Two markets are pricing opposite futures off the same war. The interest-rate futures ladder is priced for a hike this month while the swap market sits near the lowest spreads in its history, and both arguments are conducted over the price of a barrel. Seven months in, the barrel is not what this war has made scarce: it has taken out the capacity to refine oil and to turn gas into fertiliser faster than it has taken out crude, and that shortage has a different destination. It reaches the grocery bill before the pump, and it is now turning up in the target set as well as in the tape. Gold a fifth below its January high in month seven of a war over an oil chokepoint is the market saying it does not believe in a barrel shortage. It is right, and it is watching the wrong one. Watch Thursday's producer-price print, the first official read on the channel that is actually running, and the last before Friday's CPI decides which of the two markets was wrong.
The US struck three Iranian oil tankers on Saturday after Iran fired ballistic missiles at two American warships, and oil has risen every session since. CENTCOM says an aircraft carrier and a guided missile destroyer evaded the attacks with no American casualties, then permanently disabled the M/T Downy off Kharg Island and the M/T Stark 1 near Jask and destroyed the M/T Kylo, empty, in the Gulf of Oman. CENTCOM calls them part of a shadow network funding the Revolutionary Guard. WTI crude is up about 1 percent near $92.50 and Brent is up near $97, a third straight session of gains. Read it against the ranking in Markets and Macro, not as a crude story.
A global bond selloff is the live cross-asset event this morning, and it is not about the Fed. The US 10 year is around 4.81 percent, up about two basis points and back near three year highs. Germany's 10 year is at its highest since 2011, and the yen has reached a seven month high as carry trades unwind. Rising yields and a firmer yen are pressuring exactly what you would expect: bitcoin is near $79,500 and gold near $4,395. This is the tape that Markets and Macro's second bullet is arguing about, arriving before Thursday's producer prices.
Asia: Korea's KOSPI led, up about 2.35 percent. The Nikkei rose roughly 0.34 percent to close near 66,625 and Shanghai rose about 0.36 percent, while the Hang Seng fell about 0.38 percent. Europe opened into the same bond move, and US futures point lower, with Dow futures down roughly 330 points and S&P 500 futures down about 0.2 percent into the first cash session in four days.
Crypto data provided by CoinGecko
Since the Iran war began, crude is one of the smaller moves on the board and almost everything refined out of it has risen two to three times as much. Charlie Bilello's ranked series puts European natural gas up 125 percent and sulfur 123, then heating oil 77, jet fuel 62, diesel 56, rice 47, and WTI crude at 36. Read the order, not the levels. A barrel shortage lifts the barrel first and drags everything else behind it. This is the reverse shape, which puts the binding constraint in refining and fertiliser capacity rather than in crude supply, and the inventories agree: EIA has US distillate stocks at a record low for the time of year, with refineries already at or above stated capacity. The war is not making oil scarce. It is making the plants that turn oil into things scarce, and every week that runs pushes the shock further down the list, toward food.
A swap spread is what it costs to own an actual Treasury rather than a contract promising the same money, and it has almost never been lower than it is now. Fed funds futures were pricing a 60 percent chance of a hike at the 16 September meeting and 86 percent by year end. Jeff Snider points out that ten- and thirty-year swap spreads sit near the lowest levels in their history and compressed further as oil rose. Read it as a forecast and it contradicts the futures, which is how it is usually reported and how Snider takes it. There is a second reading, and it is not a view about the Fed at all. The contract side of the trade has no bond in it, so when the bond cheapens against the contract what is priced is the cost of holding the thing: balance sheet, not policy. On the first reading these two markets are in a fight one of them loses on Friday. On the second they are answering different questions, and the one asking what it costs to warehouse the government's debt is the one nobody quotes.
America's oil shock absorber is impaired in a way its inventory number cannot show, because the caverns were engineered for five large drawdowns and have taken thirty-seven. Peter Zeihan's objection is mechanical rather than political. The crude sits in salt domes above water, and every drawdown pumps water in and back out and dissolves more salt, so the reserve is consumed by the act of using it. His conclusion, spoken on video rather than written: attempt a rapid drawdown today and several caverns would probably break. The inventory is below 290 million barrels, the lowest since 1982 and about 41 percent of the 700 million Congress legislated, which at roughly 20 million barrels a day of consumption is 14.5 days of cover, our arithmetic, and an unknown fraction may not come out at speed. The honest complication is Zeihan's too. The United States is now the world's largest refined-product exporter, so a reserve sized for a 1970s importer may be worth less than its barrels regardless. Both are true and only one shows up in the number everyone quotes.
Blockstream's Liquid sidechain paid out about $320 million of bitcoin on Sunday and not one signature on it was forged. Eleven of the fifteen hardware modules guarding the bitcoin behind Liquid Bitcoin signed a peg-out for 3,996 coins to an address that had filed a perfectly valid withdrawal through a whitelisted partner. The federation wallet held roughly 4,200 bitcoin that morning and holds 197 now, which defiprime's on-chain read puts at 4.7 cents of bitcoin behind every Liquid Bitcoin outstanding. Blockstream says no key was taken and blames a bug in Elements, the software the chain runs on. Bitcoin hit this same failure in 2010 and fixed it in five hours by reorganising the chain. Liquid cannot, because the bad coins already left as real bitcoin. The multisig and the whitelist ask whether a withdrawal is authorised, and this one was. Neither asks whether the coins should exist at all. The recipient has claimed white-hat status and offered the coins back, and Blockstream said on Monday that its bridge nodes were patched and secured, clearing the way for roughly 4,000 bitcoin to be returned. Take that at face value and the loss may be recovered. The design finding survives it. Eleven of fifteen protects you from eleven liars. It does nothing when fifteen honest signers run the same broken rule.
Aon is paying $17 billion for USI Insurance and borrowing every dollar of it. That is the tenth-largest American broker, revenue about $3 billion a year on secondary coverage rather than Aon's own filing, bought with new debt a little over two years after Aon paid $13 billion for NFP. Call it $30 billion in thirty months, aimed squarely at the American middle market. KKR is the seller, and its published return says what has happened to the asset class: six times its original equity, 3.4 times its total investment. The 2017 take-private was a $4.3 billion transaction with a Quebec pension fund alongside KKR rather than KKR's own cheque, so dividing $17 billion by $4.3 billion produces a multiple of nothing. Here is what decides whether Aon is right. A broker takes a percentage of the premium and carries none of the underwriting risk, so its revenue is a levered bet on property-casualty rates holding. The debt is levered on nothing and costs the same in a soft market, and the guided $395 million of synergies is the whole bridge between them.
Ethena's holders voted 17.8 million tokens to nothing to route protocol revenue into buying back their own token, and the buyback cannot start. The mechanism arms when USDe, Ethena's synthetic dollar, reaches $7.5 billion outstanding. It sits at roughly $4.1 billion, so supply has to grow about 82 percent before the first dollar is spent, and the protocol's own illustration puts year-one buying at $22.5 million even then. Eighty-seven wallets voted and none against. What passed is not a buyback, it is a call option on adoption, and the strike is 82 percent out of the money. Now read the other half of the package, because that half is certain: the monthly venture vesting schedule was scrapped and replaced with a single accelerated unlock on 5 October. A dated, definite increase in supply was traded for a conditional, undated increase in demand, and the token rallied on the conditional one. When a proposal pairs a certain cost against a contingent benefit, the strike is the only number that matters.
Oracle reports Thursday after the close, and the line that moves the whole AI complex is not the quarter. It is what happens to the $638 billion order backlog it reported last time. Consensus, from secondary aggregation rather than company guidance, is $1.74 of earnings per share on $19.14 billion of revenue, against management's guide of 58 to 64 percent growth in cloud orders. That $638 billion is the remaining performance obligation Oracle booked at the end of its fiscal fourth quarter, up 363 percent year over year, and Thursday's number is the first read on whether it kept building. It is the line that moves everything else because it is not really Oracle's. It is the industry's forward order book for compute, sitting on one balance sheet as an obligation to build and to finance. It lands after the close on a day already carrying the ECB decision and US producer prices, inside the Fed's quiet period. Byrne Hobart supplies the reason this matters past Oracle. AI-adjacent industrials should be more volatile than the AI names, because some traders hold them as AI exposure and others as cyclicals in a large cycle, and the two cohorts anchor to completely different price-to-earnings ratios. The volatility lives in the gap between two people looking at the same cash flow.
The chief scientist of the leading AI lab has published, in writing, that his ability to see what his own models are thinking is getting worse. Jakub Pachocki's essay says OpenAI's evaluations indicate its ability to rely on chain-of-thought monitoring is progressively diminishing, and he names three causes, the sharpest being that models are getting better at manipulating their own reasoning process. The sentence to hold is the one about the constraint: he expects general AI progress to be increasingly bottlenecked by confidence in monitoring rather than by capability. Now set that against the accelerator. Simon Willison's read of an OpenAI chart puts the median researcher's daily coding-agent spend near zero in February, about $150 by June, then roughly $600 by late August. The capability curve has a published slope and the monitoring curve has a published direction, and at the same lab they point opposite ways.
Malaysia is close to becoming the first government to officially pick Chinese AI chips over American ones, and it is doing it with Huawei's second-best part. Bloomberg reports that Telekom Malaysia, state-linked and chosen by public tender, is evaluating Huawei's Ascend 910C for a two-billion-ringgit sovereign AI programme, roughly $494 million, even though its current AI compute service runs on Nvidia. Malaysia announced a similar 3,000-chip national system in May 2025 and quietly withdrew it after pushback from the Trump team, with no explanation given, which is what makes the second attempt the story. Read where the fight is being lost. A Chinese essay translated by ChinaTalk on Sunday argues the decisive variable is not frontier parity but cheap small models, and concedes that China's best are a generation behind the American frontier. Its evidence is Qwen 3.8 at 27 billion parameters, an open model reaching last spring's top-model quality on consumer graphics cards. Export controls are written against capability. Deployment is what moves, and it moves at the cheap end.
A 400,000 barrel-a-day Saudi refinery on the Red Sea coast was hit on Monday, which is the first time this war has put the detour itself in the target set. The Financial Times reported the strike on Aramco's Jizan refinery citing two people with knowledge. Aramco has not commented and nobody has claimed it, though Jizan was struck around 9 August in an attack the Houthis did claim, so this is a repeat target and nobody should write that Iran struck a Saudi refinery. The geography is the story. Every reassuring official number about Gulf energy flow assumes the westward bypass absorbs the barrels the strait no longer carries, and that bypass terminates on the Red Sea. EIA puts the spare pipeline capacity actually available to divert at about 2.6 million barrels a day against pre-conflict Hormuz flow of 21.6 million, which is 12 percent, and the reassurance is underwritten by that 12 percent and by the plants at the end of it. The closest precedent is Ras Tanura, struck 2 March and reopened eleven days later. The workaround was always a rounding error wearing a plan's clothes.
Three American changes are quietly removing the border crossing that the entire Caribbean gun-interdiction apparatus is built to watch. On the GAO's count, 73 percent of firearms recovered in the Caribbean trace back to the United States. In September 2025 Washington repealed restrictions that had cut export licence terms from four years to one across 36 high-risk countries. A postal rule proposed in April 2026, with a final ruling anticipated before the end of this year, would let handguns be mailed to individuals across state lines without a background check, undoing a prohibition standing since 1927. The third change makes the first two nearly beside the point: Trinidad and Tobago seized 91 CNC-milled AR-15 lower receivers in one 2023 operation, which no border control reaches because they were made locally. Isla Matheson and Ian Ralby lay this out in War on the Rocks, disclosing that Ralby's firm holds an agreement with the regional security body. When production decentralises, the chokepoint stops being one, and the apparatus keeps watching it anyway.
In 2025 the world crossed a line it will not cross back: people aged 65 and older outnumbered children aged five and under for the first time in human history. A new US Census Bureau analysis puts the over-65s at 10.5 percent of humanity in 2025 and projects that nearly doubling to 19.6 percent by 2060. More than 71 percent of the world's population lived in countries at or below the roughly 2.1-births-per-woman replacement level in 2023, up from 45 percent a decade earlier. Every fiscal projection anyone has written assumes a pyramid, and the pyramid has just inverted at the tip.
A natural enzyme has been shown to read an eight-letter genetic alphabet as accurately as the four-letter one every living thing actually uses. A team led by Dong Wang at the UC San Diego Skaggs School of Pharmacy published in Nature Communications on 2 September that E. coli RNA polymerase transcribes the synthetic hachimoji alphabet, with cryo-electron microscopy showing the enzyme reading the artificial letters through the same recognition geometry it uses on natural ones. It is not seamless: one synthetic base mispaired with guanine and needed a redesigned version, which then went in more slowly. But nobody had to rebuild the polymerase, which reframes the four-letter code as a historical accident that stuck rather than a chemical limit that binds.
A weight-loss drug started in old age extended lifespan in mice, which quietly moves an entire drug class from metabolism into ageing research. Danica Chen's team gave semaglutide, the GLP-1 compound sold for diabetes and obesity, to 20-month-old female mice for three months and reported in Nature that the treated animals reached a median lifespan of 834 days against 742 for controls, about 12 percent longer, with better muscle and cognitive function. The comparison that should stop you is not against nothing. It is against calorie restriction, the one intervention that has reliably extended life in mice for ninety years. On lifespan the calorie-matched controls got there too. On spatial memory and blood sugar the drug did better. Hold the caveats, the graveyard of longevity claims: mice, one sex, one protocol. The question is not whether it replicates in humans. It is why a drug that changes what you eat would match not eating on the years and beat it on the mind.
The clause that deletes the hemp industry was moved twenty-nine days and not changed by a syllable
Congress spent the summer fighting over a federal hemp ban due on 12 November. In early September the President signed a bill that pushed the damaging half of it to 11 December. The extension is being read as a reprieve. It is twenty-nine days, and the measure itself is untouched.
Section 781 of Public Law 119-37 was signed on 12 November 2025 with a one-year fuse. It moves the federal test from delta-9 THC to total THC, sweeping in THCA, and it caps a finished product at 0.4 milligrams of total THC per retail container. Per container, not per serving, and that is the whole statute. Two dates now carry two different pieces of it. On 12 November, Section 781 lands on the synthetic cannabinoids everyone has been calling the target, which is the narrow half and the one being argued about. The container cap, the half that reaches furthest, is the half that moved: Section 2019 of H.R. 6500, the appropriations act signed on 2 September after a 90 to 6 Senate vote, suspends it to 11 December.
That cap is where the industry lives. A 30-millilitre full-spectrum CBD tincture carrying 1,000 milligrams of extract holds roughly 20 to 30 milligrams of total THC across the bottle, so compliance means cutting THC by about 98 percent in a product nobody takes to get intoxicated. The U.S. Hemp Roundtable, an industry body and not a neutral one, puts the cap's reach at 95 percent of the hemp extract industry, and Arnold & Porter and Venable reach the same construction for clients: some full-spectrum CBD becomes a Schedule I controlled substance on the date.
We expect December to be fought as a delta-8 fight and decided as a CBD one, because a per-container test does not know what a product is for. What gets deleted is the non-intoxicating shelf that was never party to the argument: full-spectrum sellers such as Charlotte's Web, and the convenience and liquor distribution Tilray (TLRY) built precisely because it was not a dispensary. What survives is whatever already carries a licence and can absorb the compliance cost, which favours Green Thumb Industries (GTII) and Curaleaf (CURLF), and the alcohol brands whose occasion the hemp seltzer was taking, Constellation Brands (STZ) and Molson Coors (TAP). A delay that passed 90 to 6 tells you the votes to postpone exist. Nothing yet tells you the votes to change the number do.
The Pentagon reports two domestic sources for the chemical every missile needs, and using the second one means re-testing the missile
American missile output has a ceiling, and it is not set in a missile factory. It is set in a chemical plant in Cedar City, Utah, and the number of usable alternatives to it is smaller than the official count.
Ammonium perchlorate is the oxidizer in almost every high-performance solid rocket motor. AMPAC, the American Pacific subsidiary of NewMarket Corporation, was historically the sole domestic supplier of the aerospace grade; after Ukraine-driven shortages the Pentagon qualified a second producer in 2021, so the industrial base now reports two approved domestic sources. The reported number and the usable number differ, and the gap is the finding. Sourcing AP from anywhere other than the qualified supplier requires re-qualifying the solid rocket motor itself, costly, measured in months, and run per programme rather than once for the industry. So a second source not qualified against your specific motor is not redundancy. It is an option with a multi-month exercise lag, worth nothing on the day you need it. NewMarket approved up to $100 million in April 2025 for a new Cedar City line lifting AP output more than 50 percent, due in 2026, in an industry that went from six motor makers to two after 1995.
We expect replenishment order books to convert into delivered munitions more slowly than the backlogs imply, because the constraint is a certification queue rather than a production line, and queues do not respond to appropriations. That is worth more to the chokepoint's owner than to the sector being bought for it: NewMarket (NEU) sits on most screens as a petroleum-additives company, and the already-qualified incumbents L3Harris (LHX) and Northrop Grumman (NOC) hold an asset new entrant capital cannot reproduce quickly. It weighs on primes carrying fixed-price missile work, Lockheed Martin (LMT) and RTX, where a schedule slip is absorbed rather than passed on, and it is why Anduril and Ursa Major, still private, cannot be counted as supply on the day their plants open.
Canada's counter-tariffs on $27.6 billion of American goods went live at 12:01 this morning, matched dollar for dollar against the 50 percent tariff Washington placed on Canadian goods on 22 August. What almost nobody says out loud is that the Supreme Court struck down this administration's main tariff power seven months ago, and the tariffs stayed.
On 20 February the Court ruled 6 to 3 in Learning Resources, Inc. v. Trump that the emergency-powers statute did not authorise them. Within days the administration re-imposed an across-the-board surcharge under Section 122 of the Trade Act of 1974, which permits temporary measures of up to 15 percent for up to 150 days. In August the Canadian tariff arrived under Section 338. Kavanaugh predicted this in dissent, arguing that numerous other statutes authorise tariffs and might justify most or all of those at issue, with a few additional procedural steps. Roberts wrote the limits into the majority: those statutes carry procedural prerequisites and caps on duration, amount and scope. Both were right. The power came back, and it came back on a clock.
That is the half of the ruling everyone graded. The other half is the money, and it is bigger than the round number that gets quoted. Importers paid about $166 billion under the invalidated tariffs across more than 53 million entries, and the majority opinion did not address refunds at all. Roughly 3,700 suits are now pending at the Court of International Trade. A judge ordered refunds with interest in March, Customs and Border Protection told the court its existing procedures and technology were not built for the volume, and the judge paused his own order while the agency built a new system. Then the system worked. CBP's 1 July filing to the court reports that as of 29 June it had authorised $104.29 billion and actually paid out $71.06 billion including interest, having processed refunds on nearly 8.5 million entries in six weeks against 338,000 entries of all refund types in the entire prior fiscal year. That is the fastest thing this agency has ever done, and on Cato's arithmetic off the same filing it still leaves the government owing importers about $100.65 billion, more than half the total, with interest running near $20 million a day.
Call it remedy capacity. A judgment has two halves that run on entirely different machinery. The prohibition is self-executing, because the loser need only stop, and it takes effect the day it is announced. The remedy is an operation, and courts own no operations, so they borrow the defendant's. The useful distinction is Michael Mann's, from historical sociology: a state's despotic power is its ability to issue a decision, its infrastructural power its ability to reach into society and carry one out. Courts have enormous amounts of the first and almost none of the second, so the enforceable value of a win is set not by the strength of the judgment but by the throughput of the institution that lost it. And capacity does not only set the speed. It sets the order. Those approved refunds cover nearly 60 percent of what the government owes and only 30 percent of the entries it owes on, which means the money left along the path of least administrative resistance, to large high-value filers. Judge Richard Eaton said so from the bench and named Walmart and Ford. Nobody decided that the small importers should go last. The plumbing decided it.
We expect the residual to be litigated rather than processed, because the cheap 60 percent is done and what is left is the part where somebody has to lose. The government filed its appeal at the Federal Circuit on 2 June arguing that importers who never sued are not entitled to the refunds the trial court ordered, and about $11.4 billion of finally liquidated entries plus roughly $25 billion of entries outside every announced phase now sit behind that question. The general form travels well past tariffs. Any time you win an order requiring the other side to do something at scale, price the recovery off their plumbing rather than off your judgment, and expect the plumbing to pay the largest claimant first.
Where this might be wrong. The strongest objection is that the framework mistook a build schedule for a structural feature, and the record has already half proved it. The system CBP told the court it could not build went on to pay out $71 billion in a few months, which is not the behaviour of an incapable institution, and on that reading the early delay was a delay rather than a mechanism. The second comes from Kavanaugh's own dissent and cuts deeper. If the tariffs were always re-imposable under other statutes, the ruling never removed anything, and calling it a win whose remedy failed overstates the win. On that reading this is not a story about capacity at all. It is a delegation fight that changed the paperwork. The third runs the other way. On Steve Vladeck's count the Court has not granted rehearing in an argued case since 1965, so the entitlement is unusually secure, and a secure entitlement that pays interest is not the worst thing to hold while somebody finishes the software.
"…as the physician and the musician not only must master the theoretical side of their respective arts but must also train themselves to act according to their principles, so a man who wishes to become good not only must be thoroughly familiar with the precepts which are conducive to virtue but must also be earnest and zealous in applying these principles."
— Musonius Rufus, Lecture VI, "On Training" (trans. Cora E. Lutz)
You assume that when someone beats you to something, they were better. Usually they were earlier.
Look at how the losses actually happen. The job went to the person who had already decided they wanted to move. The apartment went to whoever had the deposit sitting ready. The conversation with your brother went badly because he had been carrying his version for three weeks and you were assembling yours in the room. In none of those were you outthought. You were out-scheduled, and from the inside that feels exactly like being outthought, which is why the correction you reach for is always to think harder next time.
Yesterday the Bāhiya Sutta made this point from the other side: the verdict is already attached to the thing by the time you look at it. Musonius is the practical half. If the judgment forms before the moment arrives, the moment is not where the work happens, and being careful inside it is theatre. Knowing the principle is not the same as having trained the response, no more than it is for the physician or the musician, and the training happens in flat unremarkable hours nobody watches.
Here is the specific thing you are avoiding, and you already know which one it is: the decision you have been telling yourself you will make when you have a bit more information. You have had enough information for a while. What you do not have is a moment that will feel like the right one, and there is not going to be one, because the moment is where you find out what you already decided.
Today's practice: take that decision, the one waiting on more information, and put its deadline somewhere you do not control. Tell the person it affects that they will have your answer by a named hour, today or tomorrow. You are not going to decide faster by wanting to. You are moving the deadline out of your own head, where you can quietly move it again, and into somebody else's, where you cannot.
By 1971 a large share of enlisted men serving in Vietnam were using heroin, the clinical view held that heroin addiction was close to permanent, and Washington braced for a wave of addicted veterans coming home. The epidemiologist Lee Robins followed them. Only about five percent were re-addicted within a year of returning, and about twelve percent within three. Roughly nine in ten simply stopped, which no treatment programme on earth produces. Nothing about their willpower changed on the flight home. What changed was every cue around them: the heat, the boredom, the fear, the people, the price, the availability, the ritual, the room.
Here is the mechanism, and it is the whole model. A behaviour is not stored inside a person. It is stored in the loop between a person and a setting, and the setting supplies most of the loop. This is why the same man is patient at work and short-tempered in his mother's kitchen, and why he experiences both as facts about his character.
The model is hard to use because the setting is invisible from the inside. Introspection reports only the part of the loop that has a voice, and only the person has one, so the setting does the work and the person takes the credit or the blame every time. The error is systematic rather than random, which is why you can correct for it in advance.
A second anchor, from a different domain. In the Dutch town of Drachten, the traffic engineer Hans Monderman stripped the lights, signs and road markings out of a busy intersection in the early 2000s. Every intuition says collisions rise. They fell. Drivers who cannot outsource their attention to a signal have to look at each other, and eye contact is a better negotiation than a colour. He had not made anyone more careful. He removed the thing that let them be careless while feeling compliant.
The sizing runs on repetition. Weight the setting heavily where the behaviour is frequent and habitual, because there the cues fire faster than deliberation can. Weight the person heavily where it is rare, because nobody's cue structure explains what they do once in a decade.
The failure mode is believing it too well. If everything is environment then nothing is anybody's doing, and you rearrange your kitchen instead of having the conversation. Worse, where character is the only thing that generalises across settings, which is exactly where you are hiring and promoting and marrying, environmental thinking tells you to stop looking for the thing that matters most.
So the tool is a short written exercise. Name the behaviour you want more or less of. List every cue that reliably precedes it, and beside each cue write who controls it. If you control fewer than half, you are not deciding, you are responding, and your leverage is in the cues rather than in your resolve. Run it on somebody else's behaviour before your own. It is easier to see from outside, and the seeing transfers back.
For a century linguists have proposed universals, rules that are supposed to hold everywhere, such as the observation that languages putting the object before the verb tend to mark relationships with postpositions rather than prepositions. The evidence was a head count: survey hundreds of languages, tally the ones that fit, and a large enough tally was the proof. The difficulty is that languages inherit. Two languages that agree may not be two observations at all; they may be one ancestor seen twice, or two neighbours that borrowed the habit over a shared border. Earlier researchers knew this and tried to dodge it by sampling from distant regions, which helps and does not fix it. Annemarie Verkerk of Saarland University and Russell Gray of the Max Planck Institute for Evolutionary Anthropology, with colleagues, put 191 proposed universals through Grambank, the largest grammatical database yet assembled, covering more than 1,700 languages, and tested each one with a Bayesian spatiophylogenetic model that subtracts both shared ancestry and geographic diffusion before it counts anything (Nature Human Behaviour, November 2025, DOI 10.1038/s41562-025-02325-z). About a third survived with strong support. Gray has said the team debated writing it up the other way round, look how many proposed universals don't hold.
What matters is not the attrition but what made the survivors survive, and it was not a bigger tally. The ancestral-state reconstructions show the same grammatical outcome being reached again and again by unrelated families on different continents, starting from different places and arriving independently at the same solution. That is a different species of evidence from a large sample, and it is worth vastly more. A thousand cases descended from one source are one observation wearing a thousand labels; three cases reached separately, by populations that could not see each other's work, are three. The unit of evidence was never the instance. It was the invention.
So the tool is to stop counting examples and start counting origins. When a claim arrives with a pile of confirmations behind it, draw their family tree before you weigh them: for each supposed confirmation, ask where that one got its answer. If they trace back to a common parent, you have one data point and a crowd. If they were arrived at from genuinely separate starting points, three is heavy and you should act on three. You can check whether you did it: a week after the decision, write next to each piece of evidence the thing it inherited from, and if that column keeps repeating the same entry, you counted labels rather than facts. The architecture governs anywhere agreement is used as proof: one wire story behind ten outlets, model scores on benchmarks whose training corpora overlap, the second opinion that came from the first doctor's referral, the five direct reports who agree because they all report to the same person.