The density read — six pieces, read rather than counted read · 2026-09-16

You did not trust the code's density table, so this page is a reading: three Ink pieces and three from the bank, matched on length, read end to end and compared by hand. It reaches a different answer from the code, and it explains why the code's answer was wrong.

the answer in one paragraphDensity is not facts per word. It is how many times a piece says its point. The three bank pieces state their claim once and then every paragraph adds a different support; one of them has no claim at all and is the densest thing in the set. The three Ink pieces state their claim and then re-dress it three, four and six times. Worse, the facts Ink carries are almost all lifted from the one source page it was written from, so a reader who has seen the source learns nothing new and then reads the same thought six times in fresh clothes. That is your "wrote a lot, means little", and it is also why the source page reads normal to you: the source states each fact once and never repeats itself.

1 · The six

piecewordswhat it is
Ink · Karpathy — "Anthropic Blocked the Paragraph, Not the Pipette"995Notes, 09-12, on the Anthropic misuse report
Ink · Kurosawa — "The Thieves Take What They Cannot Look At"735Essay, 09-13, on the Renoir theft
Ink · MrBeast — "GTA 6 Just Turned New Zealand Into the Most Powerful Country on Earth"650the Algorithm A side of pair 12, the piece you called empty
New Yorker, Talk of the Town — Sarah Larson on Alexandra Grant844a gallery opening, reported
New York Times, Opinion — Allison Sweet Grant, "This Is Why Women Are Scared to Tell Their Doctors the Truth"866first-person argument — the closest in kind to what Ink attempts
The Guardian — Agustin Guilisasti on London's ebike parking861an op-ed with a policy ask

2 · The finding: how many times the piece says its point

Short answer: the majors say it once. Ink says it four to six times, never in the same words, which is why no word-counting measure can see it.

pieceits one claimtimes statedparagraphs that add something new
Ink · Karpathythe model supplies the text; the hard part happens with hands in a room4 + the title7 of 9
Ink · Kurosawaan image exists only when it is seen; hiding it kills it35 of 5
Ink · MrBeastbeing first is the product; sell the head start6, the last two near-identical4 of 10
New Yorker · Talknone — it makes no argument at all06 of 6
NYT · Opinionwomen understate their pain to play the perfect patient112 of 12
Guardianthe problem is parking, not the number of bikes16 of 6

The New York Times piece is the sharpest comparison, because it is the same job Ink is asked to do: a first-person essay with a view. It states its claim in the third paragraph and never states it again. What follows is eight different supports, each a new mechanism or a new piece of evidence: children taught to placate parents; doctors as the highest-status authority; a study of two thousand patients where diagnostic errors doubled the odds of stigmatising language in the chart; the four words that get written there, Exaggerating. Resistant. Noncompliant. Hysterical.; the parallel with "bossy" and "uppity" women; the research showing women wait longer for a diagnosis and get less pain medication; the extra weight on poorer and minority patients; what a doctor could do instead. Then it closes by repeating the first thing she said in the chemo chair, I'm sorry. I'm fine. — the only repetition in the piece, and it is the point of the piece.

Set that beside the MrBeast piece, which says "being first is the whole product" in paragraphs three, five, six, eight, nine and ten, and whose last paragraph restates the one before it almost word for word. Nothing is added after paragraph four. The rest is the same thought with a new prop each time: an island, a submarine, a DLC, an auction.

3 · Where the specifics come from

Short answer: the majors' facts are earned by being somewhere. Ink's facts are borrowed from the page it was written from.

piecethe specifics in itwhose they are
Ink · Karpathythe report's five case studies, the 30-day window, chikungunya, orthopoxviruses, Klein's quotes, Sanders, Trumpall from the source story. His own: a list of lab steps, and "I'm on the pretraining team"
Ink · KurosawaCagnes-sur-Mer, three minutes, 88,000,000 euros, Amore, Marinello, Brand, Caviezel · then Heigo the benshi, dead at 27, and the calligraphy ink in the rain tanksthe first half from the source; the second half his own, and it is the best writing in the three Ink pieces
Ink · MrBeastthe 19 hours, the region switch, $80 · counting to 100,000 for forty hours, editing with an IV in his armthe launch facts from the source; the boasts from the profile
New Yorker · Talkabout forty proper nouns and a dozen quotations from five people: mermaid teal booties from a Berlin shop, tombstone wax, a yellow school bus to Avenue C, the mother saying "Oh, there's Keanu"the writer was in the room for three hours
NYT · Opinionfive hours into the infusion, the semireclined vinyl chair, the P.A. summons, the Benadrylthe writer was the patient
Guardian3m rides in July, 450km of Cycleway, one car space holds ten bikes, half of London households own no car, 15% conversion would make room for a millionthe writer runs an ebike company and says so

This is the answer to your other question. You said the source subpages read normal and the essays do not. The source carries each fact once, in the order it happened, and stops. Ink carries the same facts, adds no new ones, and then adds the repetition. So against the source, an Ink piece has negative information: everything the reader already had, plus a thesis stated six times.

The exception proves it. Kurosawa's brother Heigo and the black ink poured into the rain tanks are the only passages in the three Ink pieces that a reader could not get from the source page, and they are the only passages that read like writing.

4 · What the code got wrong

Short answer: the code measured new words and called them new information. Saying the same thing in fresh vocabulary scores as novel.

the code saidthe read says
novelty 95% for Ink, near the bank's 97% — "almost every sentence brings new content words"Meaningless here. An island, a submarine, a DLC and an auction are four new nouns carrying one old idea. The metric rewards exactly the habit we are trying to catch.
quotations: zero per thousand words in InkTrue at the median — half of Ink's pieces contain no quotation mark of either kind — but it under-counts pieces like the Karpathy one, which quotes seven times in single quotes. The gap is real; the number was measured with the wrong mark.
names 3.2 vs 7.4 per hundred wordsHolds, and the read sharpens it: the count is not the point, the source of the names is. Ink's names are the source's names.
hollow verbs 30% vs 23%Holds and matters. The majors' paragraphs are things happening; Ink's are things being.
compression and content variety — pair 12 the most repetitive text in the tableHolds. This was the one code measure that found what you found by reading, and it only works at the whole-piece level.

5 · The "I" test

All six are first person. The difference is what the "I" is for.

the "I" as evidence — NYT"When I was coherent again, the first words out of my mouth were 'I'm sorry.' Immediately followed by 'I'm fine.'" — the writer's own behaviour is the thing being argued about
the "I" as credential — Ink, Karpathy"I'm on the pretraining team at Anthropic — I work on the runs that give Claude its core knowledge — so I read that line with a stake in it." — the writer announces standing, then offers a view

The Times writer also takes a real risk in her own piece: I can't help worrying that having this article published will offend someone and earn me a label of my own. Nothing in the three Ink pieces costs the writer anything.

6 · What I would change

proposed · awaiting your yes 1. Count the restatements, not the words. The one measure worth building from this read: how many paragraphs restate the piece's claim. It can be asked of a reader model per piece, cheaply, and it is the number that separates every bank piece from every Ink piece in this set. It also catches pair 12, which the per-paragraph ruler cannot. 2. Retire "novelty" from the density page. It measures vocabulary, not information, and it flatters us. 3. The commission asks for the wrong thing. Two of the three bank pieces earn their density by the writer having been somewhere; the third, Talk of the Town, has no argument at all and is the densest of the six. Our brief asks for a stake and a view and gets a view restated. The next loop should ask for the reported piece: someone named, someone speaking, something that happened, and the view left to arrive on its own. 4. One fact the reader could not get from the source. The Heigo passage is the proof. Make it a requirement of the commission that the piece carry at least one thing the source page does not: a memory, a number the persona worked out, a person they know.
limitsSix pieces, chosen for length match, three of the bank's fourteen sources. The Talk of the Town piece is reported on assignment, which Ink cannot do; it is here for contrast, not as a target. The restatement counts are my reading, not a measurement, and another reader would score Kurosawa's three as two or four.
Read · 2026-09-16 · six pieces, by reading · related: information density (the code's table) · the model's read · Algorithm A, the read · the writing gap · what the writer reads