Stocks closed Friday at a record on their best week in months, and the number underneath that week was 23,000 jobs lost. What stands out is where this week's prices came from: almost every one was computed from a comparison rather than from a fundamental. Berkshire restarted its buyback because the index ran past it, not because its own price fell. An index weight stepped up off a lockup expiry rather than off a buyer. A Bitcoin fork died because no market would quote it at all. The things that are actually scarce, meanwhile, moved without repricing anything. A Saudi export refinery was hit twice in thirteen days, and 518 sanctioned tankers sit outside the one sanctions regime that measurably works. Watch Wednesday at 8:30 Eastern. July CPI is the first number this month that is not a comparison to something else, and it arrives with the bond market and the policy market holding incompatible positions that only one print can settle.
Crypto data provided by CoinGecko
The ten-year breakeven closed Thursday at 2.26 percent, which is below the core inflation consensus for a print that has not happened yet. The Federal Reserve's H.15 release last Friday puts the ten-year constant maturity at 4.69 percent and the ten-year inflation-indexed note at 2.43, and the difference is what the bond market expects inflation to average for a decade. Wednesday at 8:30 Eastern the Bureau of Labor Statistics publishes July CPI, and Reuters consensus is 3.4 percent headline and 2.5 percent core, against a June print of 3.5 percent. So the inflation market is pricing a decade at a level the next single month is not expected to reach, while the policy market spent Friday cutting September hike odds to 42 percent. The honest caveat is that a breakeven is not a clean forecast, because it also carries an inflation risk premium and the liquidity discount on index-linked paper, and both push it below the true expectation. That objection has a limit: neither moves 124 basis points, which is the gap to June's headline. The asymmetry is the part nobody prices. A soft core print vindicates the breakeven and leaves the curve alone, while a hot one has to move 2.26 percent and 42 percent in opposite directions at once, and that trade has no crowded side.
Michael Pettis argues that cutting interest rates in China mechanically shrinks the household share of the economy, which means the rebalancing everyone wants requires higher rates rather than lower ones. His mechanism is a balance sheet, not a forecast. Chinese banks borrow from households and lend to businesses and to the state, so the deposit rate is household income and the lending rate is corporate cost, and cutting both transfers income from the first group to the second. The consumption share falls as a matter of arithmetic. July's data is what that produces: a $112.5 billion trade surplus on exports up 23.9 percent, against consumer prices up 0.5 percent from a year earlier and down 0.1 percent on the month. Against the argument, honestly: producer prices are up 3.5 percent year over year, which is not deflation, and the 2000s precedent Pettis leans on ran alongside a credit boom this one lacks. But every model treating Chinese easing as stimulus for global demand has the sign backwards, and $687 billion of surplus through July says the adjustment is being exported.
Greg Abel restarted Berkshire's buyback at 1.50 times book, above the ceiling Warren Buffett spent two decades refusing to cross, and the stock never fell to get there. Berkshire repurchased about $4.5 billion of itself in the second quarter, roughly $4.8 billion for the half, after buying none at all between May 2024 and March 2026. Cash fell from a record $397 billion to $365.5 billion, and the same quarter ended fourteen consecutive quarters of net equity selling with about $20 billion of net purchases. Price the disclosure rather than the size. Buffett deleted the 1.2 times book cap in July 2018 and replaced it with his own estimate of intrinsic value, which converted every repurchase into a published opinion. Shareholders' equity of $747.9 billion over 1.43 billion Class A equivalents puts book at $348.26 a B share against Friday's $521.80 close. That is 1.50 times, above a ten-year median of 1.42. The stock did not get cheap. The index ran past it, and Abel bought the relative discount his predecessor would not have called one. Tim Cook did the same thing in March 2012, ending Steve Jobs's no-payout policy with a $10 billion authorisation against $98 billion of idle cash, and that program has since retired more than 40 percent of Apple. Against all of it: $4.5 billion is 1.2 percent of the cash pile, which Berkshire's own operations refill inside a quarter.
The Bitcoin fork that failed last week needs roughly twenty-five years to reach its first difficulty adjustment, because the mechanism that would rescue it is indexed to the throughput that collapsed. BIP-110's mandatory signaling window closed at block 961,632 with fifty-one blocks in support, 2.53 percent against a 55 percent threshold. Supporters split onto a minority chain that mined two blocks and effectively stopped. Michael Saylor put the consequence in the only unit that matters: difficulty re-prices every 2,016 blocks, so the adjustment that would make a starving chain's blocks cheaper cannot arrive until it produces the blocks it can no longer produce. Effort does not close that gap. Only hashpower does, and hashpower is what left. Bitcoin Cash in 2017 is the control, because it forked with real hashpower and exchange listings on day one and reached nearly a quarter of Bitcoin's market capitalisation that November. The tell came before the vote and it was not a poll: Stephan Livera notes no fork futures or prediction market ever developed real volume, and a market that declines to quote you has already answered. This brief called the failure on July 13. The caution on our own call is that the fee-revenue problem BIP-110 was raised to solve is untouched, so the next attempt arrives with more capital, not less.
Steve Yegge's team is now committing code about 3.6 times faster than its own pipeline can verify it, which means continuous integration has stopped being a gate and become a queue. He published the figures on Saturday and the multiplication is ours: 175 real commits a day this month, peaking near 250, against a build gate that takes roughly thirty minutes, which demands 87.5 hours of verification per 24-hour day. His merge queue hit 166 deep, and the response was to abandon bisection and land megabatches of 120 to 150 commits, then diagnose the wreckage in parallel. The economics underneath are the part nobody else has. About $87,000 a month of API-equivalent token burn, roughly 69 billion tokens in July at 96 percent cache hits, delivered for about $2,800 out of pocket by rotating thirteen subscription seats, a thirty-fold discount to list. The counter is in his own numbers: this is one engineer with an unusual tolerance for chaos, and he says maintaining the rig eats a fifth to a quarter of all the work it does. The generalisable part is the pigeonhole arithmetic. Once commit rate outruns build slots, one commit per green build is not hard, it is impossible, and verification becomes the binding constraint on agent-written software.
In a blind test of 1,053 paying developers, only 13.6 percent refused a dangerous command that Claude Code's automatic permission mode would have blocked 89 percent of the time, and Anthropic makes that mode the default on Friday. Simon Willison published the numbers Saturday. Each session had one permission prompt quietly swapped for a clearly harmful command, and roughly six in seven humans approved it. A third-party evaluation by Trajectory Labs ran 72 held-out indirect prompt-injection scenarios, 720 attack attempts, and reported none succeeded against three Claude models in that mode as of mid-July. Willison carries both caveats and so should the reader: 89 percent leaves 11 percent unblocked, and he says plainly he wants independent confirmation of the 720. The finding that matters is not that the machine did well. It is that the human review requirement, the control every enterprise AI policy is built on, has now been measured at 13.6 percent effectiveness against the exact failure it exists to catch.
Saudi Arabia's Jazan refinery has been struck twice in thirteen days, which converts Amin Nasser's repair-speed claim from a talking point into a public experiment. Houthi forces claimed a drone attack on the 400,000 barrel-a-day facility on Sunday, hours after the kingdom said a fire there had been brought under control at dawn. Aramco had shut the same plant on July 27 after a strike damaged its gasification complex and tank farm, with restart targeted for mid-August. On last week's earnings call Nasser said Aramco can restore damaged facilities six times faster than the industry average and return to pre-conflict production within days. Nobody has been able to test that on a repeat strike until now. The squeeze is two-ocean: Bab al-Mandeb transits fell 24 percent the week the embargo landed, to 269 from 354 on Lloyd's List Intelligence numbers, and the damage concentrates in tankers, where mainstream transits dropped 42 percent to 53. Kpler puts Saudi crude moving south off the west coast at roughly 1.5 million barrels a day since the July 22 attacks, about half its pre-attack run rate near 3 million. The corridor is thinning rather than closing. The honest limits are real: no injuries, the fire out by dawn, the cause undisclosed. But the number that matters is the restart date, not the damage. Aramco has a published repair claim and a second strike inside two weeks turns that claim into the market's working estimate of how much spare Saudi export capacity actually exists. If mid-August slips, the buffer the entire crude curve is priced against is a talking point rather than a barrel.
There are 518 Russian shadow-fleet tankers the EU and the UK have sanctioned and the United States has not, and the measured efficacy gap between those two regimes is 86 percent against 50. Robin Brooks and Ben Harris published the count at Brookings on Thursday from the vessel database they maintain: the EU has designated 671 ships, the UK 621, the United States 216 and nothing new since the last administration ended, with 43 covered by all three. Brooks's earlier measurement is the number that makes this bind. Activity drops about 90 percent for ships sanctioned jointly by all three, 86 percent for American designations alone, and about 50 percent when only Europe and Britain act, because the fear of secondary sanctions is what actually moves a Greek charterer. Against it, and it is a real objection: pushing volume off dark tonnage may simply return it to compliant ships inside the price cap rather than reduce it. The Senate's sanctions bill goes to a House that reconvenes on August 31. The list is the cheaper instrument and it is sitting there.
A supermassive black hole has been found 30,000 light-years from its galaxy's centre, and there is no galaxy visibly around it. University of Maryland astronomers reported it in The Astrophysical Journal Letters on July 27, catching a dormant object roughly the mass of the Milky Way's own central black hole as it shredded a passing star 9.3 kiloparsecs out in the suburbs. It was found by an AI classifier turned loose on sky-survey data in August 2025 that hit within three months. Dormant black holes are invisible by definition, so the population we have counted is the population that happened to be disturbed. That is the shape worth keeping: a detector that only fires on disturbance will always undercount, and the undercount is not random.
Human language diversity peaked between three thousand and one thousand years ago, and the collapse began with states and empires, not with European colonialism. A team led from Pompeu Fabra in Barcelona, with Yale and UT San Antonio, reconstructed global language counts across the Holocene and published in Science on July 23. Roughly 4,500 to 6,200 languages existed 12,000 years ago; the peak reached tens of thousands; about 7,600 survive. The uncomfortable part is the timing. The collapse predates the era we habitually blame for it by thousands of years, which puts the agent inside the state rather than outside it: the same consolidation that builds an administration is what thins the languages underneath it. Whatever else empires were doing, the most effective monolingual policy in history was never written down as one.
LiDAR over southwest Amazonia found more than twenty thousand pre-Columbian earthworks, and the population implied is one and a quarter to three million people. The survey, published in Nature on July 29, flew 4,430 kilometres of laser transects and detected roughly five times more construction than satellite imagery finds, extrapolating to 23,760 earthworks across 183,000 square kilometres, about triple the previously assumed extent. The monuments run from 600 BC to AD 850, with the population peak around 100 to 300 AD. The consequence is not archaeological. The pristine-rainforest baseline that soil, biodiversity and some climate models are calibrated against is a measurement taken on an abandoned garden, and a control group stops being one the moment you learn somebody was already running the experiment.
Every connected product sold into Europe is about to carry a printed expiry date for its security, and the first binding deadline is five weeks away.
The EU's Cyber Resilience Act applies in full on 11 December 2027, which is exactly why almost nobody is watching 11 September 2026. On that date any manufacturer of a product with digital elements placed on the EU market must report an actively exploited vulnerability to ENISA and the relevant national CSIRT, with early warning inside 24 hours, full notification inside 72, and a final report within 14 days of a fix. It binds what is already in the field, not only what ships next.
The reporting clock is the visible half. The structural half is the support period: at least five years, or the product's expected use time if shorter, during which the manufacturer must handle vulnerabilities, and must state that period in the product information. A door lock or a building controller is sold on a twenty-year field life and booked as a single delivery. The Act does not touch the hardware. It shortens the declared economic life and staples a warranty-shaped tail to the sale, and nobody reserves against that tail, because nothing has ever required one.
The split is not who runs a product-security organisation, because all of them do. It is duration, and who pays for it. Schneider Electric (SU.PA) and Siemens (SIE.DE) sell into twenty-year industrial lifecycles and can price a long declared support period as a specification. Logitech (LOGI) already publishes a five-year end-of-support policy for its CollabOS devices, but the back half of it runs only under an active paid service plan, and a period the Act requires you to declare on the product is not one you can keep behind a subscription.
Watch: the Q3 and Q4 2026 calls of EU-exposed connected-hardware makers, for the first unprompted use of "support period" as a cost or a reserve rather than a compliance checkbox. If it lands in operating expense rather than capitalised programme spend, 2027 margin guidance across the complex is stale, and the bill-of-materials tooling the law mandates (JFrog, FROG) becomes a required line item.
Washington is withdrawing the rule that was going to force American medical sterilisation off ethylene oxide. The switch is happening anyway, and you can read it inside the one company that owns both sides.
Ethylene oxide sterilises roughly half of all sterile medical devices in the United States, about 20 billion units a year on the FDA's count, and across contract sterilisation facilities gamma irradiation takes about 40 percent and electron beam under 5. The EPA's 2024 emissions standard would have cut EtO emissions from commercial sterilisers by about 90 percent, and industry estimated the retrofits would strip 30 to 50 percent of national sterilisation capacity while facilities were offline. In March 2026 the EPA proposed reversing it, and a final rule is expected in the second half of this year.
The consensus read is that a reversal restores the status quo. It does not, because the migration was already bought. Validation packages, materials-compatibility testing and the FDA's radiation sterilisation master file pilot are sunk, and sunk does not reverse. A withdrawn rule leaves behind a permanently cheaper exit ramp off ethylene oxide for everyone who comes next.
The evidence sits inside Sotera Health (SHC), which reported on 6 August. Its Sterigenics segment runs contract sterilisation across EtO, gamma and e-beam, and its Nordion segment supplies the cobalt-60 that gamma consumes. Second-quarter Sterigenics revenue was $212 million, up 7.0 percent in constant currency; Nordion was $49 million, up 16.7. Across the first half that gap ran about fourteen points, inside one company, on one customer base. Cobalt revenue is lumpy, so a single print proves nothing, which is precisely why the third one matters.
Watch: Sotera Health's third-quarter report in early November 2026, read against the EPA's final rule. If Nordion's constant-currency growth holds more than ten points above Sterigenics' for a third consecutive quarter while the reversal is in hand, the shift has decoupled from the regulator and the binding constraint becomes cobalt-60 rather than an emissions standard. Steris (STE) wins on either modality. Sotera (SHC) is where the shift is legible, because its two segments sit on opposite sides of it.
Supply-Indexed Demand: when an index computes a company's weight from its free float, the float stops being only a count of shares available to sell and becomes an instruction to buy, so the moment insiders are freed to sell is the moment the index is required to bid.
On May 1, 2026, after a February consultation held with the SpaceX listing in prospect, Nasdaq replaced the Nasdaq-100's 10 percent minimum-float gate with a schedule: below 33⅓ percent free float, a company's index weight is computed on a fraction of its market capitalisation that rises proportionally with float, roughly three times float, rather than on the full number. SpaceX listed June 12 at $135, joined the index July 7 at roughly 1 percent weight, and QQQ bought about $4.3 billion of it by the prior close, funded by proportional trims of every existing constituent.
On August 6 the first lockup expired. Up to 911.5 million insider shares came free, lifting float from 4.9 percent to 11.8 percent of shares outstanding. Every outlet asked the same question, which was whether a billion shares would break the stock. It closed up 6.1 percent, and the verdict was that the overhang had been priced in.
That reads one side of the ledger. Under the multiplier, float going from 4.9 to 11.8 percent takes the market cap the index counts from roughly 15 percent of the company to roughly 35 percent, about a 2.4 times step in weight, arriving with no bid from anyone who wanted the stock. For any constituent inside the 33⅓ percent band, a lockup expiry is a supply event and a demand event with the same trigger and the same sign. The calendar is dense: roughly 319 million more shares as early as August 12, further tranches in September and October, more than four billion potentially free by year-end.
The projection. Nasdaq's own stated purpose for the multiplier is to add low-float names "in multiple tranches over time." Taken at face value, that makes float the schedule. The call: by the Nasdaq-100's December 2026 annual reconstitution, SPCX's index weight is at least double its July 7 entry weight, 2 percent or more, and rises in at least one month in which the share price falls. That second clause is the whole mechanism, isolated. Weight up while price down cannot come from the price, so it has to come from the float.
The reason a published formula does not simply get arbitraged flat is that preannouncing an order defuses it only by marking it informationless, which is what strips the adverse selection component out of the quoted spread. Publishing a size does not do that on its own. A market maker looking at a stock with tranches landing on some dozen dates between now and December cannot tell, on any given morning, whether the buyer in front of it is a formula or someone who knows something.
Where this breaks. The strongest objection is that the price impact simply is not there any more. Greenwood and Sammon (Journal of Finance, April 2025, 80(2): 657 to 698) find the abnormal return on an S&P 500 addition fell from an average 7.4 percent in the 1990s to 0.3 percent over the past decade, a collapse that happened while index-linked assets grew. Their mechanism is Admati and Pfleiderer's (Review of Financial Studies, 1991) sunshine trading: preannouncing an informationless order coordinates the supply of liquidity and cuts the announcer's cost. A weight step computed from a published formula is about as preannounced as an order can get. August 6 is itself evidence for this reading, because the most telegraphed supply event of the year produced a 6.1 percent gain. The objection that most threatens the call is arithmetic. Weight is the multiplier times market capitalisation, and SPCX has lost more than half its value from the June 16 intraday peak of $225.64. A 2.4 times multiplier against a market cap that has halved nets to roughly nothing, which is flat weight, no forced buying, and a test that fails for reasons unrelated to the mechanism. Dated falsification: if SPCX's Nasdaq-100 weight has not exceeded 1.5 percent by the December 2026 reconstitution despite float passing 20 percent, the multiplier is not applied on the cadence this argument assumes, and Supply-Indexed Demand is an accounting identity rather than a flow.
The reusable part is not about SpaceX. Any rule that computes an entitlement from a quantity a third party controls has handed that party the trigger. This builds on August 2's Optioned Demand, where a backlog number blended a firm order with a written option, a single published quantity carrying two opposite meanings with the market pricing only one of them. Float is now that number. It says how much can be sold and how much must be bought, and those are not the same claim.
"This Strife is wholesome for men. And potter is angry with potter, and craftsman with craftsman, and beggar is jealous of beggar, and minstrel of minstrel."
— Hesiod, Works and Days
You would assume envy toward a peer is the thing to suppress. Hesiod says the opposite. He had written that there was one Strife. There are two, and the second is worth having.
Take them separately, because they arrive wearing the same feeling. The first Strife wants the other person to have less. It is satisfied by their setback and requires nothing of you, which is why it is comfortable: the only ambition that can be fully realised while you sit still. The second wants your own work to be better. It is satisfied only by your effort, never for long, and it does not care what happens to the other person.
The test that separates them is in Hesiod's list, and it is easy to read past. Potter against potter. Craftsman against craftsman. Beggar against beggar. Every pair is matched, and nobody in the list is measuring themselves against a king. The useful Strife only fires against someone close enough that their work is legible to you, near enough in method and level that you can see what they actually did differently. Against a rival that specific, envy stops being a mood and becomes information.
Which produces an uncomfortable inventory. The person at your exact level doing your exact job, whose work you stopped reading about eight months ago, and whose name now comes out of your mouth only inside an explanation of why their approach would not work for you. You did not lose interest. You made the first Strife easier by removing the evidence.
Set that against Montaigne, a week ago. The unwitnessed hour, the thing you make and show nobody, builds the person. Hesiod's potter is the other half, and it builds the skill. You need both, and the one you are avoiding is this week's.
Today's practice: name the one peer at your level whose work you have been avoiding, spend twenty minutes actually inside it, and send them one sentence naming the specific thing they do better than you do it. If you looked and sent nothing, the practice did not happen.
The boats are going under. On the far bank of the Zhang River the cooking cauldrons are being smashed, one after another, by the army that owns them. Each soldier is left carrying three days of dry rations and nothing else.
This is 207 BC, the Chu army under Xiang Yu, as Sima Qian tells it. The popular reading is that this is a story about motivation: no way home, so fight harder. That reading is why the move keeps getting copied badly.
The real problem Xiang Yu faced is that a claim about your own future behaviour cannot be checked in advance. It costs nothing to make, so it gets discounted to nothing, and the other side stops listening to what you say and starts acting on what your situation makes profitable. Game theory calls the useless version cheap talk. There are exactly two ways out and both are purchases. You can commit, by destroying your own option to do the thing the other side is counting on, irreversibly and where it can be seen. Or you can signal, by paying a price now that only someone telling the truth could afford.
Now notice what he destroyed and what he kept. The cauldrons are the sharper move, because boats can be rebuilt but an army without cooking pots physically cannot conduct a siege, and that is a fact about its logistics rather than a claim about its resolve. He kept three days of food. He removed the option to wait. He did not remove the option to eat.
A second domain. A goldsmith in medieval London has the same problem and cannot solve it, because anything he says about his own metal is cheap talk and anything he could do to prove it, he could also fake. So England built the escape hatch. From a statute of 1300 the Guardians of the Craft went shop to shop to assay silver and gold and strike the leopard's head, and from 1478 the testing was pulled into Goldsmiths' Hall, which is where the word hallmark comes from. The test is destructive: a scraping is cut from the piece and melted in a bone-ash cup until the base metals burn away and only a bead of the real metal remains. The mark is credible for two reasons. Passing it costs metal, and the maker cannot apply it himself.
Mechanism. The fix is to convert the claim from a sentence into a fact about the world. Credibility is not a property of your character. It is a property of your situation, and situations can be changed.
Sizing. Too little and every promise you make is quietly discounted, and you pay that discount in worse terms from everyone, forever, without ever being told the price. Too much and you have bought belief with flexibility you will want back, in exactly the futures where you would have wanted it. The rule is on the riverbank. Commit only as far as the specific belief you need to move. Smash the cauldrons and keep the rations.
Failure mode. Four ways this model makes things worse than ignoring it. The act is invisible or quietly reversible, so you pay the whole cost and buy nothing, which is the resignation you can withdraw and the deadline only you know about. You commit against the wrong belief: they never doubted your resolve, they doubted your competence, and no burnt boat has ever fixed that. The world moves and the commitment still binds, which is not a malfunction but the price you agreed to. And the expensive one: you commit against something that has no beliefs. A disease, a season, a market, an impersonal process. There is nobody on the other side to convince, so the destroyed option is pure loss dressed as resolve.
The decision tool: price the credibility before you buy it. Three questions, in order, for any future situation. First, whose belief exactly? Name one person and write the single sentence about your future behaviour they currently do not believe. If you cannot write that sentence you do not have a credibility problem, and any commitment you make will cost you optionality for nothing. Second, can they see it and can I undo it? Two binary tests. Not observable to that person, not a signal. Quietly reversible, not a commitment. Anything failing either test is theatre you are paying for. Third, what is the smallest option I can destroy that makes the sentence true? Buy exactly that much and keep the rations. And if the honest price turns out to be higher than the belief is worth, stop trying to be believed and go find a hallmark: a third party whose test is expensive and whose mark you cannot apply to yourself.
This year researchers described a structure inside a cow that no one had ever seen. Reporting in Science on the back of 450 newly sequenced ciliate genomes, they named it the hydrogenobody: an organelle inside the single-celled ciliates that live in the rumen, wrapped in a single membrane rather than the double membrane of the hydrogenosomes it superficially resembles, which suggests it arrived by an entirely different evolutionary route. Its work is to pour out hydrogen, and living inside those same ciliate cells are methanogens, for whom hydrogen is food and methane is breath. So the gas the world files under "cattle" is manufactured by a microbe inside a microbe inside the animal, running on the output of a structure that had no name until 2026.
For a decade the effort to suppress that methane has aimed at the methanogen, with additives that block the enzyme performing the final step. The new picture casts the methanogen differently: not the engine, but the drain for a hydrogen supply that keeps arriving whether or not the drain is open. Close a converter and the substrate it was consuming does not vanish. It accumulates, and the pressure finds another route or another organism. Cut the supplier and the converter has nothing left to convert. The general form is uncomfortable, because it inverts what an investigation naturally turns up first: the thing visibly performing the act is the easiest to identify, the easiest to name, and usually the least durable place to intervene.
So here is the test. When a fix keeps needing to be reapplied, the same argument every quarter, the same incident every release, you targeted a converter, and before the next attempt you should be able to write one sentence naming what supplies the input to the behaviour you want stopped. If you cannot write that sentence, you are not buying a solution, you are buying a subscription, and the price is the repetition. The architecture is identical in an organisation, where the recurring meeting is the converter and the unresolved ambiguity upstream of it is the supplier that keeps refilling the room.