Three instruments, run over the same two corpora: 88 English Ink pieces (79,502 words, every public edition 09-04 → 09-10) against the 267-piece majors bank (338,621 words). One: disproportion finders over everything — every word, two-word phrase, sentence shape, opener and title scored by how far its rate in Ink sits from its rate in the majors, nothing named in advance. Two: an LLM counting ten habits with quotes over 33 Ink pieces and 30 majors, so a habit code cannot see gets a number anyway. Three: four editors reading 40 matched human/Ink pairs with no list given — just「name every difference a reader would feel, with the sentence that shows it」— and their notes merged afterwards into one catalogue of 29 habits.
Three of the four readers opened with the same habit, and the code found the same thing from the other end. The register is present-tense, definite and certain: twice the is, half the but, half the was — Ink states where a person narrates and concedes. The single most over-used sentence in the language of Ink is the short copular declarative with the, it or that as its subject —「That is the whole game.」— at seventeen times the human rate, and the human sentence it replaces opens on a named subject, runs past twenty words, and is allowed a dash or a question.
Two things follow from that, and they are the two the readers felt first. Every paragraph must land. No Ink paragraph stops on a fact, an image, a half-thought or a shrug; the last sentence is always a short, symmetrical, quotable restatement of the paragraph above it, so the reader is never left to draw the inference. Over a whole piece it is a metronome with a cymbal on every bar, and the reader stops believing any of them. And nobody real is in it: no room, no hour, no weather, no body, no overheard speech. Sources arrive with full institutional titles, lifted from the news brief; supporting characters are categories rather than people. No Ink persona ever stands next to anyone.
Two corrections to the first cut, both from the code: paragraph endings are not a tell by length (30% of Ink paragraphs end on a sentence of ten words or fewer, against 31% of the majors' — what differs is what the short sentence does), and humans use more hard metaphors than Ink, not fewer. Both are in §4.
Each reader ranked what they found over their own ten pairs; a habit's count is the sum across the four, so the maximum is 40. The same habit under four names has been collapsed into one row. Ranking is by how strongly the readers said the habit marks a piece as machine-written, then by pair count.
| # | the habit | /40 | Ink | a human instead | code? |
|---|---|---|---|---|---|
| 1 | Every paragraph ends on a line that sums it up | 38 | 「A crate of vodka is not a lever.」— Feddo | 「Denver's stadium costs are not yet known.」— Bradbury | yes |
| 2 | The piece announces its own moves | 34 | 「Scorecard first. That's the honest order.」— Phil Knight | 「Probably for the same reason that few hackers take up painting.」— Paul Graham | yes |
| 3 | Nobody real is in it | 35 | 「Dr Elver Luiz Mayer, of the Federal University of the São Francisco Valley, examined the remains.」— Darwin | 「Temp workers: We don't need them. Wait a day for turnover.」— Osterman, quoting a manager | partly |
| 4 | It ends by telling you what to do | 33 | 「simply ask to see his test results.」— Buffett | 「I'm keeping my torch, thank you very much.」— Steinem | yes |
| 5 | Not this, but that | 38 | 「There is no ladder. There is simply a stretch of water.」— Orwell | 「sleep is not a neat off-and-on switch but a fully human and fiendishly manifold activity」— Gopnik (once, in 827 words) | yes |
| 6 | Nothing is wasted — no joke, no digression, no aside | 29 | 「each was an engine wound up by the other's tariff, and the gears did the rest.」— Mark Twain | 「a small makeshift lab on a ship in the Antarctic, which likely smells of penguin guano to this day.」— Youngflesh | partly |
| 7 | Intensifier wash | 26 | 「relentless, rigorous, invisible care across every single millimeter」— Jony Ive | 「We now live in a country with more than seventy million completely insane people.」— Ian Frazier | yes |
| 8 | The writer is never wrong and nothing costs them anything | 23 | 「I have learned to trust that intuitive reaction far more than any technical document.」— Jony Ive | 「I can't prove that one should do this.」— Paul Graham | weak |
| 9 | Every paragraph the same size | 22 | 「Both are legal. Both are rational. Only one of them builds.」— Phil Knight | 「I watched another. And another. And then it hit me.」— Tim Urban (his paragraphs run 6 to 120 words) | yes |
| 10 | The same point, five times | 20 | 「tennis is won in the head, not in the racket.」— Mouratoglou (the same claim closes five paragraphs) | 「Indeed, the first version of this principle was to have good new ideas. But that didn't seem general enough.」— Paul Graham | yes |
| 11 | Lists of three or four, always at the end | 20 | 「the splice becomes the event, the caption becomes history, and the feed becomes the country.」— Didion | 「Your bed linens? Your sugar bowl? Your pubic hair?」— Ruttenberg (the escalation is the joke) | yes |
| 12 | The credential paragraph | 20 | 「I've spent the better part of a century trying to evaluate uncertain business propositions.」— Buffett | 「I am a labor economist.」— Osterman (five words, subordinated to the work) | partly |
| 13 | Ceremonial hedging — the lane disclaimer, the refused verdict | 18 | 「Then there is the money, which is outside my lane, so I will not pretend to parse every nuance of it.」— Kumar | 「My hunch is it's not terminal.」— Toynbee (five words, mid-sentence) | yes |
| 14 | Costume words, and no contractions | 18 | 「Apple introduced a folding apparatus」— Rams, who never once writes「phone」in 1,144 words | 「hire professional fakers and let them drive the company into the ground.」— Paul Graham | yes |
| 15 | Explaining what the reader already knows | 16 | 「Extradition, to put it super simply, is just one government formally asking another country to surrender a suspect.」— Kardashian | 「Real macros need to operate on the parse tree of the program.」— Paul Graham, writing to programmers, no gloss | yes |
| 16 | Figures handled as equipment: stacked, announced, then explained | 16 | 「The analogy I keep returning to is a perfect spell-checker.」— Terence Tao | 「This results in the 'gold-plating' of facilities.」— Bradbury (one figure in 1,331 words) | partly |
| 17 | Numbers recited, not chosen | 15 | 「Lucky was 38.1 percent. Worked hard was 35.7 percent. Task was easy was 21.4 percent.」— Chestnut | 「swept away 70 percent of the natural killer cells」— Gopnik (two numbers in the whole essay) | yes |
| 18 | Abstract things doing human things | 15 | 「the ice says otherwise; the ice wins」— Wolfram | 「Dead worms on the side walks fry and sizzle」— E. B. White | partly |
| 19 | The reader is put in the lesson | 12 | 「You have never consciously considered the switch beside an entrance.」— Rams | 「Tell me how I'm supposed to handle this fucking story.」— Tim Urban (the reader is asked, not instructed) | yes |
| 20 | It opens on the persona reading the news | 12 | 「I read the headline twice. The first reading was hope, which surprised me more than the news did.」— Plato | 「At a YC event last week Brian Chesky gave a talk…」— Paul Graham | yes |
| 21 | The persona quotes its own greatest hits | 11 | 「Weniger, aber besser: less, but better.」— Rams, closing on his own most famous line | 「Some of this peak teen-age energy I smuggled into 'White Teeth'」— Zadie Smith, self-citing only to say life was worse | yes |
| 22 | Titles all cut the same way | 10 | 「Britain Pays 5.82% and Calls It Prudence」— Keynes | 「What to Do」— Paul Graham | yes |
| 23 | The trailing「-ing」clause | 8 | 「sell the single moment when the machine dissolves, leaving only what you are trying to create」— Steve Jobs | 「Impacting lives quietly along the way.」— Morgan Housel (as its own sentence) | yes |
| 24 | The piece runs to the assigned length | 8 | 「That is just human nature. We took a video game and turned it into an extreme world race.」— MrBeast, final paragraph | 「Unnamed sources say that the Commander-in-Chief is evaluating a potential 2028 bid for God.」— Patricia Marx, 566 words, stopping where the escalation stops | yes |
| 25 | The both-ways ending | 7 | 「If the claim is correct… If it fails in some subtle way… Either way, the honest assessment today is that I simply do not know.」— Tao | 「So is it a good idea, or a bad one? We still don't know. That's how little we know about founder mode.」— Paul Graham | yes |
| 26 | The assignment is visible in the text | 7 | 「The question I was handed was where the First Amendment line falls…」— Kumar | 「I've been thinking about teen-agers」— Zadie Smith | yes |
| 27 | Emotion named instead of shown | 6 | 「What I actually feel is a certain curiosity.」— Terence Tao | 「The tears will smudge this if I don't take care. No matter, I will type it so it is legible.」— Patsy Clarke | partly |
| 28 | Numbers spelled out in words | 5 | 「it occupies one hundred and seventy-four minutes」— Sherlock Holmes | 「capping bus fares at £2」— Toynbee | yes |
| 29 | The seam shows | 3 | 「the magistrates expressly decline to conjoin those二人 with the ruine of the ground」— Isaac Newton | no counterpart exists in the 40 human pieces — the comparison is simply numbers that hold | yes |
The Ink sentence on the left with its persona, the human sentence on the right with its author, and one line under each on what the difference is.
The instrument is the log-odds ratio with an informative prior (the「fightin' words」method), z-scored so a rare word cannot win by luck. Nothing was named in advance. What came out is not a list of words; it is a register, and it reads the same whether you look at words, phrases, sentence shapes or openers.
| over-used | Ink | majors | under-used | Ink | majors |
|---|---|---|---|---|---|
| the | 753 | 503 | to | 192 | 268 |
| is | 225 | 108 | of | 188 | 258 |
| it | 157 | 97 | but | 32 | 61 |
| an | 72 | 37 | in | 134 | 183 |
| whether | 16.2 | 3.4 | my | 20.1 | 41.0 |
| upon | 10.7 | 1.2 | it's | 2.3 | 16.6 |
| says | 11.3 | 1.6 | about | 20.1 | 39.9 |
| cannot | 12.1 | 2.0 | as | 39.4 | 62.5 |
| a | 336 | 267 | all | 13.2 | 28.2 |
| real | 15.2 | 4.5 | other | 6.9 | 17.9 |
| remains | 8.3 | 1.0 | people | 17.2 | 31.7 |
| merely | 7.3 | 1.1 | some | 6.9 | 17.6 |
| market | 7.5 | 1.3 | was | 49.9 | 70.8 |
| simply | 9.1 | 2.2 | or | 26.5 | 42.2 |
| september | 6.2 | 0.9 | I'm | 1.0 | 8.6 |
| open | 8.8 | 2.2 | |||
| its | 26.5 | 13.3 | |||
| never | 17.9 | 7.6 | |||
| exact | 5.4 | 0.7 | |||
| machine | 5.3 | 0.6 | |||
| must | 11.1 | 3.6 | |||
| single | 7.7 | 1.8 | |||
| nobody | 5.7 | 0.9 | |||
| true | 8.9 | 2.5 | |||
| ten | 6.7 | 1.4 |
Two families and one leak. The intensifiers of certainty — exact · real · true · single · never · nobody · merely · simply · must · cannot — are a performed confidence, five to eight times the human rate each. The formal default — whether · upon · remains — is the machine's untuned English. And machine · market · says plus, below the cut, reader · headline · phrase · sentence · narrative are the brief's own vocabulary leaking through. The words Ink is short of are the words of a person: my · people · was · it's · I'm · some · other · all.
| over-used | Ink | majors | under-used | Ink | majors |
|---|---|---|---|---|---|
| that is | 22.3 | 2.6 | I was | 2.9 | 10.9 |
| is the | 32.6 | 8.4 | if you | 4.6 | 10.8 |
| and the | 37.5 | 16.6 | would be | 1.4 | 5.7 |
| is a | 28.5 | 11.1 | to do | 2.2 | 6.6 |
| I have | 14.4 | 4.0 | there are | 1.9 | 6.1 |
| it is | 20.0 | 7.2 | want to | 2.2 | 6.4 |
| do not | 8.2 | 1.4 | a lot | 1.8 | 5.8 |
| does not | 7.1 | 1.4 | to get | 1.3 | 4.8 |
| whether the | 5.7 | 0.2 | the best | 0.4 | 4.1 |
| and it | 8.4 | 2.4 | to have | 1.0 | 4.2 |
| is what | 5.7 | 1.1 | |||
| the whole | 6.6 | 1.6 | |||
| is an | 7.0 | 1.8 | |||
| the report | 3.7 | 0.2 | |||
| a man | 4.7 | 0.9 |
The same register one word wider: the copular claim (that is · is the · is a · it is) against the human's past-tense first person (I was) and the plain verbs of doing (want to · to get · to do). Three uses of「the best」in 80,000 words: Ink does not rank, recommend or enthuse.
A shape is opener class | length band | mark: what the sentence starts with (a name, the, it, that, something else), how long it runs, and whether it is copular (is/was/are as its main verb) or carries a dash or a question mark.
| over-used shape | Ink | majors | under-used shape | Ink | majors |
|---|---|---|---|---|---|
| 「That …」≤ 8 words, copular | 152 | 9 | a named subject, 21+ words | 311 | 636 |
| 「The …」≤ 8 words, copular | 237 | 45 | any other opener, ≤ 8 words | 70 | 239 |
| 「That …」9–20 words, copular | 121 | 10 | a named subject, 21+ words, with a dash | 12 | 127 |
| 「A …」≤ 8 words, copular | 47 | 4 | a named subject, 9–20 words | 759 | 1016 |
| 「The …」≤ 8 words | 348 | 65 | opening on in, 21+ words | 43 | 104 |
| 「The …」9–20 words | 656 | 245 | a named subject, ≤ 8 words, a question | 10 | 55 |
| 「A …」9–20 words | 233 | 67 | a named subject, 9–20 words, with a dash | 6 | 51 |
| 「It …」≤ 8 words | 136 | 27 | any other opener, 9–20 words | 97 | 169 |
| 「The …」9–20 words, copular | 290 | 117 | opening on as, 21+ words | 8 | 49 |
| 「It …」9–20 words | 175 | 65 | opening on and, 9–20 words | 35 | 83 |
This is the owner's「X is X. X isn't X」, found rather than assumed. The short copular declarative with a pronoun or the as its subject is the most over-used shape in Ink at 17×; the human sentence it replaces opens on a named subject, runs long, and is allowed a dash (10×) or a question (5×).
| over-used | Ink | majors | under-used | Ink | majors |
|---|---|---|---|---|---|
| That is | 252 | 7 | There are | 4 | 42 |
| I have | 135 | 24 | If you | 43 | 83 |
| It is | 141 | 36 | I was | 14 | 43 |
| There is | 62 | 15 | At the | 4 | 23 |
| If the | 51 | 12 | As a | 6 | 23 |
| The report | 33 | 2 | I think | 2 | 17 |
| Let me | 29 | 2 | I don't | 6 | 22 |
| I am | 49 | 13 | In the | 35 | 61 |
| Here is | 25 | 2 | It was | 39 | 64 |
| That's the | 23 | 3 | For example | 0 | 21 |
| I read | 22 | 2 | When I | 14 | 30 |
| You cannot | 20 | 2 | And yet | 2 | 14 |
Ink never says for example, in fact, of course, I think, and yet — the small spoken moves a person uses to concede, illustrate and soften. It opens on the verdict instead:「That is …」, thirty-six times the human rate. And it opens on itself: Let me · Here is · The report · I read are the stage directions of §2's habit 2, caught by the code as openers.
The disproportion finders see only what the corpus repeats. So a second instrument read 33 Ink pieces and 30 majors and counted ten habits with quotes, one at a time. Four of the ten are real tells by count. Four are not — and two of those were on the owner's own list, which is why the count was worth paying for.
| habit | Ink /1k | majors /1k | × | verdict |
|---|---|---|---|---|
| D1 · a number spelled out —「at a depth of eight metres — upwards of twenty-six feet」 | 1.74 | 0.41 | 4.2 | a real tell |
| F · stiff or certainty vocabulary —「calling the meltdown entirely beyond their control」 | 3.85 | 1.49 | 2.6 | a real tell |
| B · a thing speaks or knows —「The machine swallows all that desperate intent」 | 1.89 | 0.85 | 2.2 | a real tell |
| E · a disclaimer or hedge —「I'm no foreign correspondent, but…」 | 1.19 | 0.55 | 2.2 | a real tell |
| C2 · a metaphor that works —「the invisible traffic cop」 | 1.81 | 1.51 | 1.2 | matched; not a tell |
| D2 · a round number as colour —「another two living out of a roller bag」 | 0.70 | 0.61 | 1.2 | matched; not a tell |
| A · a flat verdict sentence —「There is zero cushion.」 | 5.37 | 4.10 | 1.3 | cleared — Graham writes「Speed defines startups.」 |
| D3 · false precision —「in five minutes, and it'll take the market years」 | 0.22 | 0.22 | 1.0 | cleared — identical rates |
| G · the explaining reflex —「Here's what civilians fail to grasp…」 | 1.56 | 1.71 | 0.9 | cleared — humans explain slightly more |
| C1 · a metaphor that makes it harder —「an unforgiving spreadsheet with jet engines」 | 1.15 | 1.85 | 0.6 | cleared — humans use more of them |
The regex first cut, kept for comparison (09-09/09-10 editions only, already on the plan page): sentence length is matched at 17.5 words against 17.9 — so length is not where the gap lives. Contractions half the human rate; questions a third; exclamations none; dashes a fifth. Metaphor nouns 2.5 per 1k against 0.9–1.6. Disclaimers 0.45 a piece against 0.15. Paragraphs 109 words against 50–69 — the wall. The two instruments agree on disclaimers and disagree on metaphor: the regex counts metaphor nouns, a vocabulary measure; the LLM counted figures that work and figures that don't, and the ones that don't are ours by a factor below one.
By close reading of every title in both corpora: 37 of 88 Ink titles (42%) turn on a negation, an absence or a failure, against 56 of 267 human titles (21%). By the wide keyword lexicon the same measure reads 28% against 10% — the lexicon undercounts Ink, because the frame is often carried by no keyword at all. The second gap is quieter and larger: a title that is simply a noun phrase — no verb, no argument, a name for the thing — 19% of Ink titles against 46% of human ones. Human titles are flat, borrowed, or a found phrase (「What to Do」;「'I hate everybody including you'」). Ink's state the argument and twist it, in one shape, over and over.
All four readers volunteered this without being asked for it.
Three things, all on the writing loop. The critic's anchors (lib/ink/critic.md) take their「this works / this doesn't」pairs from §2's top eight, so a deduction means the same thing in every loop. The writer's example sheet (lib/ink/shelf-sentences.json) takes the right-hand column — whole human paragraphs keyed to the habit they answer — because demonstration beats a rule. And the loop's code audit (lib/ink_tells.py) is §3 and §4 recomputed every loop, with the majors' rate as each finding's allowance, so the list is never hand-kept: the four cleared habits are dropped from it today, and the paragraph-ending rule replaces the flat-sentence ban. Every loop's row lands on the ledger.
scratchpad/study/catalogue.md (the four readers, merged) · scratchpad/study/discovery-1..4.md (the open notes) · scratchpad/study/study-data.json (disproportion · LLM counts · titles) · scratchpad/study/count-{ink,majors}-1..3.json · scratchpad/study/titles.json · next: the writing loop · the ledger · earlier: the first read · the register study