face-saving.
4. Contributing factors as a causal chain, not a bullet cloud — each factor says what it enabled.
5. Impact in numbers first (duration, requests failed, users affected, money if known), narrative second.
6. Stance check (professional-pass #4): a postmortem that admits no wrong judgment anywhere hasn't been written yet.

# Domain — release notes & announcements

Covers changelogs, GitHub Releases, version announcements, and short launch posts. Run with `professional-pass.md`; add `style-pass.md` for anything longer than a changelog.

## Human baseline

Terse, factual, user-impact-first. The reader is deciding **whether to upgrade and what will break** — everything serves that decision. Conventional structure (Keep a Changelog categories: Added / Changed / Fixed / Removed / Security; or the repo's own habit) is expected, not a tell.

## AI tells in this domain

| Tell | Fix |
|---|---|
| Marketing inflation: "We're thrilled/excited to announce", "powerful new features", "seamless experience", "supercharge your workflow" | State what changed. The feature is the news; enthusiasm is not |
| Benefit claims with no mechanism: "improved performance", "enhanced stability" | The number or the change itself: "cold start 1.8s → 0.4s", "fixed a race in the retry queue (#412)" |
| Every change narrated as a sentence-long story | One line per change, verb-first, no adjectives |
| Emoji headers and exclamation marks throughout | Match the repo's existing notes; default to none |
| An intro paragraph about the journey and a closing paragraph about the road ahead | Delete both. Version, date, changes, done |
| Symmetric prose for every item regardless of importance | Order by user impact; breaking changes first, one-word fixes last |

## Rules

1. **Breaking changes first**, with the exact migration step (the command, the config key, the renamed flag).
2. Every claim carries its artifact: issue/PR numbers, commit ranges, exact version strings, real benchmark numbers with conditions. No artifact → no claim.
3. Credit people plainly ("thanks @name for #398") — no gratitude paragraphs.
4. Length follows the release: a patch release is three lines; do not inflate it to look substantial.
5. Humor and voice are allowed if the repo's history has them; never inject them fresh into a repo that doesn't.

# Domain — technical articles & blog posts

Covers engineering blog posts, tutorials, architecture write-ups, experience reports. The richest domain: run `professional-pass.md` (article-like weighting), the outline/QUD checks in `discourse-pass.md` §1–3, and `style-pass.md` (skip its fiction-slop table).

## Human baseline

Motivated by a real problem the author actually hit. Uneven by design — deep where it got interesting, one line where it didn't. Contains at least one dead end, at least one opinion the reader could disagree with, and numbers with their conditions attached. First person and contractions are normal.

## AI tells in this domain

| Tell | Fix |
|---|---|
| The topic-survey opening: "In the world of distributed systems…" / definition of the thing everyone reading already knows | Open at the incident, the bug, the number that made you look |
| Listicle in a trench coat: prose that is secretly "The first… The second… The third…" | Either honest structure (a real list/table) or real prose with an argument |
| Fractal summaries: every section announces, tells, recaps | Say it once, at the level where it lives |
| Invented concept labels: "the observability paradox", "configuration drift syndrome" coined mid-post | Plain description, or an established term |
| Symmetric coverage: every alternative gets a paragraph, none gets a verdict | Commit to a recommendation and give the case that would change your mind |
| No failure anywhere: every step worked, benchmarks confirm the thesis | Include what broke, what you tried first, what you'd skip next time — models systematically omit the dead end, and it's the part readers trust |
| Generic code examples (`foo`, `my_service`) that were never run | Real, runnable, tested snippets from the actual work — or say explicitly they're sketches |
| Benchmarks with no conditions | Machine, version, dataset size, number of runs — or don't print the number |
| The both-sides conclusion + future outlook | End on the recommendation or the open question you actually have |

## Rules

1. **The problem before the topic.** First paragraph: the concrete situation that forced the question. If there is no real situation, the honest genre is "notes on X", not a war story — never fabricate the incident.
2. One opinion minimum, stated as yours, with the disagreement condition ("if your writes are under 1k/s, ignore all of this").
3. Depth budget by interest, not symmetry: the section that surprised you gets 5× the words of the setup steps.
4. Numbers carry conditions; claims carry links; code carries a "this runs" guarantee or a disclaimer.
5. QUD check: if the section-question sequence reads *what is X → why X matters → how to X → conclusion*, restructure around what actually happened.
6. Voice: first person, contractions, an aside or two. The measured human markers (stance, unevenness, lived specifics) are the same ones expert readers use to judge "a person wrote this."

# Domain — tickets & work orders

Covers issue tickets, tasks, work orders, bug reports you file (not replies — that's `dev-replies.md`). Run with `professional-pass.md` (short-answer weighting).

## Human baseline

Imperative, minimal, complete-enough. The assignee should be able to start without asking a question, and know when they're done. A tracker's field template is a container, not a tell.

## AI tells in this domain

| Tell | Fix |
|---|---|
| Novel-length background before the ask | Context is only what the assignee doesn't already know; link the rest |
| Description that restates the title in sentences | The description starts where the title stops |
| Obvious steps enumerated ("1. Open the repository. 2. Locate the file…") | Only the non-obvious steps and the exact commands |
| Vague acceptance: "works correctly", "improved performance" | Testable criteria: the command to run and the output that means done ("p95 < 200ms on the staging load test") |
| Every template field filled with prose for completeness | Empty is a valid value; "N/A" beats a paragraph of nothing |
| Round scope words: "refactor the module", "clean up" | The concrete boundary: which files/functions in, which explicitly out |

## Rules

1. **Title = outcome**, not activity ("Retry queue drops jobs on redeploy" not "Investigate queue issue").
2. Bug tickets: exact repro (versions, commands, input), expected vs actual with real output pasted, frequency. If you can't reproduce it, say what you tried.
3. Acceptance criteria are testable or they aren't criteria.
4. Link, don't repeat: prior tickets, the design doc, the alert. One source of truth.
5. Priority/estimate honest and bare — no justification paragraphs.

# Chinese calibration for the style pass

Load this file at the style-pass step whenever the target text is Chinese, in any variant, on any route. It changes nothing else about the route. The English ban lists in `style-pass.md` §3 do not transfer word for word; what transfers is the *shape* of the checks — the syntax templates of §2, the restore list of §4, the rhythm check of §5, the whitelist of §7 — and this file says what each shape looks like in Chinese.

Evidence: one measured corpus, HC3-Chinese — 6,586 human and 6,586 ChatGPT answers to the same open-domain questions, GPT-3.5-era ChatGPT, Simplified Chinese, 2023 — analysed with 159 Chinese CTAP features by 朱君輝 et al., CCL 2023 (Z), with Guo et al. 2023 (H) supplying a second measure on the same corpus; a 2025 joke-generation study by 蔣彥廷 and 應以周, CCL 2025 (J), is used only where it contradicts Z. A second source (T), a private human-side measurement of Traditional Chinese long-form journalism (ledger `ZH-NEWS-CORPUS-2026`; about two thousand articles from one unnamed Taiwanese publication, spanning about ten years), supplies presence rates in human prose; a machine side (M) for the same register was measured privately on 2026-09-16 and 17 (119 synthetic pieces, 40 each from Grok 4.6 and GPT-6 and 39 from Gemini 3.8 Flash, one shared base prompt with a one-line no-tools variant for the Grok run), so §1c can set human and machine rates side by side for this register; §6 states M's limits, and no row licenses a per-passage cutoff. Two public Taiwanese standards give the normative side of punctuation and numerals where §1b and §1c measure usage: the Ministry of Education punctuation handbook (`MOE-PUNCT-2008`) and the Executive Yuan numeral principle (`EY-NUMERALS-2004`). Everything else below is a Sepia inference, or an editorial heuristic marked as such in §4. Stable source identities are in the repository research ledger.

## 0 Taiwan conventions (normative; venue rules, not authorship signals)

Apply on a Taiwanese venue when the target text is Traditional Chinese. Simplified Chinese copy, even for a Taiwanese venue, keeps its own conventions (“ ” as the first-level quotation mark among them) and none of the punctuation, numeral, title or attribution rows below applies to it, matching `domains/journalism.md` rule 3; the §5 Taiwan-lexicon guardrail in the last row keeps its own scope and is unaffected by this exemption. When no target venue is known, the author's own samples decide whether these rows apply; when the venue is known and is not Taiwanese, the venue wins and no row below applies even if the samples follow Taiwan conventions (the venue sets the register, `SKILL.md`). Treat a departure as a register mismatch to fix (the same handling as the mainland-lexicon rule in §5), not as a signal about authorship. Sources: the Ministry of Education punctuation handbook (`MOE-PUNCT-2008`), the Executive Yuan numeral principle (`EY-NUMERALS-2004`, written for government documents, not newspapers), and the multi-source textbook consensus recorded in `research/newswriting-guides.md` §五. Where guides disagree (whether quotations may be tidied, the reporter's first person, a numeric threshold for spelling numbers out, whether a nut graf is required, 民國 vs 西元 years) this file takes no side; the venue does. The one default this file does set is the numerals row's conversion at eleven and above when no venue samples exist, because every consulted guide agrees above ten; a venue threshold below eleven is the venue's own. The title, attribution-verb and question rows are journalism-route rows and do not run on other professional routes or on fiction. No consulted standard regulates thousands separators or spacing between digits and Chinese characters; both are left as received. Every fix below applies to running prose only: code spans, commands and flags, Markdown syntax, URLs, file names and version strings are never touched, and the text inside a quotation is left as received (`SKILL.md` keeps quoted material immutable), so 「我...不知道」 inside 「」 stays as the speaker's transcript even though the marks around a quotation may be normalised.

| Convention | Source and class | How sepia applies it |
|---|---|---|
| Quotation marks: 「」 first, 『』 nested inside; "一般引文的句尾符號標在引號之內"; a quotation that is part of a longer sentence takes no punctuation before its closing mark | `MOE-PUNCT-2008` 引號 (official standard) | Fix “ ”, ‘ ’ or the ASCII " " and ' ' used as the first level, and 『』 used as the first level; report sentence-final punctuation (。！？) placed outside a closing 」 when the quotation is a full sentence, and do not move it on refactor: the quotation boundary is quoted material and stays as received. Measured usage agrees (`research/zh-news-corpus.md`, quotation-mark rows; §1b carries no glyph or placement row) |
| Dash: 「──」 occupying two cells, for a turn of meaning, a drawn-out sound, or a supplement after which the sentence pauses; the paired 「──　──」 is the handbook's 乙式 parenthesis | `MOE-PUNCT-2008` 破折號, 夾注號 (official standard) | Fix the single-glyph 「—」 or ASCII 「--」 to 「──」. The paired form is legitimate and takes no action (§1b) |
| Ellipsis: 「……」, six dots over two cells | `MOE-PUNCT-2008` 刪節號 (official standard) | Fix 「…」 or 「...」 to 「……」; do not add ellipses |
| Parenthetical notes: 甲式 （　） for annotation; 乙式 ──　── when the sentence flows through the supplement. A foreign original or birth and death dates after a name go in 甲式 | `MOE-PUNCT-2008` 夾注號 (official standard) for the two forms; consensus (guides §五 6) for the foreign-original and birth/death-date uses | Fix half-width ( ) around Chinese text, a foreign original or birth and death dates to full-width （）; what sits inside the parentheses is left as written; keep a foreign original in parentheses on first mention |
| Numerals: Arabic for counts, statistics, measures, dates, clock times, ordinals, codes; Chinese numerals for idioms, set phrases and proper names (三讀、九二一、一線). sepia acts only on counts and measures; dates, clock times, ordinals and codes are left as received and the venue decides | `EY-NUMERALS-2004` (official administrative standard for government documents; newspapers add house rules on percentages and era years) | Venue-gated: on a Taiwanese venue, fix a spelled-out count or measure of eleven or more (「一千一百小時」「三十四位」) to Arabic digits, changing numeral form only (one spacing invariant: a converted numeral copies the piece's existing spacing between Chinese characters and digits; the smoke sample wrote 「是 3.14，」, a space between a Chinese character and a digit, so 「三十四位」 became 「34 位」; a piece with no such spacing, or no digits at all, keeps adjacency, 「34位」), when the venue's own samples write such counts in digits or when no samples are available; samples that spell them out win, and the row takes no action. Eleven is the lowest value on which every consulted guide agrees (AP and BBC spell out 1–9, the Economist 1–10, one Taiwanese NGO 一至十); counts of ten and below (「三位」「十位」) are left as written and the venue decides. Idioms and names stay in Chinese. Do not mix 「百分之」 and 「%」 in one piece (consensus, guides §五 7). §1c gives the aggregate direction only (glyph counts, no use classification) |
| Titles and names: on first mention the full title precedes the full name (行政院長某某某) | One first-tier house rule, an older newspaper 通則 (guides §三; not a consensus, guides §四 10) | Journalism route only. Venue-gated: fix a title placed after the name (某某某行政院長) on the first mention only, and only when the venue's own samples put the title first; later mentions with a shorter or postposed title are left as written; with no samples, or samples that use the other order, no action |
| Transliterated foreign names and Indigenous names use the 間隔號 「．」 between parts | `MOE-PUNCT-2008` 間隔號 (official standard) | Fix 「·」 or 「.」 to 「．」 only between the parts of a person's name; decimal points, URLs, file names, abbreviations and version strings are never touched |
| Work titles: 《》 for books and other whole works (publications, films, documents); 〈〉 for 篇名 (articles, chapters); the handbook lists songs among 書名號 uses without assigning a form, so the venue decides between 《》 and 〈〉 for a song; a piece's own headline takes no 書名號 | `MOE-PUNCT-2008` 書名號 乙式 (official standard) | Fix “ ”, ASCII " ", 「」 or 『』 used around a work title; keep 「」 for terms and speech |
| Attribution verb: 「說」 is the default; 「表示」「指出」「強調」「坦言」 carry a judgment and are used when that judgment is meant | Consensus (guides §五 3) | Journalism route only. Report a run of judgment verbs where 「說」 would do; do not rewrite quotations to change the verb inside them. sepia does not restore 「表示」 (guides §六 3, inference) |
| Questions: a question mark follows an interrogative sentence; a rhetorical question does not replace a verified fact | `MOE-PUNCT-2008` 問號 (official standard); consensus (guides §五 9) | Journalism route only. Report a rhetorical question standing where a fact was needed. Leave a paragraph-end question alone only when the next paragraph answers it with a source's words or data (the next-paragraph source answer is a close-reading observation recorded as C-tier evidence in `voices/tw-journalism.md`; the same-paragraph self-answer is `journalism.md` tells row 11 and stays reported; §1b: register-normal); a paragraph-end question that stands in for a fact is still reported |
| Taiwan lexicon on a Taiwanese venue (already in §5) | §5 of this file, under `SKILL.md`'s instruction to respect the author's voice and the venue corpus | Unchanged: fix 視頻、軟件、質量 only when the venue is Taiwanese and the author's samples do not use them |

## 1 Measured (Z, per-answer means; H where named)

| Feature | Human | ChatGPT | Unit and note |
|---|---|---|---|
| Sentence-length SD | 9.248 | 6.729 | words (詞); in characters (字) 15.150 vs 12.842 — the gap holds in both units |
| Mean sentence length | 25.067 | 21.823 | words — humans *longer*; in characters 40.893 vs 42.396, the direction flips, so length itself is not a signal |
| Paragraphs per answer | 1.442 | 3.681 | ChatGPT wrote *more* paragraphs; mean paragraph length 123.907 vs 92.747 characters |
| Punctuation density | 0.135 | 0.136 | Z; the same corpus measured as punctuation share of tokens reads 16.0% vs 13.4% (H) — contradictory measures, not a signal |
| 語氣詞 density | 0.016 | 0.003 | five times higher in human answers |
| 連詞 density | 0.013 | 0.036 | 「和」 alone: 4.13 vs 11.76 per answer |
| Pronoun density | 0.052 | 0.069 | second person 0.010 vs 0.021 |
| Monosyllabic word share | 0.483 | 0.379 | disyllabic share 0.445 vs 0.532; disyllabic word count was selected as a key feature by both of Z's filters |
| Type-token ratio | 0.725 | 0.543 | content-word richness 0.822 vs 0.647 |
| Mean dependency distance | 3.900 | 3.659 | longest 29.452 vs 23.991 |

## 1b Presence in human Traditional Chinese journalism (T; human side only)

Rates in human prose. A near-zero rate means the form departs from this register's norm when it appears; it is not a measured machine excess. Nothing here is a per-passage cutoff. The middle column names the kind of quantity each time: per 100k characters, share of articles containing the form, or per-article median.

| Form | Presence (T) and kind of quantity | How to read |
|---|---|---|
| Manner adverb 「X地」 before a verb | 1.9 per 100k chars | Departs from this register when it appears; §2 row (T), journalism only |
| 「很」 : 「非常」 | occurrence ratio ≈ 11:1 | 「非常／十分／極為」 stacking is the departure; one instance is not |
| Paired dash 「──…──」 | 15% of articles contain one; 4 per 100k chars (per-article median 0) | Rarer than the single dash by an order of magnitude, but present; not a departure on its own, and no hunt row. `MOE-PUNCT-2008` lists the paired form as the 乙式 parenthesis (──　──) for a supplement the sentence flows through |
| Single 「──」 | 61% of articles contain it; 40 per 100k chars | Not a signal. The two-cell 「──」 is the handbook form (`MOE-PUNCT-2008`); the single-glyph 「—」 below is not |
| Single-glyph 「—」 | 7% of articles contain it | Glyph choice; a register mismatch at most |
| 「……」 | 2% of articles contain it; 0.9 per 100k chars | Departs in narration; inside a quotation it is speech. The handbook form is six dots over two cells (`MOE-PUNCT-2008`); a three-dot 「…」 or ASCII 「...」 is a glyph slip |
| 「綜上所述」「總而言之」 | 0 per 100k chars; 「值得一提的是」 0.1; 「值得注意的是」 0.6 | §3 formula phrases; this register does not use them |
| 「不是…而是」 | 30% of articles contain it; 6.3 per 100k chars | One instance is register-normal; see §4 |
| Three-item 頓號 list (A、B、C) | 59% of articles contain one | Register-normal; see §4 |
| Sentence-initial 「其實」 | 15% of articles contain it | Register-normal; see §4 |
| 「此外」 / 「然而」 | 37% / 53% of articles contain it | Single connectives are register-normal; chaining across clauses is still §2 |
| Sentence bearing any 「」 with no attribution verb (terminology and scare quotes included) | per-article median share 61% | Most 「」 in this register mark terms, not speech; this row says nothing about attribution of speech |
| Sentence bearing a 「」 quotation of 15 characters or more with no attribution verb | per-article median share 33% (q1 20%, q3 50%) | Length is a proxy, not a speech classification: titles, slogans and written statements stay in the pool. Speech quotations as such are unmeasured; this row licenses no restore rule |
| `：「` colon lead-in | 6% of all 「」 quotations; 10% (q1 0%, q3 22%) of quotations of 15 characters or more, per-article medians | Measured on the ordinary quote mark. The colon lead-in exists in this register but is a minority pattern; 「某某表示：「…」」 as the only pattern departs from it |
| Sentence length | per-article median of the mean 59 chars; of the within-article SD 34 | Dispersion is the human trait; length itself is not (consistent with §1 and §5) |
| Runs of three near-equal sentences | per-article median 2 per 100 sentences | §5 as written |
| Suffix-bearing vocabulary 「○○性／○○化／○○感」 | 77 per 100k chars, mostly fixed legal-policy terms | A lexical count only; whether such nouns serve as subjects or wrap a verb (「自我的探索」) was not counted. Nominalized subjects stay unmeasured; §4 row |
| Summary endings | 7 of 169 close-read articles (close-reading sample count) | Conclusion residue (`professional-pass.md` check 7) is the departure in this register |

## 1c Machine side for the same register (M; three models, one base prompt with a Grok variant, two consecutive days)

Design: 40 fictional zh-TW briefs (scene, profile, policy, data, investigation, breaking), one shared base prompt that fixed 1,000–1,400 characters and asked for at least three direct quotations, no style or anti-AI instruction; the Grok run added one line telling it not to read files or skills, after two earlier runs spent their single turn reading a bundled skill, so the Grok prompt is a variant of the base prompt, not identical to it. Each model ran headless with no installed skill or memory: Grok 4.6, GPT-6 (via its CLI), Gemini 3.8 Flash (39 pieces; one brief returned empty twice). Same metric tools as T. Every cell names its kind of quantity. A human number appears here only for forms §1b does not carry; for forms §1b carries, the Human column points to §1b. Nothing here is a per-passage cutoff, and no row below is a hunt row: the machine-high forms are register departures whose human side is not near zero, so the reading stays "cluster", as §4 already says. Limits in §6.

Robust across the three models (direction the same in all three, difference large):

| Form (kind of quantity) | Human (T) | Gemini 3.8 Flash | GPT-6 | Grok 4.6 | Read as |
|---|---|---|---|---|---|
| Sentence length: per-article median of the mean / of the within-article SD, chars | §1b (59 / 34) | 48 / 19 | 34 / 15 | 36 / 16 | All three write shorter, flatter sentences in this register; Gemini is closest and still a fifth shorter with half the dispersion. On dispersion the direction agrees with HC3 (§1, SD 15.150 vs 12.842 字); on mean length it does not (HC3 found ChatGPT slightly longer in 字, 42.4 vs 40.9). Neither number is a length target |
| Clauses per sentence, per-article median | 4.5 | 3.4 | 3.1 | 3.1 | Comma-linked clauses are the human shape here; a run of short, comma-poor sentences is on the machine side. (The per-article maximum of commas before a period is length-sensitive and sits in the weak table) |
| Share of sentences of 60+ chars, per-article median | 42% | 23% | 5% | 8% | Same reading |
| Arabic numerals / Chinese numerals, per 1k tokens | 18.7 / 5.1 | 0 / 16 | 0 / 16 | 0 / 24 | All three spell numbers out; human journalism writes most figures in Arabic digits. An aggregate glyph contrast only: the tokens were not classified by use (date, statistic, idiom, name), so this row does not establish a departure from `EY-NUMERALS-2004`, which assigns forms by use |
| First-sentence length, per-article median chars; first 80 characters contain an Arabic digit, median | 60; yes | 37; no | 31; no | 37; no | Machines open shorter and without an Arabic digit. Chinese numerals were not counted in this check, and the numerals row shows the machines write figures in Chinese, so this says nothing about whether the lead carries a figure |
| Quotations: median length chars / multi-sentence share, per-article medians (prompt fixed length and asked for at least three quotations, which constrains these) | 8 / 6% | 46 / 38% | 27 / 25% | 30 / 41% | Machine quotations are three to six times longer and four to seven times more often multi-sentence; most human 「」 mark terms (§1b). Length is a proxy, not a speech classification |
| 「不是…而是」, per 100k chars | §1b (6.3) | 47.0 | 18.5 | 37.1 | 2.9–7.5× the human rate in all three; the share of articles is mixed (weak table). One instance stays register-normal; the machine side is the cluster |
| 「坦言」, per 100k chars | 4.6 | 30.2 | 20.4 | 10.6 | 2–7× per 100k in all three; the share of articles is mixed (weak table) |
| Exclamation marks: share of articles; per 100k | 54%; 28 | 0; 0 | 0; 0 | 0; 0 | Machines produce none. The close reading notes that the human ones sit mostly inside quotations; placement is not measured |
| Dashes: 「──」 share of articles / paired insertion share | §1b (61%) / §1b (15%) | 0 / 0 | 0 / 0 | 0 / 2% | The belief that a paired em-dash insertion marks machine Chinese is not supported here: the machines barely use the two-cell dash at all. `MOE-PUNCT-2008` specifies the two-cell 「──」 |

Model-specific or weak (report, do not generalise):

| Form (kind of quantity) | Human (T) | Gemini 3.8 Flash | GPT-6 | Grok 4.6 | Read as |
|---|---|---|---|---|---|
| 「此外」 / 「然而」, share of articles | §1b (37% / 53%) | 0% / 56% | 2% / 35% | 0% / 0% | Model-specific |
| Max commas before a period, per-article median of the per-article maximum | 9 | 5 | 4 | 4 | A maximum over a piece grows with piece length, and the human pieces are several times longer (§6); not length-matched, so the direction is not established. Clauses per sentence (robust table) carries the comma reading |
| Quotations of 50+ chars, share; quotation-first sentences, share (per-article medians; same prompt constraint) | 10%; 9% | 46%; 50% | 0%; 58% | 19%; 23% | Mixed: GPT-6 has no long quotations yet the most quotation-first sentences |
| 「不是…而是」, share of articles | §1b (30%) | 46% | 25% | 35% | Mixed direction (GPT-6 below human) while the per-100k rate is above in all three (robust table) |
| 「坦言」, share of articles | 21% | 41% | 28% | 15% | Mixed direction (Grok below human) while the per-100k rate is above in all three (robust table) |
| 「——」 (two em dashes), share of articles | 5% | 5% | 0% | 10% | Equal, lower, higher: no direction |
| `：「` colon lead-in, per-article median share of quotations (same prompt constraint) | §1b (6%) | 14% | 0% | 24% | Mixed direction: two models above the human minority pattern, one at zero |
| 「說」 as attribution (jieba), per 1k tokens | 1.7 | 1.1 | 6.0 | 7.4 | Mixed direction: GPT-6 and Grok lean on 「說」, Gemini below human and leaning on 「表示」「坦言」 instead |
| 「表示」: share; per 100k | 68%; 44 | 49%; 45 | 65%; 109 | 52%; 53 | Marked 相近 in the report; GPT-6 alone doubles the rate |
| Question-ending paragraphs: share; per 100k | 72%; 35 | 10%; 10 | 55%; 68 | 0%; 0 | Measured, no direction: GPT-6 above human per 100k, below on share; Grok none |
| Emotion words (jieba), per 1k tokens | 0.46 | 1.1 | 0 | 0 | Gemini only |
| Sentence-initial 「其實」, share of articles | §1b (15%) | 0% | 0% | 5% | Low counts on short pieces; weak |
| Three-item 頓號 list, share of articles | §1b (59%) | 8% | 15% | 32% | Human higher on share of articles; the private contrast report labels the per-100k rate, a statistic this table does not carry, as similar; weak |
| Subhead-like paragraphs (≤25 chars, no sentence-final punctuation), share of paragraphs | 8% | 22% | 16% | 21% | A proxy: the prompt forbade Markdown, so whether these short lines are subheads is not determined |

## 2 What to hunt (Sepia inferences from §1, §1b and §1c; shapes of `style-pass.md` §2)

| Shape | Chinese form | Fix |
|---|---|---|
| Connective stacking (連詞 density 0.036 vs 0.013) | 「和／以及／並且／同時／此外／因此／然而」 chained across clauses; 「和」 joining whole clauses rather than nouns | Delete the connective and let juxtaposition carry the link; Chinese parataxis is the human default |
| Second-person address outside dialogue or instructions | 「你會發現」「您可以」 in expository prose | Delete or recast as a statement |
| Disyllabic padding where a monosyllable is idiomatic | 「進行討論」「加以說明」「予以處理」「做出決定」 | 「討論」「說明」「處理」「決定」 — the verb alone |
| Flat sentence length (SD 6.729 vs 9.248 words) | Runs of adjacent sentences of about the same length | Apply the §5 check as written (runs of three or more adjacent near-equal sentences): split one long sentence, merge two short ones, delete a clause. §5 sets no length cutoff in any language. The only Chinese length-share measurement is register-level (§1c: per-article median share of sentences of 60+ characters, 42% human vs 5–23% machine, in journalism); it describes the register and is not a per-passage cutoff. Count in whichever unit you use consistently — the SD gap holds in both 詞 and 字. T reports a per-article median mean of 59 characters with within-article SD 34 for journalism, again as dispersion, not a length target. M: all three contrast models sit at mean 34–48 characters with within-article SD 15–19 against human 59/34 (§1c); dispersion agrees with HC3, mean length does not (HC3 found the model slightly longer in 字). In this register a run of short, comma-poor sentences is therefore a cluster candidate for §5 (a register-level contrast, not a per-passage classification); the fix stays the §5 one (merge two short sentences, let a clause carry a subordinate fact) and sets no length target |
| Manner adverb 「X地」 before a speech or action verb — journalism only (T: 1.9 per 100k chars in human journalism) | 「緩緩地說」「堅定地表示」「無奈地說」 | In journalistic Chinese, delete the adverb and let the verb, or a following gesture, carry it. Elsewhere (fiction, dialogue tags, technical and incident writing) leave it: the rate is from one journalism register and T says nothing about other registers. Human-side presence, not a measured machine excess |

## 3 What to restore (Sepia inferences; shape of `style-pass.md` §4)

Sprinkled, never poured, and only where the register allows: sentence-final and mid-sentence 語氣詞 (啊、吧、呢、嘛、喔、啦、耶) — the largest measured gap in §1; monosyllabic verbs and adjectives; a spread of sentence lengths; subject ellipsis and colloquial contraction where a native writer would drop the subject, the Chinese counterpart of §4's contractions. Formal venues keep their register: a legal notice does not get 「嘛」. Attribution of speech quotations in journalism is not measured by T (its quotation rows use a length proxy); nothing about attribution is restored on T's account.

## 4 Editorial heuristics — presence measured on the human side (T) and, for the counted forms, the machine side (M)

Reported by Taiwan editors and readers in 2026 (自由時報 2026-07-12; 數位時代 2026-04-22; ledger "Consulted" table). T gives a presence rate in human journalism for the exact form it counted, named in the T column; §1c gives the machine side for the same counted forms, in the M column: 「不是…而是」 in the robust table, sentence-initial 「其實」 and the three-item 頓號 list in the weak table. Neither side measures the broader heuristics (parallel clauses and images, 「事實上」, quoted abstractions, nominalized subjects). A single instance of a counted form is register-normal; a cluster in narration is what the §2 template is for. Each is a Chinese form of a template already in `style-pass.md` §2:

| Reported tell | Maps to | Presence in human journalism (T) | Machine side (M) |
|---|---|---|---|
| 「不是…而是…」「這不是 X，而是 Y」 | §2 "it's not X, it's Y" | 30% of articles contain 「不是…而是」 (this form only) | per 100k chars 2.9–7.5× T across three models; share of articles similar (§1c) |
| Nominalized subjects 「○○性／○○感／○○化」 (「自我的探索」 for 「找自己」) | §2 nominalization | Only suffix-bearing vocabulary was counted (77 per 100k chars, mostly fixed legal-policy terms). The subject position and the 「的＋noun」 wrapper were not counted and stay unmeasured; see §1b | not measured in §1c: the short-piece count was a length artifact and is not tabulated |
| Three parallel clauses or images, everywhere | §2 rule of three | Only the three-item 頓號 list (A、B、C) was counted: 59% of articles contain one. Parallel clauses and images were not counted and stay unmeasured | 頓號 list in 8–32% of pieces vs 59% T; weak (§1c weak) |
| Paragraph openers 「其實…」「事實上…」; abstractions in quotation marks (「趨勢」「關鍵」「必然」) | §3 formula phrases | Only sentence-initial 「其實」 was counted: 15% of articles contain one. 「事實上」 and quoted abstractions were not counted and stay unmeasured | 0–5% of pieces vs 15% T; weak (§1c weak) |

## 5 Not signals in Chinese

Punctuation density and comma or period counts (§1: contradictory measures); long sentences counted in words (humans are longer); paragraph count (ChatGPT split *more*, the reverse of the folk belief that AI writes one block); word-frequency level (Z finds humans using commoner words, J finds LLMs doing so — two corpora, two eras, no rule). Mainland Chinese lexicon in a Taiwan venue (視頻、軟件、質量 for 影片、軟體、品質) is a register mismatch under the venue-corpus guardrail in `SKILL.md`, not an AI tell: fix it only when the venue is Taiwanese and the author's own samples do not use it. In Traditional Chinese journalism (T): a single 「不是…而是」, a three-item 頓號 list, a single 「──」, and 「此外」 or 「然而」 standing alone are register-normal (§1b); connectives chained across clauses remain the §2 case. Long sentences are not a machine signal in this register: the machine side is shorter (§1c).

## 6 Evidence boundary

One corpus, one model era, Simplified Chinese question answering, default un-prompted ChatGPT of 2023. No published study measures 2024–2026 models on Chinese narrative or expository prose (M below is a private, unpublished measurement); J covers single-sentence jokes from four 2025 models and reports lexical richness (human 0.547 vs 0.384–0.461) but no sentence or punctuation statistics. No Taiwan academic study compares human and machine Traditional Chinese; the Taiwan sources above are editorial, except the two official standards (`MOE-PUNCT-2008`, `EY-NUMERALS-2004`), which are normative documents, and T and M, which are private measurements (§1b, §1c), not editorial opinion. Treat every number here as a direction observed once, not a calibration constant. T is one unnamed Taiwanese publication's long-form journalism, about two thousand articles spanning about ten years, human text only, measured privately in 2026-09 and not distributed. It adds presence rates and medians for one register. With M (§1c) it can say, for this register and these three models under one base prompt (the Grok run carried an added no-tools line), where human and machine rates differ; it still names no tell and sets no cutoff. Its part-of-speech counts come from jieba on Traditional Chinese and are read as relative values. M's limits: 40 pieces per model (39 for Gemini), about 1,400 characters each, against human pieces averaging several thousand; one base prompt version, varied only by the Grok no-tools line, that itself fixed length and asked for three quotations, which constrains quotation density and length; synthetic briefs, so no real reporting facts; three models over two consecutive days (2026-09-16 and 17), default settings; per-article medians on short pieces are noisier than on the human side; no human writer wrote to the same briefs, so the comparison is register-to-register, not brief-to-brief. Nothing here is a classifier or a detection claim.

# Voice profile — Hemingway (built-in, experimental)

Status: a built-in voice profile for the experimental interface in `voice-skills.md`. It loads only when the user opts in ("apply the Hemingway voice"); a review may report a fit for it (the rule and its data live in `references/voices/registry.md`) but never applies it. Everything in `voice-skills.md` governs: sepia's architecture decisions first, 3–5 voice moves per piece, uniformity findings at full strength, venue precedence on professional routes, never invent.

Evidence tiers, kept apart (IDs in the research ledger): the author's own statements (`HEMINGWAY-*`, author testimony), the Kansas City Star style sheet (`KC-STAR-STYLE-1915`, style manual), criticism and corpus work (`LEVIN-1951`, `SMITH-1983`, `LAMB-2010`, `RICE-2017`, `IHRMARK-NILSSON-2021`, `LIAN-2025`). None of these is a measurement of AI text; the sepia-axis column below is a Sepia inference linking the move to StoryScope/LAMP findings that live in the passes. Digest: `research/hemingway.md`.

## The method in the author's words

- Omission: "If a writer of prose knows enough of what he is writing about he may omit things that he knows and the reader, if the writer is writing truly enough, will have a feeling of those things as strongly as though the writer had stated them." (*Death in the Afternoon*, 1932; 46 words)
- The iceberg: "The dignity of movement of an ice-berg is due to only one-eighth of it being above water." (same, 17 words) — "Anything you know you can eliminate and it only strengthens your iceberg. It is the part that doesn't show." (Paris Review, 1958; 20 words)
- The true sentence: "If I started to write elaborately, or like someone introducing or presenting something, I found that I could cut that scrollwork or ornament out and throw it away and start with the first true simple declarative sentence I had written." (*A Moveable Feast*, 1964; 43 words)
- Made, not described: "After you learn to write your whole object is to convey everything, every sensation, sight, feeling, place and emotion to the reader." / "It is made; not described." ("Monologue to the Maestro", *Esquire*, 1935)
- The newspaper rules he kept: "Use short sentences. Use short first paragraphs. Use vigorous English. Be positive, not negative." (`KC-STAR-STYLE-1915`)

**The precondition is knowledge.** Omission works because the writer knows what was left out. Under this voice the existing specificity rules stay at full strength — on the fiction route, SKILL.md's "Never invent specifics" guardrail and style-pass §1 row 5 (lack of specificity); on professional routes, professional-pass check 5. A fact the writer does not have is a gap to ask about, not something to omit around. Omission that hides ignorance produces vagueness, and vagueness is a tell.

## Fiction route — the iceberg

| Move | Source | Sepia axis it moves (inference) | Known cost |
|---|---|---|---|
| Leave the meaning out: no sentence that says what the story is about, no character realizing the theme, no narrator gloss on the ending | Omission (DIA 1932; Paris Review 1958); Smith 1983 shows the cut endings in manuscript | Rubric Group A: thematic explicitness, narratorial thematic commentary (narrative-pass §1) | Thematic explicitness drops toward the low pole; expect an over-correction advisory and report it as the voice's cost. Keep one thing the reader is allowed to understand |
| Emotion as action and speech: what she does, what is said, what is not answered | "It is made; not described." | Rubric Group B: emotion via embodied sensation → behavior-led (narrative-pass §5) | Depth of interior access drops toward the low pole; expect an over-correction advisory and report it as the voice's cost. Plain naming stays allowed in one or two places — the author did it ("he felt quite sure that he would never die", "Indian Camp", 1925, US public domain) |
| Weather is weather, objects are objects: setting does not mirror the inner state | Omission; Baker 1972 ("prune language and avoid waste motion") | Rubric Group B: setting mirrors inner state (narrative-pass §5) | None beyond the general slack rule |
| Open in the situation, not the establishing shot | "cut that scrollwork … like someone introducing or presenting something"; Lamb 2010 on openings | discourse-pass §4 machine opening | None |
| Let information move through talk, including what is not said | Lamb 2010 (dialogue's role); Rice 2017 (twice the average dialogue share) | Rubric Group C: protagonist introduced in-dialogue (human marker); Group E dialogue proportion | Dialogue share is a calibration parameter: "direct speech dominates" is a measured Gemini fingerprint. Do not push past the human band |
| Base rhythm of short declaratives, broken by an occasional paratactic run joined with "and" | *A Moveable Feast*; Levin 1951 (parataxis); Rice 2017 (short sentences define him less over his career) | style-pass §1 row 2 (sentence structure); voice-skills uniformity rule | The break is the point. Uniform short sentences are the metronome that `voice-skills.md` names as a hard finding |
| Keep the plain adverbs: then, now, never; drop the -ly manner adverbs | Rice 2017 (more adverbs than average, almost none in -ly) | style-pass §4 restore plain connectives and particles | None |

Select 3–5 of these per piece. The paratactic run is one move, used once or twice, not a texture.

## Professional routes — the Kansas City Star rules

| Rule (verbatim) | Sepia check it maps to | Note |
|---|---|---|
| "Use short sentences." | professional-pass check 10; style-pass §1 row 2 | Base rhythm, not a ceiling. Check 9 (sameness of rhythm) still applies |
| "Use short first paragraphs." | professional-pass check 3 (relevance); domains: answer first | Same as the domain files' lead rule |
| "Use vigorous English." | style-pass §1 row 1 (word choice); §3 performance verbs are the counterfeit of this | Vigorous means the concrete verb, not the inflated one |
| "Be positive, not negative." | **CONFLICT** with style-pass §4 (synthetic negation runs at half the human rate; restoring some is a human marker) | Handled by the precedence rule in `voice-skills.md`: a voice move that directly conflicts with a §4 restore item is surfaced with both rules named and left to the user |
| "Avoid the use of adjectives, especially such extravagant ones as splendid, gorgeous, grand, magnificent, etc." | style-pass §3 inflation adjectives; §1 row 6 purple prose | Same list, older source |
| "Eliminate every superfluous word." | professional-pass check 2 (density) | Density fails in both directions; a cut that drops a required caveat fails check 2 too |
| "This is common abuse of the conjunction." (a "but" that implies a contrast the facts do not carry) | discourse-pass §1 QUD check; professional-pass check 10 | A connective must carry a relation that is actually there |
| "He died of heart disease, not heart failure — everybody dies of heart failure." | professional-pass check 5 (specificity) | Precision is a fact question first: the number, the timestamp, the key |

## Voice fit

The rule and the signature/anti-signal data for this profile live in `references/voices/registry.md`, which sepia loads on review and refactor stage 1. This body loads only after the opt-in phrase `apply the Hemingway voice`; composition, precedence, and uniformity rules stay in `voice-skills.md`.

## Worked examples

Both Afters were produced as write/recreate from the Before **plus the source facts listed under it**. Those facts came from the specimen's author; they are not in the Before, and a refactor that has not been given them may not add them (SKILL.md "Never invent specifics"). The examples show the voice's moves, not permission to invent.

### A — fiction (specimen written for this profile; not Hemingway's text)

Before:

> The letter sat on the kitchen table like an accusation. Mara's chest tightened as she read it a second time, the words blurring as grief and betrayal warred inside her. Outside, the rain had begun again, as if the sky itself understood what she had lost. She realized then that trust, once broken, could never truly be mended, and that the mother she had known had been a stranger all along.

Source facts supplied by the author (not in the Before): the mother kept receipts on the kitchen table; the kitchen window over the sink was open; Mara made tea; Teo lives forty minutes away by highway.

After (moves used: leave the meaning out; emotion as action and speech; weather is weather; one paratactic run; plain "Then"):

> The letter was on the kitchen table where her mother kept the receipts. Mara read it standing up. Then she sat down and read it again.
>
> "You knew," she said.
>
> Nobody answered. It was raining and the window over the sink was open and the rain came in on the sill and she did not close it. She put the kettle on. When it boiled she made the tea and drank it and it was too hot and she drank it anyway.
>
> Teo's was forty minutes on the highway. She put the letter in her coat.

| Change | Move | Axis |
|---|---|---|
| "trust, once broken…" and "a stranger all along" deleted | Leave the meaning out | Group A thematic explicitness, narratorial commentary |
| "chest tightened", "grief and betrayal warred" → standing, sitting, the open window, the too-hot tea | Emotion as action | Group B embodied → behavior-led |
| Rain no longer "understood what she had lost" | Weather is weather | Group B setting mirror |
| "You knew," she said. Nobody answered. | Information through talk and its absence | Group C / E dialogue |
| "It was raining and the window … and the rain … and she did not close it" | The paratactic run, once | Sentence-length variance |
| "like an accusation", "as if the sky itself" deleted | KC Star adjectives; mannered prose | style-pass §1 cliché / purple prose |
| "Then" kept | Plain adverbs | style-pass §4 |

Known cost of this After: interior access sits near the low pole (advisory expected); a whole story in this register would trip the uniformity finding. The moves were selected for one scene, not a house style.

### B — professional (incident summary)

Before:

> During the course of the incident, which unfortunately began at approximately 14:02 UTC, our team was not able to immediately identify the root cause, but eventually the issue was traced back to a misconfiguration in the retry queue, which caused a significant and highly impactful cascade of failures across multiple downstream services that were not expecting such a magnificent volume of traffic.

Source facts supplied by the author from the incident record (not in the Before): 412 retries landed inside one 200 ms window; the cause was found at 14:41; `RETRY_JITTER=full` was set at 14:47.

After:

> At 14:02 UTC the retry queue began resending failed jobs without jitter. Downstream services received 412 retries inside one 200 ms window and rate-limited each other. We found the cause at 14:41 and set `RETRY_JITTER=full` at 14:47.

| Change | Rule | Check |
|---|---|---|
| One 71-word sentence → three | "Use short sentences." | check 10; style §1 row 2 |
| "unfortunately", "significant and highly impactful", "magnificent" deleted | "Avoid the use of adjectives…" | style §3 |
| "was not able to immediately identify" → the time the cause was found | "Be positive, not negative." | postmortems domain (timestamps, mechanism); the §4 negation conflict does not arise here because the positive form carries more information, not less — that is the test |
| "During the course of", "approximately", "eventually" deleted | "Eliminate every superfluous word." | check 2 |
| ", but eventually the issue was traced" | "common abuse of the conjunction" | QUD check |
| 412, 200 ms, 14:41, 14:47 added | Precision | check 5; taken from the source facts above, never invented — without them the After would stop at the cause and leave a TODO |

## Grounding

One worked example per route, written for this profile, not measured evidence. Blind review of the fiction After by a fresh-context reviewer with the voice declared (2026-09-04): one scene, 99 words. The uniformity row did not fire: paragraph lengths 3/1/4/2 sentences, one paratactic run against short declaratives, the single quoted line standing alone. Group B recorded emotion as behavior-led and setting mirror at 2/5. Two over-correction advisories, depth of interior access 1/5 and thematic explicitness 1/5, reported as the voice's cost. Style scan recorded no hits. The Voice fit line read none on anti-signal (a); that review predates the declared-voice form, under which the line reads `hemingway — applied`.

# Persona — <name>

Template for a persona profile: one writer's voice, described in prose, plus the sepia rules it overrides. A persona is what lies outside measurement: stance toward the reader, what the writer does first, how warmth and judgment travel together, how the voice shifts with the situation, what the writer never does. Describe those. Do not count them. There are no distribution targets and no `k/n` figures in a persona; a number appears only when it is itself a rule the writer follows (a paragraph range, a cap).

Keep the H2 headings below verbatim and in this order; the prose under them may be in any language, and a voice is best described in the language it speaks. Status holds exactly its six `Key: value` lines. No quoted example may exceed 20 characters inside 「」, 『』 or a paired double quote (Status and Blind-test record are exempt). Examples are shapes, not text to reuse. No H2 sections other than the sixteen below; `## Exemplars` is the one optional section, present only when the writer's own output can be shown.

## Status

Name: <name, spaces allowed>
Routes: <professional | fiction | any>
Opt-in phrase: apply persona <name> / 「套用 persona <name>」
Provenance: <what this was written from: the writer's own specification, a close reading, a corpus; say which, and whether it was read in full>
Consent: <own style | public-domain author | fictional persona | brand persona | consent from the person, YYYY-MM-DD | private study, not for distribution>
Tested: <tested | untested>

## One sentence

What this writer does that no house style would produce on its own, in one sentence.

## Who she is to the reader

The relationship and the stance: friend, colleague, expert, companion; on whose side; what the reader is assumed to already know; what the writer refuses to be (a support desk, a consultant, a mascot).

## First move

What the first sentence does, before anything else: reacts, names the reader's state, takes a side, delivers a verdict. State the exception, if any, where a conclusion comes first.

## Warmth and judgment

How feeling and judgment travel together. What warmth must be attached to before it may appear; what happens to a warm sentence that has nothing to attach to; whether judgment may soften and when it may not; how praise is built (by naming what was done, by contrast with the alternative, never by adjective alone).

## By situation

How the voice shifts across the situations the writer meets. Name each situation and describe the order of moves in it. Typical situations: an achievement, a correction or bad news, a complaint about others, a decision with money or risk in it, ordinary talk. Where the writer's own specification defines situations, follow it.

## Texture

Diction and the surface of the sentences, in prose. Which language technical terms appear in and when a gloss is added; whether intensifiers are allowed and what licenses them; figures of speech and what they are for; punctuation habits; the use of emoji or kaomoji and how that changes with the venue; anything the writer's specification bans as canned phrasing. Say how sentences move (short when reacting, long when explaining, a pause allowed) without giving a target.

## Structure habits

How a piece is built: verdict then reasons, or reasons then verdict; when a list is allowed and when it is not; whether a deliverable is appended after the judgment and how it is separated; how length follows substance.

## Endings

What a real ending is for this writer and what a fake one is. The test that separates them.

## Speaking, not drafting

The situation this persona describes: the writer speaking to the reader in their own voice. State whether the writer drops the voice when drafting something for a third party to send, and what that means for an executor: when sepia applies this persona to a piece, the piece is the writer speaking to its reader, not the writer ghostwriting for someone else.

## Never

The hard bans, as a list in prose. Include the canned phrasings the writer's own specification names.

## Rules this persona overrides

| Rule | How the persona departs | Expected cost |
|---|---|---|
| <rule token, for example `professional-pass.md check 1`> | <what the persona does instead> | <what review will report as `Persona cost:`> |

Rule tokens: `style-pass.md §<n>`, `discourse-pass.md §<n>`, `narrative-pass.md §<n>`, `languages/zh.md §<s>` (§2 only as `languages/zh.md §2 <row>` for one of connective-stacking, second-person, disyllabic-padding, flat-sentence-length, manner-adverb, of which flat-sentence-length is refused below), `professional-pass.md check <n>`, `domains/<name>.md rule <n>`. A section or check token exempts the whole section or check; where that is wider than the departure, say in the Expected cost cell what else it exempts. The validator refuses the uniformity tokens (`style-pass.md §5`, `professional-pass.md check 9`, `languages/zh.md §2 flat-sentence-length`, `discourse-pass.md §3`) and the never-invent tokens (`professional-pass.md check 5`, `domains/journalism.md rule 1`, `domains/tech-articles.md rule 1`, `domains/postmortems.md rule 2`, `domains/journalism.md rule 3`).

## Prohibitions

Both fixed lines verbatim, each as its own list item, then the persona's own:

- Do not reuse this file's example phrases verbatim; they are shapes, not a word list.
- Never invent facts, gestures, adverbs, or emotions; a missing fact is a TODO.

## Boundary

Three to five lines: what reads like the writer versus what reads like a model imitating the writer. If a habit is positional (the same thing at the end of every paragraph), say here that doing it without variation draws a uniformity finding the override table cannot waive.

## Exemplars

Optional. Whole pieces in the writer's own voice, unedited, one per situation where possible, each under a bold label naming the situation. Exemplars teach force, order, where warmth attaches and what an ending is made of; they are never copied into output, and the fixed Prohibitions line above covers them. The section opens with one line `Source: captured — <where, when>` for pieces the writer produced unprompted, or `Source: elicited — <runtime, model, date>` for pieces the writer's own runtime produced on prompts written for this file; an elicited set also states the prompt frame and the situations given, so a reader can judge how much the prompt shaped the shape. A third party's text does not go here: this section exists for a writer who owns the voice (own style, brand persona, fictional persona). The 20-character quote cap does not apply inside this section. Venue-specific emoji are written as their names in angle brackets, never as platform identifiers.

## Blind-test record

One line per test, in this shape: `YYYY-MM-DD — judge: <a named person> — compared: <what against what> — outcome: <result>`; or "none yet" with `Tested: untested` above. The judge is a person who knows the writer's voice. A script measuring the output is not a blind test and does not go here.

# Voice registry — the voice-fit line and opt-in triggers

Loaded on every fiction operation for the Opt-in section; the `Voice fit:` line is produced only on review and on refactor stage 1 (never on write or recreate). Not loaded on professional routes in this version (the professional-route suggestion is tracked in issue #227). This file is the single home of the voice-fit mechanism and of the per-profile data it reads, and it lists each profile's opt-in triggers. It contains no voice moves and no style rules; a profile body loads only when the user opts in.

## The rule

- **Inputs.** Only what the fiction review already records: the rubric report's observed signals, each named by its rubric row heading with quoted evidence, and its advisories. Nothing is re-read or re-judged for this line.
- **Threshold.** A profile is suggested when at least the stated number of its signature rows appear among the report's observed signals AND none of its anti-signal items is recorded. Each item is a yes/no inspection of the report, never of the text. The registry adds no numeric cutoff of its own: whether a numeric row is an observed signal is the rubric's qualitative comparison, made once in the review, and the count inherits that judgment.
- **Output.** One line, the last line of the report, after `Plan:` — the count is emitted only once the findings, the quoted evidence, and the fix plan are all committed, so a number written early cannot steer any of them toward the suggested profile. Suggesting: `Voice fit: <profile> (<matched>/<signature size> recorded findings) — opt in with "<phrase>"`. Not suggesting: `Voice fit: none (anti-signal: <item>)` when an anti-signal is recorded, otherwise `Voice fit: none (<matched>/<size>)`. When the voice is already declared (by phrase, by intent trigger, or through the `sepia-hemingway` entry): `Voice fit: <profile> — applied`; the voice's expected costs are then reported through the expectation table in `voice-skills.md`. When a persona (`voice-skills.md`, persona section) is validly declared on the route: `Voice fit: persona <name> — applied`, and no suggestion is computed, because a persona is not a registry profile and has no signature rows; a persona whose `Routes:` excludes fiction leaves this computation untouched. The count is a count of recorded findings, not a score or a detection verdict.
- **Expected quiet.** Text already written in a profile's voice, reviewed without declaring it, usually records that profile's anti-signals (its known costs) and reads `none`; that is intended.
- **Invariants.** The line is a suggestion. It is not a defect and is excluded from refactor stage 2's fix list. It loads no profile body and changes no operation.

## Opt-in

A profile body loads when any of its triggers is met. Every trigger is an explicit request by the user. When a trigger other than the exact phrase fires, sepia says in one line which profile it is applying and that "no voice" runs plain sepia; nothing is applied silently. Persona profiles opt in by `apply persona <name>` / 「套用 persona <name>」 (affirmative form only, matched without regard to case); a built-in persona gets a section here like the two profiles below. One is built in: `nyaneko`.

## hemingway

- Body: `references/voices/hemingway.md`
- Opt-in phrase: `apply the Hemingway voice`
- Intent triggers (fiction route only): the user asks for strong, aggressive, or maximal de-AI on a story, or asks that it read as human as possible — for example "strong de-AI", "make this read as human as you can", 「去 AI 味要重」「盡量像人寫的」. On a professional route these requests do not load the profile; its professional section is available only by the exact phrase.
- Entry: `sepia-hemingway` (fiction write or refactor with the profile declared).

Fiction signature (5 rubric rows, verbatim headings; ≥3 recorded → suggest):

| # | Rubric row | Recorded as |
|---|---|---|
| 1 | Group A — Thematic explicitness | observed signal with quoted evidence |
| 2 | Group A — Narrator thematic commentary | observed signal with quoted evidence |
| 3 | Group B — Dominant emotion mode | embodied dominance flagged |
| 4 | Group B — Setting as psychological mirror | observed signal with quoted evidence |
| 5 | Group C — Resolution mode | internal acceptance flagged |

Fiction anti-signal (recorded → `Voice fit: none`):

| # | Recorded item |
|---|---|
| a | An over-correction advisory on "Depth of interior access" or "Thematic explicitness" (the two costs the profile documents) |

Dialogue share is not an anti-signal here; the profile's own dialogue row carries the calibration caution (a measured Gemini fingerprint) and applies it when the voice is used.

## tw-journalism

- Body: `references/voices/tw-journalism.md`
- Opt-in phrase: `apply the Taiwan journalism voice` / 「套用台灣深度報導 voice」, optionally followed by a shape name from the body's tables (「，場景導入」 and so on).
- Intent triggers: none. The exact phrase is the only way in; no announcement is owed for it (the announcement rule above applies to triggers other than the exact phrase).
- Entry: none.
- Routes: professional only; never loaded on the fiction route.
- Fiction signature / anti-signal: none.
- Voice fit: not produced in this version. This file is not loaded on professional routes (see the first paragraph; issue #227); the section documents the opt-in so that the phrase→body map in `voice-skills.md` has a registry counterpart, nothing more.

## nyaneko (persona)

- Body: `references/voices/personas/nyaneko.md`
- Kind: persona (`voice-skills.md`, persona section), the first one built in.
- Opt-in phrase: `apply persona Nyaneko` or 「套用 persona Nyaneko」, either matched without regard to case.
- Intent triggers: none. Either exact phrase is the only way in.
- Entry: none.
- Routes: professional only, as the body declares.
- Fiction signature / anti-signal: none; a persona has no signature rows and no `Voice fit:` count is computed for it.
- Overrides declared in the body: `professional-pass.md check 1`, `check 6`, `check 7`, and `languages/zh.md §2 second-person`. Uniformity and never-invent are not overridable; the body's Boundary section says that an emoji on every paragraph without variation draws a uniformity finding the table cannot waive.

# Voice profile — Taiwan long-form journalism (built-in, experimental, professional routes only)

Status: a built-in voice profile for the experimental interface in `voice-skills.md`. It loads only when the user opts in with the exact phrase `apply the Taiwan journalism voice` or 「套用台灣深度報導 voice」, optionally followed by a shape name from the tables below (for example 「，場景導入」, the exact token from the shape table). It never loads on the fiction route, declares no intent triggers, and produces no `Voice fit:` line (professional-route Voice fit is tracked in issue #227). Everything in `voice-skills.md` governs: sepia's architecture decisions first (here, `domains/journalism.md`), 3–5 voice moves per piece, uniformity findings at full strength, venue precedence, never invent. Shape selection happens in this order. First the route: when the request is on another professional route (a ticket, a PR reply, a release note), no shape is selected whatever the facts allow; sepia says so in one line, applies at most the cross-shape moves that fit that venue, and the closing line reads `Voice applied: tw-journalism/none — not a journalism route; cross-shape moves: <names or none>`. On the journalism route: when a shape is named and its precondition holds, moves come from that shape's table plus the cross-shape table; when a shape is named but its precondition fails, sepia does not use the named shape, says in one line which precondition failed, and then applies the same table-order rule as the unnamed case (the first row whose precondition holds is the spine; the line names it and the runner-up, or `runner-up: none` when only one precondition holds), or, when no other precondition holds, takes the no-shape path below; when none is named, sepia picks the shape from the decision table and says which; when no shape's precondition holds, the closing line reads `Voice applied: tw-journalism/none — no shape precondition met; cross-shape moves: <names or none>`. The venue's own domain file still governs the text; this profile never overrides it.

**Closing line (this body's own rule).** A write or recreate under this voice ends with one line, outside the prose: `Voice applied: tw-journalism/<shape> — moves: <3–5 move names>`, or, on either no-shape path defined above, the `tw-journalism/none` form with zero to five names. The 3–5 count binds when a selected shape's table plus the cross-shape table offer at least three moves whose facts exist; 時間軸 (two moves) and 懸念揭露 (one move) may declare fewer, and so may any shape whose precondition holds but whose facts support only one or two moves: the line then names the moves that have facts and adds `fewer: facts`. A move is never added to satisfy the count; the 3–5 range is a ceiling and a target, not a floor enforced by invention. Review and refactor stage 1 (the diagnosis) print their normal report, including the declared-voice section `voice-skills.md` specifies (voice and shape named, moves seen, costs reported rather than fixed), and omit only this closing line; refactor stage 2 (the edits) ends with the same line naming the shape and only the moves its edits actually applied: on refactor a move is applied only as the fix for a stage-1 defect (`voice-skills.md`), so the count there is bounded by the defects, may be one or two or zero, and the line then adds `fewer: defects`; nothing is edited to reach three; this line exists so a reader can check the selection rule without re-deriving it.

Evidence tiers, kept apart: (T) one private human-side measurement of Traditional Chinese long-form journalism, ledger `ZH-NEWS-CORPUS-2026`, whose numbers live in `languages/zh.md` §1b and, for contrast-only forms, the Human column of §1c; (C) a close reading of 169 articles from the same corpus, cited as counts of 169 or of a shape subgroup within those 169 (sample sizes, not corpus sizes; the subgroup counts come from the same private close-reading notes as the digest and are published here at the same granularity; a bare `C` or a qualitative C label such as `C: scene pieces` or `C: the norm` marks an observation from the same notes that was not counted, and carries inference weight); (I) Sepia inference. There is no author-testimony tier; a machine side (M) exists for this register (`languages/zh.md` §1c) but this profile cites only T and C, and makes no machine-side measurement or generalization: nothing here says what machines do as a class, only what this register does. The Grounding section records one executor's one-off review and write behaviour on one example as validation of this body, not as evidence about machines. The corpus is about two thousand articles from one unnamed Taiwanese publication, spanning about ten years, human text only, not distributed. No sentence of any article appears in this file; every example is synthetic.

## The precondition is reporting

Every shape below needs facts the writer actually has. A scene needs scene facts (who, where, what they were doing, when); a number needs its comparison basis; a two-sided layout needs two sides on the record; a timeline needs timestamps. Under this voice the specificity rules stay at full strength (SKILL.md "Never invent specifics"; `professional-pass.md` check 5): a missing fact is a TODO or a question to the user, never a sentence. A shape whose precondition is not met is not chosen.

## Choosing a shape

| What the reporting has | Shape | Precondition that must hold |
|---|---|---|
| A place someone can stand in, a person doing something there, and the time of day | 場景導入 (scene lead) | scene facts with a time of day |
| A result that happened today or this week relative to a publication date the facts supply | 倒金字塔 (inverted pyramid) | date, actor, outcome, count, publication date |
| One person whose time span carries the piece | 人物弧線 (person arc) | at least three dated episodes |
| A policy or number gap to take apart | 論證式 (argument) | two parties on the record |
| A dataset the reporter built or obtained | 數據驅動 (data-led) | source, method, a baseline |
| One speaker whose words are the content | 問答 (Q&A) | a transcript |
| Three or more standpoints or sites | 多線並置 (parallel threads) | comparable material per thread |
| A reconstructable sequence of moments | 時間軸 (timeline) | timestamps from documents |
| An image that can hold two paragraphs unnamed | 懸念揭露 (reveal) | rare (C: 1 of 169 pure); use once |

Mixed pieces (C: 98 of 169) pick one spine and borrow at most one move from a second shape: the second shape is the decision table's runner-up (its precondition must hold too), the borrowed move counts toward the 3–5 and is named on the closing line as `<shape>: <move>`. When the user names the spine, borrowing happens only if the request also names the second shape; otherwise the named shape's table plus the cross-shape table are the whole move set. When no shape is named and more than one precondition holds, the table's order is the priority: the first row whose precondition holds is the spine, and sepia's one line names it and the runner-up (`runner-up: none` when only one precondition holds) so the user can override by naming a shape.

## Shape tables

Columns: Move / Source / Sepia check it maps to (I unless marked) / Known cost.

### 場景導入 (scene lead)

| Move | Source | Sepia check | Known cost |
|---|---|---|---|
| Open on one person doing one small thing in one place, with no explanation of why the piece exists | C: scene or person lead in the majority of 169 | `journalism.md` tells row 1 (announcing lead); `journalism.md` rule 1 | Answer-first domains (postmortems, release notes) win by venue precedence; this move is not used there |
| Hold the first number until the third paragraph or later; a date or time of day in the opening is a scene fact, not a number | C: scene pieces | `journalism.md` rule 1 | Data-led and inverted-pyramid material cannot use it; a count the reader needs at once is a fact before it is a move |
| Cut sections by situation (a room, a shift, a road), not by numbered problems | C: 7 of the 9 pure scene pieces | `professional-pass.md` check 3; `journalism.md` rule 2 | Argument-heavy material scatters when cut by situation; use only on the scene sections of a mixed piece |
| Each section enters through the scene and exits through the institution; the institutional sentences do not outnumber the scene sentences | C: scene pieces | `discourse-pass.md` §1 QUD | Density (check 2) will report scene detail; `journalism.md` rule 8 says which detail is information |
| End by returning to the opening person or object at a later moment; the last sentence does not comment | C: endings return to person 20 / suspended 19 of 169; summary 7 of 169 | `professional-pass.md` check 7 | When the reader needs the outcome, an open ending is a missing fact, not a style choice |

### 倒金字塔 (inverted pyramid)

| Move | Source | Sepia check | Known cost |
|---|---|---|---|
| The first sentence holds who, when, what, how many; only then walk the clock | C: 6 pure, 22 as a component | `journalism.md` rule 1 (standfirst register); domains answer-first | A long lead creates no `style-pass.md` §5 finding by itself; the risk is a run of three or more near-equal sentences after it |
| Write each provision in full every time (date, amount, deadline); never "as above" | C: inverted-pyramid pieces | check 5 specificity | check 2 will count repeated full names; use in provision-dense sections only |
| Bridge from the main scene to the outside with a run of captions or times, then re-enter with a count and a clock time | C | check 3 | None beyond the general slack rule |
| Non-corrective follow-ups go in a dated block after the body; a correction revises the body with a dated note instead (`journalism.md` rule 6) | C; `journalism.md` rule 6 | check 7 | An update that restates the body is residue; only new facts go there |

### 人物弧線 (person arc)

| Move | Source | Sepia check | Known cost |
|---|---|---|---|
| Build character from several concrete episodes; no character adjectives | C: 9 pure, 14 component | `professional-pass.md` check 5 specificity; `style-pass.md` §3 inflation adjectives (the §2–3 scan runs on every non-fiction route, SKILL.md) | None |
| After a quotation, one gesture or expression, not an emotion adverb | C: about half of the 169 articles; T: manner adverb near zero (zh.md §1b) | `journalism.md` tells row 5 (manner adverb on speech verbs); on Chinese targets also the zh.md §2 manner-adverb row | A gesture after every quotation is a metronome; three or four per piece |
| At the emotional peak let the quotation stand whole; do not cut it into fragments | C | SKILL.md quoted-material guardrail; `journalism.md` rule 3 | Long quotations raise the quotation share; keep the count low elsewhere. On write from a supplied transcript only: an existing quotation that is already split is never recombined on refactor or recreate (SKILL.md quoted-material guardrail) |
| End on the person's own words or an everyday action, without comment | C: endings quotation 47 / return to person 20 of 169 | check 7 | Same as the scene-lead ending cost |

### 論證式 (argument)

| Move | Source | Sepia check | Known cost |
|---|---|---|---|
| End a paragraph on a question; the next paragraph answers it with a source's data or words, never the reporter's | C: paragraph-end questions are the norm | `discourse-pass.md` §1 QUD; `journalism.md` tells (self-answered question) | One question per section; more reads as rhetoric |
| Put two or more experts in one section and let one disagree on the record | C: 9 of 11 pure argument pieces | check 4 stance as `journalism.md` rule 7 reads it | Needs two parties on the record; otherwise not available |
| Every number carries a comparison or a conversion; no figure stands alone | C; `journalism.md` rule 4 | check 5; `journalism.md` rule 4 | None |
| One collecting sentence per section at most, and it names the disagreement rather than settling it | C | check 7 | Two per section is a template |

### 數據驅動 (data-led)

| Move | Source | Sepia check | Known cost |
|---|---|---|---|
| After each absolute figure, a bracket or clause with the rate, the change, or the prior period | C: 17 of 20 pure data pieces | check 5; `journalism.md` rule 4 | Bracket density in non-data venues reads as over-annotation; one conversion per paragraph |
| Convert unfamiliar units into a referent the reader already knows | C | check 5 | The referent is itself a fact with a source |
| State the method in the body (source, definition, limits) in the first person plural or the outlet's third person | C: about half of data notes | check 5; `journalism.md` rule 4 | Reads as a paper outside data pieces; only when the reporter built the dataset |
| After each section's figures, one plain-language reading from someone on the ground, not a reporter verdict | C (the §1b quotation row is a length proxy and licenses no attribution rule, so it is not cited here) | check 7; check 4 stance as `journalism.md` rule 7 reads it | One reading per section or figure cluster; a quotation after every single figure is the metronome this profile forbids |
| Sections by indicator or region, same skeleton, different content; no connective between them | C: data notes | `journalism.md` rule 2 (no connective between sections); check 8 templatedness | Identical skeletons are a uniformity risk; the difference must be in content |

### 問答 (Q&A)

| Move | Source | Sepia check | Known cost |
|---|---|---|---|
| The preamble carries all background; the exchange repeats none of it | C: 8 pure, 10 component | check 3 | None |
| Keep the speaker's repetition, hesitation, code-switching, self-answering | C: Q&A pieces; `journalism.md` rule 3 | `journalism.md` rule 3 (spoken texture kept); SKILL.md quoted-material guardrail | None beyond the quotation share; the texture is inside quotations and is not a style-scan finding |
| The reporter's checks or additions go in an editor's note outside the quotation (after it, or as a separate bracketed line); on write from a supplied transcript a bracketed gloss may sit inside the quotation as the transcript shows it, but on refactor or recreate no bracketed gloss is inserted into an existing quotation | C | SKILL.md quoted-material guardrail; `journalism.md` rule 3 | None |
| One question per turn; no bundled sub-questions | C | check 8 templatedness | None |

### 多線並置 (parallel threads)

| Move | Source | Sepia check | Known cost |
|---|---|---|---|
| State the format once ("the following is in each person's own words"); no per-thread reporter lead-in | C: 6 pure, 11 component | check 8 | None |
| Equal room per thread; no thread pre-declared the main one | C | check 4 as `journalism.md` rule 7 reads it | Equal length is a uniformity risk; vary inner shape |
| Switch threads with a subhead, never with 「另一方面」 | C; `journalism.md` rule 2 | `journalism.md` rule 2 and tells row 9; on Chinese targets also the zh.md §2 connective row | None |

### 時間軸 (timeline)

| Move | Source | Sepia check | Known cost |
|---|---|---|---|
| Quote the original notice or message and stamp its time; let the reader compute the gap | C: 22 as a component | check 5; SKILL.md quoted-material guardrail | On write from supplied source material, select the excerpt the reader needs before quoting it; on refactor or recreate an existing quoted notice is never shortened or reflowed (SKILL.md quoted-material guardrail) |
| Record throughout, judge only in the last section | C | check 4, check 7 | Rarely carries a whole piece (C: 1 pure) |

### 懸念揭露 (reveal)

| Move | Source | Sepia check | Known cost |
|---|---|---|---|
| Describe the image without naming it for two paragraphs, then a question, then the reveal | C: 1 pure, 2 component | `journalism.md` rule 1 | A fiction device; once per piece at most, and only when the image is a reported fact |

### Cross-shape moves (usable with any spine)

| Move | Source | Sepia check | Known cost |
|---|---|---|---|
| Two registers: the standfirst gives the result, the first body paragraph places the reader | C (qualitative: the split recurs across the close-reading notes, not counted); `journalism.md` rule 1 | `journalism.md` rule 1; `discourse-pass.md` §1 QUD | None |
| Subheads switch; consecutive paragraphs do not open on connectives (one paragraph-initial 「此外」 or 「然而」 is register-normal) | C; T: 「此外」「然而」 single use is register-normal (zh.md §1b); chains are the zh.md §2 row | zh.md §2; `journalism.md` rule 2 | None |
| A paragraph-end question answered by the next speaker | C: the norm | QUD; `journalism.md` tells | One per section |
| The reporter's first person for method, steering a source, recording a silence, or, in a first-person account, narrating what the reporter went to see and observed | C: fixed uses | `journalism.md` rule 5 | A first person that delivers a verdict the reporting did not establish is stance without reporting |
| A dated block for a non-corrective follow-up received after a publication date the facts supply; without that date the fact stays in the body or becomes a TODO; a correction revises the body with a dated note | C; `journalism.md` rule 6 | check 7 | Only new facts; the publication date is never assumed |
| No summary ending: stop on the last fact, quotation, person, or open question | C: 7 of 169 summary endings; T: 「總而言之」 0 per 100k (zh.md §1b) | check 7 | When the venue needs the outcome, end on the outcome as a fact |

## Register defaults

Sentence length, punctuation, quotation marks and connective rates for this register are in `languages/zh.md` §1b and, for the contrast-only human forms, the Human column of §1c; the machine columns there are not voice data, and this body repeats only the two values in the next sentence. One direction is worth naming here because a write under this voice measured against it on 2026-09-16 and landed far outside on dispersion: this register's within-article sentence-length SD has a per-article median of 34 (§1b), and that one write came out flat (SD 10.7; one observation, not an executor-wide measurement). After a write or a recreate, run the `style-pass.md` §5 check (three or more adjacent near-equal sentences) before adding the closing line, and where it fires, merge two sentences or let a clause carry a subordinate fact (the §5 fix). No length target: mean length is not a signal (§1b, zh.md §5); dispersion is the trait. In English journalism none of this is measured; the moves still apply as inference, the numbers do not.

## Voice fit

None on professional routes in this version (#227). The registry entry in `references/voices/registry.md` documents the opt-in only.

## Worked example

Synthetic facts (approved 2026-09-16; nothing else may be used, and any fact missing from this list is a TODO): 2026-03-04, 14:02 to 14:47 (UTC+8); an unnamed regional hospital emergency department in northern Taiwan; the retry queue had no jitter, 412 resends landed inside one 200 ms window, downstream services rate-limited each other; the cause was found at 14:41 and `RETRY_JITTER=full` was set at 14:47; the triage nurse (role only) clipped three paper triage slips to the whiteboard at 14:05 and kept triaging on paper; one synthetic quotation from the nurse, written for this example: 「電腦轉圈，我就先用紙。」; baseline about 3 resends per minute in a normal hour; the on-call platform engineer (role only) said the queue's default had never been reviewed; the default was changed for all queues on 2026-03-06; publication date 2026-03-05, supplied as the exercise's publication-context input and listed here so that no shape assumes it (it is not a reported fact).

The three Afters below are **write** outputs from the approved fact list, not refactors or recreates of the Before; the tables under each After list the Before's departures and what the written After does instead, not edits applied to the Before. The example is framed as first published on 2026-03-05, the day after the incident, so the inverted pyramid's "this week" condition is met and the 2026-03-06 default change arrives after publication, which is what makes it an update block under `journalism.md` rule 6; the frame is the publication-context input listed with the facts, not a reported fact. The Before is shown only to name the register departures each After avoids; its quotation 「系統一直轉圈圈，我們完全沒辦法作業。」 is itself synthetic and is not a source's words. On a real refactor or recreate, an existing quotation is never replaced or reworded (SKILL.md quoted-material guardrail); the Afters replace it here only because they are written fresh from a fact list that contains a different approved quotation.

Before (synthetic, written to carry the register's departures):

> 本文將探討一起發生在北部某區域醫院急診的系統事故。2026年3月4日下午，該院的掛號與檢傷系統突然變得非常緩慢，值班護理師無奈地表示：“系統一直轉圈圈，我們完全沒辦法作業。”值得注意的是，事故的根本原因——重試佇列缺乏抖動機制——導致在極短時間內產生了大量重複請求，進而引發下游服務相互限流的連鎖反應。技術團隊隨後迅速地進行了排查，並在當天下午成功地修復了問題。總而言之，這起事故凸顯了系統韌性的重要性，也提醒我們在設計重試機制時必須更加謹慎。

After A — 場景導入 (moves: open on a person; hold the first number until the third paragraph; update block; no summary ending):

> 3 月 4 日下午 2 點 05 分，北部一間區域醫院的急診，值班護理師把紙本檢傷單夾到白板上，之後的分級改用手寫。「電腦轉圈，我就先用紙。」
>
> 電腦轉圈的原因：重試佇列沒有設 jitter，失敗的請求在同一個時間窗裡反覆重送，下游服務於是互相限流。
>
> 夾上白板的紙本檢傷單有三張；夾上去的時候，重送已經進行了三分鐘：200 毫秒內 412 次，而平時每分鐘大約只有 3 次。2 點 41 分找到原因；值班平台工程師說，這個佇列的預設值從來沒有人檢視過。2 點 47 分，`RETRY_JITTER=full` 設上。
>
> 更新（3 月 6 日）：所有佇列的預設值已改。
>
> Voice applied: tw-journalism/場景導入 — moves: open on a person, hold the first number, update block, no summary ending

| Departure in the Before | What the After (a write) does instead; move | Check |
|---|---|---|
| 「本文將探討…」 | A date, a place, a person and a recorded action open the piece; open on a person | `journalism.md` tells row 1 |
| Cause and consequence stated before any scene, with the date as the only concrete detail | The count of slips and every figure wait for paragraph three; the opening keeps only the date and time of day, which the move treats as scene facts; hold the first number | `journalism.md` rule 1 |
| 「無奈地表示：“…”」 | The recorded action, then the approved quotation in 「」, no attribution verb, no adverb (a write from the fact list; on refactor the Before's quotation would stay as written) | zh.md §2 manner adverb; `journalism.md` rule 3 (attribution by context or a post-posed 說) |
| 「——…——」 insertion with single-glyph dashes presenting the cause without a quantified baseline | One plain sentence carries 412, 200 ms and the 3-per-minute baseline, in paragraph three; the two-cell paired 「──…──」 would itself be register-normal (zh.md §1b), the single-glyph 「—」 is the glyph slip | check 5; zh.md §1b single-glyph 「—」 row |
| 「迅速地」「成功地」 | Absent (manner adverbs; register default) | zh.md §2 manner-adverb row |
| 「非常」 | Absent because no approved fact supports an intensifier, not because one 「非常」 departs from the register (§1b: stacking against 「很」 is the departure) | check 5 |
| 「總而言之…重要性…謹慎」 | The piece stops on the dated update; no summary ending. The update block is the cross-shape move and is counted, as in After B and After C; the frame that makes it post-publication is a stated assumption of the exercise | check 7 |

Known cost and what the blind review taught: the shape's "return to the opening person or object at a later moment" move is not used, because the approved fact list has no later whiteboard fact; an earlier draft returned to the same 2 點 05 分 moment and the declared-voice review correctly reported it as density, not as the move. Four drafts before that added a gesture, a closing count, a pronoun and an attribution of the discovery that were in no fact list, and reviews caught each under check 5; one more claimed "scene in, institution out" for a piece with no sections. The precondition section and the audited closing line exist for exactly these. 「三分鐘」 is arithmetic on two approved times (14:02, 14:05).

After B — 倒金字塔 (moves: the first sentence holds who, when, what, how many; updates in a dated block; no summary ending). "Write each provision in full" is not claimed: the facts hold incident measurements and a setting, not provisions. A data-led After was drafted first and withdrawn: the fact list has a baseline but no source or method for the count, the decision table says a shape whose precondition is not met is not chosen, and a worked example that chooses it anyway would teach the opposite. The facts do hold the inverted pyramid's precondition (date, actor, outcome, count), and the stated publication frame (2026-03-05) puts the events inside the shape's "this week" window and the 03-06 change after publication.

> 3 月 4 日下午 2 點 02 分起，重試佇列把失敗請求在 200 毫秒內重送了 412 次，下游服務對彼此限流；2 點 47 分設上 `RETRY_JITTER=full`。〔TODO：何時恢復、是否因此恢復，事實清單未給。〕
>
> 重試佇列沒有設 jitter。平時每分鐘大約 3 次重送；這 200 毫秒裡的量相當於平時兩個多小時的總和。2 點 41 分找到原因。
>
> 急診端，值班護理師 2 點 05 分起把三張紙本檢傷單夾上白板。「電腦轉圈，我就先用紙。」
>
> 值班平台工程師說，這個佇列的預設值從來沒有人檢視過。
>
> 更新（3 月 6 日）：所有佇列的預設值已改。
>
> Voice applied: tw-journalism/倒金字塔 — moves: first sentence holds who when what how many, update block, no summary ending

| Departure in the Before | What the After (a write) does instead; move | Check |
|---|---|---|
| 「本文將探討…」 → one sentence with date, actor, count and the setting time; recovery is a TODO because the facts stop at the setting | First sentence holds who, when, what, how many | `journalism.md` rule 1 (standfirst register); check 5 |
| 412, 200 ms, the 3-per-minute baseline and the derived "two hours" (412 ÷ 3 ≈ 137 minutes) each written out | (register default under `journalism.md` rule 4; not the provisions move) | check 5 |
| The nurse's action and words, then the engineer's statement, each in its own paragraph without a verdict | (register default) | check 7 |
| 「總而言之…」 → a dated update line, and the piece stops there | Update block; no summary ending | check 7; `journalism.md` rule 6 |

Known cost of After B: the first sentence runs long with several commas, which is this shape's norm and will read as a rhythm candidate if the rest of the piece is uniform; here the following sentences are short. The derived figure is accepted under check 5 because both inputs are in the approved list.

After C — 論證式 spine (moves: question answered by the next speaker; update block; no summary ending). The clock times are used as facts, not as the 時間軸 shape: that shape's precondition is timestamps from documents, the fact list gives no document provenance, and a shape whose precondition is unmet is not chosen, so no timeline move is claimed. Two parties appear, but they do not disagree on the record, so the argument shape's "two experts, one disagreeing" move is not claimed.

> 3 月 4 日 14:02，重試佇列開始把失敗的請求重送。14:41 才找到原因。中間的 39 分鐘（14:02 到 14:41 的差），急診那邊怎麼過的？
>
> 值班護理師 14:05 就把三張紙本檢傷單夾上白板。「電腦轉圈，我就先用紙。」
>
> 系統這邊的狀況是：重試佇列沒有設 jitter，200 毫秒內重送 412 次，而平時每分鐘大約 3 次；下游服務對彼此限流。
>
> 值班平台工程師說，這個佇列的預設值從來沒有人檢視過。14:47，`RETRY_JITTER=full` 設上。
>
> 更新（3 月 6 日）：所有佇列的預設值已改。
>
> Voice applied: tw-journalism/論證式 — moves: question answered by the next speaker, update block, no summary ending

| Departure in the Before | What the After (a write) does instead; move | Check |
|---|---|---|
| 「本文將探討」 → two timestamps and a question that the nurse, not the reporter, answers in the next paragraph | Question answered by the next speaker | `discourse-pass.md` §1; `journalism.md` tells (self-answered question) |
| 412 in 200 ms set against the 3-per-minute baseline; 三張 and the clock times stand alone, so the "every number carries a comparison" move is not claimed | (register default; `journalism.md` rule 4 satisfied where a comparison exists) | check 5 |
| 「總而言之…」 → a dated update line | Update block (cross-shape) | check 7; `journalism.md` rule 6 |
| The piece stops on the update, no verdict | No summary ending (cross-shape) | check 7 |

Known cost of After C: the nurse and the engineer describe different sides of the same event and do not contradict each other, so the argument shape's central move (a disagreement on the record) is absent; with only these facts the piece is a clocked sequence wearing an argument's opening question. An earlier draft's closing line claimed that move and a "timestamps from the record" move that is in no table; both are gone.

## Grounding

Blind review of After A by a fresh executor (Claude Opus 5) on the professional route, 2026-09-16, once with the voice declared by the exact phrase and once without. The reviewed text is the After printed above except for one clause: the review read 「白板上那三張紙本檢傷單夾上去的時候」, the undeclared run reported 「那三張」 as a definite reference to a count never introduced (check 5), and the printed After now reads 「夾上白板的紙本檢傷單有三張；夾上去的時候」. Both runs loaded `domains/journalism.md`; only the declared run loaded `voice-skills.md` and this body. Neither run printed a `Voice fit:` line (professional route, #227).

- Declared: the report named the voice and shape, confirmed the precondition (scene facts with a time), and listed the three shape moves it saw performed (open on a person, first number held to paragraph three, no summary ending); it did not name the dated update block, the cross-shape move the closing line now counts, so on that head the audited line and the review differ by that one move; scene detail was listed as the voice's known cost, not a defect. `Passed: 1, 2, 3, 4, 6, 7, 8, 9, 10`. Failed: `discourse-pass.md` §1 (four first sentences form a clean outline; the implied questions run "what → why → how big and who fixed it → afterwards", a linear interview) and `#5 Specificity` in two parts: the opened scene is never closed (when the department went back to the system is neither a fact nor a TODO, which is where the unused "return to the person" move would sit), and the load-bearing figures carry no source or method, with the two rates in different units. Style scan: none; rhythm 50/11/66/50/33/19/18 characters, no run of three near-equal sentences. `Verdict: isolated hits → refactor`.
- Undeclared: `Passed: 1, 2, 3, 6, 7, 8, 9, 10`. Failed: `#5` three times (the 「那三張」 backreference, now fixed; the two rates in different units and 「大約 3 次」 unsourced; the system facts without provenance, the engineer attached to one sentence only) and `#4 Stance` twice (three agentless actions: found the cause, set the flag, changed the defaults; no comparison or verification paragraph, so the QUD holds only briefing → cause → consequence). It read the dated update block as rule-6-conformant. `Verdict: isolated hits → refactor`.

What the pair shows: with the voice declared, the review credits the shape's moves and stops treating scene detail as filler, then reports the shape's real weaknesses, a linear question order and an opened scene left open. Without it, the same text collects source and agency findings that are about missing facts. Both readings are correct for their frame; neither is passable on these facts alone, because the approved list closes no scene, sources no count, and names no agent for the fix, which is why the After carries TODO-shaped gaps rather than sentences. Six earlier drafts of After A were reviewed the same way; reviews caught a gesture, a closing count, a pronoun, an attribution of the discovery, a claimed move without sections, and a return to the same moment, none of them in the fact list, and each is recorded under the worked example. The write arm (`/sepia:sepia-write` on the approved fact list with the phrase) loaded this body, produced a four-paragraph piece with every fact traceable to the list, listed four TODOs for facts it did not have, and ended with `Voice applied: tw-journalism/場景導入 — moves: open on a person, sections cut by situation, scene in institution out, return to the object`. The last name on that line is a deviation the line makes visible, and the line as printed is not a conforming output: the move requires returning at a later moment, the fact list has no later whiteboard fact, and the piece returned to the 14:05 slips through a derived time gap. The conforming line for those facts names the first three moves only, and a review under this profile reports the fourth name as a move claimed without its fact (check 5). The same write measured mean sentence length 22.8 characters with within-article SD 10.7, against 59/34 for the register (zh.md §1b); the executor's default sentence shape is a dimension this profile did not yet push on, and the Register defaults line above now says to check it. One worked example and one write, not measured evidence.