Portal · a combine across eight layers · ground-truth seam
The Date That Can Only Say No
A dated attestation is a one-sided instrument. It can prove a claim false, and it can never prove one true. Eight layers of this archive walk that single inequality: six collisions, a limit case, and one foil. Sort the collisions by the gap and the gap turns out not to matter: two years kills a claim exactly as dead as four hundred and ninety.
There is a move that historians, chemists, lexicographers and astronomers all make, and each of them thinks of it as belonging to their own field. A conservator finds cobalt blue in a painting that claims to be from 1665, and the painting is finished. A lexicographer finds the word boredom in a newspaper from 1829, and Dickens is no longer its inventor. In both cases the entire argument is one comparison of two integers.
The Latin name for it is terminus post quem, the limit after which. It is one of the oldest tools in scholarship and one of the least examined, because its power is so obvious that nobody stops to notice its shape. The shape is this: it works in exactly one direction. An attested date can tell you that a claim is false. Nothing you do to it will ever make it tell you that a claim is true.
This is a portal. It builds no new fact: every date below is already published and machine-checked inside one of the eight layers it walks, and the bench re-derives each of them from that layer's own numbers. What the portal adds is the sorting, and what the sorting shows.
The bench · two lamps and an empty socket
Pick a case. The gold line is the attestation: the year by which the thing is proved to have existed, fixed by evidence rather than by the argument it is being used in. It can still be wrong, and one of these six says so about its own. The white line is the claim, and you can drag it anywhere you like. Watch what the bench is able to say.
⌖ the one-sided bench
Drag the claim across the attestation. There are three sockets. Only two of them have wire behind them.
The third socket is not a joke about scepticism. It is a fact about the instrument. An attestation is a report that something was somewhere at some time. From that you can derive that a claim requiring its absence is false. You cannot derive that any claim is true, because the evidence you have is compatible with the thing having existed for any length of time before the moment it happened to be written down. The floor you have found is a floor under your searching, not under the world.
Six collisions, sorted by the gap
Below are the six layers where a claim collides with an attested date, ordered by the signed gap: the attested year minus the claimed year. Negative means the thing was already there before its supposed beginning. Positive means the thing was too new to be there at all.
Read the column of gaps and the point arrives on its own. Nothing in the verdict depends on the size of the number. Lemaître beats Hubble to the law by two years, and that is as final as the four hundred and ninety years between the Strasbourg dancers and the first person to put a number on their dead. A one-sided bound does not return a quantity. It returns a verdict, and the verdict has no volume knob.
This is the exact inverse of the move another portal in this archive makes. The Number That Ends the Argument collects results where you should stop arguing about mechanism and size it, because the magnitude settles what the reasoning could not. Here the magnitude settles nothing, and the reason is structural: an inequality between two dates carries only its sign.
The limit case · a claim with no year in it
Every case above has a year on both sides. The seventh does not, and it is the one that shows what the instrument is really made of.
What the First Sleep Was First Of
A claim with no date in it, refuted by a single dated thing.
Janine Rivière, quoted in the live dispute over whether pre-modern Europeans slept in two shifts, writes that segmented sleep “is not discussed in manuals of health or discussions of the proper regimens of sleep.” There is no year in that sentence. It is a claim about the whole genre, forever, and it cannot be dated because it asserts an absence rather than a beginning.
An absence claim has no year, so it has no gap. What it has instead is a target the size of the entire surviving record, and exactly one way to fall: find one. That layer swept the Text Creation Partnership corpus, 61,287 hand-keyed books and 1,547,130,448 words printed between 1473 and 1810, and found an English health regimen printed about 1490 that keys sleeping posture to a completed first sleep:
“slepe fyrst on thy syght side for that is kyndely for thy dygestio… And after thi fyrst slepe turne on thy lifte syde… And whan thou hast layen theron a good while and slept turne ayen on thi ryght side and ther slepe all nyght forth” Gouernayle of Helthe, London, ca. 1490 · TCP A01993, as transcribed
One book out of sixty-one thousand ends the sentence. And that layer is careful about what it bought, in a passage worth quoting because it is the honest version of every move on this page:
“That is a real counterexample to the sentence quoted above, found by a search rather than by memory. But it prescribes turning over, not getting up, and it says nothing about an hour of wakefulness.” What the First Sleep Was First Of, on the 1490 regimen
Now turn it around, which is what makes this the limit case. Suppose the sweep had come back empty. It would have proved nothing whatever, and that layer says so in four separate places rather than one: the corpus is “a transcribed subset of what survives”; no list of spellings catches a paraphrase, so “every second-sleep count on this page is a floor”; an argument from silence “cannot distinguish a thing nobody did from a thing nobody bothered to mention”; and the transcribed books effectively stop in 1800, which puts the most interesting half of the claim outside the window entirely.
That last item is the whole portal in one sentence. The search has a terminus of its own. Sixty-one thousand books is an enormous denominator and it is still a denominator, and the layer prints it rather than leaving the reader to imagine one. Almost nobody does. A dictionary's first citation is the earliest quotation its editors happened to be sent, with no denominator at all, which is exactly why the boredom floor has already been lowered twice on the record: 1852 to 1844 to 1829, and its own page refuses to call even 1829 the bottom.
Searching harder can lower a floor. It can never raise one. That is the same one-sidedness as the bench, wearing different clothes: more evidence can only ever move the verdict toward no.
The foil · what the third lamp costs
The one member whose lamp turns green
Every Tree Wrote the Same Bad Year
Dendrochronological crossdating does the thing the whole rest of this page says cannot be done: it hands back a calendar year. Not a floor, not a range. In that layer's own words, of the beam pulled from a wall in Show Low, Arizona: “a stretch of prehistory acquired exact years. Not a range. Years.”
It can do this because it is not arguing from earliest existence at all. It is arguing from a match: a run of wet and dry years is printed simultaneously into every tree on a plateau, so a piece of wood carries a barcode that can be aligned against a chronology built from trees it never met. That layer proves the transfer rather than assuming it, and its numbers are on its own page: all 48 Arizona cores land on their published year against a Colorado chronology 187 km away that shares not one tree, the weakest genuine match still reaching t = 7.21; against an Alaskan control 3,718 km away, zero of 48 land correctly.
So there is a third lamp, and here is its price. State it carefully, because the tempting version is false. A one-sided date can certainly be mistaken: an attestation can be misdated, a pigment misidentified, a document forged, a transcription misread, and then the red lamp lights on nothing. The boredom layer carries exactly that exposure and prints it, flagging its own 1829 antedating as reported by a dictionary publisher and not inspected at the original. What separates the two instruments is not immunity to error. It is where the error is allowed to live.
A collision needs only that the two dates be right and the thing be what it is said to be. A match can be wrong when every measurement is right. That second sentence is not a worry, it is a measured quantity, and the tree-ring layer measured it: 23 of those 48 certainly-wrong Alaskan matches clear the discipline's conventional threshold of t ≥ 3.5. Roughly half the cores produce a “significant” match at a date that cannot be right, from ring widths that are all perfectly correct. That page names the cause precisely: it is “what happens when you take the best of several hundred tries and then apply a cutoff designed for one try.” The separation between true and false is real (best false 5.17 against weakest genuine 7.21, with clear air between) but it is a separation somebody had to go out and measure, and the conventional cutoff sits inside the noise. A refutation by collision has no such residue: once the two dates hold, there is nothing left over for chance to do.
And then the foil folds back into the spine. Ask what a tree-ring date dates, and that page answers with a ladder. Bark still on the sample gives the felling year. Sapwood without bark gives a felling range about a decade wide for oak. Heartwood alone gives, verbatim, “A floor and nothing else: the tree cannot have been felled before this. There is no upper bound at all.” The exception contains the rule. And the sentence that layer writes next is the sharpest statement of this portal's whole subject that the archive contains:
“The bottom row is the one that gets flattened in retelling, because ‘dated to 1486’ reads the same on the page whether it means a felled year or a floor.” Every Tree Wrote the Same Bad Year, §10, “What a date actually dates”
That is why this move is worth naming. A one-sided date and a two-sided date are written down identically. Four digits, no marking, no apparatus. The asymmetry is invisible in the number and survives only in the sentence around it, which is the first thing retelling throws away.
The join
Put the eight together and each says something none of them says alone.
One. The same inequality is doing the work in a paint cross-section, a newspaper column, a manuscript, a relativity paper, a Louvre press clipping and a city chancery, and then again in the limit case, over a corpus of hand-keyed books. Six fields on the axis, one instrument, and each field thinks of it as its own.
Two. The instrument is one-sided in both of its directions. Whether the attestation lands above the claim or below it, the verdict is refutation or nothing. The two families look like opposites and they are the same lamp.
Three. The gap carries no weight. Three years and four hundred and ninety years do identical work, which is the tell that you are not looking at a measurement at all.
Four, and this is the one the foil pays for. Neither kind of instrument is safe from bad evidence, but only one of them can be defeated by good evidence. A collision fails only if a date is wrong; a match can fail with every measurement correct, because a finite series can align by luck. So every field that wants a positive date has to buy it, and the currency is a second independent line of evidence plus a measured error rate. Crossdating bought it, and its own page shows both the receipt and the bill.
The practical residue, for a reader with no interest in any of these eight subjects: when someone hands you a date as proof, check which way it points. A date that refutes has done all the work by itself. A date that appears to confirm is either resting on a second instrument nobody mentioned, or it is not confirming anything.
Apparatus · what is checked, what is asserted, what is not
The portal law, and where this page sits inside it
- No new fact. Every year, count and quotation above is already published inside one of the eight member layers and covered by that layer's own committed verifier, except where this section says otherwise. The portal's own checker,
research/the-date-that-can-only-say-no/verify.mjs, re-derives every gap from the member figures, re-reads them out of this page's shipped HTML, and then re-runs all eight members' verifiers to confirm the borrowed numbers still stand. - The join is a reframing, not a theorem. That these eight are “one instrument” is a family resemblance argued on this page, not a result proved anywhere. The four numbered claims under The join are asserted by this page and are not machine-checkable; the dates, gaps, directions and the exhaustive sweep are.
- Not every member states the asymmetry. Four of the six collision members do, in their own prose, and the table above quotes them. Two do not. The Redshift Before the Law lays out the chronology and splits the credit three ways without ever stating a rule about what dates can and cannot do. A Minor Leonardo also does not: the sentence quoted in its row (“counts are a noisy lower bound”) is the second of that layer's four honest limits, a caution about newspaper OCR, and it is one-sidedness about an instrument rather than about what a date can do. It is quoted because it is the right caution to carry, not as evidence that the layer states this portal's thesis. For both of those layers the one-sided reading is this portal's, applied from outside. Saying so is the point of saying it.
- Nor does every member machine-check its own dates, and the column says so in three values. For each of the six collision members it reports whether the date fact carrying its verdict is an executable assertion in that member's own verifier. Four say yes, and the portal's checker opens those four verifier files and confirms the assertion string is really in them rather than trusting a flag typed on this page. One says partly: the dancing plague pins both anchor years and asserts the toll is the last tier of source to appear, which is what its verdict rests on, but it never subtracts one year from the other. One says no: The Redshift Before the Law asserts quantities (velocities, rate constants, 24 of 46 nebulae with individual distances) and never compares two years, so its priority chronology is carried by its timeline and its primary sources. That is a real gap in a layer this portal leans on, and it is named here rather than smoothed over. An earlier draft of this bullet said “five of seven” against a table with six rows, and the check behind it compared two numbers typed side by side in the same object, which is not a check.
- The sweep is exhaustive, not a sample. The bench's sweep runs every integer year from 1400 to 2026 inclusive against all six collision cases and reports the tally. The portal verifier runs the identical sweep offline and asserts the confirmed count is zero. This is a property of the rule, not a discovery: it is stated as a check that the page's own instrument cannot quietly acquire a third verdict.
- Only one of the three “too new to be there” cases is a physical impossibility. Cobalt blue could not exist before it was synthesised, so that painting's verdict is an impossibility in the strict sense. The other two are earliest-attestation floors wearing the same label, and they inherit the bench's shared verdict wording. This matters most for the Mona Lisa: the coverage instrument its layer built runs from 1905 to 1920, so 1911 is a floor under a fifteen-year search window and cannot see 1550 to 1904 at all. What actually carries the pre-1911 half of that claim is positive evidence rather than corpus silence, the Louvre guidebooks in which two rivals outranked it, and that layer is careful to say its counts are a lower bound and that “not famous” would be a lie. Reading a floor under a short search as an impossibility is precisely the error this page is about, so it is named here. The same, more mildly, for the dancing plague: 2008 is the earliest attestation of the toll as text, not proof that nobody died, which is the distinction its own layer insists on.
- The claim years are reconstructions of a popular claim, not quotations of one. “Marie Antoinette could first be quoted in France from 1770” and “the Mona Lisa held supremacy from 1550” are the earliest dates at which the folk story could be true, chosen as the reading kindest to the claim, in the same spirit as the forgery layer's rule that it uses the earliest defensible availability year because “if even the earliest date is after the claim, the anachronism is certain.” Each is anchored to a year its member layer publishes and checks.
- Why the redshift row uses 1927 and not 1913. Slipher published the first galaxy radial velocity in 1913, sixteen years before Hubble, and that is the tempting number. It is the wrong one. That layer's own heading is “One measurement is not the discovery”, its headline object is blueshifted, and it says Slipher read his shifts as ordinary Doppler motion rather than as space stretching. A velocity catalogue attests a technique, not an expanding universe. The datum that actually collides with “discovered in 1929” is Lemaître's 1927 derivation of the velocity-distance law, which that layer prints as “two years before Hubble”, so 1927 is what the row uses. It is also the choice that argues against this page's own interest, in the same way the brioche layer takes the latest end of a disputed range so its refutation survives the worst case.
- Two members carry a softer edge than the table shows. The boredom layer flags its own 1829 antedating as soft, reported by a dictionary publisher and not inspected at the original; its solid floor is Poe's 1844, checked verbatim against the full text at the Edgar Allan Poe Society, which still refutes the Dickens story by eight years. That matters to the sort as well as to the verdict, because 1829 is the row that sets the left end of the axis: at the solid 1844 its gap is not -23 but -8, and the spine reorders. The row is marked, the bench uses the figure the layer machine-checks, and the portal's own verifier asserts that the verdict is REFUTED under both years, so nothing on this page rests on the soft one. The Mona Lisa layer measures newspaper coverage, and states that OCR counts are a noisy lower bound and that “not famous” would be a lie: its claim is a shift in degree, which is why its own chart never touches the floor. Both cautions belong to the members and are repeated rather than removed here.
- Divergence, declared. The Number That Ends the Argument sorts by magnitude and finds that size settles the question; this sorts by a gap and finds the size irrelevant, which is the inversion named in the spine section. The Auditor That Wants Nothing ranks checks by how little the checker shares with the mind that wants the claim true; this ranks nothing by independence and asks instead which verdicts an instrument is capable of returning. Repeated Until True and The Anatomy of Error catalogue record-corrections by the discriminator that kills them and by why anyone believed them; a date is one of the five discriminators those portals count, and this page takes that single discriminator apart instead of counting cases. What the Ground Can't Say collects questions where the check comes back empty; here the check comes back full and still cannot say yes, which is a different kind of silence.
- Member overlap, corrected. An earlier draft of this page claimed no member had been walked by any other portal. That was false, and it was false because the check behind it looked in the wrong place: the archive's portals build their links at runtime from a slug list, so the string
/strata/boredom-word/never appears in their source and a search for it comes back clean. Five of the eight have appeared before, as single cards: boredom, brioche, the redshift and the Mona Lisa in Repeated Until True, and the dancing plague in What the Ground Can't Say. Those portals sort by which discriminator kills a claim, and a date is one of the five they count. This one takes that single discriminator apart, so no member is walked on this axis before. The verifier now searches the slug lists too, and the honest statement is the one in this bullet rather than the one it replaced.
Run it yourself
node research/the-date-that-can-only-say-no/verify.mjsre-derives every gap, re-reads this page, runs the exhaustive sweep, and then executes all eight members' verifiers, nine scripts in all, end to end. It needs no network and no dependencies.