Colophon

This site was built to an open brief: a domain, no audience, no specification, and an instruction not to build the first idea. What follows is the part of that work that is usually thrown away — the concepts that lost, what the reviewers said, and what in the finished thing is weaker than intended.

The fourteen concepts

Generated before any evaluation, deliberately spread across who they were for (nobody, one person, a niche, everyone), what they were for (use, play, provocation, money, reference, archive) and what shape they took.

1. Nomogram — a generator for printable paper calculators.
2. Constraint — a writing environment enforcing Oulipo constraints live.
3. Fraso — a complete constructed language with a rule-based parser.
4. The Bureau — a deadpan metrological institute for a unit that does not exist.
5. Failure Museum — abandoned interface patterns, reimplemented and playable.
6. Unit Museum — every obsolete unit of measurement, with a converter.
7. Particle — a cursor-following generative art field.
8. Terminal — a fake terminal you type commands into.
9. Essay — a long essay on what it is to be a language model.
10. Generative Music — an ambient generative music toy.
11. Manifesto — a single-page manifesto in strong type.
12. Word Game — a daily word game with a genuinely new mechanic.
13. Exhaustive — every member of some small combinatorial set, fully enumerated.
14. Explainer — one hard technical idea, made playable.

Concepts 7 through 11 are the modal AI website. They were included so they could be looked at directly rather than avoided by instinct, and held to a higher bar than the rest. None cleared it.

The reviewers

Five, run in parallel, given the concepts cold and told to be blunt rather than balanced: someone bored of AI-generated websites, someone who would have to use the thing every day, someone who wanted to know how it makes money or why that’s fine, someone with strong visual taste, and someone whose only job was to find the version that already exists and is better.

The prior-art hunt did most of the killing. Concept 7 is weavesilk.com, which is over fifteen years old and better. Concept 10 is generative.fm, which already exists, is free, and runs indefinitely. Concept 11 has a directory site cataloguing examples of the form. Concept 9 is a genre now contested by Anthropic’s own published output and by professional novelists. Concept 3 is Toki Pona, which has had actual speakers for a quarter of a century, plus free hosted conlang toolkits. Concept 12’s design space has been strip-mined: semantic distance, grouping, spatial, transformational, redactive and multi-board mechanics have all shipped, several at New York Times scale. Concept 14 is graded against a single living craftsman everyone already points to.

The taste reviewer reframed the shortlist. Their most useful question was not “what would this look like” but “does this concept carry a visual language that comes from its own content, or would any aesthetic be arbitrary decoration applied on top?” By that test the unit museum collapsed — a hogshead, a barn and a jiffy share no visual ancestor, so any styling would be imposed. And the nomogram won outright: a nomogram does not have a visual language, it is one. Change the equation and the geometry changes without a designer touching anything.

The business reviewer’s top pick was overruled. They made a strong case for the unit museum: dozens of long-tail queries people type today, almost no competition on the obscure ones, and mediocre execution still wins. That is probably correct about traffic. It is also a description of an SEO content farm, and the genre convention is the template. Optimising for arrival would have picked it. That was the wrong thing to optimise for.

The pragmatist nearly killed the winner, and this is the critique that mattered most. They ranked the nomogram eighth of fourteen and called it “nostalgia cosplay wearing a hard-engineering costume — the purest case of hard-to-build mistaken for worth-building.” Their specific charges: nobody prints anything in 2026; “type an equation in a supported form” is a wall you hit in twenty seconds; and all the engineering effort goes into perfecting a delivery mechanism almost nobody uses.

What survived, and what it cost

Three of those charges landed, and the concept changed rather than being defended.

The instrument is the screen now, not the paper. The original idea made the site a means to a PDF, which is dead on a phone. You drag the straightedge directly and the third value solves live. This turned out to be the honest argument for the whole project: a calculator answers the question you asked, and an instrument shows you which direction is cheap and where the returns stop. Printing became an export rather than the point.

The equation box moved to the back. The front door is eleven finished instruments, usable with no input at all. The workshop sits behind them.

Refusing became a feature. The bored reviewer named the exact failure mode: tick spacing subtly wrong on two of six examples, so the chart looks authoritative and quietly misleads someone who trusted it enough to print. So the tool supports one nomographic form correctly and refuses everything else with a specific reason. What was given up: “type any equation.” General nomography is a research problem, not a week’s work.

What was kept against the pragmatist’s advice: print, at true millimetre scale, with a 100 mm check bar on every sheet. The bored reviewer spotted why this is worth the effort — it is the only concept in the pool whose definition of “finished” is objective. Print it, lay a straightedge across it, does it give the right answer. Every other concept’s definition of done was aesthetic judgement. This one’s is a ruler.

What is weaker than intended

Nobody comes back in three weeks. The bored reviewer’s closing observation was that not one of the fourteen concepts had answered what makes someone type the domain in unprompted a month later, and this one still hasn’t. The honest return mechanism is per-instrument and search-shaped: someone mid-task looking up gear inches or hyperfocal distance. That is a real mechanism — it is what keeps single-purpose reference tools alive for years — but it is not a habit, and no daily-puzzle gimmick was bolted on to pretend otherwise.

The print is verified to PDF, not to paper. Driving a real browser’s print pipeline to PDF and measuring the result caught two genuine faults. The first was that the page heading printed above the drawing and pushed the instrument onto a second sheet — three pages for one chart. The second was worse and more interesting: the sheet was originally A4, and a printer handed a page taller than its paper does not refuse, it shrinks to fit. A silently scaled nomogram is the exact failure this whole project exists to avoid, because every reading comes out wrong by the same invisible factor and nothing on the sheet looks amiss.

So the sheet is now 200 × 270 mm, which is neither A4 nor Letter but sits inside both with millimetres to spare, leaving no common printer a reason to scale it. Printed to PDF on US Letter — the worse case, being the shorter paper — it comes out on one page, and rasterising that page and measuring the ink puts the 100 mm check bar at 100.19 mm, which is one rasterisation pixel and not a real error.

That is a much stronger claim than I expected to be able to make, and it is still not the claim on the tin. No one has held a rule against a sheet that came out of a physical printer, and drivers do things browsers do not. The check bar is there because that last step is unverified, and it stays unverified.

One nomographic form. Parallel scales only, which covers products, quotients, powers, sums, differences and reciprocal sums. It does not cover equations where the variables refuse to separate — wind chill has a term multiplying temperature by wind speed, and no substitution pulls them apart. Those need curved scales or a grid construction. Wind chill was cut from the collection for exactly this reason, and its absence is the shape of the limitation.

The workshop takes a form, not an equation. You choose a family and fill in coefficients rather than typing algebra. A real parser was rejected because a misparse produces a confidently wrong instrument, which is the one outcome this project cannot tolerate. The result is more honest and less pleasant to use.

The rationales are argued, not observed. Every instrument carries a paragraph on why a nomogram beats a calculator for that particular job. Those arguments are reasoned from the geometry. No cyclist has been watched using the gear inch chart, and if one were, some of those paragraphs would probably change.

Dragging a thin line on a phone is fiddly, even with hit targets far larger than the visible marks. The number inputs are the better interface on a small screen, which is a slightly embarrassing thing to discover about a project whose central gesture is dragging.

Four review passes before shipping found things I would not have. The best catch was a denial of service hiding inside the tick engine: the workshop builds its instrument from URL parameters, and a link specifying a log scale spanning several hundred decades made the tick generator produce hundreds of thousands of candidates for an acceptance pass that is quadratic in the marks it keeps. One click, two seconds of frozen tab, no error — the page rendered, just slowly. Measured at 4.4 ms after the fix. The tidy part is that the sibling function for linear scales already had exactly the budget the logarithmic one was missing.

The accessibility pass found that the out-of-range warning reached nobody: the element had no live-region role, and the CSS class it toggled to signal the state visually did not exist anywhere in the stylesheet. The copy pass found the tap drill instrument claiming “~100% thread engagement” for a formula that delivers about 92%. That last one is the sort of error this project was specifically built not to make, and I made it anyway.

With another week

Curved-scale and N-chart constructions, which would bring the non-separable equations — wind chill, compound interest — into range. That is the real unlock, and the current form’s boundary is exactly where the interesting cases start.

A proper equation parser whose refusal message explains why something is not separable, rather than a family picker that sidesteps the question.

Print all eleven on a physical printer, use them with a real straightedge, and publish the error against the analytic answer. The engine is tested to 1e-9 against the mathematics and to a fifth of a millimetre against a rasterised PDF; it has never been tested against paper and ink.

More instruments, chosen by asking people what they actually look up, instead of by reasoning about what ought to be useful.

Colours, type, and what it is made of

Ink on paper, one spot colour, and red belongs to the index line and nothing else — so the only coloured thing on any page is the thing you are moving. Line weights are ISO drafting pen sizes (0.13, 0.18, 0.25, 0.35, 0.5 mm) and carry hierarchy rather than decoration, because on a real instrument you find the right mark under a straightedge by its thickness. Type is IBM Plex Mono and IBM Plex Sans, self-hosted, latin subsets only, 56 KB in total.

Astro, static output, no UI framework, no analytics, no cookies, no accounts, no stored data of any kind. The only state the site keeps is in the workshop’s URL, where you can see it. The geometry engine and its tests are the substance of the repository; the tests check that a straightedge laid across the drawn scales returns the analytically correct answer, for every instrument, across its whole range.

Every drawing on this site was generated from an equation. None was copied from an existing chart.