The institutions that could act yesterday acted loudly and on the same day. Central Command began striking Revolutionary Guard targets inside Iran at noon Eastern, the second strike in three days. A sitting Federal Reserve governor conditioned a September hike on the inflation data. And yet not one of the prices that actually moved was set by any of them. Seven sovereign bond markets repriced together with no identified seller, six of them to multi-year highs, four fifths of the data-center power American developers newly asked one utility for quietly failed to become a contract anybody signed, and twenty-one banks announced a dollar token whose defining feature is a rate it is forbidden to pay. The decisive party in each case did nothing, and an absence generates no data, which is why the institutions with the loudest instruments keep aiming them at the variables they can see. Watch the gap between French and German ten-year yields: it is the cleanest live reading of whether the missing buyer is a Japanese story or a European one.
Crypto data provided by CoinGecko
Six sovereign bond markets hit multi-year highs in one session, the American ten-year went with them, and the best-argued explanation on the tape involves nobody selling anything. Japan's ten-year yield crossed 3.00 percent for the first time since 1996, and the UK ten-year gilt rose seven basis points to 5.223 percent, its highest since June 2008. Reuters put seven rates strategists on the record within hours and they split on the mechanism. Vasu Menon of OCBC gave the consensus version: higher domestic yields pull Japanese money home out of foreign bonds. Masahiko Loo of State Street gave the harder one. Japan, he argues, is not repatriating so much as gradually ceasing to be the marginal buyer of foreign bonds at all, which makes this a buyers' strike rather than a sellers' panic. We think Loo has it, and the reason is uncomfortable. His version requires no Japanese selling whatsoever, only Japanese absence. A seller shows up in the flow data. Somebody who simply stops arriving at the auction does not show up anywhere.
France borrows more expensively than Italy, and the interesting number is not that one, it is the distance to Germany. At 10:45 Central European Time on Tuesday the French ten-year traded above 4.215 percent, its highest since November 2008; the Italian ten-year sat slightly lower at 4.188 percent; the German ten-year Bund was above 3.36 percent, a fifteen-year high of its own. Euronews, which reported all three, also says French borrowing costs have exceeded Italy's for much of the summer, so treat the ranking as background rather than news. Those two levels, on a day both made multi-year highs, are the gap worth watching. The stated cause is not solvency. Analysts told Euronews the Bund has moved broadly in line with global benchmarks while France carries an additional risk premium for its political and fiscal outlook. The debt stock does not obviously settle it either. The IMF has French gross government debt at 118.4 percent of GDP this year and 120.5 percent in 2027, third highest in the EU behind Greece and Italy, and the Banque de France expects a 5.2 percent deficit this year. Robert Timper of BCA put it plainly, saying a large political majority is necessary for reform or a bond market riot will force it. The market has stopped pricing how much France owes and started pricing whether anyone in Paris can pass a budget.
The August ISM price index did not move at all, and the version of that number most people read today is false. The Institute for Supply Management reported Prices at 71.1 in August, exactly its July reading, a twenty-third consecutive month of increase and a change of zero. FXStreet's preview and the search summaries say it fell to 71.1 from 73, and any read calling August an easing is wrong on the release's own words. This matters because of who is watching it. Governor Michael Barr said on Tuesday that if inflation appears not to be moderating sufficiently, the Fed should act decisively to raise rates. That is a sitting FOMC voter conditioning a hike, not a cut, on precisely this evidence. What the panel said is worse than the index: pricing volatility drove 57 percent of the negative comments against 29 percent for tariffs, new orders fell three points, and the only subindex that improved was supplier deliveries, which improves when the supply chain slows down.
DNO agreed to buy Capricorn Energy for about $396 million, and it is simultaneously trying to buy Genel Energy, the rival it just beat to Capricorn. DNO's inside-information release on 1 September puts the recommended cash offer for Capricorn at $5.214 a share, a 45 percent premium to the undisturbed price. Read the bidder list rather than the premium. On 7 August DNO approached Genel at 69p a share, valuing Genel near £202 million, and Genel refused. Genel then bid for Capricorn itself and lost to DNO by roughly $36 million. So three Kurdistan-exposed oil producers spent four weeks bidding for each other. Capricorn's own shareholders killed an all-share Tullow merger in September 2022, then killed the NewMed deal that replaced it. Four years on they took cash. Where the barrels are real and the export route is political, the scarce asset is consent, and the only way to buy consent is to buy the people who can withhold it.
Twenty-one banks are launching a dollar stablecoin that the law forbids them to pay a cent of interest on. Bank of America, Citi, Goldman Sachs and Wells Fargo are among twenty-one institutions that said on 1 September they will form a company this half to issue a dollar token in the first half of 2027, built for the GENIUS Act and MiCA. Start with what the statute forbids. GENIUS bars a payment stablecoin issuer from paying holders any interest, and the OCC has proposed extending that to affiliates and third parties. A token that cannot legally pay a rate cannot take balances from the deposit beside it, which makes this an option written against deposit flight, priced like insurance rather than like a business. The honest counterweight is that the ban is why issuing is profitable at all, since the issuer keeps the reserve income. Against a market Token Terminal counts near $300 billion, the bank-owned Zelle network moved more than $1.2 trillion in 2025 by being the default button. Placement will decide this, not float. The number to watch is how many of the twenty-one wire the token in as their own default.
Ethena is paying a six percent dollar savings rate because its dollar is not legally a stablecoin. EthenaPay launched on iOS on 1 September, live rather than waitlisted, with what Ethena says is 5 percent cashback and a 6 percent dollar savings rate. Read it against the statute rather than the rate. USDe is a synthetic dollar backed by hedged basis positions and is not a GENIUS payment stablecoin, so the interest prohibition above does not reach it. The product lives entirely in that gap, and it advances Ethena's own position: the 28 August governance vote turned on a circulation trigger only distribution can reach. The arithmetic is the risk. Cashback at five percent funded by a six percent engine leaves almost nothing when the basis compresses, and a funding rate is not a contract. BlockFi paid up to 8.6 percent on crypto dollars, settled with the SEC and 32 states for $100 million in February 2022, and was bankrupt by November. On consumer yield the regulatory clock has always run faster than the credit clock.
Korean chip exports rose 209 percent in August, and Korea's own trade ministry says the reason is price. The Ministry of Trade, Industry and Resources reported semiconductor exports of $46.65 billion in August, an all-time high, taking chips to 47.5 percent of everything South Korea sells abroad. Total exports rose 68.7 percent to $98.25 billion and the trade surplus reached $34.75 billion. The ministry's own explanation is the part worth keeping: it attributes the milestone to the continuous expansion of AI infrastructure investment by hyperscalers including Google and Amazon, which led to price hikes in memory chips. That sentence changes what the number is. The cleanest publicly available thermometer for global AI capex is substantially reading a memory price rather than a memory quantity, which means the same figure would collapse on a price break without a single order being cancelled. A thermometer that reports what buyers paid rather than what they took delivery of is measuring the seller's leverage, and leverage is the thing that goes first.
About four fifths of the data-center power developers newly asked one utility for never became a contract anyone signed. Writing in War on the Rocks on 1 September, Javaid Iqbal Sofi reads a funnel the utility published itself. Developers requested more than 30,000 megawatts from AEP Ohio. After paid load studies that fell to roughly 13,000. Signed binding tariff commitments came to 5,642 megawatts as of 12 February, an attrition of about 81 percent. Read the cohort before you read the ratio, because AEP Ohio filed those same figures with Ohio regulators under the headline that its tariff is working: the 5,642 megawatts are new commitments under the tariff, on top of 12,219 megawatts already contracted before it took effect. The attrition is a fact about new requests, not about the utility's book, and new requests are exactly what capacity forecasts are built from. The gap is already priced: PJM's 2027 to 2028 capacity auction cleared $16.4 billion, and the market monitor's counterfactual removing data-center load from the peak forecast comes out $6.5 billion lower, of which $6.2 billion is forecast load rather than operating load. Consumers are paying today for megawatts that failed a credit check.
SK hynix broke ground on an American memory plant that will not make a chip until the second half of 2029. SEMIVISION's account of the 27 August groundbreaking at West Lafayette, Indiana puts the investment above $4 billion, the cleanroom opening in October 2028 and mass production of next-generation high-bandwidth memory in the second half of 2029, a date Semi Doped independently confirms. The detail that matters is what does not move: leading-edge wafers will still be fabricated in Korea and shipped to Indiana for packaging and test, which is selective localization of a few strategically legible steps rather than reshoring. Nothing announced this year relieves the 2026 and 2027 memory shortage. Semi Doped's Vikram Sekar argues the base die is migrating to logic designers and can cost three to four times the stacked memory sitting on it, which pushes the memory itself back toward being a commodity. The scarce thing is quietly moving out of the part everyone is building.
American forces began striking Revolutionary Guard targets inside Iran at noon Eastern, the second strike in three days, and the tempo is the news rather than the target list. Two tankers, Bahri's VLCC Sidr and the Senegal Prosperity, were hit by projectiles minutes apart near Khasab on outbound Hormuz transits. Central Command said its strikes followed attempted Guard attacks on commercial shipping and on American service members, and Iranian media placed them along the southern coast from Bandar Abbas to Chabahar. The 30 August strike on Larak Island was the first inside Iran since 29 July. Axios reported on 31 August that the White House was weighing periodic limited strikes to stop Iran rebuilding the radar and missiles it needs to hit ships. One American official's phrase for that was mowing the lawn. A retaliation is an event. A mowing schedule is a running cost. Central Command finished clearing the strait's international lanes only last week, which is what makes re-mining the cheap move and re-clearing the expensive one. The market is already paying for the schedule. Marsh told S&P Global on 22 July that war-risk cover has gone from 1 to 3 percent of hull value to between 7.5 and 10, which on a $90 to $100 million tanker is $6.75 to $10 million per single transit. Nobody buys that twice for a war anyone expects to end.
The Shanghai Cooperation Organisation signed twenty-eight documents in Bishkek and did not mention Ukraine once. The declaration condemned military strikes on Iranian territory as a UN Charter violation and offered sincere condolences on the death of Supreme Leader Khamenei. Ukraine appears nowhere in it, confirmed not only by Al Jazeera but by TASS, whose own headline reads "SCO declaration omits Ukraine conflict." A Russian state wire volunteering that omission is about as strong as corroboration gets. What it produced was plumbing: a protocol amending the 2002 charter, a countering-threats centre in Tashkent, a logistics roadmap to 2030. What it did not produce is the thing that would matter. The SCO Development Bank was deferred again with no founding agreement, no capital commitment and no launch date, and Kazakhstan and Kyrgyzstan are now arguing over where to put the headquarters of an institution that does not exist. A bloc holding more than 40 percent of the world's people and about 25 percent of its output adjourned after 2 days without committing a dollar, a year after the Tianjin summit announced the same institution. The members' trade with each other still clears through the correspondent-banking rails the bank exists to replace.
Cappadocian obsidian travelled a thousand kilometres and the habit of carving it never left home. Of roughly 740 arrowheads excavated at Tepecik-Ciftlik, a Neolithic village in central Anatolia occupied from about 7100 to 5800 BC, exactly one had been recorded as carved. Alice Vinet of Aix-Marseille University went back through the collection and found eleven more, reported with Denis Guilbeau in PLOS One on 28 July. The incisions were deliberate, cut at two different stages, before the arrowhead was finished and after, and nobody knows what they meant. The genuinely strange part is the map. Cappadocian obsidian was traded as far as the southern Levant, and not one carved arrowhead has ever been found outside central Turkey. The material walked a thousand kilometres and the practice stayed in the village. Trade routes move things much more easily than they move meaning, which is why importing somebody's tools has never once imported their reasons for using them.
Chemists turned the world's least recyclable plastic into engine oil by removing the thing that made it unrecyclable. PVC is the hard case because of its chlorine, which wrecks catalysts and varies with every manufacturer's additive mix. A Virginia Tech group led by Guoliang Liu, publishing in Nature on 5 August and announced by Virginia Tech on 28 August, submerges waste PVC in solvent with aluminium trichloride and alpha-olefins and cooks it at 158 degrees Fahrenheit for three hours. The reaction strips off the chlorine and clips the long polymer chains into short slick ones, which is the definition of polyalphaolefin, the synthetic base stock in premium engine oil. The team reports performance competitive with commercial products. The elegant part is the symmetry: chlorine is why nobody could recycle PVC, and taking chlorine off is most of what you have to do to make the lubricant. The property that made the waste worthless is the property that made the product possible.
Using the Strategic Petroleum Reserve physically destroys it, and no agency will say how much is left. Brian Potter's 27 August account of the reserve's engineering is a lesson in what a physical asset actually is. The oil sits in salt caverns, and drawing it down means pumping water in, which dissolves the salt: roughly fifteen barrels of salt per hundred barrels of water. A full drawdown enlarges its own cavern by about fifteen percent, and creep destroys as much as 2.4 million barrels of capacity a year. Sandia National Laboratories simulates a maximum of five drawdowns per cavern; Peter Zeihan counts thirty-seven in the reserve's history. Potter's finding is the absence: no report by the Department of Energy, the GAO, Sandia or any other agency names a fill level below which damage occurs. Three experts asked separately give 150 to 160 million barrels, 170 million, and 250 to 300 million. The reserve's insurance value is denominated in remaining cycles, not barrels, and nobody outside knows the count.
The infrastructure law expires this month, and the fight everyone is watching is about the wrong number
The Infrastructure Investment and Jobs Act's program authorities lapse on 30 September 2026. The House Transportation and Infrastructure Committee approved a five-year, $580 billion replacement on 22 May by 62 votes to 2; the Senate committees with jurisdiction have released no text at all. Anyone holding aggregates or highway contractors is about to read a headline saying the cliff was avoided. It will be accurate, and it will not decide what these companies build in 2028.
IIJA put $567 billion through the Transportation Department across fiscal 2022 to 2026, and only could because $118 billion of general revenue was injected once into a trust fund whose own receipts do not support that level. Congress's own research service puts the gap for a five-year bill beginning in fiscal 2027 at $166 billion, and nobody has introduced a way to raise it. Short extensions at flat nominal dollars are the path of least resistance, because they require no member to vote for a tax.
Flat nominal funding in a market whose costs are rising is a cut nobody has to author, and it gets announced as a preservation. The Federal Highway Administration's construction cost index rose 26.4 percent year over year into late 2022 and was still adding 6.5 percent in the third quarter of 2024, the most recent reading published, in May 2025. That vintage is the mechanism, not a footnote: the index that decides what a flat dollar buys reports about eighteen months behind the appropriation being argued over.
We expect an extension rather than a lapse, and expect it read as a resolution when it is a deferral. States obligate against carryover contract authority, so nothing visibly stops on 1 October; the mark arrives in state bid lettings twelve to eighteen months later, by which point it will be filed as a soft cycle. The exposure sits with the most public-mix names: Vulcan Materials (VMC), Martin Marietta (MLM), Construction Partners (ROAD), Granite Construction (GVA). Sterling Infrastructure (STRL) and AECOM (ACM) take more work from data centres and utilities than from this trust fund, which makes them a hedge on mix, not a beneficiary.
Europe is campaigning for the world's capital and, in thirteen months, raising the cost of holding European assets from anywhere else
On 11 October 2027 the United Kingdom, the European Union and Switzerland all shorten securities settlement to one business day after the trade. If you hold European stocks or bonds from outside Europe, that raises your cost of holding them, and it lands in operations rather than in a fee schedule, where nobody looks.
What a settlement cycle actually supplies is not time to settle. It is slack, and two adjacent businesses were living in it. Funding a cross-border purchase requires an FX trade, now done on trade date and often before the securities trade is confirmed; the CLS deadline falls a couple of hours after the equity close, and currency that misses it settles bilaterally. Lending the same stock out means recalling it in time to deliver, and that window loses a day.
America ran this experiment on 28 May 2024. Beforehand, industry work warned European managers might see 40 percent of daily FX trades settle outside CLS; CLS's own estimate was roughly 1 percent of its $7 trillion of average daily settled value. The outcome landed much closer to the small number, and lending revenue did not collapse. That is being read as reassurance, and it is the wrong lesson. The American test ran in the easy direction: a European desk trades New York in its own afternoon. The European move asks an Asian desk to complete European FX and European recalls in the middle of its night, and there is no operational fix for a time zone.
We expect the cost to surface as pre-funded cash sitting idle and as more of the European FX leg settling bilaterally, visible first in 2028 custody disclosures rather than in the price of anything. It is paid either way by the layer doing the work, BNY (BK), State Street (STT) and Northern Trust (NTRS), and paid once by Asian-domiciled managers running European mandates and the lending revenue line at BlackRock (BLK). It is small per trade and invisible per fund, which is why nobody prices it. Anyone citing America's quiet 2024 transition as evidence this one will be fine is citing the easy version of the experiment, and a continent advertising this hard for foreign money will not notice it has billed them for the privilege.
Beijing spent a decade forcing its rare-earth industry down from dozens of producers to two state groups, and finished the job in 2024. The stated reason had nothing to do with America: Chinese rare earths were selling too cheaply, and Beijing wanted pricing power. One year after the structure was complete, it became the export weapon that will be sitting on the table when Xi meets Trump in Washington on 24 September.
Call the mechanism enforcement reuse. A rule is only as real as the machinery that can watch and punish the people it binds, and that machinery is the expensive part, slow to build and completely indifferent to which rule it ends up serving. Stigler's 1964 account of oligopoly makes the point exactly: a cartel's binding constraint is never the agreement, it is the detection of secret defection, which is why collusion is viable only where sellers are few and output is observable. An export-control regime has the identical requirement. So an industry restructured to enforce a price is, by construction, restructured to enforce an embargo, and the second rule arrives looking like a decision when it is really an availability. Where 11 August's Enforcement Incidence asked where the cost of using an instrument lands, this asks what it costs to have one at all.
The receipts run in one direction. In March 2021 Xiao Yaqing, then China's industry minister, complained that rare earths were selling "at the price of 'earths' rather than the price of 'rare'" because of "vicious competition and competitive price-cutting." That December, Beijing merged the rare-earth assets of Chinalco, Minmetals and Ganzhou into China Rare Earths Group for pricing power. Guangdong was folded in during 2024, leaving two groups holding the national quota. Beijing stopped publishing the quota numbers in 2025. On 4 April 2025 it put seven heavy rare earths under export licensing. Xiao himself was removed in July 2022, well before his margin problem became Washington's supply problem.
What the standard method misses is that it reads intent. Analysts ask which sectors Beijing means to weaponise and price the summit. Intent is unobservable, announced late, and frequently absent. Capability is observable and announced early, in a document nobody files under national security: a merger approval. The controlled comparison already exists inside China. The six largest polysilicon producers proposed raising roughly 50 billion yuan to buy out and shut about a third of the industry's capacity; on 9 January 2026 China's own antitrust regulator suspended the plan on monopoly grounds. Same country, same year, same anti-involution logic, and that ruling is the difference between a sector that will be weaponisable in 2029 and one that will not. Nobody at the regulator was adjudicating export leverage, which is precisely why the ruling carries information.
We expect the next Chinese export-control action to land in a sector where a state-blessed consolidation has already cleared, and not in solar, where it has been blocked. Steel, cement, petrochemicals and battery materials are where the anti-involution campaign has produced real capacity discipline; solar has produced a blocked deal and roughly 1,200 GW of annual module capacity against about half that in global demand. The instrument worth reading is the antitrust docket, not the trade ministry's press office.
The case against starts with the source that supplies the timeline. Aqib Zakaria and Nicholas, writing in ChinaTalk, argue the moat is not structural at all but facility-specific tacit knowledge, a secret roasting recipe, a mixer-settler bank tuned to one ore body, held by people, and conclude that American firms "would need to poach entire teams to extract sufficient know-how, though even that may not be enough." On that reading, consolidation is bookkeeping, the real weapon dates to Xu Guangxian's countercurrent solvent-extraction work in the 1970s, and enforcement reuse explains the timing of an announcement and nothing about the leverage behind it.
The second objection is the least welcome. Export leverage requires the buyer to have no substitute, which is true of heavy rare earths and false of nearly everything else on the anti-involution list. China is the low-cost seller of solar modules and lithium cells into a world that wants them, so an embargo there taxes Chinese exporters and subsidises foreign entrants. The machinery may get built in a dozen sectors and be usable in one. It is also leakier than the frame requires: by those same authors' arithmetic, the roughly 260,000 tonnes of NdFeB magnets China made in 2024 imply about 1,750 tonnes of dysprosium against national quotas under 900, and the holes that leak the quota leak the embargo.
The framework is wrong if China's next significant export restriction lands on a fragmented industry with no state-directed consolidation behind it. In that case enforcement capacity was never the binding constraint, intent was, and the merger docket is noise.
"The blunders are all there on the board, waiting to be made."
— Savielly Tartakower
A club player loses the same way for years and calls it bad luck. It is not bad luck. It is that the openings he plays produce four or five recurring positions, each of which offers a specific, limited menu of ways to go wrong, and he keeps picking from the menu because he has never once read it. This is the useful half of Tartakower's line, and it is easy to miss under the wit. He is not saying mistakes are inevitable. He is saying they are already present, in the position, before anyone touches a piece, and that a position offers a small number of them rather than an infinite number.
Your week is a position. It offers you a handful of specific errors and not many more. The reply you write at eleven at night, the second glass you actually decide on while pouring the first, the conversation you keep resolving by moving it. Those are not character failings in the abstract. They are the particular blunders your particular schedule and particular relationships make available, and someone who knows you well could probably list them faster than you could.
The reason to write them down is that a mistake you have named in advance stops arriving as a surprise and starts arriving as a decision. What is hard at eleven at night is not the willpower. It is that the blunder does not feel like a blunder while you are making it. It feels like a mood, and nobody argues with a mood.
Today's practice: write down the three blunders your position actually offers this week. Not the abstract ones. The ones your calendar and your kitchen and your inbox make easy. Then, the next time one of them is in front of you, say its name out loud before you act: this is the second glass, this is the eleven o'clock reply. Naming the state while you are in it is the whole intervention, because a feeling you can name has stopped being weather and become a move you are choosing.
Walk across any university lawn and you will find a path nobody designed. It cuts the corner the architect asked people not to cut, and by the second summer it is bare dirt six inches wide. No committee approved it. Each walker simply chose the easiest line, and worn grass is easier than unworn grass, so each crossing made the next crossing more likely.
Two different mechanisms are running there at once, and most people collapse them into one word. The first is a network effect, which is about value: each additional participant makes the thing more valuable to everyone already using it, so value grows faster than membership does. The second is emergent behavior, which is about order: local rules with no coordinator produce a global pattern nobody chose. The footpath is both.
Language is the purer case. A word is worth exactly what other speakers take it to mean. You can decide privately that "literally" will keep its old sense, and you will simply be misunderstood, because the value of your vocabulary is entirely a function of who else shares it. And meanings drift without anyone convening: no committee voted that "awful" would stop meaning full of awe, yet it moved somewhere no participant selected.
Now the part that decides whether this is useful. Size the thing. Below a critical mass, a network-effect product is not weak, it is worthless: an empty restaurant is not a slightly worse restaurant, it is a signal, and this is why these businesses either ignite or die rather than growing steadily. Above a point the same feedback runs against you. The path becomes a mud channel. The word gets used for everything and means nothing. Congestion is the identical loop with a saturated resource in it.
The failure mode is the important part, because following this model badly is worse than ignoring it. It is assuming a network effect exists because something is popular. A million people using the same toothpaste does not make the toothpaste clean your teeth better. A network effect requires that other users improve the product for you, and most scale does not. Companies that mistake the two defend a moat that is not there, then are astonished by how fast the exit runs. It runs fast because it is the same loop with the sign flipped, and it is not proportional: losing the first ten percent does not cost ten percent, it costs whatever those ten percent were worth to the ones who stayed.
The decision tool is one question, asked twice. First: if half the other users vanished tomorrow, would my experience get worse? Not the company's revenue. Mine. If yes, it is a network effect, and how much worse is how deep the moat is. If your experience would be unchanged, you are looking at popularity, and popularity defends nothing. Second: name the smallest group whose departure would start the unravelling. In almost every real network it is not the largest group. It is the one the others came for.
Metal gets stronger when you make its crystal grains smaller, and the reason is unusually specific. Metal bends because line defects called dislocations glide through the crystal; grain boundaries obstruct them; more boundaries means more obstruction. That is the Hall-Petch relation, and for half a century it was one of metallurgy's most reliable levers. Then Jakob Schiøtz, Francesco Di Tolla and Karsten Jacobsen simulated nanocrystalline copper under load and reported in Nature in 1998 that past a certain fineness the metal got softer; Schiøtz and Jacobsen's larger 2003 Science simulations, running grains out to 48.6 nanometres, captured both regimes and located the maximum between them. Below roughly ten to thirty nanometres, depending on the metal, a grain is too small to hold a pile-up of dislocations at all, so glide stops being how the metal deforms. What takes over is grain-boundary sliding: enormous numbers of uncorrelated slips, a few tens of atoms at a time, past each other. Every boundary you added to obstruct a dislocation is now a surface that slides. The lever did not weaken. It began operating a different mechanism, in the opposite direction.
The instinct is to file this under diminishing returns, and that is the wrong drawer. Diminishing returns means each increment buys less; you can watch the curve flatten and stop at the top. Here the output falls, and it falls because the intervention consumed the structure the mechanism needed: obstructing dislocations requires grains large enough to contain one, and grain refinement is the thing that removes them. That makes it invisible from the output alone. Strength on the way up and strength on the way down are the same measurement, and nothing in the number tells you which side of the maximum produced it. You cannot detect a mechanism handoff by watching the quantity the mechanism was supposed to move. You can only detect it by asking what is currently doing the work.
So this week, when a lever you have been pulling stops paying, do not choose between pulling harder and stopping. Write one sentence naming the mechanism you believe the lever runs, then name the smallest unit that mechanism has to act on, and check whether your own pulling has made that unit smaller than the mechanism requires. Run it on something you have already been intensifying, splitting teams further, shortening review cycles, adding checkpoints. If you cannot name the unit, that is the finding, because a lever whose mechanism you cannot state is one you have no way of knowing the sign of. Team-splitting has the identical architecture: it reduces coordination cost right up to the point where a team is too small to hold the context the work needs, after which every split adds handoffs instead of removing them.