Nothing in the last four days moved because the world ran out of something. The fertilizer bill for next year's harvest is being written by a Chinese export licence and a Russian export ban while the energy input that was supposed to drive it posts its steepest annual fall on record; a lending protocol lost $8.5 million to a vote every line of its code obeyed; an airline can price each seat by algorithm because airlines are exempt from the pricing rule that covers everyone else; and a British power station was worth attacking precisely because it sat under the legal threshold for reporting the attack. In all four the price was set by whoever controls the permission rather than by whoever controls the material, and permission is the cheaper of the two things to corner. Watch China's phosphate export restrictions, which run only through the end of this month: whether Beijing renews them will do more to set next year's food bill than anything the oil market does.
Washington and Ottawa stopped negotiating. Talks collapsed late Friday, the US put 50 percent tariffs on some Canadian goods into effect on Saturday, and Mark Carney said Canada will match them dollar for dollar from 8 September, targeting dairy, appliances, agricultural equipment, pulp and paper and electronics. His account of the American offer was that "they asked too much and offered too little." The retaliation is dated rather than immediate, which leaves two weeks in which the tariffs are real and the answer to them is still a threat.
Scott Bessent unveils the Iran sanctions package at 2 PM Eastern today, running alongside the naval blockade already in place in the Gulf of Oman. He wrote in the Financial Times on Sunday that "at dawn begins an economic D-Day, the single greatest financial offensive ever marshalled against an adversary," and described it to CNBC as "a one-two punch." Tehran dismissed the threat. The Dashboard below has what crude did with the news, and Geopolitics has the other half of this story, the half conducted through a power station.
Asia: Tokyo flat at 66,075, Shanghai off 0.4 percent, Seoul lower on profit taking, no halts or limit events anywhere in the session. Europe: London up 0.6 percent, Frankfurt flat, Paris up 0.4 percent.
Crypto data provided by CoinGecko
The one input that would carry an energy shock into food prices just posted the steepest annual decline ever recorded for it, so the food-inflation story people are telling from the oil price has no transmission channel left. Urea, the workhorse nitrogen fertilizer, ended the week of 17 August at $390.00 a tonne, down 11.51 percent on the year, which the tracker publishing the series calls its steepest annual fall on record. Natural gas is 70 to 80 percent of the cost of making nitrogen, so if energy were pushing up the price of dinner, urea is where it would show first. What is rising is not energy. Phosphate at $795.00 a tonne is up 3.58 percent on the month on Chinese export curbs, and sulfur is up 261.63 percent on the year on a Russian export ban locked in through the end of 2026. We expect next year's fertilizer cost push to come from phosphate and sulfur rather than nitrogen, because both are set by export policy with a stated end date and neither is an energy price. An input stops being a commodity when the only two variables that move it are a Chinese licence and a Russian ban.
The inflation number the Federal Reserve actually targets is 3.3 percent, the one almost everybody quotes is 2.5 percent, and Wednesday's release is the first one. Core consumer price inflation ran 2.5 percent in the year to July. Core personal consumption expenditures inflation, the Commerce Department series the Fed's two percent target is written against, ran 3.3 percent in the year to June. Eighty basis points separate two measures of the same thing, and that gap is the difference between an economy where inflation is handled and one where it sits 130 basis points above target with the funds rate already at 3.50 to 3.75 percent. It also explains why almost every public argument about what the Fed should do next is conducted in the index the Fed does not target, which is how one set of facts supports a cut and a hike at the same time, depending on nothing but which series the speaker happened to read. Wednesday's release, on 26 August, is the July reading of the series the target is actually written against, and it is the first reading of that series since June. Before you take a view on it, say which index you mean, because the two do not describe the same country.
A former chief executive of Lazard spent the weekend publicly arguing that France should cancel the fifth of its national debt the European Central Bank happens to hold, and a former ECB economist answered that doing it would end the euro. Matthieu Pigasse posted on Friday evening that the ECB holds roughly 20 percent of French debt, that this debt can be cancelled at no financial cost, and that "there is no economic constraint, there are only political choices." Luis Garicano replied on Sunday that the demand now runs through the French far left, the far right and more vaguely the centre, and that "it will be, if it happens, the end of the euro." France runs a deficit near 5 percent of GDP, the Banque de France cannot intervene in its own sovereign market and the ECB can, and the ECB's instrument for exactly this situation is conditioned on fiscal compliance France does not meet. So the one institution able to act is the one being asked to write the debt off. We expect the argument to gain respectable French sponsors through the autumn budget, because the tool designed to make it unnecessary is unavailable on the terms that would make it needed.
Term Finance lost about $8.5 million on Sunday and nothing broke. The attacker started with two ether laundered through Tornado Cash, quietly bought up the protocol's thinly held governance token until it controlled every vote in four of five USDC strategy vaults and roughly 91 percent of the Ethereum meta vault, then voted the contents out: 2,843 ether, reported at about $6.87 million, and 1.68 million dollars of USDC swapped straight into DAI. PeckShield and CertiK confirmed it. The vaults complied because complying is what the code says to do when a vote passes. Set the two numbers beside each other, because the ratio is the item: two ether in, worth under five thousand dollars at Sunday's price, against $8.5 million out, roughly seventeen hundred to one. A protocol's security is not the strength of its code. It is the cheaper of two numbers, the cost of breaking the contract and the cost of buying a quorum, and only the first one ever gets audited. Beanstalk Farms lost about $182 million the same way in April 2022, relaunched, and never got its scale back.
Compound's token holders voted $52 million into an institutional lending push on 10 May, 1.88 million COMP in favour and not one against, and it took another ninety-nine days to name the people who would spend it. On 17 August the Compound Foundation announced Aaron Schnarch, formerly chief executive of Coinbase Custody, as executive director. Several wires ran that announcement as though the money were the news. The money was May's. Only $14 million of the $52 million was ever released; the other $38 million sits behind a five-of-seven multisignature wallet and unlocks against milestones, a staffed engineering team, a production integration kit and a public testnet, every one of which requires the executives hired last week. So the holders wrote a release schedule gated on a hiring process they never voted on. Compound holds about $1.49 billion in total value against Aave's $17.6 billion, an order of magnitude behind. Term's vote moved $8.5 million in an afternoon; Compound's has moved nothing for three months. Governance is fast at saying yes and slow at everything else, and only one of those speeds gets exploited.
Providence Equity bought control of ATG Entertainment for about $500 million in 2013 and is selling it to Ari Emanuel's Mari for $6 billion, roughly twelve times on a control position over thirteen years, in an asset class where nobody can build more supply. ATG runs about 70 venues across Britain, the US, Germany and Spain, seven of them on Broadway and ten in the West End, and there will not be an eighth Broadway house. The price is the least interesting number here. What matters is what Mari is assembling: it bought TodayTix, the ticketing platform, in October 2025, and it is now buying the venues those tickets admit you to. Own the ticket and the door and you are no longer a promoter. You set the fee, you see the buyer, and the artist negotiates with the only counterparty that controls both. That is the Live Nation and Ticketmaster structure, blessed by the Justice Department in a 2010 consent decree and sued by that same department in May 2024 to break it apart. Mari is building the identical shape from the other end while the precedent sits in front of a judge.
Nvidia has told its largest customers that servers built around its chips cost more than 15 percent more from early next year, which is a 75 percent gross margin company declining to absorb an increase it could easily have swallowed. Bloomberg reported the notifications on Saturday. A Blackwell NVL72 rack has priced around $2.8 million to $3.4 million, so 15 percent is several hundred thousand dollars a rack. The cause is memory: contract DRAM prices rose 90 to 95 percent quarter on quarter in the first quarter. The power moved to Samsung, SK Hynix and Micron, and custom silicon is no escape, because Trainium, TPU and MTIA buy from the same three. The closest historical match is 1988, and what is different about it is the point. Then, 256K DRAM went from $2.95 at the start of 1988 to $12.45, partly because fabs had converted lines to one-megabit chips they could not yet yield and partly because the 1986 US and Japan semiconductor agreement had raised Japanese chip prices by design, and the price broke the following year once the yields arrived. This time the capacity went to high-bandwidth memory under contracts SK Hynix reported sold out for all of 2026 back in October 2025, so the correction has a much longer fuse. We expect Nvidia to hold its margin and the hyperscalers to absorb the increase. Wednesday's third-quarter guidance is the test: at or above roughly 74.5 percent it held, below that the incidence landed on Nvidia.
Delta intends to expand the share of its domestic fares set by an artificial intelligence pricing system from about 3 percent today to 20 percent, and the reason it can is not the model. Delta's president Glen Hauenstein described the expansion to shareholders; the airline works with Fetcherr, an Israeli pricing firm. Delta has specifically denied the framing that went viral alongside it, stating that "there is no fare product Delta has ever used, is testing or plans to use that targets customers with individualized prices based on personal data." Take the denial at face value and the story is unchanged, because the enabling condition sits a level above the algorithm: airlines are exempt from Federal Trade Commission oversight of pricing practices. That exemption is why the experiment runs here rather than in ridesharing, where Consumer Reports found Uber and Lyft quoting fares 42 percent apart for the same trip at the same moment. The model is a setting. The exemption is the rule, and the rule is the layer that decides who captures a technology's gains.
Open-weight models went from 28.4 percent of tokens to 62 percent on Vercel's AI Gateway between 24 June and 22 August, on the platform's own billing telemetry, published by its chief executive. Guillermo Rauch posted both endpoints with both dates, which is rarer and more useful than a survey: this is one company's actual routing rather than a stated intention. He puts the runway ahead down to enterprise adoption still being early and to the tooling layer needing to be rewritten to be model-agnostic. One caveat keeps it honest: gateway traffic skews toward developers who chose a router precisely because they wanted to switch models. But it lands in the same weekend DeepSeek ended weekend peak pricing for its API customers, which is the same signal arriving from the seller's side: the firms selling open-weight inference are competing the price down rather than rationing the supply. What the share number measures is switching cost, and switching cost is the thing a router exists to destroy. Once the harnesses and command-line tools above the model are model-agnostic, changing a frontier model stops being a migration and becomes a change of setting, and a supplier a customer can swap in an afternoon is not really a supplier.
A British power station was reportedly shut down for four days by hackers linked to Iran, and the government's own defence is the most revealing sentence in the story: the plant was too small to be legally required to tell anyone. The Telegraph reported it on Saturday evening, an attribution no wire service has matched; the UK government confirmed an incident at what it called "a small-scale energy generator" and pointedly declined to attribute it to Iran, so treat that half as one newspaper's reporting. What is not in doubt is the official explanation, from an unnamed government source: "We have thresholds for important generators to legally notify us of cyber activity, and this site is nowhere near." The corroborated version is larger: the New York Times counted roughly 100 American water facilities targeted across at least twelve states from 26 July, and Britain's Cabinet Office put the odds of a serious successful attack on domestic infrastructure at 5 to 25 percent in a July assessment. The reporting threshold is the attack surface. An adversary optimising for maximum signal and minimum escalation aims at exactly the rung your statute declined to count, which is why escalation ladders built from legal categories keep missing the step that gets taken. The precedent is not Stuxnet but Shamoon: a 2012 wiper destroyed roughly 30,000 Saudi Aramco workstations, and Aramco's emergency replacement buying absorbed so much of the world's hard-drive supply that winter it moved global prices.
Trump said on camera on Friday that "the ultimate intervention is our military," and by Sunday three careful commentators were arguing about a sentence he did not say. A reporter asked him on the tarmac whether he had discussed "another type of intervention." He answered, "We have many types of intervention. That's one," and then, "The ultimate intervention is our military. And if we have to use that, we will." The words "bond market" appear in the reporter's question. They do not appear in his answer. Within a day Daniel Drezner had written that Trump "actually suggested he'd use military force to intervene in the bond markets," and Rabobank's Michael Every had built an argument about sovereignty and force on the same premise. Keep the scale in view too: the Treasury operation they are all discussing adds at most $4 billion an auction against a federal debt that crossed $40 trillion this month, about one hundredth of one percent. It took under twenty-four hours for a reporter's noun to become the thing thoughtful people believe a head of state said, on camera, with a transcript anyone can pull.
The two best-sourced readers of Chinese economic policy in English read the same two People's Daily commentaries this weekend and reached opposite conclusions. Both were signed 钟才文, a pen name for the Central Financial and Economic Affairs Commission, which makes them the closest thing to an official signal before the August Politburo meeting. Bill Bishop read Sunday's, on China as an anchor of world economic stability, as saying the year's targets will be met and therefore "probably a sign no stimulus moves like in August 2024." Michael Pettis read Saturday's, which argues against focusing on the growth rate and in favour of the quality of economic operation, as suggesting the opposite: that hitting the target is becoming hard. Both readings are defensible from the text, which is the point of writing it that way, and the data does not settle it: the official seventy-city new-home price index fell 3.20 percent in the year to July after 3.30 percent in June, while a vice finance minister told Xinhua a new package to expand domestic demand is coming. The commentaries were built to be readable both ways, and that is itself the information: a government that had settled on a large package would not need a text that survives two opposite glosses. We expect this week's Politburo meeting to produce something smaller than the vice finance minister's language implies, because the case for it is still being argued under a pen name rather than in a number.
A humanoid robot ran the hundred metres in 9.39 seconds in Beijing on Saturday, faster than the 9.58 seconds Usain Bolt ran in 2009. The sprint got the headlines; the standing high jump is the better number, at 2.88 metres against a best humanoid mark of 0.95 metres at the same games a year ago. That is a category change in twelve months, not an improvement. The second World Humanoid Robot Games ran 51 events with 666 teams and 2,056 robots inside the speed skating oval built for the 2022 Winter Olympics. Two lab demonstrations are an anecdote. Six hundred and sixty-six teams with a fixture list is an industry deciding what to optimise for.
Swiss watch exports rose 9.6 percent in July to 2.63 billion francs, and the interesting part is where they went. Shipments to the United States were up 26.5 percent, a third consecutive month of double-digit American gains, while China fell 18.5 percent. Almost nobody outside the trade reads the Federation of the Swiss Watch Industry's monthly release, which is a shame, because a mechanical watch is close to a pure read on discretionary wealth: nobody needs one, everybody who buys one has surplus, and the order takes months. Right now the series says the marginal buyer of an expensive Swiss watch is American and the marginal cancellation is Chinese.
House prices inside a single American state now disagree by about nineteen percentage points. On Zillow's city-level data compiled on Saturday, Miami sits roughly 3 percent below its high while Cape Coral sits roughly 22 percent below. Those two cities share a state government, an insurance market, a hurricane season and a mortgage rate. What they do not share is how much of their 2021 and 2022 building was bought by somebody who intended to sleep in it. A statewide housing statistic in Florida has stopped describing anything: it averages a supply-constrained market with a speculative one, and the average is true of neither.
Context signal: Merck is moving Keytruda patients off the drip, and the copy that arrives in 2028 cannot follow them there
Keytruda loses its main US patent in 2028, and a franchise annualizing above $33 billion on the second quarter's run rate is standing in front of that date. What Merck is doing about it is not a legal defense. It is a change of delivery form. Keytruda Qlex, the version injected under the skin in minutes rather than infused over half an hour, booked $463 million in the second quarter against $8.37 billion for the whole Keytruda family, about 5.5 percent of family revenue worldwide, and separately, what Merck calls a double-digit share of its US Keytruda business. Management targets 30 to 40 percent adoption by the end of 2027, and a permanent J-code in April 2026 cleared the reimbursement friction holding it back.
Why that is more than a convenience story: the biosimilars queued behind 2028 are copies of the infused product. A biosimilar licensed against IV pembrolizumab is not automatically substitutable for a subcutaneous formulation carrying its own formulation patents and its own delivery enzyme; a challenger who wants that share has to build the injectable version and run its own program. Every point of conversion completed before 2028 is a point the generic wave arrives too late to take.
The resistance will not come from patients. Infusion is how oncology practices and hospital outpatient departments earn chair time and administration fees, and a two-minute injection deletes that revenue line while leaving the drug cost in place. We expect conversion to land at the low end of Merck's range rather than the high end, because the people who administer the drug are paid by the method being replaced. That makes the 2028 cliff shallower for MRK than the patent date implies and steeper than the guidance implies, and it puts the cost of the difference on the biosimilar programs at Amgen and Samsung Bioepis, underwriting a share that is shrinking before they are allowed to compete for it.
Europe grandfathered every substation already in the ground, and revoked the grandfathering the moment you expand one
Sulfur hexafluoride is the insulating gas inside most of the world's high-voltage switchgear, and Europe is taking it away on a schedule that runs to 2032. The part almost nobody has priced is that the rule attaches to growth rather than to time. Regulation (EU) 2024/573, in force since 11 March 2024, bars F-gases in new switchgear at 24 kV and below from 2026 and works up the voltage classes until it reaches the high-voltage fleet in 2032. Equipment already installed may keep running, but only while it stays where it is and stays the size it is. Move it or expand it and it counts as new.
That clause lands on precisely the thing utilities are being asked to do right now. A substation that gets a new bay because a data centre or a fab connected to it is an expanded substation, so an interconnection budgeted as a bay becomes a technology change sourced from a vendor list four or five names long. California arrives at the same place from a different direction: CARB's phase-out of SF6 in new gas-insulated equipment began 1 January 2025 and runs to 2033 by voltage class.
Then the second clock, which is the one that should worry an asset owner. From 2035 the gas may be used for servicing only if it has been reclaimed or recycled, while the switchgear being energised this year is built for forty years of life. The asset stays legal for its whole life; its refills do not. We expect SF6-free share of European transmission orders to move faster than the 2032 date implies, because the binding trigger is expansion and the expansion is happening now. That pays the vendors already shipping SF6-free portfolios, GE Vernova, Siemens Energy, Hitachi, ABB and Schneider Electric, and it lands as interconnection cost and schedule risk on the load queuing for capacity behind National Grid and on colocation developers like Equinix and Digital Realty.
Anthropic put what the independent spend data calls the most performant model on the market into metered release, and in July American businesses bought six percent of their Anthropic tokens from it. OpenAI's flagship, at roughly half the price, took twenty-five percent of its own vendor's tokens in the same month.
Both numbers come from one page: the Ramp AI Index for August, published 12 August by Ramp's lead economist Ara Kharazian, drawn from the token-spend product Ramp sells to companies on its cards. Fable 5 was six percent of Anthropic's tokens and 11.4 percent of the dollars, and the gap between those two figures is the price, roughly ten dollars per million tokens, twice GPT-5.6 Sol. Kharazian's methodology note says the sample skews tech-heavy, so real adoption is probably lower than his estimate, not higher.
Ramp reads this as a ceiling on willingness to pay. Its own words: "we've found a new upper bound for how much businesses are willing to spend on AI." The piece is titled Cracks in the AI Thesis. That is a fact about a level, and the level is the smaller half of it.
Call the larger half the Dispersion Switch: an advance that arrives CHEAPER is absorbed by every buyer without a decision, so it raises the floor; an advance that arrives DEARER has to be allocated, and allocation is a firm-specific skill, so the identical advance widens the spread instead. The variable that flips the sign is the price path, not the capability. Drew Breunig named the precedent the day the Ramp read circulated: Herb Sutter's "The Free Lunch Is Over" (Dr. Dobb's Journal, March 2005). While single-thread performance doubled every eighteen months, optimising code was irrational, because waiting meant the hardware fixed the problem for you. When it stalled, developers had to think about what work went where. Breunig's version of the same sentence: before Fable, improving your harness felt silly, because a new model would arrive at the same price or cheaper and paper over the problem.
What has to be built instead is the part economists have already measured. Brynjolfsson, Rock and Syverson's productivity J-curve shows a general-purpose technology's returns are gated by intangible complements, the process redesign, workflow co-invention and human capital that firms fund unevenly and that the national accounts cannot see. Routing, evals, harnesses and context design are that asset class. And path dependence is what makes the switch durable: once the routing layer exists, it is not unbuilt by the price cut that would have made it unnecessary.
The spread is already visible on the input side, in Ramp's own July numbers. The top one percent of businesses spent a median $7,400 per employee on AI. The top ten percent spent $650. The median firm spent $11.95. That is dispersion in inputs, not yet in output. But Syverson's canonical finding is that inside a single four-digit US manufacturing industry the ninetieth-percentile plant already makes about 1.92 times the output of the tenth-percentile plant from the same measured inputs, before anyone adds a routing decision on top. This is the second answer to the question 24 June left open, when the end of complimentary frontier access was called the first. We expect Fable 5 to still be under fifteen percent of Anthropic's tokens when Ramp publishes its December index, because the binding constraint is price rather than unfamiliarity.
Where this might be wrong. The strongest fact against this read is on the same page it rests on: in the same July, Anthropic's share of US businesses paying for AI rose 1.1 points to 43.5 percent, its fastest month of the year, while OpenAI's rose 0.23 points to 39.7. A vendor whose premium tier is repelling buyers should not be the vendor gaining buyers fastest, and that is difficult to reconcile with a demand story about price. The lag explanation is live too, and Kharazian raises it himself: this is one month, on an instrument that sees card and token-management spend. The buyers most likely to pay for the expensive tier are the ones who negotiate committed-spend contracts directly and never touch a card, the segment the measurement is least able to see. If that is the answer, six percent is an artifact of a young month and there is no regime here at all. Breunig's own objection, which he raises and does not settle, is that falling inference prices will pull everything back through the largest models; his rebuttal, that the same efficiency gains lift K3 and Qwen too, is a conjecture rather than evidence. And the precedent cuts against the read. Sutter's 2005 forecast was that the end of free performance would force developers into concurrency, and for most of them it never did: cloud elasticity, managed runtimes, JIT compilers and GPUs restored the free lunch by another route, and the dispersion he predicted concentrated in a narrow band of systems programmers rather than spreading through the economy. The same escape hatch is open here. An orchestration layer sold as a product converts the skill into a purchase, and Stripe agreeing to pay more than seven billion dollars for OpenRouter is a wager that it will. This read fails if Fable 5 clears fifteen percent of Anthropic tokens in Ramp's December index with no price change, or if Anthropic prices Fable within twenty-five percent of GPT-5.6 Sol before then.
"All those who wish to change something in themselves must learn … to inhibit their immediate reaction to any stimulus to gain a desired end."
— F. M. Alexander, The Use of the Self (1932)
Alexander was a Shakespearean reciter in the 1890s who kept losing his voice on stage and could get no doctor to explain it, so he set up mirrors and watched himself speak. What he found was that the instant he decided to recite, before a sound came out, his head pulled back and down and his throat shortened. The hoarseness was not something happening to him. It was the first thing he did.
He could see the movement and still could not stop it, because by the time he had decided to speak the pattern had already run. Trying to say the line correctly produced a more careful version of the same pull, since the correction was being carried out by the habit it was meant to fix.
What worked was refusing to speak at all for a moment after deciding to. Not relaxing, not adjusting, not doing it better. Just declining to begin. He called that inhibition, and he called the opposite end-gaining: going straight at the result and taking whatever means the body reaches for first.
This is why so much self-improvement is effort without movement. You add a technique on top of a pattern and the pattern executes the technique. The person who resolves to listen better interrupts more attentively. The one who decides to be calmer holds the tension somewhere quieter. Nothing was inhibited, so nothing changed, and the effort is real, which is what makes it so hard to see.
Between the stimulus and the first movement there is a gap that belongs to you, and it is the one part of the sequence your habit has not already claimed.
Today's practice: pick one physical thing you do twenty times a day, standing up from a chair, reaching for your phone, starting to type. Once today, insert a two-second stop between deciding to do it and beginning to move. Correct nothing during the stop. Just do not start yet, then go, and feel what your neck does.
For thirty years the most-cited proof that human intuition is broken was a basketball study. In 1985 Thomas Gilovich, Robert Vallone and Amos Tversky went through the Philadelphia 76ers' shooting records and found that a player who had just made a shot was no more likely than usual to make the next one. The hot hand was a story fans told themselves, and it became the textbook case of a cognitive bias.
Then in 2018 Joshua Miller and Adam Sanjurjo showed in Econometrica that the way the original study counted was itself biased. Take a finite string of coin flips, keep only the flips that came immediately after a head, and average those. You do not get fifty percent. You get less, because conditioning on "the previous one was a head" is a selection, and the selection drags the average down. Correct for it and the 1985 conclusion reverses. The flagship example of human irrationality contained an error of exactly the kind it accused people of.
That is the mechanism, and most people carry the wrong version of it. A systematic error is not a mistake in a head. It is a mistake in a procedure. The giveaway is that it repeats in the same direction. A random mistake shrinks as you gather more data, which is why more data feels like safety; a systematic one does not shrink, it gets more confident, because every new observation arrives through the same crooked pipe. Which is why the dangerous errors never feel like guessing. They feel like measuring.
A case from as far from basketball as you can get. In June 1985 Alvan Feinstein, Daniel Sosin and Carolyn Wells compared lung cancer patients treated in 1977 with patients treated between 1953 and 1964 at the same hospitals, in the New England Journal of Medicine. Survival was better for the whole group and better inside every single stage, and nothing about the treatment explained it. Better imaging had found small metastases that used to go unnoticed, so patients once called early-stage were now called late-stage. They left stage one as its sickest members and joined stage three as its healthiest, and the average survival of both groups rose without one patient living an extra day. Feinstein called it the Will Rogers phenomenon, after the line about Oklahomans moving to California and raising the average intelligence of both states.
Run some of this, not all of it. Any comparison can be attacked by asking who got selected into it, and someone who demands a clean selection mechanism before believing anything ends up believing nothing and calling it rigour. You have found a bias only if you can name the selector and say which way it pushes.
The tool. Before accepting any comparison, finish this sentence out loud: what decided which things I am comparing was ______, which pushes the answer ______. "The follow-up survey only reached people still using the product, which pushes satisfaction up." If you can complete it, you can size the error and often correct it in your head. If you cannot, you have found the one thing to check before you decide anything on it.
Split a large group into two piles so that both match on every attribute at once: trivia players across categories of knowledge, patients across baseline characteristics. Perfect balance is almost never available, so the real question is how much imbalance you are stuck with, and the intuition is wrong. Assign each item by coin flip and the leftover imbalance grows steadily as the pile gets bigger. In the early 1980s the mathematician János Komlós conjectured that the imbalance never has to exceed a single universal constant, no matter how many objects you have and, astonishingly, no matter how many attributes you are balancing at once. The conjecture is still open, but the ceiling moved: Wojciech Banaszczyk got it to the square root of log N in 1998, where it sat for nearly thirty years, until Nikhil Bansal of the University of Michigan and Haotian Jiang of the University of Chicago announced a bound of log N to the one-fourth power in the autumn of 2025. That is 1 for ten objects and about 3 for ten-to-the-eighty-first, roughly the count of atoms in the observable universe, which Yale's Daniel Spielman put plainly: in your life you will not meet a number for which it exceeds 5.
The mechanism is worth more than the number, and it is not "try harder on each attribute." Earlier algorithms tracked one quantity, the total imbalance piled up so far. Bansal and Jiang tracked a second they called dependency: nudge one attribute, and how far does the imbalance in all the other attributes move? They then chose each step to make that joint disturbance as small as possible, because the attributes only look inseparable. You cannot fix the convertibles without upsetting the colours, and yet there turn out to be directions in which they barely bother each other. What decides whether you land near the bound, then, is not effort per criterion. It is whether you are scoring your moves on the wrong axis.
So, for this week: the next time you allocate across several criteria at once, a team across skills, a week across obligations, a study across arms, rank each candidate move by how much it disturbs everything else rather than by how much it fixes the dimension currently bothering you, and take the smallest-disturbance move even when it is second-best on the problem you sat down to solve. You can check whether you did it. A week after the decision, write down the dimensions you were not thinking about and measure how far each drifted. If the ones you ignored degraded by more than the one you fixed improved, you ran the coin flip and paid for it. The same architecture is why a reorganisation that repairs reporting lines can quietly destroy institutional memory, and why the hard part of a clinical trial is the randomisation rather than the treatment. Both are discrepancy problems, and the cost lives entirely in the dimensions nobody was watching.