The first read — what the owner found in the box editions recorded 2026-09-06fixes shipped 2026-09-07 · unread

The Ink deploy plan ended with「the owner reads the first brews, then the timers」. This is that read: the 09-04, 09-05 and 09-06 editions on chat.xbbapp.com (34 pieces, 32 in English), read on the phone. Each defect below carries the owner's words, the measurement over every English piece on the box, the line in the brew that causes it, and the fix owed. The daily-brew timers stay off until these are fixed and re-read.

Too calm, too third person, too indifferent. Obscure, difficult language that keeps the reader from understanding. Assumes the reader is an expert and doesn't care to explain. Overall, it doesn't feel that the writer is alive. 太文人的酸臭和装逼。the owner, 2026-09-06, after reading the box editions

1 · The defects

#	Defect	The owner's words	Pieces named	Measured
D1	The model's metaphor lexicon	AI uses words humans don't use that often — door / gate, ledger / account — as metaphor.	The Sky's Ledger · The First Door · Churchill's gate ×3 (08-31 edition)	machine 25 · engine 21 · room 16 · wall 15 · arithmetic 14 · door 12 · map 10 · register 7 · account 7 · ledger 6 — across 32 pieces. Median 6.7 such nouns per 1,000 words; the Nadal piece runs 1.2.
D2	The bystander tone — reads like literary analysis	The writer writes in a bystander tone, which makes the article read like a literature analysis.	The Slow Burn of the Brain · The Sky's Ledger. positive The Williams Show Returns	Median first person 1.1 per 100 words (Nadal 2.6). 24 of 32 titles begin with「The」; 26 of 32 are a noun phrase with no verb (The Arithmetic of Defeat · The Cave of the Feed · The Cartographer's Decagon · The Committee's Red Line).
D3	Obscure wording	「To the Engineer Who Cut the Power」is technical.	To the Engineer Who Cut the Power	Long words (10+ letters) 8.7 per 100 — the highest on the box (median 3.7); the piece says「denominated」,「motive power」,「a store and a mill」. The Sky's Ledger runs 6.0 with「verily」,「peruse」,「occultation」unexplained.
D4	No explanation; assumes an expert	It lacks easy-to-understand explanations and analogies to help the user get the concept right. The topic is fine and interesting if the writer actually spends time explaining it well, and then gives his comment.	The Committee's Red Line (CFIUS)	Explaining moves (that is · for example · think of · like a · a parenthesis) 1.4 per 1,000 words; CFIUS is never expanded; the first paragraph already runs on「the factor the Committee weighs」. Nadal explains the scoreline before commenting: 6.2.
D5	No emotion	「信用的成本」is good because I know the background. I wish it were more 劲爆 and blunt; Li Ka-shing should be very frustrated. I don't see the emotion.	信用的成本 (Li Ka-shing, zh original)	The piece states its case as observation (「我不辩护」「我已退休，不在决策席上，不评论这笔交易本身」); the one heat it allows is a proverb-shaped close. In English: 31 of 32 pieces have no exclamation mark at all; feeling words median 1.6 per 1,000 (Nadal 5.0).
D6	文绉绉 — the point hidden behind the prose	A good topic for the parallel between film and novel, but the writing is too 文绉绉. It reads like the author has a point but wants to hide it behind obscure text so only the patient and tolerant find it.	What the Page Keeps from the Camera	Longest sentences on the box (21.9 words a sentence, median 13.3); the claim arrives in the last paragraph;「the time a reader takes to let desire become real inside a sentence」in paragraph one.
D7	Too calm, too cool — no stake in the subject	The point is clear but the tone is too calm and too cool, as if there is a job-seeking phenomenon and the writer only wants to analyse it, with zero emotional concern about the topic.	The First Door	Metaphor nouns 19.6 per 1,000 (door · threshold · room, again and again); first person 1.0; feeling 2.0; a 162,000-jobs peg turned into「the self each graduate dares to bring into the room」.

2 · The measurement — every English piece on the box

Regex counts over the body text, per piece. meta = the concrete-noun metaphor set (door, gate, ledger, register, machine, engine, map, mirror, threshold, arithmetic …) per 1,000 words. I = first-person pronouns per 100 words. feel = stated feeling or bluntness (angry, absurd, wrong, love, afraid, frankly …) per 1,000. explain = explaining moves per 1,000. 10+ = words of ten letters or more per 100. s.len = mean words a sentence. The Nadal piece, the owner's positive example, is the shaded row; the pieces the owner named are in coral. Corpus: the edition JSONs served at /ink/edition-<date>.json.

Date	Author	Format	Title	Words	meta/1k	I/100	feel/1k	explain/1k	10+/100	s.len
08-31	Michael Moritz	Essay	A Company Town, Again	702	7.1	2.4	0.0	4.3	4.3	14.9
08-31	Nir Eyal	Essay	The Setting My Daughter Would Have to Set Herself	614	6.5	3.6	1.6	0.0	5.2	11.6
08-31	Marques Brownlee	Review	The Spec That Stopped Being a Feature	654	4.6	2.4	0.0	0.0	4.6	17.7
09-04	Winston Churchill	Essay	The Arithmetic of Defeat	1096	10.0	1.2	4.6	5.5	2.4	11.8
09-04	Mark Twain	Shouts	To the Person Who Wants to Regulate Everything	617	1.6	0.8	1.6	0.0	3.7	19.9
09-04	Warren Buffett	Notes	The Beige Book and the Machine	911	18.7	1.1	3.3	1.1	2.4	14.9
09-04	Anthony Bourdain	Dispatch	The Ink-Stained Lunch Counter	1017	2.0	1.0	5.9	3.9	2.1	13.7
09-04	Dan Wang	Notes	Blasting Through	908	8.8	1.0	0.0	3.3	5.5	11.5
09-04	Michael Moritz	Review	The Underwriters Are Waiting	645	6.2	0.3	0.0	3.1	4.5	9.5
09-04	Plato	Essay	The Cave of the Feed	980	16.3	1.1	8.2	3.1	1.3	8.4
09-04	Banksy	Overheard	Two Names, One Hanger	461	8.7	4.1	0.0	2.2	2.8	12.8
09-04	Gary Chapman	Advice	When Your Spouse Works from a Screen	749	1.3	1.7	20.0	0.0	4.1	14.1
09-04	Leonhard Euler	Riddle	The Bridges of a Small World	207	38.6	1.4	0.0	0.0	0.5	9.9
09-05	Geoffrey Hinton	Essay	The Leash We Were Told to Build	903	1.1	0.4	1.1	11.1	4.4	9.9
09-05	Jeff Koons	Shouts	The New Patronage, Itemized	672	17.9	0.7	0.0	3.0	5.1	12.2
09-05	Charles Darwin	Essay	The Missing Middle	1117	1.8	2.4	3.6	3.6	3.3	10.4
09-05	Abraham Lincoln	Letter	The Shape of Your Love	673	1.5	3.9	8.9	3.0	0.9	15.6
09-05	Stephen Hawking	Notes	The Cartographer's Decagon	881	1.1	0.5	2.3	2.3	4.3	16.9
09-05	Christopher Nolan	Review	The Long Exposure	901	6.7	0.7	2.2	3.3	3.7	15.0
09-05	Charlie Munger	Notes	When Lithium Bends	724	4.1	0.1	4.1	6.9	2.8	12.5
09-05	Zaha Hadid	Dispatch	The Mud Remembers the Plan	901	6.7	1.6	1.1	0.0	2.0	8.9
09-05	Akira Kurosawa	Overheard	The Crew Meeting That Could Have Been an Email	525	3.8	0.0	1.9	1.9	1.5	9.9
09-05	Franklin D. Roosevelt	Essay	The Bond Market's New Deal	965	10.4	0.4	1.0	3.1	3.3	12.1
09-05	Dieter Rams	Kicker	The Extra Button	183	16.4	1.1	0.0	0.0	1.1	9.2
09-05	Ada Lovelace	Dialogue	The Oracle Is an Engine	1078	19.5	1.3	8.3	1.9	6.6	12.0
09-06	Thomas Feddo	Notes	The Committee's Red Line	698	4.3	0.1	1.4	1.4	5.9	18.3
09-06	Rafael Nadal	Review	The Williams Show Returns	802	1.2	2.6	5.0	6.2	2.4	14.6
09-06	Ada Lovelace	Letter	To the Engineer Who Cut the Power	782	11.5	3.7	1.3	2.6	8.7	19.0
09-06	Isaac Asimov	Essay	The Slow Burn of the Brain	971	9.3	0.6	1.0	3.1	4.8	20.2
09-06	Leonardo da Vinci	Review	The Sky's Ledger	1029	18.5	1.3	4.9	2.9	6.0	17.4
09-06	Ryusuke Hamaguchi	Review	What the Page Keeps from the Camera	942	3.2	1.6	1.1	8.5	2.5	21.9
09-06	Brené Brown	Essay	The First Door	1022	19.6	1.0	2.0	4.9	7.9	20.0

Medians: meta 6.7 · I 1.1 · feel 1.6 · explain 3.0 · 10+ 3.7 · s.len 13.3. Exclamation marks in 32 pieces: one. The positive example sits on the far side of the median on every column that matters: more I, more feeling, more explaining, fewer metaphor nouns, plain words.

3 · The causes — where the brew asks for this

None of the seven is the model disobeying. Each is something the brew demonstrates, asks for, or selects for. The lines are in lib/ink_brew.py and lib/ink_score.py.

Cause	Where	What the line says	Feeds
Demonstration. The house's own dialect is in the commission.	`FORMATS["Notes"]`	「numbered or sectioned notes on a development — the ledger, not the sermon」. The only Notes piece in the 08-31 edition wrote「written in every ledger of every war」. Same law as the chat panel's tone study: the model copies the analyst's vocabulary.	D1
The vantage is「the record」.	`BOUNDARY`	「Write from the record as if that were the natural place to write from.」「Memory, the public record and stated distance are yours; invented testimony is not.」 A writer told its place is the record writes as a reader of the record — the bystander. Nothing in the rules says the writer has a stake, an opinion it feels, or a reader it wants to move.	D2 D7
The reader is never described.	the writer's prompt	The word「reader」reaches the writer only as「the reader has not seen any image」. The slate says「write for that reader」(the 07:30 reader) and stops. No line says the reader is outside the field, on a phone, and owed a gloss for every term of art.	D3 D4
The judge is the opinion editor of a serious daily.	`EDITORIAL_RUBRIC`	「Score this piece as the opinion editor of a serious daily would — the Pulitzer test.」 Five lines: clarity · reasoning · specificity · position · carry. No line asks whether a non-expert could follow it, and no line asks whether a person is in it. The gate drops to the spares by this score, so the survivors are the ones an Economist leader-writer would keep.	D2 D4 D7
The sentence rule for reported pieces is an analysis shape.	`RHYTHM_RULES["reported"]`	「A sentence carries its thought to where it leads, the second clause giving the cause, the condition or what follows.」 Tuned 09-04 to close the length gap with the majors; it also asks every sentence to reason.	D2 D6
The exemplar banks carry the century's diction.	`exemplars()` + the voice pass	「take your words, your images and your turns of mind from these」, and the editor's voice pass moves the wording toward the bank. Ada Lovelace's bank is 1843 prose, Leonardo's is a translation of the notebooks — hence「denominated」,「verily」. The Sudowrite finding cuts both ways.	D3
The title has no shape rule.	the writer's JSON	「final title (short, specific, no colon-subtitle)」. Left alone the model's default is「The [abstraction] of the [noun]」— 24 of 32 begin with The.	D2
Restraint is the last thing the writer reads.	the writer's JSON	「self_check: which rule you were most tempted to break, and how you didn't.」 A compliance frame closes the brief; stance is asked for after the piece, as a field, never as the reason to write.	D5 D7
Chinese originals get no register line.	`LANG_RULE["zh"]`	「用中文写……按中文的思路和节奏下笔」— language only. The rendition brief (a different path) names 财新 / 三联, the calm periodicals; the Li Ka-shing original is the model's own default calm, with a profile that keeps its distance. Nothing asks for the frustration the owner expected.	D5

the shape of itEvery one of the seven is a register defect, not a fact or a structure defect: the facts are sourced, the notebook rule holds, the lengths are right. The brew asked for restraint in six places and for a pulse in none, then scored with a rubric that cannot see the difference. The tone study found the same on the chat side: prompting sets intent, the net sets the floor, and a demonstration beats a rule.

4 · The fixes shipped 2026-09-07

the owner's rulings, 09-07Subtract, then add. A model must not score a piece: it can count long words, long sentences and pronouns, it cannot judge quality. Too many instructions fight one another and narrow the creative space. The readers are average people who know nothing of CFIUS, dark matter or the Analytical Engine; they want vivid, clear, fluent language, never a paper. Examples must be real, good essays. And no「you write, I score」loop: the owner has no time for one. So the changes below went in as one batch, unmeasured, for the owner to read on the next brew.

what shipped (lib/ink_brew.py · lib/ink_lint.py · lib/ink/register.json) Subtracted from the writer's house rules: the「never announce」rule, the voice negatives (no throat-clearing, no summary, no moral), the per-class rhythm line, and the self-check field. The format briefs lost the house dialect (「the ledger, not the sermon」). Added, positive: two bullets naming the reader (an ordinary person on a phone, not in your field: explain first, then judge) and the stake (you are in it, first person, blunt, the way you would tell a friend). The JSON now asks for the stake first, before the body, and for a title with a verb in it. The persona's own exemplars come with「today's English, never your century's diction」. Beside them ride three real openings from the shelf: Paul Graham, Tim Urban, Morgan Housel, Scott Galloway and the Nadal piece, in a small bank the box can override. Chinese originals get one register line: plain words for the layman, blunt allowed, not a paper. The gate: the model score is out of the brew. Survivors are ranked by the deterministic register floor (metaphor nouns per 1k · long words per 100 · first person per 100 · mean sentence length · the「The X of the Y」title), fewest flags first, ties in the editor's order. On the 09-06 edition the floor gives the Nadal piece 0 flags and the pieces the owner named 2 to 4. Allowances are provisional, set from the Nadal piece and the box's medians, to be re-set from a human corpus. Not done: the two rubric lines (the rubric no longer gates, so they are moot); the majors-set allowances; the length-gap risk (the rhythm line that closed 12.7 → 21 words a sentence is gone — the shelf's examples now carry rhythm, and the floor flags a mean over 20).

The list as it was owed, kept for the record. Bracketed: the defects it targets.

Name the reader. One line in the commission: a curious reader on a phone who is not in your field; every term of art gets its plain gloss the first time, and an example or a comparison a stranger could follow; explain first, then judge. [D3 D4]
The stake before the piece. Move stance from an after-the-fact JSON field into the brief: before writing, the persona says in one line what it feels about the event and why it cares; the piece has to carry that line. The gate reads for a pulse: a piece with no first-person judgment and no stated feeling goes back. [D2 D5 D7]
Retire「write from the record」as the vantage. Keep the boundary (no invented testimony) as a boundary; the vantage line becomes a person with a stake reading today's news, allowed to be angry, delighted or worried on the page. [D2 D7]
Strip the house dialect from every writer-facing line (ledger, door, seam, floor, receipt …) and add a metaphor-noun detector to ink_lint with an allowance per thousand words set from the majors, never zero. [D1] shipped — and, from the second read, the real measure: 隐喻率, §6 below.
A title rule and a title lint. The title says what happened or what you think, with a verb in it; the owner's positive is「The Williams Show Returns」. A title of the shape「The X of the Y」/「The X's Y」is sent back. [D2]
Today's English for historical figures. One line beside the exemplars: your ideas and your images, in the English of today, not your century's diction; gate the long-word rate. [D3]
Two lines in the rubric — could a reader outside the field follow it; is there a person in it (a stake, a feeling stated, a judgment owned) — and the owner's blind pairs as the judge's audit, per the loop. Until the judge can see the defect, the gate keeps selecting for it. [all]
A register line for Chinese originals, distinct from the periodicals' calm: blunt is allowed, 劲爆 is allowed, the writer may be frustrated on the page. [D5]

5 · The yardstick — what the Nadal piece does

「The Williams Show Returns」is the owner's positive example and the outlier on every column. What it does, in order: it gives the scoreline in the first sentence, says「for me, I want to keep another thing in mind」in the second, explains what the super tiebreak was and why forty-six and forty-four matter, and ends with「So my personal feeling is clear」and「I would be happy to watch them fight for as long as they choose to play.」 Plain words, a person in the room, the fact before the feeling, the feeling stated. That is the register the fixes above are aiming at.

6 · 隐喻率 — the metaphor rate built 2026-09-07

When I read our articles I also found the writer used too many metaphors, making concepts even more difficult to understand. The 隐喻率 is a good bench.the owner, 2026-09-07, after Manus's Research Bench write-up (胡迪Hoodie via 今天很好学, mp.weixin.qq.com/s/2CXssAsQMxdPn_5aLcaijw)

Manus measures a research report on two signals: 信息量, the number of fact points a model can extract, and 隐喻率, live metaphors per 100 sentences, annotated by a model under two guards — dead metaphors and technical terms are excluded, and every hit must come with a literal replacement, because a hit with no plainer wording is not a metaphor. Their scale on research reports: a restrained model 1.4, the most florid 9.6, whose images run to bodies and organisms. Their point is the owner's: a metaphor is meant to make the strange familiar, and at high density it makes the abstract more abstract, and it is a lossy function besides. This is a count with a quote and a replacement per hit, which is the kind of model use the owner allows: checkable by a reader, never a taste score.

What was built. A metaphor seat in the brew (DeepSeek Flash by default, Gemini as the console's alternative), the article's prompt with the two guards plus one of our own: the author is named, so their own coinage reads as a term (the live test had rewritten Chapman's「love languages」). Every survivor of the gate is counted. A piece over the allowance gets one surgical pass: the flagged expressions are swapped for their plain wording and nothing else is touched, length-guarded to ±15%, then it is counted again. The rate rides the piece file and adds one flag to the register ranking. Never a score, never a drop.

The scale is the annotator's, not the article's. A loose prompt counted「the East End burned」as a metaphor and read 139 per 100 sentences on Churchill; the conservative prompt read 35 on the same text, and its list was real (raging furnace · a lottery of survival · a Roman candle · the coin a tyrant respects). The final calibration, with the final prompt on the production seat (DeepSeek Flash, temperature 0):

Corpus	pieces	median	top	The florid tail
the shelf's real openings (Graham · Urban · Housel · Galloway · the Nadal piece)	5	29	50	Galloway 50 in ten sentences; the Nadal piece 0
09-06 edition (old brief)	8	26	63	Brené Brown (The First Door) 63 · Leonardo (The Sky's Ledger) 42 · Hamaguchi (文绉绉) 40
09-07 edition (new brief)	12	26	94	Chapman's scoreboard advice 94 · Hinton 40; Churchill's Blitz reads 26 once his coinages are spared

Per 100 sentences on this annotator. The typical piece sits where the human openings sit; the pieces the owner named are exactly the tail. So the allowance is set just above the human control's top, 40, a floor on floridity that leaves a human-level piece alone. A different judge is a different scale: re-measure the shelf bank before moving the seat. ⚠ The annotator is noisy on a sustained image: Chapman's piece read 34 in one run and 94 in the next, the difference being whether every recurrence of the scoreboard was counted. The floor is coarse on purpose; a piece near it may get the pass one day and not the next.

The live test, Chapman's piece: 156 → 14 per 100 sentences on the loose prompt, 792 → 789 words, the scoreboard surviving as「competition」and「You know the ledger」becoming「You know the list」. What the number cannot tell yet is whether the owner prefers the plain version; the next brew on the box carries it.

not built, on purpose信息量 (fact points per piece) — the notebook rule already fixes the facts a piece may use, so the count would measure the notebook, not the writer. And no allowance for Chinese originals yet: the two zh pieces read 9 and 14, well under the floor.

7 · Read 1 — the owner's thumbs and notes on the box 2026-09-07 and 09-08

Pulled from the box's feedback tables: 13 thumbs and 17 notes across the 09-07 edition (the new brief) and the 09-08 edition (Phases 0–2). The tally is the finding: 09-07 read 5 up, 2 down; 09-08 read 2 up, 4 down. Every meter improved on 09-08 and the read got worse. Shape is necessary and not sufficient.

Group	The notes	Where it stands
1 · Wording too serious, not 口语化	Dan Wang「too serious, need 口语化」· Churchill「use close-to-life words」· FDR「the problem is still the wording」· Tao「can he just talk like a normal man?」· Rothman「a ton of legal writing」· Asimov「can be more fun」· Chapman「more 有人情味」	planned Phase 3 (the profile cut: the analyst's prose is 72% of the prompt) · Phase 4 (the writer model). No 口语化 meter exists yet.
2 · 不知所云: no thread, broken flow, self-talk	Einstein 09-07「awful… broken flows… 故弄玄虚」· Munger「看上去写了很多，不知道他在说什么」· Lincoln 09-08「支离破碎、不知所云」· Banksy 09-08「不知所云」· Stan Lee「自说自话、唠唠叨叨」	planned in part Phase 5 (the editor stops pre-writing the argument) · Phase 3. The desk (Phase 1) removed the empty-notebook cause and these pieces still lost the thread. No coherence test exists, and a model judge would not be one.
3 · Pointless pieces	Dylan 09-08「came from nowhere, 100% unnecessary… reduce this kind」· Twain's fable「for primary students」	no design The slate policy fills four light slots a day by quota and lets them peg to「the day itself」. A policy change, not a prompt: a light piece needs a why-now like any other.
4 · Under-explained for laymen	Feynman「spend more words to explain the path integral, as to laymen」· Munger「普通读者会觉得天书」· Rothman	partly the reader line (Phase 0) and real material (Phase 1); no further item.
5 · The Chinese renditions	Feynman「直译 not 意译」· Dylan 09-07「the Chinese translation ruined all these」· Banksy「中文翻译也很奇怪」· Kurosawa: translate a name once, then use the original	no design outside the upstream plan, which is about the English originals; deserves its own read on the rendition page.
6 · The piece as a task	Brownlee「no one asked him… it should feel like true motivation」· Churchill「he mentioned he is 80, which is a lie」	the distance tic fixed (0.6, on main); the motivation itself is Phase 5's question (a subject and a question, not a thesis). The age slip: no design.
7 · Format and tics	Banksy's Overheard「应该以对话形式给出」· Stan Lee's「Excelsior」	no design
What read well	Kurosawa「very good」· Feynman「good」· FDR「great content and opinion」· Chapman「very good content」· Dan Wang「explained Hugging Face, first person, his domain」· Bourdain 09-05「inspired by a topic, not a paraphrase of it — the exact style I like」	content and opinion are the strength; the wording is the gap, consistently.

what the read says about the planThe next step should start from the down-votes, not the meters. Two groups have no design at all — the pointless light pieces and the Chinese renditions — and one has no meter, the colloquial register. The two planned phases that speak to the largest groups are 3 and 5.

statusRecorded 2026-09-06 from the owner's read; the §4 batch shipped 2026-09-07 and the 09-07 edition (12 pieces) was brewed on it; the owner's second read added 隐喻率 (§6), built the same day. The daily brew stays OFF until that read says so. Siblings: the brew · the quality score · the loop · AI tone (the chat side's version of the same finding).

recorded 2026-09-06 · corpus = the three box editions served at /ink/edition-*.json · siblings: the brew · the quality score · the brew log · the loop