# 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. Loads per `SKILL.md` Routing. 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; how they are attributed is rule 3's editorial guidance, not a measured baseline. 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 the user supplies a transcript, 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 guardrail) | 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 |
| Figures the reader cannot size, without a comparison or a local referent | Give each one a baseline, a prior period or a referent the reader knows, in the figure's own unit or with the conversion written out; rule 4 says which figures need one | close reading: data pieces pair their figures 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 with a plain attribution verb (Traditional Chinese: 「說」, the `languages/zh.md` attribution-verb row; English: "said"), placed before or after the quotation as the venue places it; terminology in quotation marks is not speech and is not counted as a quotation; the speaker's words are never rewritten, except from a transcript the user supplies (the quotation rows above; SKILL.md quoted-material guardrail). 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 whose size the reader cannot judge on its own (412 resends) carries a comparison (baseline, prior period, or a referent the reader knows), stated so the reader does not convert: in the figure's own unit, or with the conversion written out; a figure the reader can size at sight (three slips, 14:05) needs none. When the reporter built the dataset, the method is 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 (`professional-pass.md` Domain weighting), then rules 7 and 8 as the venue's reading of checks 4, 2 and 3.
