Wednesday did not have one story, and forcing it into one would flatter it. The de-escalation trade held its bid without extending it: volatility came out, the Dow made a second straight record, and crude round-tripped a Houthi missile strike on a Saudi tanker inside a single session. Read the rest as separate tracks. An eighth tanker hit since July 22 at the chokepoint nobody is watching. US services prices charged rising at their fastest pace in fourteen months. A Republican governor freezing data-center grid approvals in the state that calls itself the epicenter of AI. And the sharpest development of the day sitting outside all of it, in four names that walked out of Google. Watch Alphabet's next disclosure of cloud backlog and related-party arrangements against the reported Mechanize agreement. If the money keeps flowing toward the door, the four percent the market took off Alphabet was priced on the wrong event.
Korea's Kospi fell about 4.6 percent in Thursday's session and the exchange tripped its sidecar, the automatic brake on programmatic sell orders. In that same Thursday session SK Hynix closed down about 10 percent and Samsung Electronics about 6, on renewed doubt about who pays for AI capacity. This is a new move rather than the one reported earlier in the week: Korea closed down 5 percent on Monday August 3 and rallied about 4 percent midweek, and neither of those sessions tripped a brake. Thursday's did.
Asia: Nikkei off more than a percent to 65,538, Hang Seng down 1.8 percent, Shanghai flat. Brent held near $79 through the Asian session, which is the same indifference the Red Sea strike got in Geopolitics below.
Asia's chip complex is trading the AI capital question from the other end of AI & Tech below. Stephen Innes of SPI Asset Management reads the selloff as profit-taking and risk reduction ahead of Friday's payrolls rather than a verdict.
Crypto data provided by CoinGecko
Two services surveys landed Wednesday morning pointing in opposite directions, and only one of them publishes the series that matters. S&P Global's US Services PMI came in at 54.6 against 51.2 in June and a 53.6 flash, a nine-month high. ISM's separate services gauge printed 54.1 against 54.5 expected, a miss. Read as activity, they cancel, and activity surveys disagree constantly. But S&P Global's release carried a line the wire tweets dropped: tariff- and energy-related costs pushed input inflation to its highest since May 2025, driving the sharpest rise in prices charged for services in fourteen months. Prices-charged series are what a central bank actually acts on, and this one is running at that high in the sector that is two-thirds of the economy. Neel Kashkari, one of three regional presidents who dissented at the July 29 meeting for a quarter-point rise, told CNBC the same day that the increase could begin as early as September. He now has the number he was voting without.
American gross investment has not moved in seventy years, and the part that adds to the capital stock has halved. Eric Basmajian's NIPA work puts gross investment at about 18 percent of GDP this decade against 16.5 percent in the 1950s, while net investment fell from 8 percent to 3.8. The gap is depreciation, which went from 8 percent of GDP to roughly 14. Of every gross investment dollar, 48 cents added to the capital stock in the 1950s and 21 cents does today, so maintenance went from half the bill to nearly four-fifths. The mix explains it: intellectual property products went from 1.3 percent of GDP in 1960 to 5.5 percent while nonresidential structures fell from 3.6 to 3.2, and a dollar of prepackaged software is mostly gone in three years against a working lifetime for a house. Jerusalem Demsas found the same shift in payrolls Wednesday: more Americans now work for remodeling companies than for single-family and multifamily developers combined. Record investment and a shrinking capital base are not a contradiction when the assets keep getting shorter.
Bending Spoons is paying about $2.25 billion for Airtable and roughly $965 million of it is Airtable's own bank account. The release gives both numbers and the subtraction is theirs: a $1.285 billion enterprise value, and an equity value of about $2.25 billion arrived at by adding "Airtable's current net cash-and-cash-equivalents balance." So the operating business, 500,000 organizations and about $480 million of annual recurring revenue still growing more than 20 percent, is priced at $1.285 billion, or 2.7 times recurring revenue. Airtable raised more than $1.4 billion across nine rounds, so the business is changing hands for less than the money that went into it, and two-fifths of the price is that money coming home. Against the $11.7 billion mark of December 2021, $2.25 billion is nineteen cents on the dollar. Bending Spoons has done this before with Evernote, AOL, Vimeo and WeTransfer. A late-stage venture round is not a valuation. It is an option on one, and the strike expires.
Six Ethereum researchers proposed on Tuesday to burn validator rewards toward zero as staking rises, and the number that makes it urgent is not in the proposal. Ethereum's staking ratio hit an all-time high of 34.4 percent this week. Its DEX spot volume, per CryptoRank, fell to $29 billion in July, down 76 percent from an August 2025 peak of $122 billion. Record capital securing less than a quarter of the volume it secured a year ago, in a token that opened Wednesday at $1,868.36, roughly half where it stood then. EIP-8361 puts permanent consensus yield at 1.2 percent against about 2.6 percent now, which against 41.4 million staked ether removes on the order of 580,000 ether of annual issuance and hands it to holders who staked nothing. The researcher Izzy makes the objection that lands: at zero yield the expensive decentralized operators are priced out first and cheap centralized ones inherit the set, so the moral hazard is not removed, only moved somewhere less visible. EIP-1559 burned user fees five years ago Wednesday and nobody objected. Nobody had booked those fees as income.
McDonald's US comparable sales rose 0.8 percent against 2.5 percent a year earlier, and every point of it came from check and mix while guest counts went negative. That is the shape of a discount that did not work: the same customers trading down, and fewer of them. July US comps have already turned slightly negative. Chris Kempczinski called swapping the buy-one-add-one-for-$1 program for the Under $3 menu "a bad trade," and changed the US president the same morning, elevating Skye Anderson after twenty-six years inside. McDonald's has been here. Campaign 55 launched in April 1997, dropping a featured sandwich to 55 cents each month to mark the chain's founding year; sales ran as much as 6 percent below the prior year, the company killed it in June rather than run the planned full year, and in July the US president, Ed Rensi, was replaced by Jack Greenberg. The operating detail underneath is the real one, and it is the part 1997 did not have: only 60 to 65 percent of the system had implemented the Under $3 menu as designed, so roughly a third of the restaurants were not running the national offer. A 170-basis-point comp deceleration against a system doing more than $139 billion in global systemwide sales is real money, and it lands on franchisees just as the ten-year remodel cycle reopens and beef costs climb. An execution problem has an owner. An infrastructure problem has a bill and a date.
Alphabet fell about 4 percent on Wednesday because four people quit, and the detail nobody has priced is that Google is funding where they went. Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le left to found Discovery Loop, a public benefit corporation building model stacks to automate scientific experimentation. Dean was Google's thirtieth employee and drove the Tensor Processing Unit. Demis Hassabis moves to Alphabet chief scientist while Koray Kavukcuoglu takes day-to-day control of DeepMind, with the flagship Gemini model still unshipped against an expected June launch. Now the part that inverts the story: Google invested in Discovery Loop and signed a cloud partnership with it, and is separately negotiating a deal worth more than $1.5 billion with the coding startup Mechanize to license technology and bring in people. Google is paying on both sides of the same door. That is not a talent drain, it is a company converting employees into suppliers and customers, and the market marked it down about four percent for the privilege.
Washington drew its frontier-AI perimeter around the license rather than the capability, and a Shanghai lab spent this week showing what that costs. Per Axios, the framework produced under Executive Order 14409, signed June 2 with a sixty-day clock that ran out August 1, defines a covered frontier model as closed-source with state-of-the-art capabilities and national security risks, and defines neither term. Covered models are asked to submit to a pre-release government review of up to thirty days, with high-security storage and access logs. Participation is voluntary. Open models are excluded, and the framework says explicitly that nothing in it restricts them once released. The framework itself will not be published. Two identically capable models therefore receive opposite treatment based purely on distribution choice, which converts open-weight release into a thirty-day regulatory arbitrage on every launch. The mirror arrived the same week: MiniMax released an open-weights model whose license bars use or deployment by anyone in the United States, the EU, the UK or South Korea. Washington treated open weights as the category too free to govern. A Shanghai lab just used one as a border.
The first hard gate on the AI buildout arrived this week, and it was a Republican governor rather than an environmental lawsuit. Greg Abbott ordered the Public Utility Commission of Texas and ERCOT to stop approving new data-center grid connections pending an audit of every applicant, and ERCOT said in its August 3 market notice that it will not issue batch-zero large-load classifications by the scheduled August 7 deadline and will ask the Public Utility Commission for a timeline exception at its August 20 open meeting. Read the disclosure list, because it is the tell: tax breaks received, power use and generation, water use and cooling, community-impact mitigation, facility ownership. That is a transparency demand, not a capacity demand. The scale makes it binding rather than symbolic. ERCOT's interconnection queue holds more than 1,800 projects totaling over 474 gigawatts, about ninety percent of it data centers. Abbott called Texas the epicenter of AI, which is exactly why the freeze matters. Size the audit against what it can actually stop: those requests run against an all-time peak demand roughly one-fifth their size, which makes the queue an option book rather than a forecast of demand, and most of those options were free to write. An audit that asks who received tax breaks and how much water a site cools with is the first thing that has ever charged for one. The thesis breaks if the commission clears the backlog inside a quarter with the ownership, water and subsidy fields quietly dropped, so watch the August 20 commission meeting for whether the batch-zero timeline restarts with that disclosure list intact. The constraint on this buildout is arriving as a legitimacy problem years before it arrives as a physics problem, and legitimacy has no queue position.
A second chokepoint escalated to a sunk vessel on Wednesday and the oil market treated it as a single session's news. UKMTO issued Warning Incident 105-26: a vessel nine nautical miles southwest of Al Mukha, Yemen, attacked by an uncrewed surface vessel, crew rescued, ship reported sunk. Separately, Houthi spokesman Yahya Saree said the Saudi product tanker Wafa was struck by several ballistic missiles off Yanbu, the eighth tanker targeted since the movement declared a maritime blockade of Saudi Arabia on July 22. Saudi authorities have not confirmed. Saree gave the mechanism on the record, and it is the part that should worry a shipowner: the strikes followed the kingdom diverting ships away from the Bab el-Mandeb chokepoint. Brent went back above $80 on the news and gave it up the same afternoon as Hormuz optimism reasserted. Rerouting is not risk reduction when the adversary re-targets the route you rerouted to.
The replenishment schedule for American air defense is a public document, and it runs years past any ceasefire signed this month. CNN reported Wednesday, sourced to officials, that the US military has expended nearly 80 percent of its interceptors for one key missile-defense system. CSIS's count across the war puts roughly two-thirds of Patriot interceptors depleted. Now the arithmetic, which nobody prices: FY27 budget requests for Tomahawk, THAAD and Patriot surge by as much as 927 percent, and Patriots ordered in FY27 are unlikely to arrive for another three and a half years. The shortage is the headline. Its publication date is the finding. A request that size tells a counterparty exactly how long the constraint lasts, which converts a magazine problem into a schedule the other side can wait out, and it puts the refill revenue at the interceptor primes, Lockheed Martin and RTX, on a delivery clock measured in years, not negotiating sessions. A shortage you can conceal is leverage. A shortage with a delivery date attached is a calendar, and both sides are reading it.
A two-centimetre animal that died 518 million years ago just turned out to be carrying the first draft of a spider's fangs. Researchers from Yunnan University and the University of Leicester reported in Nature on July 1 that X-ray tomography of Urokodia, from the Chengjiang fossil beds in southern China, revealed soft anatomy mummified inside the rock, including a pair of pincer-like appendages behind its stalked eyes. Those are chelicerae, the defining hardware of the group that became spiders, scorpions and ticks, now more than 100,000 described species. The paper landed on the forty-second anniversary of the site's discovery, a quiet argument for going back to old rocks with new instruments.
Mathematicians have proved that a perfectly fair election is impossible, and they mean it in a specific and unusually useful sense. Sebastian Holdum of Copenhagen and Frederik Ravn Klausen of Cambridge published an impossibility theorem in Annals of Operations Research on August 2: once enough parties compete, no system can simultaneously guarantee that every local winner takes a seat, that national seat shares match national vote shares, and that the parliament stays a fixed size. One of the three has to give. Germany let its parliament swell from 598 seats to 736 in 2021 to hold the other two, then in 2023 gave up the guarantee that local winners get in. Britain gives up proportionality: Labour took 411 of 650 seats on 33.7 percent of the vote in 2024 while Reform UK's 14.3 percent bought five. As Klausen puts it, it is not corruption or conspiracy, it is simply mathematics.
The Shroud of Turin has now been sequenced, and what came back is a guest list rather than an answer. An international team led from the University of Lancashire reported in Scientific Reports on August 4 the first PCR-free metagenomic sequencing of the official 1978 samples. They found multiple human mitochondrial lineages, skin microbes, fungi, salt-adapted archaea, and DNA from wheat, maize, bananas, peanuts, cattle, dogs, cats and Mediterranean red coral. The authors are explicit that none of this dates the cloth or settles its authenticity. Maize and peanuts are New World crops, which tells you the linen has been in rooms nobody recorded. The object stopped being evidence about one night a long time ago. It is now an archive of everyone who ever picked it up.
Every published elevation in America is about to be restated while the ground does not move an inch, and the deadline that binds is a survey-submission cutoff in January 2027 almost nobody outside surveying has on a calendar.
NAVD 88, the datum every US elevation is quoted against, was finished in 1988 and built by walking spirit levels between physical bench marks. The National Geodetic Survey is replacing it with NAPGD2022, defined instead by a gravimetric geoid from the GRAV-D airborne survey. Federal Register notice 89 FR 81891 sets the sequence: a Federal Geodetic Control Subcommittee approval vote in mid-2026, then a months-long transition. What makes this economic rather than technical is that heights change. NGS's 2010 floodplain pilot with FEMA in North Carolina found the new surface differing from NAVD 88 by roughly half a foot in the state's southeast and more than a foot in the north and west. Nothing physical moved. The record did. Flood Insurance Rate Maps, elevation certificates and highway design surfaces are all written in the retiring datum and each converts on its own budget, so the danger is a mismatch window: a survey delivered in NAPGD2022 and read against a map still in NAVD 88 is wrong by the geoid difference, and neither document says so. The forcing function is statutory: the Geospatial Data Act of 2018 gives anyone spending federal geospatial money five years from adoption to move, while several states must first amend their datum laws. Expect an unbudgeted reconciliation spend booked as revenue by the conversion economy, Trimble (TRMB), Hexagon (HEXA-B.ST), NV5 Global (NVEE), Bowman Consulting (BWMN), and as cost by parties mostly not listed: state DOTs, FEMA-mapped municipalities, utilities. No clean short exists. Watch: the Federal Register for NGS's post-approval notice, against NGS's own January 13, 2027 cutoff for submitting survey projects into the current system. If that publishes while the window closes and most states are still unamended, the yardstick is changing faster than the records written against it.
Washington just cut the price of American farm labor by rewriting one formula, and the first casualty is not a worker. It is the payback math under every harvesting robot sold in the last five years.
The Adverse Effect Wage Rate is the wage floor for hiring H-2A guest workers, and for decades it came from USDA's Farm Labor Survey, one high regional number that rose almost every year. On October 2, 2025 the Labor Department replaced it by interim final rule at 90 FR 47914, keying the rate to the Bureau of Labor Statistics' Occupational Employment and Wage Statistics survey and splitting it into two skill tiers. Most H-2A jobs fall into Skill Level I, set at the 17th wage percentile for those occupations. Michigan State University Extension puts the saving for that state's H-2A employers at $24 million to $53 million, 11 to 24 percent against the 2025 rate of $18.15 an hour. It survived its first test on May 14, 2026, when the Eastern District of California denied the United Farm Workers a preliminary injunction, finding the claimed harm speculative. There is still no merits ruling and no final rule. The interim rule is simply operating. The second-order effect no desk has assembled: every business case for agricultural automation is a spread against that floor. A half-million-dollar harvester is underwritten on labor at $18 and climbing; pin the floor to a 17th percentile that no longer tracks a rising survey and the payback stretches past the machine's useful life. Cheap labor is the oldest killer of mechanization, and this cut came by decree. If the rule survives merits review while certifications keep growing, expect margin relief at labor-intensive US growers, Limoneira (LMNR), Mission Produce (AVO) and Calavo (CVGW), and a stalling order book at the vendors whose pitch is labor substitution: Deere (DE), AGCO (AGCO), and harvest-robotics firms, nearly all private. Watch: the Labor Department's quarterly H-2A disclosure file and the United Farm Workers v. DOL docket. If certified positions keep rising through fiscal 2026 while average wage offers flatten or fall, the substitution is running backwards, and the automation thesis reprices without a product announcement.
A positive-skew return distribution is investable only where the population producing it is finite and expensive to join. Remove the entry gate and the identical statistic reverses its instruction.
Blockworks Research has published an unusually clean count of liquid-token outcomes: of roughly two thousand tokens that have crossed $50 million in circulating market cap since January 2020, 4.1 percent beat bitcoin, and the median lost 97 percent from the day it qualified. The comparison its own author reached for is the right one: Bessembinder's study of 25,967 US stocks from 1926 to 2016 found 4.3 percent of them account for the entire net gain of the American stock market over Treasury bills.
Two markets, the same number, produced by the same mechanism: positive skewness in the cross-section of returns. The aggregate gain comes from a handful of extreme right-tail outcomes compounding, not from the typical member, so mean and median point in opposite directions. The error is assuming it carries the same instruction.
It does not, and the reason has nothing to do with crypto. Bessembinder's result became the argument for indexing because in equities own the whole population is a purchasable instruction: the population is enumerable at any instant, and joining it costs millions of dollars and years of audited operating history. That gate does the work, and it appears nowhere in the statistic. A token enters the same distribution at a $50 million market cap that can be manufactured in an afternoon, and the tail refills faster than anyone can buy it: the hundredth-largest entered at $47 million in June 2020 and roughly $900 million by December 2024. Skew you cannot own in full is not diversification. It is a subscription to dilution.
So the prescription inverts. Where entry is free, the answer to a concentration statistic is not breadth but a selection test on the one property free issuance cannot fake: cash returned. In Blockworks' own cut, exchange tokens beat bitcoin at a 26 percent hit rate against roughly 1.2 percent for everything else, and the lone decentralized-exchange token in that winner set shares their one feature: fee revenue routed to recurring buybacks.
This is the mirror of August 1's Effective-Coverage Collapse: that argued the numerator was failing because searchers stopped being independent; this is the denominator, minted without limit.
The call: through June 30, 2027, that 4.1 percent share does not double. It stays below 6 percent even if bitcoin sets a new all-time high inside the window.
Where this breaks. The objection that damages it most is that the gate sits inside the control sample. Bessembinder's data opens in July 1926, before the Securities Act, before the SEC existed, when a US listing needed an underwriter and a share register but no federal registration at all. Near-free entry is in the dataset that produced the 4.3 percent, and indexing still won. That demotes the claim from "different regime" to "same regime, extreme parameter," and it should be demoted: the honest version is that creating a listed company in 1926 still cost months and thousands of dollars, so the gate survives in degree rather than in kind.
Second, the minus-97 percent median has an endogeneity problem. The observation window opens at qualification, an event the issuer chooses and times, so the distribution is conditioned on the moment of maximum narrative rather than drawn from an exogenous calendar. That mechanically manufactures a terrible middle, and the statistic this argument leans on inherits the bias. Measured from fixed dates it improves materially.
Third, and worst for the thesis: if a single publicly visible feature lifts the hit rate more than twentyfold, the universe is trivially screenable, and "the index is unownable" is the wrong conclusion. The right one would be the far duller "the index is unownable and the filter is one line of code."
Falsification, dated: if the share beating bitcoin clears 6 percent by June 30, 2027, or if a quarterly-rebalanced capitalization-weighted alt basket beats bitcoin over that window, the enumerability argument is wrong. The winners were mechanically capturable without picking, and the gate was never the variable.
"The normative method for distinguishing right from wrong in historical information on the grounds of (inherent) possibility or absurdity is to investigate human social organization, which is identical with civilization."
— Ibn Khaldun, The Muqaddimah (1377), trans. Franz Rosenthal
Ibn Khaldun wrote that in a fortress in what is now Algeria, in the introduction to a history he never finished. The problem he set himself is the one you had this morning: most of what you are told is false, and you must decide which parts before lunch.
You would assume the way to do that is to vet the source: who reported this, who told them, have they been reliable. He did not throw that test away. He said it is the weaker one, and leaning on it is how careful people believe impossible things, because an honest man can hand you a number that could not have happened.
The stronger test asks whether the reported thing is consistent with how that kind of thing works. Al-Masudi put the Israelites in the wilderness at six hundred thousand fighting men, and Ibn Khaldun refused it, not because he doubted al-Masudi but because the land could not have fed a force that size and the terrain could not have let one march. Nobody in that chain had lied. Everybody in it had been checked.
Notice how much of your verification is chain-checking. Who sent it. Where it ran. Whether the account has been right before. That tests the transmitter, and its appeal is exact: you can run it without knowing anything about the subject. The other charges admission. You need a model of the thing before you can say whether a claim about it is possible, which is why what gets past you comes from the fields where you have none.
Take one claim you accepted this week because of who said it, and give it ten minutes of the other test. Work out what it implies about a size, a rate or a timeline, write that implied number down before you look anything up, then check it. If it survives, you have converted a borrowed belief into one you own. If it does not, the correction is the small prize. The bigger one is the name of the person you would still have believed.
In the eleventh century a merchant in Fustat, in what is now Old Cairo, handed a cargo to an agent in Sicily and simply waited. No court had jurisdiction over both ends. Avner Greif reconstructed how it held from the Cairo Geniza documents, and the answer was not trust in any soft sense. These traders, whom Greif calls the Maghribi coalition, ran a rule with teeth: an agent who cheated any member would never again be employed by any of them. Cheating paid once. Honesty paid indefinitely.
What did the work was two numbers, not goodwill: the probability that two parties meet again, and the speed at which what one does today becomes known to everyone who might deal with him tomorrow. Where both are high, cooperation is the profit-maximizing play and no virtue is required. Where either collapses, defection is, and no amount of virtue reliably survives it.
The same two numbers govern a reef. Cleaner wrasse eat parasites off larger fish that could easily swallow them, and Redouan Bshary found they would rather eat the client's mucus, which is more nutritious and hurts the client, so honest cleaning costs the cleaner. Clients enforce it: they chase cheaters and switch stations. In a 2006 Nature paper with Alexandra Grutter, Bshary showed cleaners behave measurably better when another client is watching. A fish is running reputation management. That is reciprocity, and it needed no brain that understands ethics. It needed an audience.
Sizing follows. The more repeated and observable the game, the more weight reciprocity carries and the less you need written enforcement. One-shot and unobserved it carries nothing, and relying on it is not generosity but negligence. Most situations sit between, so ask where on that line you are, not whether the other party seems decent.
The failure mode comes in two kinds. The first is noise: strict reciprocity in a garbled channel turns one misread into an unending feud, because each side is correctly retaliating against what it thinks it saw. Every long grudge between two reasonable people has this structure. The second is what happened to the Maghribis. Their enforcement worked beautifully inside the group and could not extend outside it, so as trade grew past the coalition they had no way to deal with strangers. Greif's comparison is the Genoese, who built slower, colder legal enforcement that worked with anyone.
The decision tool: before you rely on reciprocity, price it with three questions. How many more times will we interact, honestly counted, not hoped? Who else finds out what happens here, and how fast? And what does defection pay once, against cooperation paid out over the interactions actually remaining? If the answers are few, nobody and a lot, you do not have a relationship but a single transaction, and you should write it down and get it signed. If they are many, everyone and not much, stop negotiating and start building. The question is never whether someone is trustworthy; it is whether their situation makes trustworthiness pay. And then the sting the Maghribis found: the conditions that make trust work, few players and everyone watching, are the conditions that keep a network small. Any group good enough to run on reciprocity alone has, for exactly that reason, already chosen a ceiling it cannot see.
Show a child and a chimpanzee the same puzzle box, opened by a demonstrator who taps the top three times with a stick and then slides a bolt. The taps do nothing. Victoria Horner and Andrew Whiten reported in Animal Cognition in 2005 that when the box is opaque, both copy the whole sequence. Make the box transparent so the uselessness of the taps is visible, and the chimpanzees drop them and go straight to the bolt while the children keep tapping. Derek Lyons and colleagues could not get the children to stop in PNAS in 2007: they kept the useless taps under time pressure, and after being told some of what they saw was unnecessary. The behaviour has a name, overimitation. It shows up across cultures, from Western labs to the Aka and Ngandu of the Congo Basin, and a PLOS ONE study found adults doing it without knowing they were in an experiment. The animal that prunes correctly is the chimpanzee. The animal with cumulative culture is us.
Which suggests the filter we admire is the one that ends inheritance. Nearly every technique a person receives is cognitively opaque: whoever hands it over cannot say which steps carry the load, and the receiver cannot either, because the cost of dropping a live step arrives late, or rarely, or in somebody else's ledger. A copier that keeps only what it can see the reason for can never acquire anything it did not already understand, and its toolkit stops growing at the edge of its own comprehension. Indiscriminate copying looks like superstition and is in fact the mechanism that lets a practice carry more knowledge than any of its carriers hold. Strip the ritual and keep the substance is the chimpanzee's strategy, and the chimpanzee's toolkit does not accumulate.
So take one inherited process you did not design and mark each step JUSTIFIED or UNEXPLAINED, then resist cutting the unexplained ones, because inability to defend a step is the exact condition under which faithful copying is correct policy. Promote one unexplained step to a test instead: drop it for a bounded run, write down in advance what failure would look like and how long it takes to surface, and keep the step if you cannot see the mechanism and cannot afford the observation window. Watch, too, who prunes fastest. On an opaque process the quickest pruner is the least reliable transmitter. The same architecture runs wherever competence outruns explanation: the configuration line no engineer can justify that turns out to be load-bearing, the contract clause nobody remembers negotiating. Understanding is a filter, and a filter passes only what it already recognises.