The register study — what the popular titles actually read like, in two bands phase 2 · 2026-09-08

the Chinese findingThe same shape as the English one, and one more. 财新 runs forty characters a sentence; Ink's Chinese runs thirteen to twenty-six, so the periodical is the longer-sentenced writer here too, as the owner's ear said on 09-04. But 财新 has almost no person in it: 我 and 我们 at 0.06 per hundred characters, against 1.2 in 信用的成本 and 2.8 in the Confucius riddle. That is exactly why 财新 is the wrong yardstick for what the owner asked of 信用的成本: its register is institutional calm, and the blunt, first-person column the owner wants is a different animal. Ink also leans harder on 四字格 than the periodical does (0.9–2.4 against 0.9), the Chinese cousin of the English metaphor-noun density. The floor is set from the columnists, not the periodical. The owner named two 公众号 readers actually finish, 六神磊磊读金庸 and 兽楼处, both with tens of millions of readers, both known as 不装逼 and 真性情, and both famous for being critical and humorous at the same time, where 财新 is the serious, high-quality journalism. That is the Chinese version of the two English bands in one line: the periodical is the shape of seriousness, the columnists the shape of being read. 46 of their articles were found through Sogou's article index and measured the same way, and their openings now ship as the Chinese shelf, so a Chinese commission is shown its own writers and never 财新's calm. Both run twenty-five characters a sentence, against 财新's forty-one: the writing people finish uses sentences half the periodical's length, which reverses the 09-04 reading of「compact Chinese wants longer sentences」, a reading taken against 财新. Both carry more person than 财新 (我 0.2–0.7 against 0.06), and 六神磊磊 leans hard on 四字格 (1.85 per hundred), which is his subject's idiom. The Chinese floor, ink_lint.REGISTER_ALLOW_ZH, takes the columnists' 75th percentiles as ceilings (28 characters a sentence · 四字格 2.0 · 文言 1.75 per hundred) and their 25th as the person floor (我 0.2 per hundred). Ink's own Chinese pieces sit inside it on sentence length and person; the fables and the riddle run over on 四字格.

the findingThe reading level is not where Ink leaves the shelf. The density of images is. The majors read at grade 10–11 with 20–23 words a sentence: harder and longer than Ink's own pieces, which read at grade 7.5–10 with 12–17 words a sentence. The old「long sentences over 20 words」flag was stricter than the standard it was meant to enforce. But the majors' metaphor nouns run 0.2–0.6 per thousand words, and Ink's run 2–6: ten times the density of doors, machines, ledgers and maps. The owner's complaint (obscure, like a paper, hiding the point) was never about syllables; it was about pictures standing where plain words should be.

the third finding, with the New Yorker inThe literary band is not the harder one. It is the more personal one. 125 New Yorker pieces were read through the owner's subscription on 09-08 (Talk of the Town, Shouts & Murmurs, Personal History, Annals of Inquiry) and the band grew from 38 pieces to 161. The New Yorker reads at grade 8.5 to 11, Shouts at 8.7, Personal History at 8.5: no harder than the newsletters, and easier than the Guardian. What sets it apart is voice. First person runs 1.7 per hundred words across the band and 5.0 in Personal History, against 1.1 in the plain band and 0.3 at the Conversation. Craft, on the page, is a person talking, not syllables. The old assumption that the literary band would be「looser where craft is expected」was wrong on grade and long words and right only on images.

the targets, set 2026-09-08Each band's ceiling is the bank's 75th percentile on each meter, so a piece is flagged when it reads harder than three quarters of the real pieces in its band:

Band	grade ≤	words / sentence ≤	long words / 100w ≤	metaphor nouns / 1k ≤	first person / 100w ≥
plain (n = 106)	13.3	22.9	7.4	1.2	0.3
literary (n = 161)	12.1	23.8	6.5	2.3	0.6

In ink_lint.REGISTER_ALLOW; the mapping of formats to bands is ink_brew.FORMAT_BAND. One caveat remains: the Conversation reads at grade 14 with academic authors, which pulls the plain band's grade ceiling up; a general-reader newsletter (Axios, Morning Brew) would pull it down, and both refused the crawler.

the mitigation, shipped 09-08Three moves, one per family. The echoes are fixed at the source: the brief no longer says「plain words」or「genuinely」; it says「everyday words」and「what you know」, because a word in the brief is a word in the piece. The nouns and the verbs are counted and handed to the editor: the lint's new ai_echo and ai_nouns categories ride the plain-words pass, which is told to replace a figurative use and leave a literal one (a real door stays a door). The antithesis turn is chased into the title and the dek, where the body lint never looked:「The Anthropic Settlement Is Not a Victory. It Is a Tollbooth.」and「a mirror, not a map」are now caught by antithesis_hits and the piece is retitled on the writer seat; and the plain「not X, but Y」in the body is a tell at one occurrence, not two. The no-announcing-distance rule, cut in Phase 0, is back in positive form, since three of sixteen pieces opened with「I was not in the room」the day it was gone. The per-word rates above are the regression test: re-run this table after a brew and the ×-column should fall.