Information density — how much a piece says per word, Ink against the bank measure · 2026-09-16
You read the MrBeast pair and said he wrote a lot and meant very little. That is a different fault from the tells: a piece can be clean of antithesis and still be empty. This page measures it, by code, on the same ruler for the 142 published English pieces, the two loops, and the 341 bank pieces, with the MrBeast pair called out. Five proxies, none a judgment; a full essay of the aesthetic kind is allowed to sit low on some of them, which is why the bank is split into its bands. run 2026-09-16 17:03 · medians
the answer in one paragraphInk's density gap is people, not facts. A bank piece names someone every 14 words; an Ink piece every 31, and it quotes no one at all: the bank carries three quotations per thousand words, Ink none. Numbers are the same on both sides, so the source dump is not the cure and the reporting is not the shortage. The second gap is the verb: three in ten of Ink's verbs are is, has, does, against two in ten in the bank, and in the light formats nearly four in ten. That is the verdict register measured from a different side: states and rulings instead of things happening. The MrBeast piece is empty in a third way. It carries more names and numbers than almost anything in the bank and still says little, because it says it again: the most compressible text in the table, and the lowest variety of content words. Density there is lost to repetition, not to vagueness.
1 · The five proxies
compress — bytes after zlib compression per 100 characters of the lowercased text. Repetition compresses; a piece that keeps saying the same thing scores lower. Longer pieces compress a little better on their own, so compare within a length. specifics — names, numbers and quotations per 100 words: a name is a capitalised word not at a sentence start; a number is a digit run or a spelled number; a quotation is one pair of quote marks. The editors' "not specific enough" fault, counted. novelty — the share of sentences that bring at least two content words the piece has not used before; a sentence that only rearranges earlier words scores no. hollow verbs — of everything that looks like a verb, the share that is is, are, was, has, does, do: a state or a ruling rather than an event. function — the share of words that are function words (the, of, that, it…); the higher, the less each word carries. content TTR — distinct content words over content words in the first 600 words, so length does not decide it. All figures are medians per group.
2 · The table
| group | pieces | words | compress | specifics /100 | names | numbers | quotes /1k | novelty % | hollow verbs % | function % | content TTR | words / sentence |
|---|
| Ink · published (en) | 142 | 860.0 | 45.8 | 5.3 | 3.2 | 1.9 | 0.0 | 95.4 | 29.9 | 46.2 | 76.0 | 15.9 |
| Ink · Essay | 33 | 1036 | 44.2 | 4.6 | 3.0 | 1.6 | 0.0 | 96.0 | 28.7 | 46.3 | 75.9 | 16.1 |
| Ink · Notes | 28 | 873.0 | 45.0 | 5.8 | 3.3 | 1.9 | 0.9 | 94.5 | 29.9 | 46.9 | 75.7 | 17.4 |
| Ink · Scorecard | 7 | 832 | 46.2 | 9.0 | 3.2 | 4.9 | 1.0 | 96.2 | 25.2 | 44.2 | 75.3 | 15.1 |
| Ink · Review | 24 | 935.0 | 45.8 | 6.8 | 4.2 | 2.0 | 2.2 | 97.2 | 26.2 | 45.6 | 76.8 | 17.1 |
| Ink · Dispatch | 6 | 969.0 | 45.8 | 4.3 | 2.8 | 1.4 | 0.0 | 97.8 | 27.2 | 47.0 | 81.4 | 15.5 |
| Ink · Shouts | 7 | 698 | 45.9 | 8.3 | 6.3 | 2.0 | 4.7 | 92.7 | 37.7 | 45.3 | 72.5 | 12.2 |
| Ink · Kicker | 10 | 185.0 | 52.6 | 4.7 | 2.4 | 1.8 | 0.0 | 100.0 | 37.2 | 49.0 | 85.2 | 9.6 |
| Ink · Letter | 12 | 678.5 | 46.9 | 4.3 | 2.3 | 1.6 | 0.0 | 95.6 | 32.0 | 48.1 | 75.7 | 13.9 |
| the bank · all | 341 | 973 | 46.0 | 9.9 | 7.4 | 1.6 | 3.0 | 96.9 | 23.4 | 44.9 | 79.0 | 18.8 |
| the bank · plain | 180 | 1327.0 | 43.6 | 8.4 | 6.4 | 1.6 | 2.5 | 95.7 | 23.1 | 44.5 | 74.8 | 18.6 |
| the bank · literary | 161 | 582 | 49.8 | 11.1 | 8.5 | 1.7 | 4.7 | 100.0 | 23.8 | 45.7 | 82.1 | 19.0 |
| the bank · reported | 257 | 1160 | 44.8 | 9.9 | 7.4 | 1.6 | 3.1 | 96.6 | 22.7 | 44.2 | 77.8 | 19.0 |
| the bank · voiced | 59 | 475 | 50.2 | 9.0 | 6.6 | 1.9 | 2.7 | 100.0 | 28.6 | 48.8 | 82.5 | 17.7 |
| the bank · light | 25 | 700 | 48.2 | 12.6 | 10.0 | 1.0 | 2.8 | 96.0 | 21.5 | 44.0 | 80.7 | 17.7 |
| loop · the pick (L21) | 12 | 1083.5 | 45.2 | 7.8 | 4.0 | 3.0 | 1.9 | 91.7 | 33.2 | 47.6 | 77.1 | 16.2 |
| loop · Algorithm A (L22) | 12 | 959.0 | 44.8 | 7.0 | 3.9 | 2.3 | 2.7 | 92.2 | 32.0 | 49.5 | 75.8 | 16.5 |
| pair 12 · left (L21) | 1 | 1080 | 41.4 | 16.8 | 12.7 | 3.9 | 1.9 | 75.0 | 36.5 | 44.3 | 60.9 | 9.2 |
| pair 12 · right (L22) | 1 | 650 | 44.8 | 7.5 | 5.7 | 1.8 | 0.0 | 82.0 | 27.1 | 42.2 | 65.3 | 11.0 |
medians per group · Ink = the box's published English pieces 08-28 → 09-13 · the bank = the majors, plain and literary bands, by class · the loops = the twelve fixed assignments on 09-14 · pair 12 = the MrBeast pair on the read pages
3 · What the table says
Short answer: three gaps, and each names a different fix.
- Half the specifics, and none of them people. Names run at 3.2 per hundred words in Ink against 7.4 in the bank; quotations at zero against three per thousand. Numbers are level. The bank's lightest pieces, the Shouts and the Onion, are the most specific of all, because a joke needs a proper noun. Ink's own Shouts come closest to the bank on names and are still short of it. Nobody in an Ink piece is named and nobody speaks, which is the reader's "nobody real is in it" from the writing gap, now as a count.
- The hollow verb. Three in ten of Ink's verbs are states, against 2.3 in the bank. The light formats reach 37 percent. It is the flat verdict measured on the verb rather than the sentence: X is Y, that is the, it has. The two loops sit higher still, at 32 to 33 percent, so the repair pass did not move it, which is expected: the ruler counts the gavel sentence, not the verb.
- The MrBeast piece is repetition, not vagueness. Both sides of pair 12 carry more names and numbers than the Ink median, the left side more than the bank's literary median. What they lack is variety: content TTR of 61 and 65 against 76 for Ink and 79 for the bank, novelty of 75 and 82 percent against 95 and 97, and the left side is the most compressible text in the whole table. "New Zealand", "launch", "head start", "first" come round again and again. That is what you felt as writing a lot and meaning little.
an Ink sentence, typicalThe whole thing is a bet on attention, and attention is the only currency that matters now.
a bank sentence, typicalAt Les Halles in 1997, on a Friday in the second seating, Jorge cut the gas and said "we finish on the flat-tops," and we did.
Count them: the first has two hollow verbs, no name, no number, no one speaking. The second has two names, a year, a quotation and two events. Same length.
4 · Where this sits with the tells
Short answer: density is the other half of the machine sound, and the ruler does not measure it yet.
The tells are how the sentence is dressed; density is whether there is anything under it. A repair pass that only removes tells can produce a clean, empty paragraph, which the paragraph-selection page named as its first risk and the fact-count key was meant to catch. This table says the fact count is the wrong key: Ink's numbers are already at the bank's rate. The keys that would move the writing toward the bank are a name, a quotation, and an event verb, and the key that would have caught pair 12 is repetition, which no per-paragraph ruler can see.
for Ink · proposed
1. Add two whole-piece meters to the gate's report, never its verdict: names per hundred words and hollow-verb share, with the bank's medians beside them. 2. Replace the pick's fact-count tie-break with names plus quotations, since numbers are not the shortage. 3. A repetition meter on the whole piece, content TTR on the first 600 words, as a key in the pick between drafts: the draft that repeats itself least wins a tie. 4. The commission: this table is the second argument for asking the writer for a reported piece, someone named, someone speaking, something that happened, rather than a stake and a view.
limitsNames are counted by capital letters, so a piece about Apple and Samsung scores higher than one about a fever; the persona's own name is not excluded. Quotation marks count quoted titles as speech. The verb proxy is a suffix guess, not a parser, and counts every -ing and -s word that is not a verb; it is the same guess on both sides. Compression rewards length a little. None of this changes the direction of the three gaps, all of which are two-to-one or more.