Releases 1.232.1 – 1.266.0

Newest first. Three numbers, and which of them moves says what happened to you rather than how much work it was.

The first moves when something that worked stops working: a query that used to answer and now does not, a saved list that will not load, a link you shared that no longer opens what it opened. It has never moved, and that is the whole of what it is for.

The second moves when the page answers something it could not answer before, or answers it better than it did. New grammar is the usual reason. A correction is the other, and a correction is not a break even though the answer changes underneath somebody: "days be christmas" used to say 154 days and now says nothing, because it was reading Belgium out of the word "be", and taking that away left the page more right rather than less useful. An answer that gains a drawing of what it just said counts here too.

The third moves when nothing you type comes back any different.

Numbers get skipped when it takes more than one commit to introduce a release, so only the release from the last commit goes out.

1.266.03 August 2026

A reader that says what kind of line it wants

Nothing you type answers differently. This is the last of the floor being replaced: what changes is how a question finds the reader that can answer it.
Ninety readers were called in an order worked out by hand, and that order was the grammar. A reader placed one line too late never sees the query it was written for, and adding one meant guessing where it goes, with fifteen thousand tests as the only thing checking the guess.
A reader now says what a line must hold for it to be worth asking, next to where it is written: a fuel economy on both sides of a joining word is the economy reader's line, and no order in a file has to say so. The question is read once into what it is made of, and the readers whose claim fits are asked, most specific first.
Six have moved so far and the suite counts the rest, so the chain can only get shorter: eighty-four still placed by hand, from ninety.
Registering the paper sizes was the first thing the new layer found: the page knew the word "paper" and not one of the things it names, because the sizes lived inside the pattern that reads them and nothing else could ask.

1.265.03 August 2026

A page that answers a question can now say what its words are

Nothing you type answers differently today. This is the floor being replaced under the house, and it is worth a release note because of what it makes possible next.
The page reads "how far is new york from tokyo" and, until now, could not tell you what "how" was, or "far", or that "new york" was one name rather than two words. Those things were spelled out letter by letter inside forty separate patterns, which is how a page ends up knowing a sentence without knowing any of its words. A module now declares a vocabulary once and gets three things from it: what a reader matches, what the suggestion list offers, and what the page can say a word is.
Saying "you can ask me about this" and "this is a word I know" turned out to be the same sentence said twice, in two registries built a day apart. They are one now, which named a hundred words on its own.
A name is a phrase as often as a word, so the longest run that means something is read first: New York is a place, not two strangers.
The suite counts what is left, over every question it can answer: four hundred and nine words the page reads and cannot name, down from six hundred and forty-three. It only comes down, and a reader added with its own private vocabulary raises it.

1.264.03 August 2026

A name one typo away comes back

distance from tokyo to lodnonlodnon isn't a place I know9,564.34 km, reading it as london
how far is toyko from londontoyko isn't a place I know9,564.34 km, reading it as tokyo
100 usd in eruERU isn't a currency I have a rate for87.90 EUR, reading it as EUR
1 kilometr in mileskilometr isn't a unit I know0.621371 miles, reading it as kilometre
A refusal that says "I don't know that word" is standing next to the table it is refusing from, and until now it would not look. The tables can be listed, so the nearest spelling is found, the corrected question is run, and its answer goes underneath. Nothing is acted on: what was typed is what stays in the box, and the reading is offered, not taken.
Two letters the wrong way round count as one mistake rather than two, because that is what typing them is. It is the commonest typo there is, and counted the usual way "toyko" is too far from Tokyo to be worth mentioning.
Underneath, three things this page did not have. A module can now say what a word is, not just what it means: the converter used to ask its own table whether "mpg" was a unit, get a no, and print it, on a page that answers in mpg. A refusal can name the readings it would accept instead and have them run before they are shown, which is the rule the suggestions have always lived by. And a word can be measured against every spelling of its kind.
The suite counts what is left. A hundred and twenty-eight of three hundred and six refusals now send the reader somewhere, and that number only goes up: a new dead end fails the build.

1.263.03 August 2026

Every word this page has, offered wherever a word can go

3pm innothing offeredtokyo, london, new york
3 days beforenothing offerednow, today, christmas
57 mpg in lmpg isn't a unit I knowl/100km, offered as you type
57 mpg in l/90kmcan't read that yet3.71391 l/90km
blood mooncan't read that yetSun, 31 Dec 2028 · total
sec(0.5)can't read that yet1.13949392732
Every gap in the suggestion list had been found the same way: somebody typed a question, got silence back, and it was fixed where it was found. The suite now types every question it knows one word at a time and counts the ones still met with silence. There were sixteen hundred. There are twenty-five, and each of the rest is a number being typed rather than a word.
What closed them is one rule instead of five: every word this page has a word for, tried wherever a word can go, and kept when it answers. Nothing about it knows the grammar, so it cannot be wrong about the grammar. A city goes after a clock, a feast after a count of days, a calendar after "in the hebrew", a decimal after "50 percent as a", and none of that is written down as a rule anywhere.
Each part of the page now says what it can be asked about, in the file that knows. Fuel economy had been read for a year by a page that would not offer it, because the offering lived in a file that only knew the unit table; pace was the same story told twice. Both are offered now, and so is whatever gets added next, without anybody remembering to add it in two places.
The hundred in l/100km turns out to be a convention rather than a fact about the unit, so the distance is read off the question: l/90km, l/km, litres per 50 km. And half a mileage says which half is missing. "57 mpg in l" used to come back "mpg isn't a unit I know", which is false on a page that answers in mpg.
Eclipses. "lunar eclipse" is what an almanac calls it, "eclipse" is what somebody types who has one kind in mind, and "blood moon" is folklore for the colour of a total one. All three are one question. The sun's is refused plainly, with the next lunar one underneath it: the earth's shadow on the moon is the same for everybody who can see the moon, and the moon's shadow on the earth is a track a couple of hundred kilometres wide.
The three reciprocal trig functions, whose absence was conspicuous next to the tangent: the derivative of tan is printed as 1/cos(x)^2, which is sec(x)^2 and had no name here to be called by. With them the hyperbolic three, their inverses, and the arc- spellings.

1.262.02 August 2026

One answer, and then the next one

next two full moonsnothingWed, 29 Jul 2026 · 10:37 am · Fri, 28 Aug 2026 · 12:19 am
next two leap yearsnothing2028 · 2032
last two leap years before 2090nothing2088 · 2084
next two holidays in the usnothingLabor Day · Columbus Day
how many full moons in 2027nothing12 full moons
Reported, five times over, and it was one thing rather than five. Every reader here can name the next one of the thing it knows about and not one of them could name two: "next full moon" answered and "next two full moons" did not, and so it went for the leap year, the holiday, the equinox and the Friday. It looked like a handful of missing phrasings and it was a missing idea.
Each of those readers works out "the next one" inside its own arithmetic, as a loop from today that stops at the first hit, and then prints the answer as a sentence. The instant is thrown away at the moment it is formatted, so nothing outside that reader can step along the series it just walked one pace of, and a second one had to be written into every reader by hand. That is the same code fifteen times over and fifteen chances to disagree about what "next" means.
There is no table of things that recur in this release. There are two older facts about the page and finishing them was the whole of the work. An answer now says when it is: `at` is the instant an answer names, a field the "how many days until" reader invented and only a few readers filled in, and every reader that prints a date fills it in now. And there has only ever been one clock, which every reader reads the time through, so pinning it moves the present for the whole page at once. "next two full moons" asks the page for the next full moon, takes the instant off the answer, pins the clock past it and asks again. Whatever answered the first time answers the rest, in its own words, and there is still exactly one definition of what the next full moon is.
Two things came free. An anchor, because "last two leap years before 2090" makes 2090 the present for the length of the question, and a reader with no notion of an anchor answers about one correctly without knowing it was asked for. And every reader written after this one, which gets counted and anchored questions the day it says when its answer is. The suite counts that last part rather than trusting it: an answer whose headline prints a date and does not say which day it is fails the build. It stood at 85 of 304 when this went in and it stands at all of them.
What it will not do is invent. A step that does not move ends the walk, so "last two full moons" stays quiet rather than printing one answer twice: the moon reader has no reading for the one behind us, and a list is not the place to add one.

1.261.02 August 2026

Three quarters of the name it could not find was the question

los angeles is what timezonelos angeles is what isn't a place I knowPacific Daylight Time
tokyo is in what timezonetokyo is in what isn't a place I knowJapan Standard Time
what time zone is it in tokyoit in tokyo isn't a place I knowJapan Standard Time
what timezonewhat isn't a place I knownothing, which is a line still being typed
Reported. The question word can lead or it can trail, because the place is what the sentence is about either way: "what timezone is los angeles" and "los angeles is what timezone" are one question asked by two people, and only the first was read. The second matched the plain shape of "tokyo timezone" with the whole of the asking swallowed into the place, so the page replied that "los angeles is what" is not a place it knows. It is not, and nobody typed it. Three quarters of that name is the question and one quarter is its subject, and a reader told to check their spelling has nothing there to check.
The same fault in the shape that puts a pronoun where the place goes. "what time zone is it in tokyo" is how the question is spoken at least as often as it is written the tidy way, and the "it" was being read as the first word of the city's name. Bare "it" already means this clock here, so only the pair comes off and the place behind it answers.
And the question with no place after it at all now says nothing rather than blaming the word "what" for not being a city. Every keystroke gets an answer here, which means every half-typed line gets one too, and a refusal aimed at a word the reader has not finished typing is noise at exactly the moment it is least wanted.

1.260.02 August 2026

A clock somebody owns, which is the only kind anybody has

my 9am in los angelesnothing6:00 am in Los Angeles
time tomorrow at 9am my time in los angelestomorrow at 9am my time isn't a place I know6:00 am in Los Angeles
my time in la + 43 hoursnothing12:00 pm, Sat 25 Jul
5 hours after 3pm in tokyo3pm in tokyo isn't a time I can start from8:00 pm
what is my time in lanothing5:00 pm in LA
Reported, three times over. "9am in los angeles" has two readings and this page has always taken the first of them, which is my nine o'clock shown there, and the person who wants to be understood says which nine they mean. Saying it was the one way to be told the question was not a question. So the possessive is now filler in front of a clock, the same as the "the" of "the meeting at 3pm" has been, and it is dropped rather than acted on, because the reading it asks for is the reading taken anyway: the clock a conversion starts from is your own unless you name another. Only "my" and "our", which can mean nothing else. "the 9am in tokyo" is the 9am flight and means the opposite thing, so it is left alone rather than answered wrongly. The first of these went out in 1.259.0 with no entry to its name, which is what this paragraph is for.
The second one failed a reader further along. "time ..." is a question this page owns whole, and it hands whatever follows to the place resolver unless it can see a clock in what is left. It saw a clock by reading a line that opens on one, so a day in front of it put the whole tail beyond recognition, and the answer named a place nobody had typed: the day, the clock and the zone gathered into one imaginary city, on a line where all three had been understood separately. It now reads a day in front of a clock as readily as a clock, and hands the rest to the reader that has moved a day behind a clock since meetings went in. A misspelt place in one of these is blamed by name again, so "time tomorrow at 9am in narnia" is Narnia's fault and not the sentence's.
The third is arithmetic on a clock named by its place. "my time in la + 43 hours" is what it says there now, and 43 hours on from that, and the shift reader could start from three things only: the word "now", a clock reading, or a bare city. The phrase was none of the three, though the question reader has read "time in la" and "la time" since the day it went in. It hands back an answer rather than a place, so nothing else could use it, and that is now a reading of its own that both of them share. The same seam fixes the place that still carries the word introducing it, which is why "5 hours after 3pm in tokyo" had been refusing a line it had just read the whole of.
One guard had to be narrowed for the last of those. A number, a unit, an "in" and a word is the shape of a conversion, and the clock declines that shape at the door so the converter gets its turn: "5 hours in minutes" is not a question about a city. But "5 hours after time in tokyo" is that shape too, and it was turned down here and turned down by the converter for having no unit in it, so the line said nothing at all. It is let through now only where the front of the line is a length of the clock, which leaves "100 usd in eur from 1 jan 2020" the money question it is.

1.259.02 August 2026

Every rule is local, which is why the answers read like that

derivative of sin(x)*cos(x)cos(x) × cos(x) + sin(x) × (-sin(x))cos(2x)
derivative of x*x*x(x + x) × x + x^23x^2
derivative of sin(x)^2+cos(x)^22sin(x) × cos(x) + 2cos(x) × (-sin(x))0
derivative of x^2*x^32x × x^3 + x^2 × (3x^2)5x^4
What makes a differentiator trustworthy is that every rule in it is local: each one is a line out of a table and none of them has to know what the rest of the expression looks like. It is also exactly why the answers read the way they did. The product rule on sin·cos hands back cos·cos + sin·(-sin), which is right, and is not what any person would write down. So the tree is now taken apart into terms -- a sum of products, each of them a coefficient and a set of factors with powers -- and the ones that are the same thing add up. Most of what that fixes is not trigonometry at all: x·x is one factor squared, and the derivative of x·x·x used to come back as three terms every one of which was correct and no two of which could be added.
Three identities sit on top of that, and only three: sin²+cos² is one, cos²-sin² is cos of twice the angle, and twice sin times cos is sin of it. They are the ones that make an answer shorter rather than merely different. Every other identity in the book trades one form for another of the same size, and a calculator that applies those hands back an expression nobody asked for and nobody can predict. The angle is carried through rather than assumed, so the derivative of sin(3x)cos(3x) is 3cos(6x).
The guard is the half of this that keeps it honest. Collecting is nearly always an improvement and is not always one, so both printings are produced and the shorter wins: a rewriter with no way to decline is one that will eventually hand back something longer than it was given and call it simplified. Coefficients are kept as a numerator and a denominator rather than as a number, or the derivative of sqrt(x) prints as "0.5 / sqrt(x)" -- the same value, a worse answer, an exact half turned into a decimal for nothing. And a negative power prints as the division it is, which matters twice, because "-2x^-3" is also the shorter string and would otherwise have won.
Two things the collecting found. A constant folded into a division only where it comes out whole, since "-1 × (cos(2x) / 2)" is not -0.5cos(2x) to any reader. And nought on the bottom now refuses: the quotient rule over a constant zero comes out nought over nought, tidies to nought, and "the derivative of 1/0 is 0" is a confident answer about a thing that does not exist. Every rewrite is checked against a central difference at three values of x, because a simplifier that changes the value is the one failure here that would never look like one.

1.258.02 August 2026

The first reader here that works on the sum instead of the answer

derivative of x^2nothing2x
d/dx sin(x)nothingcos(x)
integral of x^2nothingx^3 / 3 + C
integral of tan(x)nothingI cannot integrate that
Reported. Every other reader in this file works on numbers, and you cannot differentiate a number: the arithmetic engine evaluates as it parses and hands back a total, and by the time it has finished the expression is gone. So this one has a parser of its own that builds a tree instead of working anything out. It is not a flag on the other one, because that reader's entire shape is "compute as you go" and threading a second mode through it would leave two behaviours tangled in a function that currently has one.
Differentiating is the complete half. Every rule is local, so anything built out of the operators, the brackets and the named functions can be taken apart and done: the product rule, the quotient rule, the chain rule every time, and the power rule in all three of its cases, including the one where x is in the exponent. The answers are checked against a central difference rather than against this file's own output, which is the only check on a differentiator worth having -- a symbolic answer that is wrong is wrong in a way no amount of reading it will reveal.
Integrating is deliberately not complete, and the refusal is the point of it. There is no general method: plenty of perfectly ordinary expressions have no antiderivative that can be written down at all. What this knows is the standard table, sums of it, constant multiples of it, and a linear argument inside a function, which is the substitution everybody does by eye. Everything else says so out loud. A reader that always produces something is a reader that is sometimes making it up, and this file would rather be silent than plausible.
Two bugs found on the way, both of the kind that look right. Implicit multiplication was binding tighter than the power, so "3x^2" gathered "3x" and squared the pair and the derivative came back 18x -- true, of a different expression. And the rules generate a great deal that is correct and unreadable: "3(x ^ 2)", "-1sin(x)", "e ^ x × ln(e)", "(x + 1 - x) / (x + 1)^2". An answer nobody can read is a wrong answer for this purpose even when every symbol in it is right.

1.257.02 August 2026

Eight words it would print at you and two it would answer to

first quarter in tokyo in 1996nothing12 first quarters
first quarternothingWed, 19 Aug 2026 · 10:47 pm
waxing gibbousnothing21 Jul → 29 Jul 2026
quarter in cm0.635 cm0.635 cm
"moon phase" has printed eight words since the moon went in and would answer to two of them. The namer decides between the eight by how much of the disc is lit, so the page would cheerfully tell you the moon was at first quarter while nothing in the file had ever worked out when first quarter is: the phase series took a boolean and computed the new moon and the full. The quarters are their own series, with one term added at the first and taken off at the last, the two not being symmetric about the full moon because the moon does not travel at one speed. The four shapes between the four instants are not instants at all -- a crescent is the stretch between a new moon and a quarter -- so those answer with both ends and how long it runs. A name a page prints and cannot be asked about is a name it only half knows.
The reported query failed three readers away from the moon. "in" is the inch, so the rewriter that turns fraction words into numbers read "quarter in" as a quarter of an inch and handed the readers "first 0.25 in tokyo", which nothing wanted. That rewrite is right far more often than it is wrong -- "quarter in cm" really is 0.635cm -- and what tells the two apart is not the fraction or the "in" but the word after it, which now has to be something convertible for the inch to be the reading taken.
One word is deliberately left where it was. "last quarter" on its own is a calendar quarter, which is what this page has answered since the periods went in and what most people asking mean by it; the moon takes it only when something says the moon was meant, by the word, by a place, or by a window to look in. "first quarter" and "third quarter" have never answered as anything here, so they were free.

1.256.02 August 2026

A year is not a day, and thirteen of a thing is not one of it

full moon in tokyo in 19966 Jan 1996, alone13 full moons, all of them
second full moon in tokyo in 1996nothingMon, 5 Feb 1996 · 12:58 am
last full moon in 1996nothingTue, 24 Dec 1996
20th full moon in 1996nothingthere are only 13
Both reported, and they are the same fault twice. The date engine has always known whether what it read was a day, a month or a year, and says so on every date it hands back. Nothing that reads the moon ever asked. So a year collapsed to its first instant and "full moon in 1996" answered with the 6th of January: a full moon, in 1996, and not the question. There are thirteen of them. Picking the first and printing it alone answers a question about a year as though it had been asked about a day, and the giveaway is that it never said thirteen — a reader had no way to tell the answer was one of a set unless they already knew. Now a window with more than one phase in it is reported as what it is, and a month, which usually holds exactly one, still answers as one thing and says so.
And once a window can hold several, they can be counted through: "the second full moon in 1996", "the last", "the 13th". Counting is the only way to name one of thirteen otherwise identical things. Asked for the twentieth of thirteen, the answer is that there are thirteen, rather than the last one handed back with nothing said about the difference, which would be a lie of omission about the one thing actually wrong with the question. The place still moves which day each lands on, so that January full moon is the 6th in Tokyo and the 5th in New York, exactly as it was two releases ago.

1.255.02 August 2026

Where the moon is, and not merely when it is full

supermoonnothing24 Nov 2026 · 360,768 km away
lunar eclipse in tokyonothing28 Aug 2026 · below the horizon there
lunar eclipse visible in tokyonothing7 Jul 2028 · 12° up
lunationnothinglunation 1,281 · 29d 8h long
blue moonnothing31 Dec 2028
2026 moon phasesnothing13 full moons and 12 new moons
Everything the moon could be asked before this was about when a phase falls, and the series for the phases answers that without ever knowing where the moon is in the sky or how far away it is. Three of these need both. A supermoon is a full moon near perigee, so it is a question about distance; whether an eclipse can be seen from a particular city is a question about whether the moon is above that city's horizon, so it is a question about position. So Meeus chapter 47 went in: the periodic terms for the moon's longitude, latitude and distance, truncated where the terms stop mattering at this scale. Against Meeus's own worked example it lands within about four arcseconds and five kilometres of the full series, which is a hundred times finer than any question being asked of it, and every term left out is one that would otherwise have been carried in every visit to this page forever.
An eclipse of the moon is not a track across the ground the way an eclipse of the sun is. The earth's shadow falls on the moon, and everybody who can see the moon at all sees the same eclipse at the same instant, so "visible in Tokyo" reduces to one question: is the moon above Tokyo's horizon while it is happening. The two examples above are the same August eclipse from two places, up over New York in the small hours and under the horizon in Tokyo where it is the middle of the afternoon, and asking for one that can be seen from Tokyo therefore skips forward two years to the next that can. The dates, kinds and magnitudes are checked against the published record rather than against this file's own output, which is the only check on a series worth having.
The rest are smaller and were waiting on nothing but being asked for. A lunation is one turn of the cycle, new moon to new moon, and it is not a constant: the moon runs fast near perigee and slow near apogee, so one can be six hours longer than another and the mean everybody quotes is the average of a thing that is never quite that. Blue and black moons are the second full and the second new moon inside one calendar month, both of them folklore with an older seasonal definition behind them, so the answer says which reckoning it used rather than leaving somebody to discover they were told the other. And a season named at the end of a moon question is now a window to search in rather than a day to report on, which is what "new moon in tokyo in the winter" was always asking for.
Two honest limits. Neither "supermoon" nor "micromoon" is defined by any almanac, so a line has to be drawn somewhere: these draw it at 361,885km and 405,000km, say so in the answer and give the distance, so that anybody who draws it elsewhere can see where this one fell. And a penumbral eclipse of very small magnitude sits inside the uncertainty of the method itself, which computes the July 2027 one as grazing the shadow rather than just catching it; that is a penumbral magnitude of six hundredths, a dimming no eye can detect, and widening the threshold to catch it would risk inventing eclipses that are not there.

1.254.02 August 2026

One moment, two clocks, and two ways of being taught what a season is

full moon in tokyonothingWed, 29 Jul 2026 · 11:37 pm
new moon in tokyonothingThu, 13 Aug 2026
new moon in new yorknothingWed, 12 Aug 2026 · the day before
when does winter start21 Dec…and 1 Dec by the other reckoning
The moon read the end of its question as a date and only ever as a date, so "full moon in december" answered and "full moon in tokyo" did not. The reasoning that left the place out was half of the truth and the better-sounding half: a phase is a moment, the moon is full at the same instant for everybody, and where you happen to be standing cannot move it. What standing somewhere does move is the date and the clock that instant gets written down as, and that is the whole of what this prints. The two new moons above are the argument. They are the same moment. In Tokyo it is two in the morning on the 13th and in New York it is one in the afternoon on the 12th, and a reader who could not be told where you were was answering for somewhere else. It now reads that tail the way every other reader of the sky here reads one, through the same function that learned "sunrise tomorrow in tokyo" and "sunrise in june": a place, a day, either order, or neither. Nothing was added for the moon in particular, which is the point. "days until the full moon in tokyo" came along with it and cost nothing.
And a season now says both of the things it can mean. The astronomical reckoning runs solstice to equinox, so winter opens around the 21st of December; the meteorological one is three whole calendar months, so winter is December, January and February, which is how the forecasters count and how a great many people were taught. Neither is a mistake and they are about three weeks apart, which is enough to put a date in a different season depending on nothing but who taught you. The astronomical one still leads, being what this computes from the sun and what it has always answered, but the other is no longer merely named in the echo as the road not taken: it is given, with its date, underneath. Either can be asked for outright, and in the weeks the two disagree, "what season is it" says so plainly rather than picking one and being right at half the room.

1.252.02 August 2026

One problem solved six times and abandoned at six different points

days until the full moonnothingin 5 days
days until summer solsticenothingin 332 days
when is the full moonnothingWed, 29 Jul 2026 · 10:37 am
days until sunsetdays until isn't a place I knownothing
The last release taught the longest day two sentences it could not hear, and the interesting part turned out to be how many other readers could not hear them either. Two sentences can be wrapped round almost anything here without changing a single thing that has to be computed: "when X", and "how long until X". Every reader that answers about a moment was spelling both out inside its own pattern, and every one of them had stopped somewhere different. The moon took "how long until" and not "days until", and wanted a "next" in the middle of "when is the next full moon" that nobody types. The solstice took "when is" and could not be counted to at all. The seasons answered "when does winter start" and not "when winter". The clocks insisted on a verb after the word. The sun read whatever stood in front of its keyword as the name of a city, which is where "days until isn't a place I know" came from. That is not six faults. It is one fault fixed six times and left unfinished in six places.
So it is done once now. A reader that knows which moment it is talking about says so, by hanging that moment on the answer it hands back, and one pair of wrappers reads both sentences for all of them. Anything holding a date can be counted to, in any unit the rest of the grammar counts in, and asked about with or without the verb. They run after every other reader has declined, which is what makes this safe: nothing that answered yesterday can answer differently today, and the only thing these can do is turn a silence into an answer.
They also turn one kind of answer into a silence. "days until sunset" is a question with a mistake in it, since the sun sets this evening and every evening and no number of days is the reply, but it was being met with a complaint about a city nobody had named. Where the line is understood and simply has no date in it to count to, the answer is nothing at all. A real attempt at a place is still told: "sunset in springfield" says springfield is not a place it knows, which is the whole reason that silence had to be kept clear of it.

1.251.02 August 2026

The word the question is asked with, and the question somebody actually had

when longest day in parisnothingSun, 21 Jun 2026 · 16h 11m
days until longest day in parisnothingin 333 days
when christmasnothingin 154 days
how long until the longest daynothingin 332 days
Both reported, and they turn out to be two faults rather than one. The first is a missing word. Every reader here that answers "when is X" wanted the "is", and "when X" is how the question gets typed by anybody who is not writing out a full sentence: "when christmas", "when easter", "when longest day in paris". The word carries the entire question and named no part of the answer, which is how it went unnoticed for so long. What may not follow a bare "when" is a verb. "when does winter start", "when do the clocks change" and "when does the sun rise in tokyo" are other readers' sentences, and every one of them would survive having its "when" removed and arrive somewhere else as a fragment meaning something different, so a verb after the word leaves it alone.
The second is a question this could compute and could not be asked. The longest day has been answerable since the sun went in, and it answers with a date; what somebody wants to know is how far off it is. Counting lives in the date grammar, this reader is not in it, and the two had never been introduced, so "days until longest day in paris" handed the date parser "until longest day in paris" and it had nothing to say. It counts in any unit the rest of the grammar counts in, and in days when the question names none. The count runs forward without exception: asked in July, the longest day of this year is a month behind, and "days until" is transparently not asking about that one, so it steps to the next. Asked plainly it still answers about the year in hand, that being a different question and an equally fair one.

1.250.42 August 2026

An answer that was one word long and a hundred characters wide

pi to 42 placesoff the side of the pagewrapped
pi to 100 places…and the page scrolled sidewaysfive lines
Reported. Almost every answer here is short enough that none of this came up, and the ones that are not are a single unbroken run of digits: a hundred decimal places of π is a hundred-character word, and a word is exactly what a browser treats it as, so there was nowhere in it to wrap and it went out of the right-hand side. The cost was not only the answer. The document itself grew a horizontal scrollbar, which moved the nav, the examples and the footer sideways along with it, none of which were too wide for anything. Long runs now break inside themselves. Ordinary answers are untouched, since a line still prefers to break between words and only splits a word that could not fit on a line of its own.
The same number one Enter later had the same problem for a different reason. In the saved list the answer sits beside the query and was set never to shrink, which is right when it fits and is why the query is the half that gives way; when it does not fit, a thing that cannot shrink simply hangs out of the row. It may now give way too, which costs nothing on any answer with room, because an item only shrinks when there is not enough space to go round.

1.250.31 August 2026

The sum that was confident about a line it had only partly read

1+2+3+...+100106nothing
1+2+...+1013nothing
3++477
The arithmetic reader gathered up the pieces of a line it recognised and never asked what it had walked past. A character it had no token for was dropped and the rest closed over the gap, so "1+2+3+...+100" lost its ellipsis, arrived at the parser as "1+2+3++100" and answered 106. That is the one answer shape this file is built to never produce. A silence is visibly a silence and somebody types something else; a wrong number is indistinguishable from a right one at the moment it matters, and 106 is a real total of real numbers that were really on the line. Every character now has to have become a token or the line is not read at all. What it does not do is add the run up: summing 1 to 100 is a reader this could grow and would be new grammar, and not having it is not a licence to answer as though the dots had never been typed. A doubled sign is untouched, since the second one is a sign and not a hole, and "3++4" has been 7 throughout.

1.250.21 August 2026

The list that could not be scrolled, and the row that was the word you had just typed

150 kw in hnothing offeredhp · horsepower
150 kw in ↓the key did nothingbrowses the list
100 usd in euran empty box over the answerthe answer
Three faults, all reported, and all in one place: which half of the line an offer replaces. That was being worked out three times over. The list worked it out to draw itself, a pick worked it out again to take one, and the arrow keys worked it out a third time, and only the first of the three asked both of the sources a list can come from. The arrows asked the currency reader alone and gave up when it said no, so every list that came from the general one, which is every list about a unit or a half-typed question, could not be arrowed through at all. Worse than nothing happening: the key had already been claimed by the time it gave up, so it did not fall through to the history either, and the one press did nothing in both directions at once. It is one fact now, settled where the list is built and read from there by all three.
The currency list ranked an exact match first, so the word already in the box led the list that was supposed to be finishing it. A row that replaces the text it is made of changes nothing, and when it was the only match it cost more than a wasted row: the leading offer is drawn up in the line rather than in the list, and the row it came from is skipped there, so a one-row list rendered as an empty box, standing over the answer it had just hidden behind itself. Nothing typed in full is offered back now, compared against the label rather than the code so that "eur" and "euro" go the same way, and a list left with nothing to say closes instead of standing there empty.
And it closed on the one keystroke it was there for. Nothing read a partial word after a connector, so "150 kw in" offered four units and "150 kw in h" offered none, which is the moment somebody had started typing the word they wanted. The narrowing was never the hard part, since an offer that does not start with what was typed has never reached the run, and one that matches it exactly is not an offer. There was simply no reader for that shape.

1.250.11 August 2026

The one operator whose meaning is its working

4!2424, and 4 × 3 × 2 × 1
10!3,628,800…and ten numbers multiplied
0!1an empty product, which is 1 by definition
20!2,432,902,008,180,000,000…with the middle elided
"4!" came back "24" and nothing else, which is the answer and none of the reason. Reported, and it is a fair complaint about this one operator in particular: a factorial is the only mark on this page whose meaning is its working. It says "multiply this by everything under it", and anybody who has to look that up has to look up exactly the line that now gets printed underneath. Nobody needs to be told what the plus did. Elided past a dozen terms, since 20! is a screenful of multiplication and the shape is clear from either end of it. Nought and one are told apart, too: nought is the empty product and one is not, there being nothing to multiply in the first and exactly one thing in the second, and calling them both empty would be wrong about the interesting one.

1.250.01 August 2026

It knew the word you were reaching for and made you type it

150 kw incan't read that yethp · w · mw · ps
32 GB at 100 mbcan't read that yetmbps · mbytesec
1 cup incan't read that yetl · ml · tbsp · tsp
how fcan't read that yethow far is tokyo from london
The grey completion has finished a city name since it went in. This finishes the question, which is the other half of the same problem and much the larger one. Somebody who types "150 kw in" has told this page everything it needs except the one word it could supply itself: there are four units it converts a power into and no way to know from the outside which of them it spells which way. The page sat there saying "can't read that yet", which was true and was not help. One rule is the whole design: every offer is run before it is shown, and only the ones that answer are offered. Nothing in the suggester knows anything about the grammar. It proposes, the readers dispose, and a suggestion therefore cannot lead anywhere but to an answer. That is also what keeps it from rotting, since a reader that stops answering stops being suggested without anybody having to remember this exists. It is why a word typed after a number is offered across every dimension at once rather than a guessed one: "32 GB at 100 mb" could be going to mb, mbit, mbps or mbytesec, all four are run, and the two that answer are the two offered, without the suggester ever being told what a download is. The unit the line is already in is never offered back, compared by the unit each spelling lands on rather than by its letters, so "1 cup in cups" goes the same way "20 km in km" does. And "when will the sun rise in tokyo" answers now, because the suggester tried to offer it, the run said no, and the reader that should have taken it was missing a tense.

1.249.01 August 2026

The longest night, the last day of a week, and the tense a question was asked in

longest night in bostonnothingTue, 22 Dec 2026 · 14h 45m
last day of week 30nothingSun, 26 Jul 2026
what day was it 256 days agonothingMonday
what day will it be in 100 daysnothingSunday
Three reports, and every one of them is a question this page could already answer being asked in words it had no slot for. The longest night is the shortest day, which this file knew and could not be told: being told the day is nine and a quarter hours is little use when the thing being planned around is the fourteen and three quarters of darkness, so the night is now counted and printed as a night, sunset first. A week is a stretch to count days in and was the only stretch that could not be counted in, though the page prints an ISO week number on every date it gives. Which end a week opens on is said out loud, because the two halves of the world disagree: ISO runs Monday to Sunday, which is what the numbering is defined in, and an American calendar runs Sunday to Saturday, so the same numbered week ends on a different day depending on who printed it. And the pronoun: on its own "it" is the whole of "what day is it" and means today, and with a date behind it the word is left over from the sentence, so the date engine was handed "it 256 days ago" and answered nothing. English also puts that pronoun between the two halves of a verb, and there was no slot for it there either.

1.248.01 August 2026

Three answers that knew more than they were saying

0x8fb + 482,3472,347, and 0x92B
15 last primes under 282nothing199 … 281, fifteen of them
first 15 primesnothing2 · 3 · 5 · 7 · 11 …
in 18 days after halloweennothingWed, 18 Nov 2026
Three reports, and all three are the answer knowing more than it said. A sum written in hexadecimal came back in decimal and stopped, which answers a question nobody asked in that form: somebody adding to 0x8fb is working in hex and wants 0x92B. The decimal stays the headline, being the one every reader can check, and the operands go underneath in both, since a wrong-looking total in hex is nearly always one literal being read as something other than what was meant. A list of sixty primes gets topped and tailed with an ellipsis, which is right until somebody says how many they want, and "15 last primes under 282" is a question with its own answer size in it. Said at either end and in either order, and with no bound at all once a count is given, the sieve doubling its reach until it has them. And a leading "in" was stopping an anchored length dead: "in 18 days" reads on its own, so the same words with a date behind them reading as nothing was the offset reader wanting a sign and then a count, and finding that word sitting between the two.

1.247.01 August 2026

An age is reached in a year, and only one year could be named

what year was i born if i was 36 in 1996nothing1959 or 1960
born if i was 40 in 2000nothing1959 or 1960
what year was i born if aged 36 in 1996nothing1959 or 1960
what year was i born if i am 361989 or 19901989 or 1990
Working back from an age only ever counted from today, so the year the age was actually reached had nowhere to go. That year is the whole point of the question: "36 in 1996" is how anybody works backwards from a fact they already have, a photograph, a record, something somebody once said. Asked with the year in it, the line answered nothing at all. Two years still come back, because half a year of birthdays is the difference between them and one number would be wrong half the time without ever looking wrong. The wording follows the year now, though: asked about a year that has gone, both halves are settled, and saying "if the birthday has been this year" about 1996 would be the answer pretending the birthday is still coming.

1.246.01 August 2026

Which Darwin, which Valencia, which Portland

darwin9:30 am in Darwin…and Darwin, Northern Territory
valencia2:00 am in Valencia…and Valencia, Spain
portland5:00 pm in Portland…and Portland, Oregon
monaco2:00 am in Monacounchanged, being Monaco
A clock answer named a city and stopped, and half the cities anybody types are two cities. There is a Darwin in Australia and a Darwin in England, a Valencia in Spain and one in Venezuela, a Portland at each end of the United States. The answer was correct and gave the reader nothing to check it against, which is the one thing this page tries never to do. Where the place is now leads the detail line. The region where a region is how the place is actually said, since an Australian says Darwin, Northern Territory and an American says Boston, Massachusetts, and the country everywhere else, because a Spaniard does not say Valencia, Valencian Community. Nothing at all when the name would only repeat itself: Monaco is not in Monaco, it is Monaco, and "New York, New York" is a song rather than information. None of it can go stale for the reason the coordinates cannot, a city not moving and a state line moving rather less often than that.

1.245.01 August 2026

The unit on the end was asking the question and nothing was listening

istanbul to monaco kmnothing1,793.51 km
istanbul to monaco milesnothing1,114.43 miles
tokyo to london in milesnothing5,943 miles
istanbul to monaco2:00 am in Monaco2:00 am in Monaco
Two places with a length unit after them can only be one question, and it was the one shape of it that answered nothing. Reported. The trailing word was already how "tokyo to london distance" gets read, and a unit does that job and more, since it says which unit to answer in as well as which question is being asked. It is a weak keyword like the other one and safe for the same reason: it answers only when what is left of the line turns out to be two places, so a pair with nothing after it is still a clock question, "3pm new york to tokyo" is untouched, and a unit of the wrong kind puts the line back where it was.

1.244.01 August 2026

Every base converted and a letter did not

65 in asciinothingA
a in asciinothing97
hello in binarynothing01101000 01100101 …
32 in asciinothingspace
"65 in ascii" is the same question as "65 in hex" asked about a different alphabet: a number written in another set of symbols, and this set has twenty-six of them and a shift key. Both directions, because a byte in a hex dump and a letter in a table are one fact seen from either end. Said in code points rather than bytes, since that is what a character is, and the bytes get their own line and name their encoding, a byte count that does not being one nobody can check. The codes that print nothing say what they are instead of answering with blank space and looking broken. And the base rules keep every line they can read: one word or quoted, and nothing that is already a number, so "ff in binary" is still 255 written out and only "ff" in quotes is two letters.

1.243.01 August 2026

Three kinds of radiation, and none of them turns into another

1 sv in usvnothing1,000,000 usv
1 sievert in graynothingradiation dose is not absorbed dose
1 tesla in gaussnothing10,000 gauss
1 newton in lbfnothing0.224809 lbf
Four dimensions the table had no entry for, and the radiation one is the reason to be careful. Radiation is three quantities that get called one thing: how much is happening is becquerels, how much energy landed in the body is grays, and how much harm that does is sieverts. So they are three separate dimensions here and this file will not turn one into the next, which is exactly what a single lumped-together table would have done quietly. Force went in under the pressure that was already here, a pressure being a force over an area. Two clashes had to be settled and both went the same way, to whoever asks more often: "rem" stays the browser's font size and "rad" stays the angle, since a page is what this file measures most of the time, and the radiation reading is worked out under the refusal instead. "mT" stays Mountain Time, because "3pm ct to mt" is asked every day of the week and a millitesla has never once been typed into a calculator.

1.242.01 August 2026

A0 is a square metre, and that fixes all thirty-three of them

a4 in mmnothing210 × 297 millimeters
a4 in inchesnothing8.26772 × 11.6929 inches
c5 envelopenothing162 × 229 millimeters
letter in mmnothing215.9 × 279.4 millimeters
A file that converts every length there is could not be asked for the commonest pair of them, because a paper size is not a length: it is two of them under a name, and the name was the part with no reader. The A series is a rule and not a table, which is the whole reason it belongs here. A0 is one square metre with its sides in the ratio of the square root of two, and those two facts fix both sides exactly, at 841 by 1189 millimetres. Halving the long side gives the next size and keeps the ratio, which is why a folded sheet is the same shape it was and why A4 to A3 is a clean two hundred per cent on any photocopier in Europe. B is a metre across, C sits between the two, and that is what makes a C4 envelope hold an A4 sheet. The rounding has to be done the standard's way rather than the obvious way: the nought is the arithmetic to the nearest millimetre and every size after it is the one above halved and rounded down, one step at a time, off the size that was already rounded. Done in one jump instead, A5 comes out 149. The American sizes are a table, because a letter is eight and a half inches by eleven for no reason but that it is.

1.241.01 August 2026

A slash read and the word it stands for did not

6 GB at 45 mb per secondnothing17 minutes 47 seconds
6 GB at 45 mb per secnothing17 minutes 47 seconds
how long to download 6 gb at 45 mbpsnothing17 minutes 47 seconds
6 GB at 45 MB per secondnothing2 minutes 13 seconds
The rate rule took "mbps", it took "mb/s", it took "/sec" and "/second", and it did not take the word "per" spelled out, which is the only one of those anybody says out loud. So the sentence copied off a speed test answered and the sentence spoken by a person did not. Reported, and it is the same shape as most of what turns up here: the arithmetic was already right and one spelling of the question was missing. The capital is still the byte either way, so 45 MB per second is eight times the file in an eighth of the time, which is the whole reason this section exists. And the question may bring its own words now: "how long" is a date question everywhere else on this page, so naming the verb is what makes it safe to take, no date query having ever downloaded anything.

1.240.01 August 2026

It could do trigonometry and it could not do the area of a circle

area of a circle with radius 5nothing78.5398163397
volume of a sphere radius 5nothing523.598775598
area of a rectangle 10 by 12nothing120
hypotenuse of 3 and 4nothing5
Three multiplications and a constant, and it is the sum this page is asked for more than the trigonometry and the combinatorics that were already here. The arithmetic was never the hard part, which is true of most of what has been missing: there was no shape of sentence to ask it in. Only the shapes people type. A cone, a trapezium and a regular hexagon are formulas somebody looks up once and does not come back for, where a circle, a rectangle and a triangle are the three asked in a kitchen, a garden and a hardware shop. A diameter is what anybody measures and a radius is what every formula wants, so the halving gets done and then said out loud. A dimension may carry a unit, and then the answer carries its square or its cube, because an area in bare numbers is an area somebody has to remember the units of. And a measurement the shape has not got is said rather than left silent: a circle has no volume, and the line was understood, only the pairing was wrong.

1.239.01 August 2026

Nobody runs in kilometres an hour

5 min/km in min/milenothing8:03 min/mile
marathon at 5 min/kmnothing3:30:58
10k at 5 min/kmnothing50:00
pace for a 4 hour marathonnothing5:41 min/km
A pace is a speed written upside down, and it is the way the number is actually said: minutes a kilometre is what the watch shows and what a race gets planned in. This file has turned a distance and a time into a speed since the units went in, and had no way to say the same fact the other way up, so the reader asking got the one form nobody uses. Three questions, and they are the three a runner asks. What this pace is in the other unit, a mile and a kilometre being the two the sport is split between. How long a distance takes at it. And the whole thing backwards: what pace a target time needs, which is the one asked in front of a race and the one nobody does in their head, since it is a division into minutes and seconds. "10k" is ten thousand by the time the reader sees it, the size word and the unit being the same thousand, so reading the bare number as metres is right rather than lucky, and the detail line says which reading it took.

1.238.01 August 2026

Two units it already read, and a division it could not be asked for

bmi 70 kg 180 cmnothing21.6, a healthy weight
bmi for 180 lb and 6 ftnothing24.4
bmi 12 stone 5 ft 10 innothing24.1
what is my bmi if i am 180 cm and 70 kgnothing21.6
A weight over the square of a height, and both of those are units this file has read in every spelling since the day the units went in. What made it unanswerable was the shape: the two numbers arrive in either order and in either system, so the reader takes whatever quantities are in the line and sorts them by what they measure rather than by where they sit. Lengths add up, which is what makes "6 ft 2 in" one height and not two, and weights add up for the same reason, which is what makes stones and pounds work. The band goes underneath, and one clause on what the number is not: a ratio of two measurements, and no more than that. Said once and briefly, because a calculator that lectures is a calculator nobody reads twice.

1.237.01 August 2026

The number to divide a salary by is the thing being argued about

50000 a year in hourlynothing24.04 an hour
25 an hour is how much a yearnothing52,000.00 a year
60000 a year in monthlynothing5,000.00 a month
40 hours at 25 an hournothing1,000.00
This is the sum anybody does while they are looking at a job, and it was answering "there's no month 0" until the date reader let go of round numbers a few releases ago, then nothing at all. It is a division, and not one anybody can do in their head, because the number to divide by is the thing being argued about. So it is said out loud rather than assumed quietly: forty hours a week and fifty-two weeks a year is two thousand and eighty hours, which is what an hourly rate is quoted against, and it counts no holiday, so a year with a fortnight off in it is two thousand hours and a different answer. The detail line names the week it used, because an assumption nobody can see is an assumption nobody can disagree with. Any two periods from an hour to a year, either direction, and the shift multiplication that goes with them.

1.236.01 August 2026

Twenty per cent off and then ten more is not thirty per cent off

20% off 200 then 10% offnothing144, 28% off in all
20% and 10% off 200nothing144
200 less 20% and 10%nothing144
50% off 100 then 50% offnothing25, and not free
A further ten per cent off an already reduced price is twenty-eight per cent off, because the second cut comes off what the first one left, and the gap between that and thirty is the whole reason anybody asks. This file could take one percentage off a number and had no way to take two, so the question with the mistake built into it was the one question it would not answer. The rate they add up to is on the detail line beside the rate they actually come to, and the two prices are shown in the order they happen. Said in three orders, because the discounts get written before the price as often as after it, and two rates only: three is a sentence nobody says, and the second one is where the arithmetic goes wrong.

1.235.01 August 2026

A tip has been read since the first week and a tax was never read at all

20% vat on 100nothing120.00, 20.00 VAT
8.25% sales tax on 45nothing48.71, 3.71 tax
120 including 20% vatnothing100.00, 20.00 VAT in it
remove 20% vat from 120nothing100.00
Both are a percentage added to an amount, and the one that gets typed by everybody who has ever sent an invoice was the one with no reader. The sum was never the difficulty: the answer is two numbers, the total and the part of it that is the tax, and a calculator hands back one. The way back out is the half worth having. Adding a fifth is a sum anybody can do in their head, and finding what a fifth was added to is a division by 1.2 that nearly everybody gets wrong by taking a fifth off instead, so a hundred and twenty inclusive comes out ninety-six rather than a hundred. The echo shows the 1.2, which is what says why in one line. VAT, GST, HST or plain tax, and the answer says back whichever word was typed.

1.234.01 August 2026

A type size was told about beer

12 point in inchespoint isn't a unit I know0.166667 inches
1 point in mmpoint isn't a unit I know0.352778 mm
12 pt in mmpt is volume, mm is lengththe same, with 4.23333 mm under it
1 pica in pointspoints isn't a unit I know12 points
The typographic point has been in this file since the screen units went in, under the name "csspt", which is a name nobody has ever typed. Spelled the way it is written everywhere else it came back "point isn't a unit I know", from a file that knew it perfectly well. A pint is "pt" or "pint" and is never "point", so the plain word costs the volumes nothing and it is what anybody setting text writes. "pt" itself stays the pint, because a unit that is sometimes a drink is a unit that is sometimes wrong, but it no longer stops dead: the refusal is still the pint's and the type size is worked out underneath it, which is the same shape the ounce got when a drink was being measured in it.

1.233.11 August 2026

A day's food came to eight and a third kilojoules

2000 calories in kj8.368 kj8,368 kj
300 calories in kj1.2552 kj1,255.2 kj
1 cal in j4.184 j4.184 j
There are two calories and one word for them. The small one is fixed at 4.184 joules and the one on the back of a packet is a thousand of those, and this read the word as the small one, so two thousand of them came to eight and a third kilojoules: a day's food, out by a factor of a thousand, said with a straight face and no sign that a choice had been made. The capital is what tells the two apart on a label, and a search box is where a capital goes to die. So the word spelled out is the packet's now, whatever its case, and the abbreviation stays the physicist's, because "cal" is who writes it. The echo names which of the two it read, either way.

1.233.01 August 2026

Three weeks ago read and two hours ago did not

2 hours agonothing6:00 pm
in 30 minutesnothing8:30 pm
half an hour agothere's no day 07:30 pm
30m agonothing7:30 pm
an hour agonothing7:00 pm
A length with nothing on the other end of it is how the question gets asked, and the day units have read it both ways round since they went in: "3 weeks ago" and "in a week" are both on the front page. The clock units read neither. The same question answered the moment it was wrapped in something longer, so "what time was it 2 hours ago" worked and "2 hours ago" did not, and the shortest way to ask was the one way that failed. Two of them were worse than silent: "half an hour" becomes "0.5 hour" before any reader sees it, and the date engine took "0.5 hour ago" and answered "there's no day 0", about a month nobody typed. Three smaller things went with it. An article is a number, because "an hour" is how anybody says one. "ago" is not a noun, so "30m ago" stopped being thirty million of something. And the day units are untouched, since the clock reader takes hours, minutes and seconds and no other word: "in 3 days" is a date and stays one.

1.232.21 August 2026

Twelve thousand three hundred and forty five was the first of May, 1234

12345289,356 days ago, 1234-05-01nothing
45001in 903,406 days, 4500-01-01nothing
10000000there's no month 0nothing
17760704Thu, 4 Jul 1776Thu, 4 Jul 1776
The round amounts stopped naming a month that does not exist last release. The ones that are not round were the other half of it, and they were worse than an error: they answered. A five-digit run is a year with a month after it, so 12345 was the 1st of May 1234 and 45001 the 1st of January 4500, stated with the weekday and the week number and every confidence a real date gets. Writing a date with nothing between its parts is how a machine writes one, and those are written in this era, so a packed run now has to name a year in it. Any year at all is still writable with separators, and the eight-digit form keeps the full range, because 17760704 is a date and not a price.

1.232.11 August 2026

Every round number anybody writes down was a month that does not exist

50000there's no month 0nothing
250000there's no month 0nothing
300000 at 6% over 30there's no month 0nothing
100000there's no month 0nothing
A run of five or six digits is read as a year with a month after it, which is how 202607 means July. Every round amount has that shape too, and its last digits are a nought: a salary, a deposit, a house price, a loan. So typing one came back "there's no month 0", an error about a date, for a line with no date anywhere in it, sending the reader to look at the wrong half of what they typed. A bare number says nothing here, and this was the one shape that had been getting out of that rule. Digits that name no month were never a date.