Ario — Reading that adapts to you

I

The page

The three ways of setting body copy, a concrete moment, a drop cap, and the four marginal asides.

Before the first paragraph

A lede is the opening: what this is, before any structure. It holds itself to a narrower measure than the rest, because an opening reads better narrow.

A paragraph carrying the argument takes the page’s own measure instead — the same one a heading and a figure take — so that everything on the page lines up down one edge. This is the primitive the books lean on hardest.

Example

A concrete moment, labelled and set apart, so the explanation has something that actually happened to stand on.

A drop cap and a small-caps opening set the first line apart without a heading. The initial is decorative — a screen reader gets the whole phrase, not a stranded letter.

II

Passages

Two-column notes, numbered instructions, a pull quote, a sidebar.

A pair of notes

Sections are set two to a row. They are meant to be read as a pair of observations rather than a continuing argument, and the measure keeps them that length.

Not an essay

Anything longer belongs in a sidebar, or in a part of its own one level down.

1.
Numbered items carry a bold lead-in. The shape a Dummies book uses for anything you are meant to actually do.
2.
They are prose, not a figure. A `Steps` plate shows what the machine does; this shows what you do.

A lesson gets one pull quote, or it gets none.

III

Drawn plates

Sequence, cutaway, comparison, token strip, and one transformation with real values in it.

A worked example

in

the cat sat on the

scored

mat 0.58 · rug 0.21 · roof 0.09

out

mat

Values in mono, because they are data the machine handled rather than words an author chose.

A sequence, with a loop

1

Read

What is in front of it

2

Score

Every candidate gets a number

3

Pick

One is chosen

4

Append

It joins the text

Add it to the text, then repeat

A cutaway

Top

The most specific and most temporary

Middle

What sits on it

Base

The slowest, most general layer

Read bottom to top; the top layer takes the accent.

A system, as boxes inside boxes

The service

What it is told

Versioned, validated, and reusable across services

What runs it

Where it runs

Watching

Click a box to hold the accent on it

Callers

The bank

Containment is the whole vocabulary — inside means owned by, and down the page means closer to the metal. No arrows: a drawing of every path a request can take is stale the week after it is made, while what contains what survives a refactor. Hatch depth carries the level, a darker ground marks a box that is a different kind of thing, and a dashed edge means configured rather than running.

When things happened

1969
Four machines
The first network, and every host known to every other
1974
A protocol nobody owned
The design that let one network become any number of them
1983
The switch
One day, everything changed over at once
1991
Pages
The part everyone means when they say the word
2016
Encrypted by default
The last decade spent making it private rather than larger

The gaps are authored, not computed. Forty quiet years and then three that changed everything is a shape you have to be able to see, which is why this is not a sequence with dates written into it.

Quantities that are not the same size at all

One CPU cycle

0.3 ns

Reading from memory

100 ns

Reading from an SSD

150 µs

One disk seek

10 ms

10⁻¹⁰10⁻⁹10⁻⁸10⁻⁷10⁻⁶10⁻⁵10⁻⁴10⁻³10⁻²

Each mark is ten times the last

On a log axis, because the ratio is the argument. Written in a column these look like neighbours; the hundred-thousandfold between the first and the last is the thing worth seeing.

Two readings

Think of

The one that helps

Set on open hatch, so the pair separates before a word is read.

Not quite

The one to let go of

Set on stipple — quieter, but never coloured wrong.

A sentence in pieces

“The model is learning”

The·model·is·learning?

IV

Data plates

A scoreboard and a table of tones. Rank is carried by hatch density.

A distribution

The chosen candidate takes the one accent; the rest rank themselves by ground.

A table of tones

Ink density is the only variable. Blank means zero, on purpose.

V

Graphs

Four ways of drawing a measurement. One ink throughout: what is being argued about takes the accent, and everything else is the comparison it is being argued against.

A histogram, with a cut

Bins at or past the mark take the accent, so a tail can be pointed at without a second colour.

The same plate, with a band

A stretch rather than a threshold — the middle 95%, drawn as the thing it is.

A scatter, with a subset

The highlighted points are drawn over the rest, because the argument is about them.

Intervals against a truth

Ninety-five per cent of them should cover the line. The ones that miss are the point of the figure.

A trend, two series

The accented line is the one under discussion; the dashed one is what it is being compared with. Ends are labelled, so no key is needed.

VI

Reading off the page

A table you read values from, an expression with its parts named, a number set large, and a figure prose can point inside.

Data, with its values in it

Group	Sat	Passed	Rate
Treated	240	191	—
Control	238	142	—
Withdrew	22	—	—

Rate left blank on purpose: an em dash says missing, a blank says zero to anyone skimming.

Numbers go right and in mono; words go left. The accented row is the one under discussion.

An expression, annotated

x̄¹ ± 1.96² × ( s³ / √n⁴ )

the average, give or take about two standard errors

1x̄what this sample averaged
21.96how wide to go for 95% of the time
3show much the readings varied
4√nthe root of how many were measured

1 in 20

how often a 95% interval misses the value it was built to catch

by design, not by fault

A figure, pointed at

1Where the two readings first part
2The gap the chapter is about

Markers sit at percentages of the plate's own box, so they hold when it is read narrow.

A word the book is teaching is a : marked in the sentence, and opening in place rather than in a card that a phone could never show.

VII

Plates that move

Tabs whose values transition, a fold, a figure that runs, and a number the reader pushes.

A dial, and a figure that answers

Sample size40people

noisysettled

Raise the count and the spread tightens. Told, that is a claim about sampling; done, it is something the reader watched happen.

One plate, three readings

The short reading, open by default.

The longer one, which does not need to be in the way.

A figure that runs

The·cat

token 2 of 7

It starts paused for anyone who asked for less movement, on a frame that stands on its own.

VIII

Brought in

A photograph, a loop, a clip, and a door out. Not drawn by the book, so framed by it instead.

An engraved sphere, shaded by hatch density, resting on a horizon line. — Framed rather than bled. The rule, the number and the caption are the same ones every drawn figure gets.

A dusk landscape, rendered crisp. — **Crisp** only where the colour is the information. Otherwise the book takes the image into its own register: **halftone** is how a photograph got into a printed book before offset. All three apply to a loop and a clip as well.

A dusk landscape, rendered mono. — **Crisp** only where the colour is the information. Otherwise the book takes the image into its own register: **halftone** is how a photograph got into a printed book before offset. All three apply to a loop and a clip as well.

The same engraving, its hatching shifting as if the light were moving. — Starts stopped for anyone who has asked for less movement, and can be stopped by anyone else. A GIF gets a pause only because a still was supplied to freeze on.

A clip, which waits

Clip to come

Never plays on its own. Shown here without a source, which is what any media plate looks like before its artwork exists.

Elsewhere

A link inside a sentence is a Cite: ruled in grey, warming to the accent, so a paragraph does not turn blue at intervals.

IX

The close

Three ways of asking the reader to commit first, the doors a chapter leaves open, and the recap.

Check yourself

A model re-reads the whole sentence on every round. Why not just remember what it worked out last time?

Attention is recomputed from scratch each pass, and that is the point rather than the cost: the reading of “it” that was right in round three may be wrong by round forty.

Before you read on

An interval built to be right 95% of the time misses. Has something gone wrong?

Ask what the other 5% was ever supposed to be.

No. One in twenty missing is the design, not a fault in it — a method that never missed would be claiming more than the data can support.

Try it first

What is the smallest thing a model actually handles?

A token: a piece of text, often part of a word, that the model treats as one unit. “Unbelievable” is rarely one of them.

What this leaves open

If every word is scored against every other, what stops that becoming impossible for a long page?
What the model sees
A shortlist of likely next pieces is not the same as knowing the answer. What is the difference?
How it chooses

The short version

1
One ink; hatch density carries rank and series.
2
Every plate is numbered, so prose can point at it.
3
Motion only where the movement is the explanation.
4
The teaching furniture is learned once and reused everywhere.
5
Borrowed media is framed, never bled, and never moves unasked.