Starlight

Average the light of two hundred thousand galaxies and you get a colour. It is a warm off-white, and somebody reading the newspaper over a coffee named it cosmic latte.

That is the fact this page is named after, and it is not the interesting half. The interesting half is that the first answer was wrong, and the astronomers said so out loud.

The colour of the universe is a correction

The colour was never the point. Ivan Baldry, Karl Glazebrook and the 2dF Galaxy Redshift Survey team were measuring the cosmic spectrum — the volume-averaged light of present-day galaxies — to constrain how fast the universe used to make stars. What colour that spectrum would look like to a human eye was a footnote. Footnote 3, to be exact, and you can still read the original on arXiv:

It travelled. The universe is turquoise is a lovely sentence and it went everywhere, which is how it reached Rochester, New York, and the desks of people who do colour for a living. They spotted the problem in the arithmetic that turns a spectrum into a colour: the software had been handed the wrong white point.

A white point is the colour your eye has decided counts as white, and it is not a property of the light being measured — it is a property of where you are standing. Tungsten bulbs push it yellow. A monitor pushes it blue. The program had been set to a reddish white, and anything measured against a reddish white looks green by comparison. The galaxies had not moved. The observer had.

Within weeks the team put out the correction themselves. Two sentences from Glazebrook in that announcement are the reason this page exists:

Corrected, the colour is a beige-white. NASA's Astronomy Picture of the Day describes it with a care worth stealing: “a conditionally perceived shade of beige.” Not is beige. Conditionally perceived as beige — by an eye, adapted to a particular dark, standing somewhere.

Then they ran a naming contest. Skyvory and univeige were among the entries. Cosmic latte won.

So the most-shared fact about the colour of the universe is an erratum. The reason anyone can trust the number is that the people who got it wrong put the correction where the mistake had been, under their own names, before the journal version went out. It cost Glazebrook an afternoon of being embarrassed in public and it bought everyone since a usable answer.

“Sunlight is said to be the best of disinfectants”

The line everybody quotes is Louis D. Brandeis's, from a Harper's Weekly essay of 20 December 1913 that became chapter five of Other People's Money and How the Bankers Use It. Two things go missing every time it is repeated, and both are worth having back.

The first missing piece is the second clause. Sunlight and electric light are doing two different jobs — one kills what is already there, the other deters what would otherwise arrive. A public changelog is sunlight: it cleans up what we already got wrong. A public repository is the electric light: it changes what we are willing to write in the first place, because somebody can read the diff.

The second missing piece is smaller and, for a project with a fact-checking ethic, more interesting. Brandeis did not write sunlight is the best of disinfectants. He wrote “is said to be.” He was passing on a proverb and marking it as one. The most quoted sentence in the literature of transparency is itself hedged by its author, and the hedge is the first thing a century of quoting filed off.

What “open” means here, in nouns

Open is a word that survives being meant loosely, so here it is as a list of things that exist and can be checked. Every one of these is in the same public git repository as the pages themselves, at the same commit.

One thing is deliberately closed, and it belongs in the same list. Search runs entirely in your browser against that committed index. There is no server to send a query to, no analytics on it, and no log of it anywhere. We are open about our work; we are not curious about yours. Openness that only ever flows one way — institutions private, readers observed — is the ordinary arrangement, and inverting it is the point.

The machines

Six automated gates check this site. None of them was planned. Every one exists because something shipped broken and a person found it — which turns out to be the most important sentence in this section, and we will come back to it.

The gates are only half the story. What each failure taught is the other half, and those lessons are now the operating rules of this repository.

Measure, don't count

The search indexer used to check its own work by counting records: did this page produce at least as many chunks as it has sections? One page has exactly three of the blocks it was chunking on, produced three records, and shipped 12% of its own text into the index looking perfectly healthy. Now every page's records are measured against that page's actual words, and anything under 90% gets re-segmented and a remainder record appended. The run prints a percentage per page, because a page that indexes an eighth of itself looks exactly like a clean one if all you print is a count.

Wait on a condition, never a clock

Two of these tools drive a headless browser, and both used to wait a flat 1.3 seconds for a page to render before reading it. That makes every run a race. On 11 August the indexer lost it: one build produced 629 records where the builds either side produced 637 — eight records and fifteen kilobytes of text gone, with exit code 0 and no warning. Now both tools wait for the page to stop changing, then read it twice and require the two reads to agree, and a page that never settles is a hard error rather than a quiet truncation. Removing the sleep also made the whole run faster, which is the usual reward for replacing a guess with a question.

A tool's worst failure is the page it didn't look at

Zero failures out of zero elements measured looks exactly like a clean page. Worse: when a page silently drops out of a sweep, its failures leave the total, so the number goes down and reads as progress. So the contrast checker reports any page it could not measure as UNREAD, never as ok, counts those separately from the failures, and fails on them on their own terms before contrast is even considered.

The same instinct runs through the rest. Where a check genuinely cannot answer — a heading painted with a clipped gradient has no single pair of colours to compare — it says unmeasured, check by eye rather than guessing. A tool reporting “I can't check this” is reporting a gap, not clearing it. That list once ran to 196 entries and was being read as a footnote — while 90 elements across 45 pages printed invisibly, including the site's own wordmark, so the front page came off the printer with no title on it. The tool had been saying look here the whole time.

Keep every baseline at zero

All six gates sit at a baseline of zero, and that is a maintenance decision rather than a compliment to the code. The contrast gate stood at 110 failures on the morning of 13 August and was cleared that day. A gate that reports “110, same as yesterday” requires a human to remember yesterday's number; a gate at zero requires nothing. A green number you have to read past stops being read.

Two more rules earned the same way. A check that has to be rebuilt is a check nobody runs — the contrast tool was written three separate times as a throwaway scratch script before anyone committed it. And a comment is a note, not a control: one zine's source carries a warning that anything drawn past a certain coordinate lands on the words. The warning was accurate, it was correct, it was right there, and it did not stop the next page from doing it. That is the whole difference between documentation and a gate.

The first honest run is the argument for the gate

The collision checker's first real run over the finished site found seven pre-existing defects nobody had seen — four covers with a caption printed over the title, an unbreakable label overflowing its column onto the body text, and a shipped page where a reader was being shown “ight-years.” None of these was subtle. All of them had been live for weeks, on a site with five other automated checks and a fact-check ledger, being read by people.

What machines are good at, and what they are not

There is a name for the method these gates implement, and it is not ours. The previous essay takes it from Terry Pratchett: First Sight is seeing what is actually in front of you rather than what you expected; Second Thoughts are the thoughts that watch the first ones. Not doubt — supervision.

Which is the honest shape of every tool in tools/. Look again at the table above and notice what it is a list of: six faults that a person found first. Not one gate anticipated the fault that created it. The sequence is always the same, and it does not reverse.

This is not a modest framing of automation. It is a precise one, and it is what makes the machines worth having: a gate never has to notice the same thing twice, which means the noticing can move on. Two thousand eight hundred and fifty lines of checking code is two thousand eight hundred and fifty lines of attention that nobody has to spend again.

Generative AI, said plainly

This is the first page on this site to state it, which is itself a small failure of the openness it is arguing for. Generative AI facilitates a great deal of the work here — drafting, the gates and their output, index and sitemap maintenance, the ledger, source legwork, and the sweeps that touch every page at once. Star Stuff is, among other things, an experiment in whether that can be done in service to accessibility, accuracy and transparency rather than at their expense.

So, in nouns, on both sides.

The rules constraining it are the ordinary house rules, applied without an exemption: aggregators are leads, not citations; a quotation is traced to the book, essay, paper or talk; a paraphrase is never dressed as a quote; the check is logged before the piece ships. With one addition that exists specifically because of the machine — your own summary is an aggregator. A model's account of a source is a secondary source about that source, and it is not exempt from the rule just because it is the thing writing the sentence.

And the mechanism has already failed here, in a way worth publishing. This repository keeps a working-notes file that tells the AI how the site works. A sign-off sentence in those notes got absorbed as house voice, promoted to “our masthead,” and shipped in a zine — where it was quoted with the three words that repaired it removed, and then argued against. It had never been the masthead. It was a colophon on one zine, and it had corrected itself in the next breath. Ryan Boren caught it, and both the essay and the notes now say the same thing: context is not evidence. Working notes are a lead about the site, exactly like any other aggregator.

That is one of three attribution errors published in the last week alone. Another credited five neurodivergent love locutions to a single person on the strength of a citation chain that led, at every step, back to our own pages — a closed loop with no primary at the bottom of it, in which the aggregator being over-trusted was us. The third credited an organisation where it owed a person: Betsy Selvam's byline was on the essay the whole time. Each is in the changelog with what was wrong, what is right, and who caught it. None was caught by a gate.

A public entity

There is a reason a foundation whose subject is learning builds in the open rather than merely publishing when finished. Seymour Papert and Idit Harel gave it a name: constructionism, which takes constructivism's idea of learning as building knowledge structures and adds one condition.

A public entity. Not a finished one — a public one, while it is being made. That condition is the whole of this page: the repository, the ledger, the tools, the errata and the changelog are not documentation of the learning, they are it. It is what the Stimpunks Learning Space practices, and this site is the foundation eating its own cooking rather than recommending a recipe.

And it is why the errors are the most valuable thing in the repository. A private error log teaches one person, once. Fifty published fact-checks are a curriculum — in what a closed citation loop looks like, in how a hedge goes missing, in why the tidiest version of a story is the one to distrust. We would not have any of that if we had waited to be right.

Which lands where the self-portrait already had it, and where the essay before last put the observer: there is no view from nowhere, only an unstated view from somewhere. The astronomers' white point makes that literal. Every measurement carries an observer, and the honest thing to do with an observer is name them and show your working — which is the same reason the changelog says who caught each error, and the same reason this page says how it was made.

The shorter version

The average colour of the light in this universe is a warm off-white, and the first published value for it was wrong. It got fixed because two astronomers put their own erratum where their mistake had been, and because colour scientists who had nothing to do with the paper could see enough of the working to spot the flaw in it.

That is the arrangement we are copying. The pages, the tools, the reasoning, the ledger and the errors are all in one public place, at the same commit, under a licence that lets you take them. The machines here are second thoughts made durable — they hold ground that a person already took, so the person can go and look at something new. And the looking is still ours, because it has to be.

Everything we got wrong →