Feige's Conjecture

The answer had been sitting in a 1960 geometry paper the whole time. It took twenty two years, four fields and a machine to notice.

In August 2009 Peter Winkler ran a puzzle in his column in Communications of the ACM.¹ It went roughly like this. You are standing in front of a row of gumball machines. Each machine is honest, in the sense that it pays one gumball per turn on average. You play every machine exactly once. How likely are you to walk away with more gumballs than machines?

Most people say about half. It feels like a coin flip. Sometimes you do better than average, sometimes worse, and with enough machines it should settle near even.

The real answer is that it depends entirely on how the machines are built, that no way of building them gets you past about sixty three per cent, and that sixty three per cent is what one minus one over e looks like in decimal.

That last part was a conjecture when Winkler printed it. It had been posed five years earlier by Uriel Feige, a computer scientist at the Weizmann Institute who was then also at Microsoft Research, and he had not gone looking for a puzzle.² He ran into it while trying to count the edges of a graph too large to look at. He proved a version of it with a much worse constant, wrote down what he thought the right constant was on the same page, and moved on.

Then came two decades of grinding. Three separate teams pushed the constant up from 0.0769 to 0.1798, one step of it worth five ten-thousandths, and the guess stayed a guess.

In July 2026 it fell. Twice, eleven hours apart, by two groups who had never spoken to each other, reaching for the same two ingredients. Both papers credit a language model in their own abstracts. Neither proof contains a new idea about probability. The missing ingredient turned out to be a theorem about how much of a solid shape you can cut off with a flat blade, published by Branko Grünbaum in 1960, forty four years before anybody asked the question it answers.

This piece is about that collision. It is also about what happened afterwards, which is nothing, and about why nothing is the more interesting half.

The machines

Here is the puzzle with the vagueness removed.

You have n machines. Each machine pays out some random number of gumballs, and different machines can behave completely differently. The only rule is that each one averages one gumball per play. You play each machine once, add up what you get, and you win if the total is more than n.

That is all of it. No assumption that the machines are alike, no assumption that they are gentle, no cap on how much any one of them can pay. Just independence, no negative payouts, and an average of one apiece.

You get to design the machines. The question is how good your design can be.

Two things in that figure decide everything that follows.

The first is that the honest design loses every time. If a machine always pays exactly one, it cannot overshoot, and a sum of things that never overshoot never overshoots either. Reliability is fatal here. What you need is unreliability, deliberately manufactured.

The second is the shape of the winning design. It is not "make each machine a bit erratic". It is "make each machine almost always pay nothing, and very occasionally pay everything". With n machines you set each one to pay n plus 1 gumballs with probability one over n plus 1, and nothing otherwise. The average is still exactly one. But now a single machine firing wins the whole game by itself, and you win if any of them fires.

That is where e comes from. The chance that no machine fires is n over n plus 1, multiplied by itself n times. That expression is the standard way to write "no successes in n independent tries, each at odds of one in n plus one", and it is the classic route to e. If e is unfamiliar, it is a number, 2.71828 and onwards, in the same way that pi is a number: it has no tidy decimal form, and it is not a matter of convention but a value that falls out of the arithmetic. It turns up wherever something compounds in many small steps, and here it arrives upside down, because that expression settles not on e but on one over e, about 0.3679. So your chance of winning settles on about 0.6321, and never exceeds it.

Both numbers run through the rest of this piece, so let me pin down which is which. One over e, 0.3679, is the chance of losing, and it is the quantity the conjecture actually bounds. One minus it, 0.6321, is the chance of winning, which is how the puzzle is naturally posed. Whenever one of them appears from here on, the other is one minus it.

Feige had a better name for this than I do. He described the setup as gambling in parallel, and the winning design as the bold play principle from Dubins and Savage: when the target is far away and you only get one round, you do not place n careful bets. You place one enormous one, n times over, and hope exactly one of them lands.³

Both of the obvious answers are wrong

The intuition that says "about half" is wrong. So is the intuition that says "surely you can rig it to win almost always". They are wrong in opposite directions, and the gap between them is the whole problem.

Nothing gets you above sixty three per cent. That is a ceiling on the cheating, and it is not obvious that a ceiling should exist at all. Nothing stops a machine from paying a million gumballs. Nothing stops the machines from being wildly different from each other. There is no bound on how much any payout can vary. And yet the arrangement is stuck.

Turned the other way up, the same fact says this. Whatever you do, the chance that you come home with n gumballs or fewer is at least one over e. Call the total S. Then the probability that S lands strictly below n plus 1 is at least one over e, for every design and every n. That is the statement Feige conjectured, and it is what got proved in July 2026.⁴

Two pieces of precision here, and both of them bite.

The inequality is strict, and strictness is worth a factor of two. The statement is about landing strictly under n plus 1, not about landing at n plus 1 or under. If you relax the sign, the correct constant is not one over e. It is two over e. Under the winning design, the event "S is at most n plus 1" happens exactly twice as often as "S is under n plus 1", for every single n, with no error term. The reason is arithmetic rather than deep: the design pays in lumps of n plus 1, so the only way to land on n plus 1 exactly is for precisely one machine to fire, and the chance of exactly one firing happens to equal the chance of none firing.

That matters because at least one widely used textbook statement of the inequality writes it with the inclusive sign.⁵ The statement is true as written, and weaker than it looks, and anybody who pairs it with the headline that the sharp constant is one over e will be off by a factor of two. The figure above has a switch for this. Push the rigging all the way up and the two loss numbers sit in a ratio of exactly two, at every number of machines.

One over e is a bound that is approached and never reached. At n equal to 2 the true answer is four ninths, about 0.444. At n equal to 10 it is 0.3855. At n equal to 1000 it is 0.3681. The sequence falls forever and lands on 0.36788 in the limit, so no particular number of machines ever achieves it. When people say the constant is sharp, they mean sharp in that asymptotic sense, and the exact answer at each fixed n is the slightly larger number n over n plus 1, raised to the n.

One last thing before we move, and it is not a caveat but the shape of the whole problem. What Feige actually conjectured is not "the constant can be improved to one over e". It is a statement about which design is extremal, meaning which one is the actual worst case, and one over e is a consequence of it. He wrote down two specific arrangements and conjectured that one of them is always the worst case.⁶ They are worth having in mind now, because everything later either tests them or proves them.

In the first, every machine but one is a dead certainty that pays exactly one gumball, and a single machine carries all the risk. In the second, every machine is identical and maximally bold: each one almost always pays nothing and very occasionally pays the entire target by itself. Feige's claim is that the worst design is always one of those two, and which one wins depends on how far above the mean you set the target. The constant falls out of the arrangements. This distinction is why the 2026 papers describe their result as a sharpness theorem rather than a constant-improvement theorem, and it will matter when we get to what is still open.

The question came from a graph nobody could look at

Feige was not doing recreational probability. He was trying to solve a problem in algorithms, and the inequality is a tool he needed and could not find.

Here is the problem. Somebody hands you a graph with n vertices. It is far too big to read. What you are allowed to do is point at a vertex and ask how many edges touch it, and each question costs you. You want to estimate the average degree of the whole graph, and you want to do it after asking far fewer than n questions.

Sample some vertices, average their degrees, report that. It is the obvious algorithm and it is the right one. The question is how badly it can go wrong.

It can go very wrong, and the way it goes wrong has a shape.

That last beat is the reason this problem exists.

The graph that defeats degree sampling is one where nearly every vertex is small and a few vertices are enormous. The design that defeats the gumball puzzle is one where nearly every play pays nothing and a few pay enormously. They are not analogous. They are the same object, seen from two directions. Feige needed a bound on how badly a sum of independent nonnegative things can behave, precisely because the worst graph and the worst sum have the same profile.

The algorithmic payoff is specific. If you insist on asking far fewer questions than there are people, you have to accept a factor of two in the answer. That is a hard barrier rather than a failure of cleverness. Telling a graph of average degree d apart from one of average degree twice d needs roughly n over d queries, because the difference can be hidden in a handful of hubs, and if you will not pay that price then the factor of two is what it costs you.⁷

Feige's theorem is what lets you get arbitrarily close to that factor of two from above. Write epsilon for how much error you are willing to tolerate, so a smaller epsilon means a more accurate answer and a more expensive algorithm. With the inequality, the cost is about one over epsilon, times the square root of n over d, queries. Those two counts answer different questions, and the gap between them is the whole prize: n over d is what beating the factor of two would cost, while the square root of n over d is what approaching it costs. Giving up on beating the factor of two takes the price down to its own square root.

Without the inequality, using only the standard tool, the same algorithm costs a further factor of one over epsilon squared. At one per cent error that factor is ten thousand.⁸

So the conjecture was doing real work from the day it was written. But be precise about what the algorithm actually needs, because this is where the story turns. It needs the inequality to hold with some constant, any constant, and Feige's own one over 13 already bought the entire saving back in 2006. Sharpening that constant toward one over e does not make a single algorithm faster. Sixteen years of people pushing one over 13 to one over 8 to 7 over 50 to 0.1798 were not chasing a better algorithm. They were chasing a number somebody had already told them was the right one.

Where the sharpening does pay is anywhere the constant gets printed in the answer rather than buried in an exponent, and the cleanest example is not in computer science at all. In 2020, Itai Arieli, Yakov Babichenko, Ron Peretz and Peyton Young published a theorem in Econometrica on how fast an innovation spreads through a social network. The bound it ends on carries a prefactor of 7.2. That 7.2 is not a measured quantity. It is one divided by 0.14, and 0.14 was simply the best Feige constant anybody had in 2020. Substituting the now proved sharp value turns it into one divided by one over e, which is e. A published Econometrica bound became a factor of 2.65 tighter, at exactly the slack value that is now settled, on the day those preprints went up.⁹

Nothing in the standard toolbox can touch it

If you have met concentration inequalities, your instinct at this point is that this should be a routine exercise. Sums of independent things concentrate. That is the single most reliable fact in probability. Why is this hard?

Because every tool that makes sums concentrate needs something the problem refuses to give.

The central limit theorem says a sum of many independent things looks like a bell curve, and a bell curve sits below its mean half the time, which would give you one half instead of one over e and a much better answer than anybody conjectured. It does not apply, for a reason that is easy to miss: the machines are allowed to depend on n. Every time you add a machine you are permitted to redesign all of them. The winning design does exactly that, paying n plus 1 with probability one over n plus 1, so the thing you are taking a limit of keeps changing underneath you. There is no fixed sequence to converge.¹⁰

Chebyshev's inequality and the Chernoff bounds need control on the variance or on the range. Variance measures how far a quantity typically strays from its own average, and here nothing controls it. A machine may pay a trillion gumballs, as long as it does so rarely enough.

Markov's inequality applies, and gives you something, and the something is useless. It says the chance of landing under the mean plus delta is at least delta divided by the mean plus delta. But the mean here is n, and it grows. So Markov's answer shrinks to zero as you add machines, and the whole point of the conjecture is that the true answer does not.

The last one is the one worth dwelling on, because it says why the problem is genuinely hard rather than merely unfamiliar. What is wanted is a bound that does not decay as n grows: a dimension-free bound. And there is no concentration available to build it from.

You can watch that happen. The standard deviation in the figure below is the variance put back into gumballs so it can sit beside the mean on the same scale: the typical distance between what you collect and what you expected to collect.

The arithmetic behind that is one line. A machine that pays n plus 1 with probability one over n plus 1 has variance exactly n. Independence adds variances, so n machines give the total a variance of n squared, and a standard deviation of n, which is exactly the mean.¹¹

The same design also kills the next tool along. Berry-Esseen, the refinement that tells you how fast a sum approaches its bell curve, needs a finite third moment. The moments of a payout are the summary numbers you get by averaging its powers: the first is the plain average, the second is what variance is built from and measures the spread, and the higher ones are increasingly sensitive to rare enormous payouts. That last part is the problem. Here the third moment is n plus 1, squared. It grows without bound. Every technique that wants a bounded higher moment has to first mutilate the design to get one, and mutilating it destroys the rare enormous payout that made it extremal in the first place.

That is the trap the next sixteen years walked into.

Sixteen years, one method, and one theorem worth five ten-thousandths

Feige's own proof got 1 over 13, about 0.0769. He was candid about where the number came from. It is an artifact of a case analysis, obtained by transforming the variables into a shape where the cases become manageable and paying for that with tightness, and he said as much in the paper.¹² He also said what he thought the answer was, and said the technique would not reach it.

He was right. Here is what the next sixteen years looked like.

That figure is built from a table the record holders printed themselves, with the technique named against every entry.¹³ Read as a table it is a list of numbers. Read as a picture it is one method saturating.

Every rung from 2010 to 2020 is the same idea with more inputs: set up a moment problem, relax it to something a computer can solve, add a Berry-Esseen estimate, refine. The gains go 0.0481, then 0.0150, then 0.0136, then 0.0005, then 0.0046, then 0.0211. The four authors of the last paper produced four of those seven rungs themselves, in one document.

And the reason it saturates is the reason from the last section. Berry-Esseen needs a finite third moment. The extremal design has one that runs away. So every proof in this family has to truncate the large payouts and rescale, which is the authors' own description of what they do, and truncation removes exactly the structure that makes the design extremal.¹⁴ You can get close to the answer that way. You cannot arrive, because the thing you are analysing is no longer the thing that is worst.

The answer was never going to come from moments. It was a volume.

What a statistician wanted, one year later

In 2005, a year after Feige posed his conjecture, a statistician at Magdeburg named Norbert Gaffke published a paper about a completely different problem, in a journal that computer scientists do not read.¹⁵

His problem was this. You have samples from some unknown nonnegative quantity and you want to test whether its average is at most one. You are not willing to assume anything at all about the shape of the distribution: not normality, not a bounded range, not a variance, not even that all the samples come from the same place. Under those conditions, is there a test that is exactly valid at every sample size, not just valid in the limit?

He was not even first. Two years earlier, Weizhen Wang and Linda H. Zhao had published essentially the same problem statement in the Journal of Statistical Planning and Inference. Crossref records zero citations for it, against sixty eight for a paper sitting a few lines away in the same bibliography. This corner is not obscure because it is unimportant. It is obscure because it sat between two fields, and that is going to turn out to be the whole story.¹⁶

Gaffke wrote down a candidate. Here is the recipe, and the thing to hold on to is which part of it moves. Your data does not move: you have measured n numbers and they are now fixed on the page. What moves is a set of weights. Slip an extra zero in alongside your n numbers, then split one unit of weight at random across the resulting n plus 1 values, in the sense that every possible way of splitting it is equally likely. Those weights give you a weighted average. Gaffke's quantity is the probability, taken over the random split with your data held still, that the weighted average lands at or under one.¹⁷

The padded zero is the trick. There are n plus 1 weights for n numbers because a zero has been slipped in alongside them, and the slack that zero creates is what leaves room for the average to come in under the line.

Gaffke conjectured that this quantity is a genuine p-value: that the chance of it reading at or below any level alpha is itself at most alpha, at every sample size, with no assumptions. He proved it for two samples. He checked it numerically up to fifteen. Then it sat for twenty one years.

There is a name for the general object here, and it is worth having because the whole proof runs through it. A merger is a function that takes the observed numbers and hands back a single number between zero and one. What makes it a merger is a promise about that output. Feed it any nonnegative independent quantities you like, so long as each of their averages is at most one, and then pick any level alpha you please: the chance the output comes in at or below alpha is itself at most alpha. A merger is a valid p-value for the hypothesis that every mean is at most one, and Gaffke's construction was conjectured to be one.

The construction reads much better in a different disguise, and the two are genuinely the same object rather than two analogies for it. Give each of the n plus 1 slots an exponential timer, meaning a random waiting time of the kind that governs how long you wait for a bus that turns up at a steady rate. Divide each timer by the total of all of them and you have a random split of one unit in which every split is equally likely, which is precisely the weighting from the previous paragraph. So the recipe can be restated without mentioning weights at all. Take each observation, subtract one from it, and multiply by that observation's timer. Add those up. Gaffke's quantity is the chance that the sum comes out no larger than the spare timer on its own.

Read it directly. Each observation contributes its excess over one, positive if it overshot and negative if it undershot, weighted by its own random clock. The overshoots race the undershoots, and the spare timer is a handicap the undershoots get for free.¹⁸

Two facts fall out immediately and both get used. The quantity only ever goes down as you increase any observation, which is what a p-value should do. And if every observation is at most one, it equals exactly one, because then there are no overshoots to race.

Gaffke's conjecture was proved on the ninth of July 2026, by Nikos Vlassis at Adobe Research and Philip Thomas at the University of Massachusetts.¹⁹ Their argument has a shape worth knowing even in outline, because it explains why nobody had done it.

They reduce first to the simplest possible distributions: every mean-one distribution on the nonnegative numbers is a blend of distributions supported on just two values, one at or below one and one at or above it. That reduction is old and everyone in this story finds it, including Feige himself.²⁰ Then an outcome becomes just the set of observations that came out high, so the question turns into a purely combinatorial statement about subsets. Then comes the move that gives the paper its name. They build, along any chain of subsets that grows one element at a time, a probability distribution whose tail masses are precisely the values of Gaffke's statistic. Once you have that, the answer is alpha by construction, because the set of outcomes the test rejects is upward closed, and an upward closed set along a chain is a tail.

The hard part is showing that the real, independent behaviour of the observations is dominated by that artificial chain. Which is to say: independence is not the worst case here. Nested, maximally dependent behaviour is.

There is one more thing about this paper that matters later, and it is not mathematical. Philip Thomas had conjectured, in 2019, that a confidence interval he and Erik Learned-Miller had built was valid, and could not prove it.²¹ Then in October 2023 the Journal of the Royal Statistical Society Series B ran a Read Paper on estimating means of bounded quantities, with sixteen invited discussion contributions printed alongside it. Thomas contributed one. It cites exactly one reference: Gaffke 2005. Philip Stark contributed another, and his cites Gaffke too.²²

Three of the groups that published in the eighteen day burst of July 2026 were in that one discussion: the two authors of the Read Paper itself, and two of its discussants. This was not a bolt from nowhere. It was a conversation with a venue, and it had been running for three years.

Neither of the two Feige teams has any connection to it, and that asymmetry is what this piece is finally about. Hold on to it, because the other half of this story had no venue at all.

The move that removes the randomness

Eighteen days after Gaffke fell, Feige fell. Here is the bridge, and it is genuinely short.

Suppose you have a merger, any merger. Suppose you can show that it never reads above alpha anywhere in the region where the observations sum to at least n plus 1. Then you are done: the event that your total lands at or above n plus 1 is contained in the event that the merger reads at or below alpha, and the merger's defining guarantee says that second event has probability at most alpha. So the total lands under n plus 1 with probability at least 1 minus alpha.

That is three lines, and it is the pivot of the entire proof.²³ Read it again and notice what it did. Before it, you had a question about random variables with no assumptions on them, which is what made the problem hard for twenty two years. After it, there is no randomness left anywhere. You have one specific function, one specific region, and you need its largest value on that region. It is a calculus problem.

All the difficulty has been moved into the phrase "suppose you have a merger". That is Gaffke's conjecture, and it is somebody else's twenty one years of work.

So: how large does Gaffke's statistic get on the region where the observations sum to at least n plus 1?

Start with the easy part of the region, where every observation is above one. There the statistic is exactly one over the product of the observations, so making it large means making the product small. Each observation is at least one, and their excesses over one add up to at least one, and a product of terms each at least one is at least one plus the sum of their excesses. So the product is at least two, the statistic is at most one half, and you get: the total lands under n plus 1 at least half the time.²⁴

One line, no geometry, and the number it produces is exactly the value of the first of the two arrangements Feige wrote down in 2004: the one where a single machine carries all the risk. But notice what kind of answer this is. One half is comfortably more than one over e, so on this part of the region the conjecture is not merely true, it is true with room to spare. That is exactly why this part is not where the difficulty lives. A bound this slack cannot be the thing that pins the final answer down to one over e, so whatever forces the constant that low has to be happening somewhere else.

The difficulty lives outside it. The region where the total is large includes points where some observations are zero, and the closed form does not hold there, and that is where the true worst case sits.

The balanced point is nowhere near the extreme. That is a small thing but it is a real trap, and the figure exists because I would otherwise have drawn it wrong.

Which leaves the actual question. What is that function, out there in the part of the region where the closed form fails?

It was a volume the whole time

Go back to the definition. Gaffke's statistic averages the observations using weights drawn uniformly at random from every possible weighting, and asks how often the result comes in under one.

"Every possible weighting of n plus 1 things, chosen uniformly" has a name in geometry. It is the uniform distribution on a simplex: the triangle for three weights, the tetrahedron for four, and the n-dimensional generalisation after that. Note that three weights give a two-dimensional triangle, so n plus 1 weights give an n-dimensional shape, and that bookkeeping matters shortly. And "the weighted average comes in under one" is a linear condition on those weights, meaning the weights enter it only by being scaled and added, never multiplied together. A condition of that kind always has a perfectly flat boundary, so it carves the simplex with a single straight cut.

So Gaffke's statistic is the fraction of a simplex lying on one side of a flat cut. Not like a volume. A volume.

That is the reveal, and it is not an analogy that somebody constructed. It is what the definition says, once you notice that a flat distribution over weightings is a solid shape.

There is one more step to make the connection exact, and it is worth doing because it explains the one parameter I have been hiding.

Write out the condition. Landing under one is a linear condition on the weights, which is the technical way of saying that the winning region is everything on one side of a single flat cut through the simplex. So the only question left is where that cut sits. Measured from the corner, the answer is unenlightening. Measured from the centre of mass, and in the case that decides the bound, where the total lands exactly on the target, it collapses to one number: the cut sits off centre by 1 minus delta, divided by n plus 1, where delta is how much slack you allowed in the target. Feige's conjecture, as everyone states it, is the case delta equals one.

At delta equals one, that offset is zero.

Which is to say: the case everybody cares about is exactly the case where the cutting plane passes through the centre of mass of the simplex. Not near it. Through it. And delta, the slack parameter that looked like a nuisance in the statement of the conjecture, turns out to be nothing but the distance the blade sits off centre.

Grünbaum, 1960

In 1960, Branko Grünbaum published a five page paper in the Pacific Journal of Mathematics on cutting solid shapes with flat planes.²⁵ He was not thinking about probability. Nobody was going to ask this question for another forty four years.

The theorem says: take any convex solid in n dimensions, meaning any solid with no dents in it, and cut it with a flat plane through its centre of mass, the point it would balance on. However you angle the blade, the smaller piece is at least n over n plus 1, raised to the power n, of the whole. And that is the best possible bound, achieved by a cone.

Same expression. Same letters. And the two n's are the same n, which is the part that stops it from being a coincidence. Grünbaum's n is the dimension of the solid, Feige's n is the number of machines, and the padded zero is what welds them together: n machines give n observations, the padded zero makes n plus 1 weights, and n plus 1 weights span a solid of dimension exactly n. That is why the bookkeeping was worth flagging earlier. The bound on how much a blade through the middle can miss is the bound on how often you fail to beat n gumballs.

Grünbaum opens his own paper by recalling what was already known in the plane, which he credits to Neumann, Eggleston and Newman: for any convex region there is a point such that every half-plane through it captures at least four ninths of the area.²⁶ Four ninths is two thirds squared. It is the sharp answer to Feige's conjecture for two machines, and it was in print before Grünbaum generalised it, and decades before there was a conjecture to be the answer to.

The 2026 proofs also need the case where the blade is off centre, because they want the general slack parameter and not just delta equal to one. That case was not available in 1960. It arrived in 2024, when Brayden Letwin and Vladyslav Yaskin published a version of Grünbaum's inequality for planes that miss the centre of mass by a controlled amount.²⁷ Their statement carries two correction factors that account for the offset.

At delta equal to one, both correction factors are exactly one, and their theorem collapses to Grünbaum's.

So here is the whole proof of Feige's conjecture, as it was actually assembled in July 2026, with nothing left out. Feige, from 2004, equals Grünbaum, from 1960, plus Vlassis and Thomas, from eighteen days earlier.

One piece of convex geometry from before the question existed. One piece of statistics from a different field. No new probability at all.

Two papers, eleven hours apart

On the twenty seventh of July 2026, at 04:06 in the morning UTC, six authors posted a proof of Feige's conjecture to arXiv. Eleven hours and one minute later, two more authors posted another one.²⁸

They had not spoken. They used the same two ingredients. Both of them credit a language model in the abstract.

The first paper is by Weibo Fu, Yanjun Han, Guanyang Wang, Jun Yan, Peng Zhang and Zhengqing Zhou, at Princeton, NYU, Rutgers and Stanford. The second is by Zipei Nie at Illinois and Jiaye Wei at EPFL. The convergence is not a tie: the six author paper covers a strictly wider range of the slack parameter, so it contains the two author result rather than complementing it. Eleven hours is noise, neither team could have seen the other, and there is no priority claim worth making here.

What there is, is evidence, though it needs stating carefully. Two groups of strangers, working separately, asked the same question in the same week and came back with the same two ingredients. That rules out a fluke: the step from Gaffke to Feige was not one team getting lucky with one prompt. But it is not two independent confirmations in the ordinary sense, because the same model sat on both sides of it, and one instrument producing the same output twice is a fact about the instrument at least as much as about the mathematics. What the convergence does establish is that once Gaffke fell, the route was there to be found, and that finding it took eighteen days rather than another two decades.

Now put the two links of the chain side by side, because the comparison is the sharpest thing in this whole story.

Link one: prove that Gaffke's statistic is a valid p-value. Posed 2005. Proved the ninth of July 2026. Twenty one years. It carries a single unsigned footnote about AI assistance that names no model and no author.

Link two: notice that Gaffke plus Grünbaum gives Feige. Available the ninth of July 2026. Done by the twenty seventh. Seventeen days and fifteen hours, independently, by two groups who had never spoken. Both papers put the machine in the abstract.

The loud credits are on the fast half. That is not an accusation of anything, and both teams say plainly that they verified the arguments themselves. It is an observation about which piece of the work was hard, and the answer is that the piece that took twenty one years is the piece nobody is talking about.

The ladder had a shadow, and nobody was watching it

Go back to the ladder of constants: 1 over 13, then 1 over 8, then 7 over 50, then 0.1798. Four numbers, sixteen years, one field.

Now subtract each of them from one. You get 12 over 13, then 7 over 8, then 43 over 50, then 0.8202.

Those four numbers are also theorems in extremal combinatorics, about a completely different question. The same theorems, restated, rather than analogies of them.

The question is the Erdős matching conjecture, and you can state it without any machinery. Take a group of n people and form committees of k people each. You want to form as many distinct committees as you can, subject to one rule: you must never be able to seat s committees at the same time with nobody sitting on two of them. Erdős conjectured in 1965 that the largest such collection is always one of two obvious ones. It is still open.