Every result in this series so far assumed the enemy shows up. Put both forces on the field, run the battle 15,000 times, count the outcomes. That assumption does enormous, silent work. Real enemies read your order of battle first, and then decide whether to fight at all.

This post puts that decision inside the model. It is the most deflationary result of the whole project. When I gave the simulated enemy the choice of whether and how to engage, three things happened. First, a rational enemy declined roughly 40-58 percent of the engagements he would have fought against a breakable force. That range comes from a stylized model with a perfectly informed enemy, so read it as an upper bound on deterrence. Second, of the fights he still allowed, the unbreakable force won only 10-43 percent, against the 91-100 percent a naive model predicts when every fight happens. Third, when the enemy could pick the kind of fight, not just whether to fight, the measured value of unbreakable will collapsed by about 97 percent, from 46 points to 1 point on the model’s win-probability scale. The direction of that collapse survived every parameter sweep I ran; the exact size belongs to this particular setup. All three results are illustrative engagement games built on top of the validated battle engine, not calibrated history. The orderings are the findings. The magnitudes are illustrations.

The one-line version for the meeting: the enemy will not fight your unbreakable army. He will fight around it.

How to read it: Follow 100 potential fights left to right; the blue stream bends away unfought, and only the brass threads end inside the win zone, small against the dashed naive ring.
What to see: A force that “never loses” banks about 14 wins per 100 opportunities once the other side picks the fights.

The funnel’s bar widths show the midpoint scenario: 49 fights deterred, 27 percent of the allowed fights won. The printed ranges carry the uncertainty.

1. Deterrence is real, and it is the largest single effect. The selection model fits in a paragraph. Two equal forces meet repeatedly. Each encounter draws a capability ratio from a distribution centered on parity: sometimes you have the edge, sometimes he does. Your force is either normal, breaking at historical morale thresholds, or unbreakable. The enemy knows your capability and your will, and he declines any battle he expects to lose.

Against a normal force, the enemy’s decision line sits at parity. A fair fight is a coin flip, so he takes anything at even odds or better. Against an unbreakable force, his line moves dramatically. In this setup, the unbreakable force wins any fight where its effectiveness ratio is above roughly 0.44, because it converts every close fight into a grinding attrition contest the breakable side quits first. That 0.44 is not a dial; it falls out of the two sides’ breakpoint parameters, where fighting to destruction reaches even odds against routing at a quarter losses. Everything between 0.44 and parity becomes a fight the enemy would have taken against you yesterday and refuses today.

Sweep the spread of capability draws and the engagement rule, and roughly 40-58 percent of encounters come off the calendar. That number never appears in any battle record. No dispatch reads “the enemy declined contact because our will was unbreakable.” The single largest effect of the whole investment is invisible.

Political science knows this mechanism: wars are a strategically selected subset (Fearon 1995), democracies pick wars they expect to win (Reiter and Stam 2002), initiators pick targets they can beat (Gartner and Siverson 1996). This post gives that insight a combat-model engine.

Which sounds like winning. Hold that thought.

2. Now look at which fights he declines. Part four of this series found that unbreakable will pays most in the near-parity band: at heavy disadvantage it cannot save you, and at heavy advantage you did not need it. Now overlay the enemy’s decision line. The band he vacates, 0.44 to parity, is exactly the band where will is worth the most. He donates the fights that do not matter and keeps the ones he wins.

The arithmetic that falls out is brutal. A naive model, one where every possible engagement gets fought, predicts the unbreakable force wins 91-100 percent of encounters across the swept settings. The selected record shows it winning 10-43 percent of the fights it is actually allowed. Same force. Same will. The only change is who picks the fights.

How to read it: The dashed slab is the record a model with no enemy vote predicts; the brass slab is what the after-action file would actually show.
What to see: Same force, same will; who picks the fights is worth up to roughly 80 points of win rate on paper.

There is a career-review joke buried in this. A near-invincible force, assessed on its win-loss record, looks mediocre. Its after-action file reads like an officer passed over twice. Selection does that: the enemy only ever meets you where he holds the overmatch, so the record samples your worst days. One phrasing trap is worth flagging honestly. Whether the unbreakable force’s win rate in fought battles drops below a normal force’s depends on the engagement rule; under a stricter rule it stays higher. The robust claim is the survivorship gap against the naive expectation, and that gap is large under every rule I tested.

3. A deterred fight is not a won fight. Here is the part the funnel chart cannot show. The deterred 40-58 percent are not victories. They are non-events. The enemy who declines battle keeps his army, keeps his ground, and keeps his options. He waits, adapts, and presses somewhere else.

Wars need decisions. Clausewitz’s whole argument is that war is an act of force to compel the enemy to do our will; a battle that never happens compels nothing. A force that must take ground, relieve a siege, or force a political settlement cannot cash deterred engagements. It shows up ready to fight forever, and finds that forever is exactly what the enemy is happy to give it.

So the ledger reads: the biggest effect produces no decision, and the residual fights are the enemy’s picks. That is what “the enemy gets a vote” costs.

4. Let him choose the kind of fight, and will’s value nearly vanishes. The second model gives the enemy a richer menu. Instead of fight-or-decline, he chooses among three engagement types: a grinding attrition battle, a decisive first strike, or avoiding battle entirely. The payoffs come from the same engines validated earlier in the series, at near parity where will matters most.

Walk the matrix. In the attrition column, unbreakable will is magnificent: it lifts the win probability from 49% to 95%, a premium of 46 points. In the first-strike column, will is nearly worthless for either side; part four showed why, since a fight decided by the opening salvo never reaches the morale check. Win probabilities there are 0.3% and 1.5%. The avoid column pays the status quo to both.

How to read it: Each card shows the chance our side wins under normal versus unbreakable will; the enemy stamps the one column where will pays refused.
What to see: Advertising “we will grind out any attrition fight to the death” is also an advertisement of which fight never to give that force.

The enemy reads the same matrix. Attrition is the one column where your unbreakable force does well, so he refuses it and forces the first-strike fight instead. At that equilibrium, the value of unbreakable will is 1.2 points of win probability instead of 45.9. About 97 percent of its measured value is gone, and the collapse direction held across every status-quo payoff and casualty weight I swept. The magnitude is a property of this setup; the structure is not. Any enemy with one available engagement type where will is worthless will steer the war there.

Can unbreakable will go negative, an actual liability? Only under two conditions jointly: the enemy forces a first strike at low lethality, so your force survives the opening blow but is behind, and your commander weighs casualties at more than about a third of a win. Then the normal force routs at about 60 percent casualties while the unbreakable force bleeds to about 80 percent in a battle it still loses; routing saved a fifth of the force. At the pure one-shot equilibrium, will is worthless but never negative; both sides die at the same rate regardless of resolve. Fielding a force that advertises “we will grind out any attrition fight to the death” is also an advertisement of which fight to never give it.

How to read it: The brass level is will’s value when the enemy must fight your attrition fight; the thin level at 1.2 points is what remains once he chooses the engagement.
What to see: The direction survived every sweep, the magnitude belongs to this setup; in the forced low-lethality fight will can even turn liability.

5. Russia is running this playbook right now. The closest live case is Ukraine’s drone wall. First-person-view (FPV) attack drones and persistent surveillance have created a kill zone many kilometers deep where, as witnesses told British members of Parliament in 2026, per Forces News reporting, anything that moves gets targeted. That testimony described a stalemate: roughly 1,000 Russian and 200-300 Ukrainian casualties a day while the line barely moves. The wall does not tire, does not panic, and does not rout.

One attribution I will not make. The wall deters because it kills, not because it is brave. It is sensing plus fires, the capability post 3 spent its length separating from will; a jittery wall with the same kill rate would earn the same detour. So this case is evidence of the selection and adaptation mechanics, not of deterrence by unbreakable will: a rational enemy redesigning his war around a standing threat rather than feeding it.

Russia’s response was not to feed it. A Royal United Services Institute (RUSI) commentary in August 2025, and independent open-source analysts, documented a redesign of the Russian assault echelon to fight around the drone wall rather than through it: dispersal into small teams, which open-source analysts put at four to six men, instead of massed columns, spider holes and thermal camouflage, movement at night in short bounds, electronic-warfare screens, and motorcycles. Analysts tracking Russian procurement reported targets on the order of 120,000 to 200,000 bikes plus tens of thousands of all-terrain vehicles. A motorcycle crosses a 1.5-kilometer kill zone in three to five minutes, inside the reaction time of many FPV crews. Dispersion makes each drone kill one man, not a squad.

Read that against the models. The kill zone deterred the engagements it would have won: massed armored assaults into prepared drone coverage largely stopped happening. What remains are the engagements the attacker chose, shaped to the defense’s weakest sensing and reaction seams. The unbreakable wall got a vote on nothing; the breakable humans riding motorcycles picked every fight. And as part five argued, the wall’s standing threat is only as good as its wiring against jamming, one of the seams the adaptation targets.

None of this means the drone wall was a bad investment; Ukraine holds its line partly because the wall made most attacks not worth having. It means the wall’s value migrated from won fights to prevented ones, with section 3’s warning attached.

6. And when the enemy cannot decline, will pays at full face value. The mirror case proves the mechanism. Some soldiers in this war have no vote: wounded, encircled, abandoned by their units, pinned under persistent observation. Against them, the machine’s tirelessness prints exactly as advertised, and the currency it prints is surrender.

In August 2024, per Ukrainian-side reporting, a drone from Ukraine’s Khartiia brigade found a wounded Russian soldier abandoned by his unit, dropped water and written instructions, and walked him through a surrender to Ukrainian lines. In January 2026, per Forbes reporting, three Russian soldiers surrendered to an armed ground robot near Lyman. Ukraine’s digital-transformation minister Mykhailo Fedorov claimed more than 100 drone-compelled surrenders over the winter of 2025-26; that is a belligerent’s own count, so treat it as an indicator, not a statistic. Part two of this series showed the Union army at Bull Run breaking with more than nine-tenths of its men still standing; these incidents are the same morale mechanism, collected by machine.

Put the two halves together and the insight sharpens. The variable that decides what unbreakable will is worth is not the robot. It is the enemy’s menu. Where he can decline or reshape the engagement, the value collapses toward the 1-point end of the model. Where he cannot, a tireless, fearless system breaks human will retail, one pinned soldier at a time. Strategy is the art of deciding which of those two worlds you are fighting in, and the enemy is co-author of the answer.

Best Arguments Against This

The perfectly informed enemy is a fiction. Conceded, and it is why every deterrence number here carries “upper bound” next to it. Real commanders misjudge relative capability, cannot always refuse battle, and fight under political orders that ignore the odds. Every one of those frictions shrinks the deterred share and puts value back on the battlefield. The honest reading is a band running from my frictionless number down toward whatever the enemy’s actual information and freedom of maneuver allow. Russia’s motorcycle adaptation suggests a competent enemy gets usefully close to the rational benchmark, but that is one adversary in one war.

Deterrence is not a consolation prize; it is strategic value my win-rate metric undercounts. Mostly conceded. The Cold War’s central front was held for four decades by forces whose entire purpose was to make the fight not worth starting. If your mission is defending the status quo, the top stream of the funnel is not leakage, it is the mission, and an unbreakable force that gets declined is doing precisely what you bought it for. The rebuttal is narrower than the concession: the vote is still his. He declines where you are strong and presses where you are weak, on his schedule, and a force that must attack or compel cannot spend deterred fights. Deterrence is real value that accrues only to the side content with the map as it is. Honestly, I personally find this the most valuable insight out of this project and it was completely unexpected (to me anyway).

The ugliest number leans on a modeling choice. True. The 10-43 percent depends on how eager the modeled enemy is to fight marginal battles; a stricter rule shrinks the fought set and lands the win rate higher within that range. I printed the framing that survives both rules, the gap against the naive 91-100 percent, and the gap is the finding.

What Would Change My Mind

If by mid-2027 Russia has returned, at scale and sustainably, to massed mechanized assaults into established drone kill zones, the fight-around thesis fails its strongest live test.

If a spatial, agent-based replication of the selection model with realistic misperception and political pressure to fight shows the deterred share collapsing below about 10 percent, the funnel’s top stage is an artifact of the perfect-information assumption. That replication is on my open-threads list for 2027.

If wire-service-verified reporting by December 2027 documents never-routing systems compelling the surrender of intact, un-encircled units in open battle, then unbreakable forces can force decisions rather than just deny them, and section 3 is too pessimistic.

If an engagement-choice game calibrated to a real campaign, and 2022-2026 Ukraine is the obvious candidate, finds the equilibrium value of resolve above half its fixed-enemy value, my 97 percent collapse is a property of the model, not the world.

Unbreakable will is not a bigger hammer. It is a posted sign telling a literate enemy which fights to skip. He reads it, thanks you, and redesigns the war around the fights you cannot decline, while your perfect record accumulates in battles that decide nothing.

There is one more rung on this ladder, and it is the worst one. Next post: what happens when a casualty-averse democracy fields the unbreakable army against an enemy who does not count his dead, and wins every battle on the way to losing the war.

This series’ default reading of AI in warfare goes like this: robots do not panic, so automate the weapon and you get an army that cannot rout. Post 4 already narrowed that promise to fights near capability parity. This post kills what is left of it. In my simulations, WHERE the will lives determines what it is worth, and the trooper is the worst address on the list.

Three results, one ladder. First, the commander. Across 607 recorded defeats in a US Army battle database, losing sides ended the battle by a commander’s decision to withdraw 266 times, and by the physical annihilation of their troops 26 times. Ten to one. That is a historical count, not a model output, and it means an unbreakable command decision touches ten times more endings than unbreakable troopers do. Second, the domain. In the stylized joint-force model, unbreakable will buys about +50 points of win probability, but only for the force holding the objective; supporting forces get about zero at any firepower share. Third, the substrate. “Unbreakable” turned out to be a property of the electronic-warfare and cyber layer, not of the silicon psyche. Every configuration I tested had a resilience threshold, about 40% in the anchor case and 20–50% across configurations, below which a rational enemy stops shooting the bodies and starts jamming the will, and the value of unbreakability decays from 84 points to 17 to 1.

Your robot’s will is only as strong as its radio.

How to read it: Every brick is one of 607 recorded defeats in the CDB90 battle database, colored by how the battle ended for the loser; the red seam is every army that was actually destroyed.
What to see: Brass buries red 266 to 26, about ten to one; defeat is usually a commander’s decision at the top, almost never annihilation at the bottom.

One caveat travels with everything below. Except for the ten-to-one count, which is historical data, these are stylized model extensions of the validated battle engine. The orderings and directions survived every parameter sweep I ran; the exact magnitudes are illustrations, not measurements. The research data was current as of July 1, 2026.

How to read it: Climb the building from the trooper floor to the dark nation floor; each brass plate is what an unbreakable will is worth installed at that floor, and the slab under everything is the electronic-warfare foundation.
What to see: The lights are on in the middle of the building, not the barracks; the top floor stays dark on purpose, that story is the final post.

The findings

1. Battles end at the top, not the bottom. The database is CDB90 (the US Army Concepts Analysis Agency’s Database of Battles, compiled by HERO, the Historical Evaluation and Research Organization, in the 1980s and delivered as the 1990 database), covering over 600 land battles from 1600 to 1973. Each losing side carries a code for how its battle ended. I counted the endings. Withdrawal by command decision, including withdrawals under heavy loss: 266. Troops annihilated or overrun where no one was left to withdraw: 26. That is 10.2 to 1. Even counting every other ending together (repulsed, penetrated, surrendered, scattered, annihilated: 313 in all), the commander’s withdrawal is the largest single category. Twenty-eight losers carry no code.

The pattern is not my discovery. Robert McQuie, an operations research analyst at the same Army agency, published its shape in “Battle Outcomes: Casualty Rates as a Measure of Defeat” (Army magazine, November 1987). Analyzing 80 engagements from 1941 to 1982 in the same HERO data lineage, he found that defeated forces typically quit below 10 percent casualties, and that the recorded reasons for defeat ran about 64 percent enemy maneuver against about 12 percent firepower and casualties. My 266-to-26 count replicates McQuie’s pattern in the successor database of 607 defeats. Replication is the claim, and it is a strength: two counts, two datasets, four decades apart, same answer.

How to read it: Each bar counts how the 607 recorded defeats ended, by the database’s loser resolution code; brass bars are commander-ordered withdrawals, the red bar is troop annihilation.
What to see: Withdrawal endings outnumber annihilation endings 266 to 26, and even against all other coded endings combined (266 versus 313) the commander’s withdrawal is the largest single category.

Most defeats are not a dying army. They are a live army whose commander decided the price had been paid. So I gave the model two breakpoints instead of one: a troop threshold where units physically rout, and a command threshold where the commander orders the whole force out. The result is almost embarrassingly clean. The lower threshold governs. Where the commander’s nerve gives out first, which the historical count says is the common case, making the troops unbreakable bought a mean win-probability change of +0.4 points, averaged over the six command-binding cells at 5,000 runs each. Statistically nothing. The troops never got the chance to prove their steel; the order to withdraw came first. Making the command decision unbreakable in those same cells bought about +60 points and flipped the majority of contests. Where the troops break first, the mirror image holds. The ordering held in every one of the 16 cells of the 4x5 threshold sweep where the two thresholds differ; the +60 points is a stylized near-parity magnitude, not a portable constant.

I set out to model courage and ended up auditing org charts. The lever is not the trooper who never runs. It is the command decision that never flinches early. RAND’s Will to Fight framework (Connable and colleagues, RR-2341-A, 2018) is the standing treatment of what will is, a 29-factor model of the fighting unit. My ladder asks a coarser question: not what will is made of, but where removing the break pays. Jamestown Foundation analyses of PLA (People’s Liberation Army) writing on “cognitive domain operations” suggest at least one great power already reads it this way: the doctrinal target is the adversary’s decision system, not its infantry’s stomach.

2. Will pays only where the objective lives. Now widen the lens to a joint force: land, air, and maritime elements pooling fire on a shared enemy, with exactly one domain holding the objective that defines victory. I rotated which domain holds the objective, swept the firepower shares, and made each domain unbreakable in turn.

The pattern did not budge anywhere on the grid. Make the objective-holding domain unbreakable and the side gains about +50 points of win probability (51 to 55 points across all 21 tested cells). Make a supporting domain unbreakable and the side gains about zero (within 0.4 points of zero across all 42 tested cells). Firepower share is irrelevant to this. In the amphibious scenario the maritime force holds the objective with the smallest gun share, 20 percent, and its will is worth the full +50 points; the air component carries half the firepower and its will is worth nothing. Push a supporting domain to two-thirds of total firepower and its unbreakability is still worth nothing. The mechanism is blunt: supporting forces are overrun the moment the objective falls, so their endurance is never the binding constraint. Their firepower matters. Their will does not.

How to read it: Each dot is one tested cell of the joint-force model, a domain made unbreakable at some firepower share, split by whether that domain holds the objective or supports it.
What to see: About +50 points for the objective holder and about zero for everyone else, no matter who carries the firepower.

The procurement translation writes itself. Buy endurance for whatever holds the objective. Buy cheap, expendable firepower for everything else. Paying unbreakable prices for the loudest shooter is decor. This is a stylized result, near parity and symmetric by construction, but the ordering never reversed in any sweep.

3. “Unbreakable” is a property of the wiring. Here is the uncomfortable one. A human’s will is attacked with fear. An AI’s will is attacked with electrons. Jamming, spoofing, cyber, and decapitation of control nodes do to a machine exactly what Bull Run’s panic did to the Union army in post 2: they end its participation without killing it.

So I let the enemy choose. It holds a fixed effort budget and splits it between attrition (killing bodies) and suppression (attacking the will’s substrate). Our side is nominally unbreakable, protected by a substrate resilience R running from 0 to 100%. Suppression effort past R drags the effective breaking point down a linear ramp, at a fixed gain, to a floor at a normal human rout threshold. The result is a threshold, R*, in every one of the five configurations tested. Below R*, the rational enemy stops buying bullets and buys jammers: it drives our effective breaking point down to an ordinary human rout threshold while keeping enough residual firepower to exploit it. Above R*, suppression costs more firepower than the will-kill is worth, the enemy reverts to pure attrition, and “unbreakable” finally means what the brochure said. In the anchor case R* is about 40%. Across enemy jamming potency and overmatch levels it spans 20–50%, and it rises as the enemy’s electronic attack gets sharper. That range is the right way to read the result: the threshold’s existence is the finding, its location is set by the adversary’s kit, not yours.

Below the threshold, the decay is brutal. The value of unbreakable will runs 84 points of win probability with a healthy substrate, 17 points at 20% resilience, and 1 point at zero. At the bottom of that curve you have bought a normal, breakable force at an unbreakable price. And there is a second, independent way to be worthless: against threefold capability overmatch, the will premium is near zero at every resilience level, because will cannot close that gap at all. Post 4’s parity condition still binds. All of this is a stylized mapping from electronic attack to breaking point; the shapes, not the decimals, are the claim.

How to read it: Each bar is the win-probability value of an unbreakable will at one substrate resilience level; the static is the enemy’s jamming effort, and the dashed line is where jamming stops paying.
What to see: Below R* of about 40% (20-50% across configurations) the signal drowns, 84 points of win probability collapsing to 17 to 1, so “never routs” is a property of the radio contest, not the silicon psyche.

4. Ukraine is running this experiment, in both directions. The jamming direction first. Excalibur guided artillery shells reported fall from roughly 70 percent effectiveness to 6 percent under Russian GPS jamming, per Washington Post reporting of Ukrainian and US assessments in May 2024, is the decay curve with a serial number on it. GMLRS (Guided Multiple Launch Rocket System) rockets were similarly degraded by late 2023, with truck-mounted jammers reportedly suppressing satellite guidance across tens of kilometers. And the substrate war has a cyber flank: reporting in April 2025 described Ukrainian drones abandoned for capture that carried malware, infecting the Russian systems that read them and reportedly geolocating their operators. A machine’s will cannot be frightened. It can be stolen.

Now the honest countermove, because the contest did not stop there. Fiber-optic drones, physically tethered by a hair-thin cable, are immune to jamming entirely; Ukrainian sources counted roughly 40 fiber-optic models codified for its forces in 2025, with ranges that doubled within a year. Onboard autonomous terminal guidance rides through jamming the same way: no link, nothing to cut. As of the data freeze, the attack side of the substrate war was winning, so much so that countering fiber drones was the theme of NATO’s 2025 innovation challenge, and the fashionable counters are nets, radar tripwires, interceptors, and shotguns.

Read that carefully, because it is the point, not an exception to it. The model’s claim was never “jamming wins.” The claim is that the machine’s will is a ratio, R, contested ground between suppression and hardening, and the ratio moves. In 2023 the jammers were winning and Excalibur’s will broke. In 2025 the fiber spools and the autonomy stacks pushed R back up, and the counter-effort shifted to physics: nets and shot. Each swing relocates R* and repositions everyone’s “unbreakable” claim. Morale warfare did not disappear when the soldiers became machines. It moved into the spectrum, and it is being fought right now.

Best Arguments Against This

The strongest objection: these are stylized models, not calibrated campaigns. Correct, and I will not soften it. The two-level breakpoint, the joint-force fires, and the jamming-to-breakpoint mapping are illustrative mechanisms bolted onto a validated battle engine; nobody has fitted a command breakpoint to a real headquarters. What I can defend is that every ordering in this post, command over troops, objective-holder over supporter, threshold over no-threshold, survived every parameter sweep without a single reversal. The decimals are illustrations. The directions are the findings. McQuie adds an honest wrinkle here: he read his own 80 engagements as evidence that Lanchester-style attrition equations misstate what drives combat outcomes, and this series runs on Lanchester engines. My reply is that the breakpoint mechanism exists precisely because pure attrition does not end battles. And the audit prints orderings with bands, not attrition decimals.

Second, the ten-to-one ratio depends on a definitional choice. Withdrawal endings against annihilation endings is 266 to 26; withdrawal endings against all other endings combined is 266 to 313. Some coded withdrawals surely followed local routs, and 28 losers carry no code at all. I concede the framing but keep the conclusion: under the most conservative reading, the commander’s withdrawal is still the single largest way battles end, and annihilation is still the rarest. The finding needs “commanders end battles far more often than annihilation does,” and that survives any cut of the codes.

Third, a skeptic can say the substrate argument is already dated: fiber optics and autonomy are beating electronic warfare as I write, so hardened AI will is a solved problem. That reads a moving contest as a final score. Unjammable tethers spawned net fences; autonomy spawned interceptors; every countermeasure in this war has had a counter-countermeasure inside a year. If the contest ever does end permanently in the defender’s favor, the substrate finding dies. I have put a date on what that would look like below.

What Would Change My Mind

One. A re-coding of the CDB90 endings, or a better modern battle database, published by December 2027, showing that most coded withdrawals actually followed physical unit collapse. That would deflate the ten-to-one anchor and finding 1 with it.

Two. By December 2027, credible field evidence that autonomous systems held most of their unjammed effectiveness against a peer electronic-warfare opponent for two consecutive years with no effective countermeasure fielded at scale. That would indicate that the substrate contest has ended, and the resilience-ratio framing with it.

Three. A calibrated campaign study by the end of 2028 in which hardening a supporting arm flipped an outcome that the objective-holder’s endurance could not. That would break the domain gating.

Four. A spatial or agent-based replication of the two-level breakpoint model, before this series’ planned 2027 follow-on, in which the binding-threshold ordering reverses. That is the one structural assumption everything in finding 1 leans on.

The close

Unbreakable will is not a trait you install in a tactical weapons system. It is a stack: a commander’s decision that does not flinch early, sitting on the force that holds the objective, riding on a substrate the enemy is actively trying to jam, spoof, and infect. Get any layer wrong and the courage above it is decoration. The ladder’s top floor, the nation, is still dark in the elevation above, and it is dark on purpose: that is where this series ends, and it is the rung where unbreakable machines can hurt the side that owns them.

Before that, one more escalation. Everything in this post assumed the enemy shows up and fights. Next post: what happens when it simply declines. The enemy gets a vote.

Bottom Line Up Front

An unbreakable AI will is not a decisive military advantage. It is a tiebreaker. It decides fights that were already close, and almost nothing else.

The worked example is Trafalgar. I gave the entire Franco-Spanish fleet an unbreakable will, so it fights to annihilation instead of striking at 50% losses, and the British still win 83.6% of simulations (95% confidence interval 83.0-84.2, n=15,000 Monte Carlo runs). British gunnery, calibrated at about twice the Franco-Spanish rate, and Nelson’s concentration of fire decide the battle before any morale check can matter. These are attrition-model results, validated against four historical battles but derived from three in this post; the floors below are battle-specific constants, not laws of nature.

The visualization. Trafalgar, where below the capability floor the will changes the dying, not the winner.

How to read it: the aged chart shows Nelson cutting the Franco-Spanish line into three, then the split pane asks whether the van returns. What to see: give the whole Franco-Spanish fleet an unbreakable will and Britain still wins 83.6% of runs. Below the floor, resolve changes only the body count.

The deeper result is a floor. For each battle there is a capability level below which no fraction of unbreakable force wins. Not most fractions. None, all the way to a fully robotic army or navy. Above the floor, will and capability trade along a measurable frontier. Below it, resolve changes the body count and nothing else.

The companion visualization. Pickett’s Charge, with opposing sides’ capabilities at near-parity, the AI unbreakable-will flips the result.

How to read it: the left is the historical charge that breaks and falls back; the right is the same assault given an unbreakable will. What to see: near capability parity, above the floor, the AI-willed assault breaks through and wins 92% of runs. That frontier is exactly what the floor sits under, and the reason unbreakable will is a tiebreaker rather than nothing.

And the floor is the friendly case. Against an enemy that does not break, the win-probability value of your own unbreakable will is exactly zero in every battle I model. Under an enemy first strike at saturation lethality, it is statistically zero. The sentence I would carry into a meeting: unbreakable will decides fair fights. It does not substitute for firing first.

How to read it: brass wakes are the British columns punching through the enemy line; red fragments are what is left of it.
What to see: a Franco-Spanish fleet whose crews cannot break still loses 83.6% of simulations; gunnery and concentration decide before any morale check matters.

The findings

1. There is a floor under will. The simulator is post 2’s validated engine: Lanchester attrition with an explicit morale breakpoint, the casualty fraction at which a human unit quits. For each battle I varied two dials, the loser’s fighting effectiveness relative to history (rho) and the fraction of its force given an unbreakable will (phi), and mapped the boundary where the historical loser starts to win.

The boundary has a hard edge. The square law predicts a capability floor for each battle, with a closed form: rho_1 = a_W W0^2 g_W / (a_L L0^2), where a_W and a_L are the two sides’ kill rates, W0 and L0 their starting strengths, and g_W the winner’s breakpoint. The floors: 0.65 for Pickett’s Charge (closed form 0.6491; the simulation’s first win appears at 0.700, two steps up a 0.025-step grid), and 1.12 for Trafalgar, where the closed form says 1.1231 and the simulation says 1.125. That 0.2% gap is a floor-location match, the tightest analytic-to-simulation agreement in the project. Below the floor no value of phi wins.

How to read it: the vertical dial is the loser’s fighting effectiveness relative to history; the brass line is each battle’s capability floor.
What to see: below each floor the column goes black: no fraction of unbreakable force wins; Bull Run’s floor doubles because the sun sets first.

Above the floor, will buys real ground. Pickett’s Charge sat just under capability parity, which is why it is the one land battle in my set that unbreakable will flips. The charge starts winning once just over three-fifths (62%) of it cannot break, a deterministic 1%-step scan that carries no confidence interval; the honest uncertainty is the visible slope of the Monte Carlo curve around it, 6,000 runs per point.

How to read it: the curve is the Confederates’ chance to take the ridge as the unbreakable share sweeps up; the dashed line marks the flip.
What to see: the charge starts winning once just over three-fifths (62%) cannot break; below that, resolve only deepens the casualty bill.

Along the flip boundary the trade obeys a single square-law rule: the loser’s effectiveness, multiplied by a bracket mixing the unbreakable fraction with the enemy’s own breakpoint, stays constant at the floor value. In simulation that quantity holds within about 3%, and its level matches the closed-form floor within 8% for Pickett and 2.4% for Trafalgar. Those are boundary-agreement numbers, a different quantity from the 0.2% floor-location match above. One scope condition: this holds for battles that resolve a winner inside their historical battle duration. The third battle does not; its invariant runs about 1.9 times the analytic floor, a mean of 1.88 along its boundary (because it got dark and the simulated battle had to end).

How to read it: each row pairs one battle’s floor computed two ways: hollow ring the closed form, solid dot the simulation’s first win.
What to see: where the battle resolves inside its historical duration the two agree; Bull Run’s 40-hour clock displaces its floor to 1.43-1.45.

2. Bull Run adds a second floor: the sun setting. The closed-form floor for First Bull Run is 0.70. I am printing that number only to disown it. The closed form assumes unlimited time; the simulation caps Bull Run at a realistic 40-hour window, and the real battle ended at nightfall. Inside that window the simulated floor is not 0.70. It is 1.43-1.45. The out-fought Union side cannot win at any unbreakable fraction until its effectiveness more than doubles relative to the analytic floor (1.43 over 0.70 is a factor of 2.04; the 1.9 in finding 1 is the separate invariant violation), because the sun sets before attrition can finish the argument.

I could have quietly printed the prettier 0.70 and moved on. The honest number is better for the thesis, not worse. A time-capped battle needs capability even earlier, and gives will even less room to work. What 40% unbreakable Union troops buy at Bull Run is the denial of defeat: the AI contingent holds the hill into the night while the human 60% still routs on schedule. A stalemate, at every fraction, on a 1%-step scan. Never a win. Post 2 told that story from the morale side, the Union breaking at 8.3% casualties. The frontier tells it from the capability side: some battles are not yours to win, only to not lose.

3. AI at Trafalgar would just add to the butcher’s bill. The arithmetic a reader can redo. The dial rho measures the loser against its own history: Trafalgar’s floor of 1.12 means the Franco-Spanish fleet needs about 12% better gunnery than its crews actually delivered. Per my calibration it fired at about half the British rate, so the floor sits near 56% of British effectiveness; the fleet’s larger battery closes the rest through the square of numbers. The historical fleet is close to that line: at rho = 1 it sits 11% below the floor, five 0.025 grid steps on my sweep.

At its actual calibration, an unbreakable fleet converts defeat into bloodier defeat 83.6% of the time. Nelson’s two columns cut the 33-ship line in three, the van sails uselessly on, and concentrated British fire eats the two trapped fragments whether or not anyone aboard is capable of despair.

Two flip sides. First, 83.6% is not 100%: an unbreakable Franco-Spanish fleet gets a real 16% upset chance, up from essentially nothing, by staying in the fight long enough for variance to work. Second, the stubbornness is not free for the winner: British casualties roughly double, from about 28% of the force to about 50%, when the enemy fights to annihilation instead of striking. The Iwo Jima lesson, restated at sea. Below the floor, will changes the price of the battle, not its winner.

4. Your will is worth exactly zero against an enemy that does not break. Everything above holds the enemy’s morale at its historical setting. So I swept the winner’s breakpoint from very fragile to fully steadfast and measured the loser’s gain from being unbreakable, 6,000 paired runs per point.

At the steadfast end the answer is not approximately zero. It is exactly zero, 0.00%, in all three battles. That endpoint is definitional rather than empirical: unbreakable against unbreakable, the morale term cancels by construction, the fight reduces to pure attrition, and attrition was the thing the loser was losing.

The value of your own unbreakable will is single-peaked in enemy fragility, topping out at 86 points of win probability for Pickett with the enemy breakpoint near 25% casualties, and 92 points for Trafalgar near 40% (Monte Carlo standard error at most half a point; peak locations good to about one 5-point grid step; the full hump shows in two of three battles, with Bull Run’s peak at the edge of the tested range). Why a peak? Against a very fragile enemy you were going to win anyway; against a steadfast enemy it can never matter. Your will is not an asset you own. It is a bet on the enemy’s weakness.

5. In the missile age, will does not substitute for firing first. The battles above are attritional grinds, exactly where outlasting the enemy can pay. Modern naval and air combat is built around salvos and firing effectively first. So I moved the question into a stylized salvo model in the standard missile-age tradition (Wayne P. Hughes Jr., Naval Research Logistics, 1995; Michael J. Armstrong’s stochastic extension, Operations Research, 2005) and asked what an unbreakable crew is worth as per-salvo lethality rises.

Under an enemy first strike at saturation lethality, the answer is statistically zero: a premium of about 0.1 points of win probability against a Monte Carlo standard error of about 0.8 points at 4,000 trials per cell. You are crippled before your steadfastness can be spent. The conditional is joint: it takes the enemy firing first and high per-exchange lethality together. A first strike at low lethality still leaves a modest premium of 5-8 points. And in a fair duel near parity, lethality initially amplifies the value of will, up to a premium of 33-46 points near the one-shot threshold, because the unbreakable side converts mutual-annihilation draws into wins.

How to read it: extra win probability from an unbreakable crew versus per-salvo lethality; brass a fair duel, red an enemy first strike.
What to see: under a first strike at saturation the premium burns down to statistical zero; will does not substitute for firing first.

A cross-check rules out the boring explanation. Scaling both sides’ kill rates over a 32-fold range in the continuous model leaves the will premium flat at 45.2 points; the 0.6-point spread (the standard deviation across the six speed settings) is pure Monte Carlo noise; the square law makes this flatness exact in theory, so the run verifies the engine rather than discovering anything. Speed is not what erodes will. Decisiveness per exchange is, plus the first-strike race. If your procurement plan buys unbreakable resolve instead of sensing, range, and the first effective salvo, you have bought the tiebreaker and skipped the tie.

6. The 2020s weapon is not the 2030s army. I ran two counterfactual grades of AI: a 2020s version making one strategic weapon (the cannons) unbreakable, and a 2030s version making the whole force unbreakable, a robot army.

On land the 2020s weapon captures none of the robot army’s benefit: zero flips in 15,000 runs, an upper bound below 0.02%. The strategic/cannon is not the maneuver force; when the human line routs, the unbreakable battery is simply overrun, still bravely computing. At sea the same weapon-level AI captures 77% of the whole-force benefit (propagated-uncertainty interval 73-81%), because a ship is the maneuver force. Unbreakable will bolted onto one platform pays only where that platform is the force; everywhere else it has to live in the units that hold the line, and that is the robot-army problem, a 2030s problem.

How to read it: each vessel holds the full benefit of an unbreakable robot army; the brass fill is the share one unbreakable AI weapon captures.
What to see: on land the vessel is empty, under 0.02% in 15,000 runs; at sea, where the ship is the maneuver force, one weapon captures 77% of the army’s benefit.

Best Arguments Against This

Three battles is not a database. Correct, and I concede most of it. The floors are constants of specific battles under a specific calibration; three battles cannot establish how often real fights sit above or below their floors. But the population question was answered separately: the 607-battle audit in post 3 found roughly two-thirds (60.8-72.2% across threshold choices) of historical losses were capability losses. The frontier explains the mechanism; the audit shows it was the majority case.

Your floors live on grids. Also correct. The floors were confirmed at a rho step of 0.025 and an unbreakable-share step of 5% with a deterministic integrator, so every simulated floor is really a band one step wide. The gap has a verified direction: the simulated first win errs high, never low, so below the analytic floor no fraction wins in any run. The floors are conservative. The enemy-fragility peaks are grid maxima whose true values can only be higher. Nothing in the post survives or dies inside a grid step. The one place discretization could have hidden a real discrepancy, Bull Run, I printed as a discrepancy.

The salvo parameters are made up. The missile-age model is a notional 20-versus-20 duel with parameter jitter, not a fit to engagement data. What I defend are the shapes: the collapse under enemy first strike plus saturation, the fair-duel hump, and the flatness under pure speed scaling, which is analytically forced. If a data-calibrated salvo model moves the shapes, finding 5 is in trouble. The magnitudes were never the claim. My next series on AI-weaponry will dive much deeper into this; the modeling is done, the analysis is not.

Ukraine looks like drones winning a war, not tiebreaking one. The strongest current-events objection, and it cuts the other way. Reported estimates put drones at about two-thirds to 85% of battlefield casualties by 2025, figures that trace mostly to official sources and deserve that hedge, yet the front line barely moved: a kill-zone stalemate at rough parity, machine attrition setting the body count while neither side’s will flips the map. And Ukraine’s 2026 operational gains followed strikes on radars, jammers, and fires: capability degraded first. That is the model’s ordering: capability first, will as the tiebreaker at parity.

What Would Change My Mind

1. By end of 2027, a spatial or agent-based re-implementation of these battles that finds no capability floor, with winning unbreakable fractions appearing well below the square-law line.

2. Within twelve months of publication, a documented engagement, verified by two independent non-belligerent sources, in which a force clearly below capability parity won primarily because an autonomous element refused to break rather than because it out-shot the enemy.

3. During 2026-2027, a side in a missile-age exchange absorbing an effective first strike at saturation lethality and then winning on endurance.

4. Phase-separated casualty data for historical routs, fighting-phase versus pursuit-phase, showing most routed losers were at or above parity during the fighting. That would move battles into the close-fight column, where will decides.

5. By end of 2026, a naval historian or operations researcher showing my Trafalgar gunnery calibration is wrong by enough to lift the Franco-Spanish fleet above its 1.12 floor.

The close

Clausewitz made resistance the product of means and will. The temptation of AI is to read that as a multiplication you can win on either factor. You cannot. Below the capability floor the will term multiplies a number too small to save; against a steadfast enemy it multiplies by exactly zero; against a first strike it never gets to multiply at all.

Will is not a weapon. It is a tiebreaker, and tiebreakers only matter in ties.

If you are going to buy an unbreakable will anyway, where in the force should it live? The trooper is the obvious answer and the wrong one. That is post 5.

There are a good number of legal, policy, ethical, and even religious discussions asking if we should incorporate AI into weapons. I have been taking a different approach, projecting AI-characteristics onto well-studied historic battles, and using Lanchester equations to model what AI-weapons will do to battle. This series focuses on being able to embue AI weapons with an unbreakable will.

If you believe battles are decided by breaking the enemy’s will, the historical record disagrees with you most of the time. Across 607 usable defeats in a 660-battle database built for the US Army, roughly two-thirds (60.8-72.2% across reasonable threshold choices) were capability losses. The loser was so out-fought that it would have lost even fighting to the last man. And because prisoners taken in the pursuit get booked as winner effectiveness, that two-thirds is an upper bound, not a midpoint.

The true will-collapse signature, a loser at or above capability parity that broke early, shows up in about one-fifth of defeats (16-29% across the same thresholds). That is a lower bound. Correct it honestly, using measured rout casualties, and it rises to about a third (21-38%). It stays a minority under every correction I could construct.

Two-thirds of history’s losers were out-fought, not out-frightened.

How to read it: every dot is one defeat, 1600-1973, sorted into the three audit classes; the brass block is every will-collapse signature found.
What to see: two of every three losers were beaten on capability; pursuit booking makes that an upper bound.

There is a twist. The database’s compilers included a morale code, and checked against the will-collapse signature it carries essentially no information. Even the people who built the morale column could not make it predict.

The data comes from CDB90, the US Army Concepts Analysis Agency’s Database of Battles: over 600 land battles from 1600 to 1973, compiled in the 1980s by Trevor N. Dupuy’s Historical Evaluation and Research Organization (HERO), with per-side strengths, casualties, and outcomes. The research data was current as of July 1, 2026.

The findings

1. Clausewitz made will half the equation. Pop history made it the whole thing. Here is what he actually wrote: “If you want to overcome your enemy you must match your effort against his power of resistance, which can be expressed as the product of two inseparable factors, viz. the total means at his disposal and the strength of his will.” Means times will. Two factors, inseparable, multiplied. The sentence everyone repeats, that the object of war is to break the enemy’s will, is a paraphrase that dropped the other factor and grew teeth.

The paraphrase won because routs are cinematic and the winner’s version flatters the winner: we broke their spirit sounds better than we had more guns. The paraphrase is also half true, which is the durable kind of wrong. Armies really do quit long before annihilation. The Union quit First Bull Run at 2,896 casualties out of 35,000 present, about 8.3 percent; that story carried post 2. The question here is different. How often did the quitting, rather than the out-fighting, decide the result?

2. The audit is arithmetic you can redo. For every defeat in the database, take four numbers: each side’s starting strength and casualties. Lanchester’s square law (the standard attrition model in operations research, validated in this project’s first phase) turns them into one summary statistic for the loser, a capability ratio. In one line: R_cap = (loser per-man kill rate x loser starting strength squared) divided by the same product for the winner, with each side’s kill rate inferred from the casualties the other side suffered. At R_cap = 1, the loser had the fighting power to trade evenly. Below 1, it was out-matched on the day regardless of anyone’s nerve.

The classification rule is three lines. If R_cap fell below 0.70, the loss is booked as out-fought. If R_cap was at least 0.70 and the loser broke or quit before 15 percent casualties, that is the will-collapse signature. Everything else, near parity but bled past 15 percent, is booked as ambiguous: a hard repulse, not a collapse.

How to read it: the ratio multiplies inferred kill rate by strength squared, loser over winner; the cells below are the whole classification rule.
What to see: arithmetic you can redo from two columns per side, with thresholds declared before the data was scored.

The median out-fought loser had R_cap of 0.28, crushed on capability. The median will-collapse loser had R_cap of 1.24 and broke at 4.8 percent casualties. That is the textbook rout profile. Gettysburg lands in ambiguous, correctly: Lee’s army was near parity and fought to roughly 37 percent casualties. A hard repulse is not a panic.

3. Two-thirds of the losers were going to lose anyway. Of 607 defeats, 402 classify as out-fought. That is 66.2 percent, and it holds at 60.8-72.2% across the full grid of threshold choices (capability cuts of 0.6 to 0.8, break cuts of 10 to 20 percent casualties). The will-collapse class holds 130 of 607, 21.4 percent, ranging 16.0-28.7% across the same grid. The remaining 75 (12.4 percent) are the ambiguous hard repulses.

How to read it: each dot is one defeat, placed by the loser’s capability ratio (left of parity means out-matched) and its casualty fraction when it lost.
What to see: history’s losers mass far left of parity; the morale story lives in the thin brass wedge; Bull Run enters it only after the pursuit correction.

Drop any single war from the audit (there are 65) and the out-fought share moves only between 64.3 and 68.4 percent. Drop any entire century, including the whole twentieth, and it stays between 64.2 and 67.2 percent.

How to read it: each mark re-runs the classification with one war or century removed; vertical lines are the full-sample shares.
What to see: out-fought stays within 64.2-68.4% under every deletion, including the entire twentieth century.

One direction of error matters, and it runs against my headline, not for it. The database records total casualties, with no timeline inside the battle. When a loser breaks and runs, the prisoners and stragglers taken in the pursuit are booked as if the winner shot its way to them, making losers look more out-fought than they were during the actual fighting. A second channel pushes the same direction: an army that breaks early also stops inflicting casualties, which deflates its own measured effectiveness, so the upper-bound reading already absorbs it. Finding 5 measures the correction.

4. The database’s own morale column cannot find the collapses. CDB90’s compilers coded a morale advantage for many battles. If will collapse were the well-understood engine of defeat, that hand-coded column and my four-number arithmetic should point at roughly the same battles. They do not. The code flags the loser as demoralized in only 118 of 607 defeats, and of my 130 will-collapse battles it flags just 19. Agreement beyond chance, measured by Cohen’s kappa, is -0.06: statistically indistinguishable from zero (permutation p = 0.14, n = 607). It does not even flag First Bull Run, the most famous rout in American history. The field has long treated CDB90’s subjective columns with suspicion; the contribution here is measuring the failure cleanly, not discovering it.

I do not think the compilers were careless. The likelier reading is that their code captures pre-battle troop quality, green versus veteran, rather than what happened once the shooting started, and where it does fire it usually sits on top of a capability rout anyway (of the 118 flagged battles, 96 were out-fought on the numbers). But that is the point: professional military historians, building the Army’s reference database, could not make morale into a column that predicts. The database’s outcome codes, by contrast, agree with the arithmetic: all 26 losers recorded as annihilated classify as out-fought, and the will-collapse class contains no annihilations at all. Armies that collapse early do not fight to the death. The signature is coherent. The folklore column is not.

How to read it: left column, battles the database’s morale column flagged; right column, will collapses per the arithmetic; brass threads are the only agreements.
What to see: 19 threads out of 118 and 130; kappa -0.06, indistinguishable from zero. The morale column carries essentially no information.

5. The honest correction: half the loser’s casualties came after it broke. To size the pursuit bias flagged in finding 3, I collected the 11 famous routs whose published casualty figures split killed and wounded from captured and missing: Bull Run, Nashville, Missionary Ridge, Chancellorsville, Waterloo, Sedan, Rossbach, Königgrätz, Jena-Auerstedt, Blenheim, Austerlitz. These are deliberately selected severe routs, not a random sample; read every number in this paragraph as a bound, not a population estimate. In that sample, the share of the loser’s casualties that came after its line broke averaged 53% (n = 11, range 24-85%). Half the butcher’s bill was not combat. It was shooting retreating soldiers in the back. As Robert E. Lee said:

It is well that war is so terrible. We should grow too fond of it.

Remove those post-break losses and recompute. The average loser’s capability ratio roughly doubles, and 6 of the 11 battles change class, every one toward parity, none the other way. Bull Run is the cleanest case: about 45 percent of the Union’s casualty count that day was captured or missing men, and stripping them lifts its capability ratio from 0.795 to 1.42: an unambiguous will collapse at better-than-parity capability. The Union army at Bull Run was not out-fought. It was out-frightened. That is the exception, and it took the correction to make it clean.

How to read it: each row is one famous rout; the arrow is the loser’s measured fighting power once pursuit prisoners stop counting.
What to see: ratios roughly double and 6 of 11 verdicts flip toward morale; Sedan, Rossbach, Königgrätz, and Waterloo stay out-fought regardless.

Apply the measured correction to all 607 battles, reattributing casualties for losers whose recorded outcome was an actual rout or withdrawal, and the true will-collapse share rises from about one-fifth to about a third (21-38%, where 38% assumes an aggressive post-break fraction of 60%, above the measured 53% average). The control group inside the sample: Sedan’s loser moves from 0.10 to 0.37 and stays hopelessly out-fought. Sometimes an army runs because it is beaten, and the running merely sets the price.

6. A minority verdict, and it is the interesting minority. So the pop-Clausewitz trope fails its audit. In most defeats, will was not the deciding variable; capability was. In Ukraine, officials and analysts attribute about two-thirds to 85 percent of battlefield casualties to drones (the figures are mostly official-sourced; a June 2026 Oklahoma Watch fact-check rated them plausible but soft), yet after years of that attrition neither field army has collapsed into an army-scale rout. The front moves on capability: strikes, stockpiles, jamming. It is a firepower war with morale at the margins, which is what most of the 607 battles were too.

But flip the finding over. In about a third of historical defeats, at the corrected bound, the loser had the capability to trade evenly and lost anyway because somebody’s will gave out. Those are the battles where a combatant that cannot break, an AI, changes who wins rather than what winning costs. One battle in three is not a revolution. It is a large, specific, findable class of fights.

Best Arguments Against This

The anchor battle sits on the boundary you chose. Uncorrected, Bull Run’s capability ratio of 0.795 clears my 0.70 will-collapse cut by less than a tenth, and at a 0.80 cut it would classify as out-fought. If I had picked my thresholds to flatter the series’ anchor battle, this is where you would catch me. Three answers. The thresholds and the full 3-by-3 sensitivity grid are declared in the analysis script, not tuned after the fact. The population headline survives every cell of that grid. And the corrected Bull Run, at 1.42, is not sensitive to any threshold at all.

The database’s numbers are as suspect as its morale column. Fair. CDB90’s quantitative columns, the strengths and casualties this audit runs on, carry error too. Serious Lanchester curve-fitting moved to the purpose-built Ardennes and Kursk campaign databases partly for that reason (Bracken 1995; Fricker 1998; Lucas and Turkes 2004). Three things keep the classification standing on noisy inputs. It is a coarse two-way cut, not a curve fit, and it survives the full threshold grid. It survives dropping any single war or any entire century. And the class medians sit nowhere near the boundary: 0.28 for out-fought against 1.24 for will collapse, either side of a cut at 0.70. Noise would have to move battles across that whole gap, one way, to flip the shares. One caveat survives: any battle’s label is only as good as its entry. The claim is the aggregate shares, not any single battle’s coding.

Skill, not materiel or morale, decides modern battles. Stephen Biddle’s Military Power (2004) makes force employment the decisive variable, and that cuts across any raw capability-versus-will split. Conceded, and absorbed: the capability ratio is inferred from casualties actually inflicted, so employment quality is already inside it. R_cap is an effectiveness ratio, not a bean count.

Early withdrawal is not the same thing as will collapse. A prudent commander executing a planned withdrawal matches my behavioral signature, at or above parity and stopped before 15 percent casualties, without anyone panicking. Partly conceded. The class should be read as “broke or quit early,” not as a mind-reading of terror, and the outcome codes only partially separate the two. But the class has the right texture: its median member quit at 4.8 percent casualties while holding a capability edge, and it contains none of the database’s annihilation endings. Whatever mix of panic and prudence is in there, it is the ceiling on the morale story, and it is still a minority.

Eleven hand-picked routs cannot correct 607 battles. Correct. The 11-rout sample was selected on the dependent variable, so its post-break fraction of 53% is an estimate for battles like those, not for battles in general. That is why the correction reattributes casualties only for losers the database itself records as having routed or withdrawn, why the corrected share prints as a band (21-38%) rather than a point, and why the falsification list below asks for the real fix: a phase-separated re-audit at scale. Until that exists, the direction is one-way, toward morale, and the bound still leaves capability losses the majority.

What Would Change My Mind

1. By December 2027: someone runs a phase-separated re-audit, a machine-readable battle database splitting fighting-phase from pursuit-phase casualties across hundreds of battles (a digitized ACSDB or KOSAVE-class dataset would do it), and the true will-collapse share exceeds 50 percent. I would retract the headline, not just soften it.

2. By mid-2027: a replication of this audit under the linear (area-fire) Lanchester law flips the majority class on the same 607 battles. This is a weekend’s work for a skeptic.

3. Anyone produces a defensible reading of CDB90’s morale code under which it does predict the parity-plus-early-break signature, with agreement reliably above chance. Finding 4 dies, and I print the obituary. And please tell me how you did it so I learn how to better model it.

The close

A rout is not why most armies lost. It is how the losing looked. Two-thirds of history’s losers were out-fought, not out-frightened, and the most famous morale column in military-history data cannot find the exceptions. But the exceptions are real, they are about a third of defeats at the honest bound, and they cluster exactly where the capability ledger reads even.

That near-parity sliver is where an unbreakable AI stops being a rounding error and starts picking winners. Next post: what a will that cannot break is actually worth there, and why the answer is a tiebreaker, not a war-winner. The next AI-weapons series, the models are already working, focus on AI capabilities that out-fight.

The mental model most people carry from war movies is attrition: two armies grind each other down, and the one that runs out of soldiers loses. That perception is wrong for most of military history. Battles do not usually end when an army is destroyed. They end when its will breaks, and the will breaks early.

At the First Bull Run the Union army came apart after 2,896 casualties, killed, wounded, missing, and captured, out of 35,000 men present. Divide it yourself: 2,896 / 35,000 = 0.083. Eight point three percent. That figure is a historical datum, not a model estimate, so no confidence interval applies; the honest fine print is that roughly 45% of the numerator was captured or missing rather than shot, and the denominator counts men present, not the smaller number actually engaged. On an engaged basis the fraction runs higher. The point survives either accounting: over nine-tenths of the army was standing when it stopped being an army.

So I built the breaking point into the math. A Lanchester attrition simulator where every unit carries an explicit morale threshold, validated against four real battles: Iwo Jima, First Bull Run, Pickett’s Charge, and Trafalgar. Then I asked the question the whole series hangs on: what happens when one side’s threshold is removed, because we can model a soldier as an AI tactical weapons platform that does not panic, and we can model a canon or a ship as a strategic weapons platform that does not panic?

How to read it: wounded carried at the left, prisoners under guard, and the flood of unhurt men pouring toward a tiny Capitol.
What to see: the day ended in flight, not slaughter; most of the army walked away.

The visualization: First Bull Run, the rout that an unbreakable will turns into a stalemate.

How to read it: the left is history, the Union line breaks and routs toward Washington; the right is the same fight given an unbreakable will. What to see: the will stops the collapse but does not win. It grinds the rout into a bloody stalemate and raises Union losses from 8.3% to 12.5%.

What happens in the simulator? Pickett’s Charge succeeds 4.2% of the time in the baseline model (95% CI 3.9–4.6%, n = 15,000 runs). If you model the battle as if the Confederate assault was comprise of equivalent AI weapons with an unbreakable will then the assault wins 92% of the time (95% CI 91.6–92.5%). The rest are stalemates at nightfall; the AI-equivalent assault never loses outright. The caveat sits right here: these are attrition-model results with literature-calibrated breakpoints, not a spatial recreation of the battlefield, and the same experiment run on the other three battles shows the flip is the exception, not the rule.

And one of the four battles is not a counterfactual at all. Iwo Jima’s garrison actually fought without a breakpoint. In the model, that real unbreakable will roughly doubled American casualties and more than tripled the battle’s length.

The findings

1. The Union army broke nearly intact. Bull Run was the first major land battle of the Civil War, fought July 21, 1861, with spectators. McDowell’s 28,450 engaged Union troops faced 32,230 Confederates once Johnston’s army arrived by rail, so there was no Union numerical edge at the point of decision. Late in the day the Union withdrawal dissolved into a panicked flight toward Washington, entangled with the picnicking civilians, an episode the press made famous as the Great Skedaddle. The casualty ledger tells you what actually happened. Union losses: about 460 killed, 1,124 wounded, 1,312 missing or captured. Confederate losses: 387 killed, 1,582 wounded, and 13 missing or captured. Thirteen. That asymmetry, 1,312 against 13, is the signature of a one-sided rout. Both sides bled at similar rates while they fought. Then one side stopped fighting, and men who would have been shooters became prisoners.

How to read it: blue formations hold the hill, red arrows press, and the rout tapers off toward Washington; the inset bar splits the day’s casualties into killed, wounded, and captured or missing.
What to see: both sides bled at similar rates while they fought, then one side stopped, and 1,312-against-13 in prisoners is what a broken will looks like in a ledger.

2. Breaking early is the rule, and casualties are a lousy predictor of when. The foundational study is Dorothy Clark’s 1954 Operations Research Office report, which examined 43 battalion engagements in which units broke. She found casualties were probably not the primary cause of breaking, and in about a third of her cases the battalion had already absorbed over 40% casualties before it quit. Helmbold’s 1971 RAND review put historical breakpoints anywhere from roughly 10% to 70% of strength, with leadership, support, and communication mattering more than attrition. Two things follow. First, units break well short of annihilation, essentially always. Second, there is no magic number. My simulations sweep breakpoints across 8–40%, spanning the lower-middle of that observed 10–70% range; the sweep dips below the literature’s floor because Bull Run’s observed 8.3% break anchors the low end. That band is my modeling assumption, not a literature finding, and I will keep saying so. Clausewitz put the underlying idea in one sentence: an enemy’s power of resistance is “the product of two inseparable factors, viz. the total means at his disposal and the strength of his will.” The means are countable. The will is the part that fails first.

3. You can put the breaking point in the math, and the math checks out. The engine is Lanchester’s square law, the 1916 aimed-fire attrition model that operations researchers have used for a century, plus one addition: each unit routs when its casualty fraction crosses a threshold, and that threshold varies run to run. Validation is the part I care about. In 1954 J. H. Engel fit the square law to daily American casualty data from Iwo Jima and published attrition coefficients. My independent refit of the same daily series recovers A = 0.0565 against Engel’s published 0.0544. The R-squared is 0.992 on 35 daily points, but read that number gently: the fit is to the daily strength series, a cumulative and monotone curve that flatters R-squared by construction. The meaningful check is the coefficient landing within 4% of Engel’s. And the spine has survived a hostile audit before: Robert W. Samz’s 1972 Operations Research technical note revised Engel’s Iwo Jima data and found the revised data still fit. A small credibility bonus from doing the homework: a different value for Engel’s coefficient circulates in secondary sources, and it turns out to be a transcription error that has been copied around for decades. I pulled the 1954 paper scan and read the number off the page. The model also reproduces Bull Run’s rout at 8.3%, Pickett’s roughly 19% Union casualties, and Trafalgar’s zero British ships lost, before any counterfactual is run.

4. Iwo Jima already ran this natural experiment on unbreakable wills. The one force in the dataset with a genuinely unbreakable will was not an AI. It was the Japanese garrison of Iwo Jima, February 19 to March 26, 1945. Kuribayashi ordered each man to treat his position as his grave and to kill ten of the enemy before dying, forbade the banzai charges that had squandered earlier garrisons, and fought a defense in depth designed to maximize American casualties rather than to win. Of a garrison of roughly 21,000, only 216 were taken prisoner during the battle. That is a breakpoint of one: fight until dead. The model puts a price on it. Give Japan a normal human breakpoint and simulated US casualties fall from 21,431 to 10,969, a ratio of 1.95x, and the battle shortens from 36.4 days to 10.7, a ratio of 3.4x (n = 15,000 runs per condition; the Monte Carlo means are stable to a couple of men). Those are model outputs on Engel’s assault-echelon accounting basis, where the historical figure is 20,860 over 36 days; the roughly 26,000 casualties you may have read elsewhere is the all-services total for the whole operation, a different basis, and the two must not be mixed. The sentence I want a planner to carry out of this post: a real unbreakable will already existed once, and it doubled American casualties. It did not change who won.

How to read it: the cool low column at the left is the counterfactual, the towering brass-lit column the real battle, Suribachi and the landing craft between them.
What to see: the will did not move the outcome, only how long and how bloody it was. Roughly 2x the casualties across 3x the days.

5. At Pickett’s Charge, the breakpoint was the whole battle. On July 3, 1863, 11,481 Confederate infantry crossed nearly three-quarters of a mile of open ground against the Union center and were repulsed in about an hour, losing roughly half their number. In the model the charge succeeds 4.2% of the time (95% CI 3.9–4.6%, Monte Carlo); the validation ensemble puts it at 4.9%, and the 2022 continuous-flow simulation study of the charge by Poggie, Matei, and Kirchubel puts it near 6%. Attribute each number to its method and they agree: a long shot. Now move the will around. If the Union artillery had been weak-willed and fled at 10% losses, the model makes the charge a coin flip: 49.1% Confederate win (95% CI 48.3–49.9%). The historical guns did not flee. At the wall called The Angle, Lieutenant Alonzo Cushing, twice wounded, refused evacuation and kept his last guns firing until he was killed; he received the Medal of Honor for it 151 years later, in 2014 and if you’re building an AI enabled strategic weapon you should name it after him. The model says that steadfastness was worth about 45 points of win probability. And if the entire Confederate assault cannot break, it takes the ridge in 92.0% of runs (95% CI 91.6–92.5%) and never loses; the remaining 8% are stalemates at nightfall, not Union wins.

How to read it: the same field painted twice; above is what happened, below is the assault if the line could not break.
What to see: the same men with the breaking point removed cross the wall and win 92% of runs. Proportions follow the measured 92% win rate.

Between those poles sits a threshold. Sweep the AI unbreakable-will share of the assault upward and the winner flips at just over three-fifths, 62% on a deterministic 1% scan, with the Monte Carlo curve crossing even odds at the same spot. Below that share, the extra AI-resolve buys deeper casualties and the same defeat.

How to read it: Sweep the share of the assault that cannot break from none to all; the curve is the simulated chance the Confederates take the ridge, and the brass point is where the model’s winner flips.
What to see: Unbreakable AI-will is a threshold, not a dial; below just over three-fifths (62%) the added resolve only deepens the casualty bill.

6. An unbreakable AI-will is not a cheat code. Run the same experiment on the other battles and the flip disappears. Make 40% of the Union force at Bull Run unbreakable and the rout becomes a stalemate; the human 60% still breaks on schedule, the AI contingent holds the hill into the night, and no fraction up to 100% converts holding into winning at the army’s measured effectiveness. Make the whole force unbreakable and more than 99.9% of 15,000 runs still end in stalemate; no run produced a Union victory, so the residual is a statistical bound, not an observed event. At Trafalgar, even a Combined French-Spanish fleet that never strikes its colors loses 83.6% of the time (95% CI 83.0–84.2%), because Nelson’s gunnery and concentration decide the fight before any morale check matters. Why will is usually a tiebreaker rather than a war-winner is the next post but one; for now, the pattern is enough. Removing the breaking point reliably changes the body count. It changes the winner only when the fight was close on capability to begin with.

7. Men are surrendering to machines right now. In August 2024 a Ukrainian brigade used a drone to walk a wounded, abandoned Russian soldier through a surrender: it dropped water and written instructions, then guided him to Ukrainian lines. By March 2026, Ukraine’s digital-transformation minister Mykhailo Fedorov claimed drone units had compelled more than 100 Russian soldiers to surrender over the winter alone; that is a belligerent’s own count, so read it as an order of magnitude, not a ledger entry. But the mechanism is the one in this post. A man does not surrender to a drone because he ran out of bullets. He surrenders because his will broke against something that does not have one.

Best Arguments Against This

The strongest objection goes at the model’s heart: I trigger routs on casualty fractions, and the foundational literature says casualty fractions are a poor predictor of breaking. Clark found leadership, support, and communication mattered more; a third of her broken battalions had already passed 40% losses before quitting. So the model uses, as its trigger, the very variable the best empirical study demoted. I partly concede this. The defense is threefold. The threshold in each baseline is calibrated to the level at which the historical unit actually broke, so it is a summary of whatever really caused the break, not a theory of it. The thresholds vary randomly across Monte Carlo runs, encoding Clark’s finding that breakpoints scatter widely. And the counterfactual question, what changes when the break is removed entirely, does not depend on what triggers the break, only that one existed and the record shows roughly where it fired.

Three smaller objections deserve print. The model is temporal, not spatial: no terrain, no flanking, no reserve decisions, so a battle it flips might not flip on real ground; an agent-based recheck is the single most likely thing to damage these results. The calibration inherits a bias: coefficients fit to total casualties, including losses taken during the rout and pursuit, understate the loser’s fighting capability. I hedge that with effectiveness sweeps here, and a later post returns to it with corrected data. And the campaign-level empirical record for Lanchester models is contested: Bracken (1995) found the linear law fit the Ardennes campaign, Fricker (1998) refit it and found neither law fit, and Lucas and Turkes (2004) found no basic law fits Kursk or the Ardennes well. That record is exactly why this series calibrates each battle separately and prints orderings with bands, not point forecasts.

What Would Change My Mind

First, a spatial or agent-based re-implementation, with terrain, suppression, and local maneuver, that reverses the Pickett flip or the Iwo Jima cost doubling.

Second, evidence from Ukraine through 2027 of human units routinely fighting to near annihilation without machine substitution, meaning breakpoints near 100% at scale. That would say human will is far more elastic than the 10–70% historical band, and the whole breakpoint frame is miscalibrated.

The method, briefly

A battle is not a subtraction problem. It is a race between two breaking points, and the side that reaches its own threshold first loses everything it has not yet lost. Hold that model, but hold it loosely: next post, an audit of 607 usable defeats from a roughly 660-battle database asks how often morale actually was the thing that broke, and the answer embarrassed me.

So I decided to project AI weapon characteristics into the very well studied historic battles, model them using Lanchesters’ equations, and see if I could project the past into the future to answer my questions.

The seven post series makes an assumption AI weapons an be modeled like human weapons whose will never breaks. They perform as they were programed (or trained) regardless of what happens. So the series asks one question: what is a will that cannot break actually worth in war? Not as a slogan. As a number, with error bars, in a model a military operations researcher would recognize.

The compressed answer, which the next six posts earn: an unbreakable will is real, measurable, and occasionally decisive. It is also narrow, conditional, and attackable. It decides fights that were already close on capability, and almost nothing else. Its value shrinks at every step up the ladder from the trooper to the commander to the force to the nation. And at the top of the ladder it can invert, into a way of losing wars while winning battles.

The visualization that opens the series. Iwo Jima, the one battle where the unbreakable will was real.

How to read it: one continuous history with a running ledger, real unbreakable casualties climbing beside the counterfactual where the garrison broke at a normal breakpoint. What to see: here the unbreakable Japanese will was not a hypothesis. It roughly doubled the American dead, about 21,431 versus 10,969 in the model, and changed nothing about who took the island.

Now the concession, because it belongs right up front. These are attrition-model results. The engine is a Lanchester system, aimed-fire attrition plus an explicit morale breakpoint, the casualty fraction at which a unit quits. It was calibrated to four historical battles and checked at the aggregate level: total casualties, duration, outcome. Three of those four are calibration plus consistency checks; the genuinely out-of-sample test is the independent Engel refit below. It has no terrain, no perception, no fog, no time resolution below the engagement. The strategic posts in the back half of the series run stylized extensions on top of that validated engine, and those results print as orderings and directions, never as calibrated magnitudes. If you want a model of what will is, this is not it. It is a model that assumes an AI weapon can think, reason, plan and act just like a human does, but the AI’s will never fails regardless of how poorly the battle is going. That narrower thing turns out to be answerable.

Why trust the tool at all? Lineage, mostly. Lanchester published the equations in 1916. Engel validated them against Iwo Jima’s daily casualty data in 1954, and when Samz revised Engel’s data in a 1972 Operations Research technical note, the revised data still fit. My independent refit of the same series recovers his coefficients (A = 0.0565 against his published 0.0544, R-squared 0.992 on 35 daily points). The historical audit in post 3 runs on CDB90, the roughly 660-battle database Trevor Dupuy’s researchers compiled for the US Army in the 1980s. Every piece of this apparatus is standard issue in operations research. The only new part is the switch that turns a side’s breaking point off. And that models one characteristic of AI weapons.

Bottom Line Up Near-the Front

Yes. In the case of Pickett’s charge, replacing Confederate soldiers with equivalent AI solders whose will never broke probably would have changed the battle and possibly the war. How to read the above: both ribbons are the same 11,481-man assault, width = remaining strength over the same clock; the dotted hairline marks the moment history stopped. In fact, the modeled AI assault wins 92% of simulated runs and never loses.

The frame

1. Clausewitz made will half of war. We are about to automate that half. Here is the sentence, in the Howard and Paret translation, that the whole project stands on: “If you want to overcome your enemy you must match your effort against his power of resistance, which can be expressed as the product of two inseparable factors, viz. the total means at his disposal and the strength of his will.” Means times will. We will study battles where the means were extensively counted: men, guns, ships, rates of fire. The will is the factor that fails first, and it has failed in every army ever fielded. The empirical literature on breakpoints (Clark 1954; Helmbold 1971) puts the historical range of casualties-at-breaking anywhere from roughly 10 to 70 percent, with leadership and communication mattering more than the count of the dead. The range is wide. The existence of the breaking point is as close to a constant as war offers. Or at least has offered in the past.

An autonomous weapon does not have a mental breaking point. Whatever else an AI-driven force turns out to be, it is the first combatant in history whose participation in the fight does not depend on its nerve. One of Clausewitz’s two factors is becoming a design parameter, something you specify at procurement rather than cultivate in barracks. That converts an old rhetorical question into a measurable one, and it deserves measurement before the money moves. This project is my attempt.

How to read it: Clausewitz factored war into means times will; the WILL formation erodes as the modeled force nears its breaking point.

What to see: one of the two factors fails first in every army ever fielded, and it is about to become a design parameter. AI-generated illustration; the WILL erodes to the model’s 28.6% breaking point.

2. What we built, and why these four battles. The instrument is a Monte Carlo combat simulator: Lanchester square-law attrition, each unit carrying a morale threshold that varies run to run, and the ability to mix breakable humans with unbreakable machines on one side. The test set is four battles, chosen so that each stresses a different joint between will and firepower.

• First Bull Run, 1861: a pure rout. The Union army came apart at 8.3 percent casualties, 2,896 of 35,000 present counting killed, wounded, missing, and captured. The test: can the model break early, the way real armies broke?

• Pickett’s Charge, 1863: a repulsed assault at near-parity firepower, the regime where a breakpoint should matter most. The test: does the model make the charge the long shot it was?

• Trafalgar, 1805: a decisive naval concentration, where British gunnery and Nelson’s geometry did the deciding and Britain lost zero ships. The test: does the model refuse to credit will where will was not the story?

• Iwo Jima, 1945: the control case, and the reason this series is not counterfactual fiction. Kuribayashi’s garrison genuinely fought without a breakpoint: of roughly 21,000 defenders, 216 were taken prisoner during the battle. History already ran one experiment with an unbreakable will, and Engel left us the daily data on what it cost.

How to read it: each panel is one validated battle; the X is where the losing force’s will broke, the caption the historical outcome the model had to reproduce.

What to see: four very different fights, one shared failure mode, and a single experimental move: turn the breaking point off.

The engine had to reproduce all four before any counterfactual ran: the rout at 8.3 percent, the charge repulsed with roughly 19 percent Union casualties, the zero British ships, and Iwo Jima's daily casualty trajectory. The charge's 4.2 percent baseline odds (95% CI 3.9-4.6%) are what the matched engine then reports: model output, not a validation target.

How to read it: each row is one check; a dot inside its ring is a bullseye, and the short brass bar is the miss distance on that row’s own zero-based scale.

What to see: one engine lands on or beside eight very different targets across four battles, before the morale switch was ever touched.

3. The experiment: turn one side’s breaking point off. Once the engine matched history, the counterfactuals ran in escalating doses. Make one strategic weapon unbreakable AI, a battery that never abandons its guns, to approximate the 2020s approach of embedding AI into strategic weapons first. Then make a contingent unbreakable, and sweep the fraction from a few percent to the whole force: the 2030s proverbial AI/robot-army. Then stop assuming the enemy cooperates. Let him jam the machines’ wiring, decline the fights he would lose, choose the kind of engagement he prefers. Finally, put a nation behind the unbreakable army, with a nation’s very human tolerance for cost, and let the campaign run to a political ending. Each escalation is one post. Each post has one claim I would defend in a hostile meeting.

4. The map of the series. Six posts follow. Here is what each one argues.

Post 2, 91.7% of the Army Was Still Standing When It Ran. Battles end when will breaks, long before armies die: the Union quit Bull Run with over nine-tenths of its men still standing, and a robot Pickett’s Charge takes the ridge in 92 percent of simulations (95% CI 91.6-92.5%) and never loses outright. Iwo Jima’s real unbreakable will roughly doubled American casualties in the model (21,431 against 10,969 simulated) and changed nothing about who won.

Post 3, Out-Fought, Not Out-Frightened. An audit of 607 defeats in the Army’s own database finds about two-thirds (60.8-72.2 percent across threshold choices) were capability losses, and genuine will collapses were a fifth, at most a third after honest correction. Two-thirds of history’s losers were out-fought, not out-frightened.

Post 4, ADM Nelson Beats the Robots. Every battle has a capability floor below which no fraction of unbreakable force wins, and against an enemy that does not break, your own unbreakable will is worth exactly zero. Even a Trafalgar fleet that never strikes its colors loses 83.6 percent of the time (95% CI 83.0-84.2%).

Post 5, The General Quits First. Historical battles ended on a commander’s withdrawal ten times as often as on the troops’ annihilation (266 to 26 of 607 recorded defeats), and a machine’s will is only as unbreakable as its jammable, hackable substrate. Your robot’s will lives in its radio.

Post 6, The Unexpected Deterrence. A rational enemy declines roughly 40-58 percent of the fights an unbreakable force would win and keeps the ones he takes; let him choose the kind of fight and will’s measured value collapses by about 97 percent in the model. A Great General will not fight your unbreakable army. They will fight around it.

Post 7, Win Every Battle. Lose the War. The capstone: a reviewer’s attack on the model, my correction, and the enemy’s answer to the correction map four regimes, and every one is decided by what the nation can bear, blood where the enemy can reach it, treasure and time where he cannot. The series’ honest scorecard and all nine operating rules live there.

Best Arguments Against This

The strongest objection covers the whole series, so it goes here, at the front: an aggregate attrition model cannot capture what will actually is. Will is perception, leadership, cohesion, exhaustion, the rumor running down the line. My model compresses all of it into one number, a casualty fraction at which a unit quits, and the foundational empirical study of breakpoints, Dorothy Clark’s 1954 work on 43 broken battalions, found casualty levels a poor predictor of when units break. So the model’s trigger variable is the one the best field data demoted. I concede the compression fully. RAND’s Will to Fight framework (Connable and colleagues, RR-2341-A, 2018) shows what the full-mechanism alternative looks like: 29 factors, unit by unit, from leadership to cohesion to supply. Their model is a model of what will is. Mine is one threshold, because my question is not what will is made of but what removing its failure point changes.

Three things keep the results standing anyway. First, the design: the counterfactual question is not what triggers a break but what changes when no trigger exists, and for that the model needs only that the historical break happened at a measurable point, which the record supplies. Second, the validation: the engine reproduced four very different battles, including an independent recovery of Engel’s 1954 coefficients, before the morale switch was ever thrown. Third, the audit: every headline number in this series was reproduced from committed data and stress-tested before drafting, dropping one war at a time, one century at a time, rerunning under fresh seeds, re-scanning grids at finer steps. Where the audit’s numbers moved under stress, the moved versions are the ones that print. What the concession still buys the skeptic, the spatial and perceptual texture of real combat, is named below as the likeliest thing to change my mind.

What Would Change My Mind

These are the series-level triggers; each post carries its own.

1. By end of 2027: a spatial or agent-based re-implementation of these experiments, with terrain, suppression, and local maneuver, that overturns a core ordering: the Pickett flip, the capability floor, or the capstone’s never-helped result. This is the single most likely falsifier.

2. By December 2027: a machine-readable re-audit of historical defeats that separates fighting-phase from pursuit-phase casualties at hundreds-of-battles scale and finds true will collapses were the majority. Post 3’s headline dies, and much of the series’ framing with it.

3. Within a year of publication: verified battlefield evidence of an unbreakable autonomous force flipping an engagement it was clearly losing on capability, by refusal to break alone. That breaks the tiebreaker thesis at its center. Keep an eye on Ukraine.

4. Any time: an independent refit of Engel’s 1954 Iwo Jima daily series that fails to recover coefficients near his published values. That is my validation lynchpin.

The close

For two centuries strategy treated the will factor as fate: a commander could cultivate it, but nobody could specify it. That era is ending on a procurement schedule, and the right response is neither the promise that armies which cannot panic cannot lose, nor the reflex that nothing fundamental has changed. The right response is to measure and adapt.

Measured, it comes to this. An unbreakable will is not a war-winner. It is a tiebreaker, purchased at strategic risk: it decides fights that were already close, only where the enemy allows the fight, only while its wiring holds, and the bill goes to a nation whose own will remains the last, and deciding, breaking point.

Next post: the army that ran from Bull Run with 91.7 percent of its men still standing, and what happened when I built its panic into the math, then took the panic away.

A scope note, because this is not a normal policy piece

I tried to audit American over-classification the way you would audit any system you expect to have been gamed: Inspector General reports, classification-cost ledgers, appeals-board dockets, clearance-revocation decisions, federal award archives, the patent office’s secrecy-order lists, a benchmark of AI models against signed classification rules. Not opinions about secrecy. Paperwork. Evaluate the n-th order effects because it takes quite the mastermind to game the system and cook all of the side effects out of others’ books.

Halfway through I had to change my method. The security apparatus controls the records that would prove or acquit it. It does not publish the audit. It stopped publishing its own cost. Its rulebooks (security classification guides) are themselves classified even where the source documents have been declassified. That is not a research inconvenience. That is the structure of a denied area, and the right way to write about a denied area is the way the intelligence community writes about one: as an assessment, with confidence levels attached. So I have two intelligence assessment research projects in flight right now: one on finding and evaluating embarrassing data that China hides, and one on finding and evaluating data the US national security community hides. Is it embarassing to them?

So that is what this is. When I say high confidence, multiple independent lines of evidence converge and I would be surprised to be wrong. Moderate confidence means the evidence is credible but single-method or carries a known gap. Low confidence means the judgment is plausible and internally consistent but rests on something unmeasured. And here is the part policy readers should not skip: a low-confidence judgment is not a shrug. In the intelligence world it is a collection tasking order. It tells you exactly what to go collect. The end of this piece is a government collection plan to provide civilian oversight of the national security establishment.

The research data was current as of July 2026. Dollar figures are US federal budget dollars unless noted.

Bottom Line Up Front

Here is what a decision-maker should now believe.

The incentive facing every American classifier almost certainly punishes under-protection and never punishes over-protection. That is no longer testimony. I searched eight public discipline channels spanning roughly forty-seven years, including all 32,213 published industrial clearance decisions, and found zero sanctions for classifying too high, against 163 career-ending sanctions for protecting too little in the same corpus. High confidence.

When an independent, fully cleared referee re-examines a contested over-classification, the classifier loses three times out of four. The appeals panel the system built for exactly this purpose overturned 75.2 percent of the contested decisions it reviewed across twenty-two years. Contested cases are a selected sample, and I will say so again below. But the direction is not subtle. High confidence in the record; the population rate remains unmeasured.

How much of the classified pile never needed protecting? The honest answer is a range that the government could collapse any time it chooses. The provable floor is 3 to 8 percent, and that floor rests on auditors who could strictly adjudicate only one-seventh of what they reviewed. Every official and insider who has ever put a number on it, from 1956 to the ISOO (Information Security Oversight Office) director himself, said half or more. I assess with moderate confidence that the true rate sits well above the floor, and I put the documented range of expert belief, not my model, on the chart.

The cost is a slope, roughly 12 billion dollars of budget per unit of excess-classification share, and where you sit on it is a belief about the one number nobody measures. The innovation damage is not a belief: prolonged compulsory secrecy destroys 20 to 36 percent of follow-on value, rising past half for the longest orders, and that result now survives every matching design I could throw at it. High confidence.

And there is a reform whose core does not depend on measuring any of this, because its job is to produce the measurement. Most of it can be done by Executive Order and does not require legislation. I know, still a high bar.

1. The bet is lopsided, and now it is counted

The classification system asks millions of cleared people, tens of millions of times a year, one question with wildly asymmetric stakes. Leave something open that turns out to matter and it can end your career. A revoked clearance is not a bad review. It is expulsion from the field that pays you tens of thousands of dollars a year more than comparable uncleared work.

Protect something that needed no protection, and nothing happens. Not rarely. Never. The census: Merit Systems Protection Board decisions, Office of Special Counsel reports, the Inspector General corpus, Defense Department and Department of War administrative releases, 32,213 industrial clearance decisions across thirty years, Government Accountability Office records, and the sanction-reporting sections of thirty annual oversight reports. Zero recorded sanctions for over-classification, in any public channel, ever. The two formal complaints in the record were both dismissed. Meanwhile the same clearance corpus contains 163 sustained mishandling sanctions, and where the mishandling allegation was proven, 110 of 111 people lost.

I began this project believing the driver was money: security officers protect the salary by over-protecting the documents. The model says that is not it. If you sanctioned over-classification with the same career stake, the stake cancels. What is left is the difference in the odds of getting caught, and one side of that ratio is now a measured zero in every public channel. The formal rules authorize sanctions in both directions. The realized practice runs one way.

The formal overclassification correction channel confirms the flow. The rules invite any authorized holder to challenge a classification they believe improper, with formal protection from retribution. Relative to the tens of millions of classification actions a year, formal challenges are almost certainly vanishingly rare. Why is a question the public record cannot answer. I assess with low confidence that anticipated career consequences do the suppressing, the rates are unmeasured, and the protected practitioner survey that would settle it sits in the collection plan below.

2. When a referee finally looks, the classifier loses three times out of four

The system has an independent referee. The Interagency Security Classification Appeals Panel (ISCAP) is cleared, sees the underlying Security Classification Guides (SCGs), and its decisions cannot be reversed by the agency it overrules. It is the closest thing that exists to an incentive-neutral second opinion.

From 1996 through 2017 it re-examined 2,935 contested documents and sided against the classifying agency on 75.2 percent of them, releasing 31.4 percent entirely unclassified. The rate held in every era and rose in the most recent one. Even the agencies’ own internal declassification reviews, judging themselves, release about nine requests in ten at least in part.

The caveat, stated plainly: these are contested and aged documents, the cases somebody cared enough to appeal more than they cared about retribution from security staff. They do not measure the whole pile. What they measure is what happens to a classification decision when the career stake is removed and someone qualified actually checks. That is precisely the experiment the incentive story predicts, run 2,935 times.

3. Everyone who ever put a number on it said at least half

You have probably seen the claim that the government over-classifies about 65 percent of what it protects. In my own earlier work that number was an assumption. I have now done what should have been done first: collected every documented numeric estimate of the genuinely-sensitive share ever made by an official or practitioner in a position to know, twenty-seven sources, fourteen of them numeric, from a 1956 Defense Department review to the 9/11 Commission’s chairman to the man who ran the government’s own classification-oversight office, who told a Senate hearing the over-classification rate was, in his words, beyond 50/50.

Every at-creation estimate since 1956 puts the excess at half or more. The distribution is honestly bimodal: senior officials cluster near 50 to 55 percent, career insiders and document-level reviews cluster near 90. The weighted median of documented belief is 83 percent, with a 90 percent band running from 38 to 99, which is the honest width of a question nobody has measured. My old 65% sits in the saddle between the camps, and a deliberately skeptical weighting reproduces it almost exactly.

This is an elicitation, not a measurement, and I hold it at moderate confidence for exactly that reason. But the burden has moved. The claim that over-classification is rare is now the claim that every official who ever went on the record about it was wrong in the same direction.

4. Nobody can measure it, and the government has stopped trying

I hit the measurement wall four independent ways, and the wall is the finding.

When Congress ordered every classifying agency’s Inspector General to measure over-classification directly, the resulting audits could strictly adjudicate 523 documents out of the 3,797 they reviewed, one-seventh of their own sample. Their provable floor, 7.8 percent, describes that sliver; allow the unadjudicated mass to err at anywhere from zero to twice the proven rate and the floor itself swims between 1 and 15 percent. Not one of the thirty-three reports produced a representative, independently adjudicated rate. The Intelligence Community published no rate at all.

I tried to audit declassified documents against the actual signed classification guides they cite. Of 145 public documents naming their guide, the number whose guide was also public was zero. The rulebook itself is classified, so the rule can never be checked from outside.

The oversight office that once published what the classification enterprise costs stopped in 2017, after telling the Senate it was not confident in its own last number. My reconstruction puts the enterprise very likely at 23 to 29 billion dollars in 2025, a judgment I hold at moderate confidence because it is anchored to the last official figure and bridged forward, and about 92 percent of it no longer corresponds to any public budget line.

A system built to resolve every doubt toward secrecy is, by construction, a system that cannot audit itself. I assess with high confidence that the true rate cannot be measured from outside, and with moderate confidence that no agency currently measures it inside either. If a favorable representative audit existed, publishing its aggregate would cost nothing, would answer people like me, and it would reign in the security apparatus that has confused security for National Security. The audit has never appeared.

5. The bill is a slope, and here is how to read it honestly

Every cost channel scales with the same unmeasured quantity, the excess share, so the honest cost object is a slope: roughly 12 billion dollars of budget per unit of excess-classification share. The base it multiplies is no longer an assumption. The Pentagon’s own budget exhibits carry explicit classified-programs lines totaling 44 billion dollars in 2019 rising to 62 billion in 2025, a steady fifth of the Pentagon’s acquisition spending.

Read the slope at the audited floor and the excess costs about 0.9 billion a year. Read it at the modeled center and it is about 4 to 5 billion. Weight it by the documented beliefs in the chart above and the median is 9.1 billion, with an honest band of 4 to 14. Only about 60 cents of each dollar is real resource loss; the rest is transfer. A budget hawk and an economist should read different numbers off this figure, and both should say which one they are reading. To be explicit about my own confidence: the slope I hold at moderate confidence, the level at low, because the one number that picks your point on the line is the number nobody measures. Low confidence on the level is the data collection tasking.

I will also say what did not survive my own checks. Clearance-required Requests for Proposals (RFPs) solicitations do not draw measurably fewer bidders at advertised competitions. Classified-intensive weapons programs show no average performance penalty across 651 program-years, though that sample by construction excludes the most classified programs. What survives is narrower: work that migrates into closed task-order channels carries a 6 to 9 percent price premium, a premium that could partly reflect vendor specialization rather than classification since that separating test remains unrunable, and the failure tail is ugly in a way averages hide.

6. Long secrecy destroys most of what it delays

One channel I can defend with real identification. Under the Invention Secrecy Act, patent applications judged sensitive are barred from disclosure. I matched 3,141 patents whose secrecy orders were later rescinded against roughly 49,700 controls granted the same year in the same technology class, and compared forward citations from the moment both were public.

Secrecy orders shorter than three years are a built-in natural experiments, since ordinary patent pendency ran two years and a short order delayed nothing. The natural experiment shows zero deficit, in every design. Then the response climbs: roughly 25 percent of follow-on value gone at three to five years under order, 40 percent at six to ten, 60 percent beyond ten. Pooled, the deficit runs 20 to 36 percent depending on how strictly you match, and matching patents by the actual text of their abstracts lands at 33 percent with the treated patents proving topically ordinary. Knowledge released late is not delayed value. Most of it is destroyed.

I hold this at high confidence as a well-identified association. And the institutional trend runs the wrong way: the stock of secrecy orders is at an all-time high, while rescissions have collapsed from hundreds a year in the early 1990s to a few dozen.

7. The security pocket veto is real; its frequency is the intelligence gap

What security never says “yes” it becomes a “no” whether that is a deliberate pocket veto or just an overworked security team is something we can look at. There are six documented episodes that pass a strict four-part test: an upward security action, an incompatible channel, a linked operational effect, and a later competent review that judged the block wrong. The Army cleared a book called Operation Dark Heart; a late intelligence-agency override led the Pentagon to buy and destroy 9,500 printed copies; the Pentagon’s own re-review then lifted 198 of the 433 disputed redactions. Another manuscript sat frozen until two weeks after the author sued, at which point 85 percent of it was released. The pattern repeats back to the Pearl Harbor intercepts.

The silence channel is now quantified too. In 1982 the CIA answered manuscript reviews in 13 days on average under a 30-day clock. By 2016 its own inspector general projected book-length reviews at more than a year. Submissions grew from 43 a year to more than 8,400. The clocks remain legally non-binding, so silence still functions as a denial that no official ever has to own.

Existence: documented, moderate confidence. Prevalence: unmeasured, low confidence, and that is a tasking, not a concession. The registry that would count these events is specified below.

8. What AI failed at, but it failed forward in a way we can use

I benchmarked four AI models as derivative classifiers against 879 signed classification-guide rules. All fail the accuracy bar the reform’s own math demands, the best at 94.4% against a required 98.6%. All are anti-calibrated, meaning they’re about as confident wrong as right which kills model confidence as an “escalate to a person” trigger.

Two deeper results matter more than the failure. Judging blind, the models classified 42 to 64 percent of a genuinely all-unclassified guide’s topics. Handed the correct security classification guide, that reflex collapsed to essentially zero, in every model. It could decide unclassified. An uncertain adjudicator hedges toward secrecy exactly like the human the incentive model describes, and the right classification guide removes the hedge.

But hand it the wrong classification guide and the reflex averages 73 percent and never falls below about 53, and grounding in an all-classified guide is worse than no guide at all. The AI models anchor on the label distribution they are shown. So the honest engineering claim narrowed: the tool disciplines the reflex only when retrieval hands it the correct SCG, which means guide retrieval itself must be part of any accreditation.

The critical insight is that the models’ disagreement is informative even though their confidence is not. If you make the AI models think about it and estimate a probability that they’re right, you the average their own probabilities, and if there isn’t enough consensus in confidence you elevate the decision to people. The system auto-decides 86 percent of the workload while wrongly answering 2.16 percent of what it decides, escalating the hard 13.78 percent to people. Truth in statistics: that 2.16 carries a confidence interval of 0 to 4.93, so a guarantee that holds at the interval’s edge costs roughly 60 percent escalation on my benchmark, and the whole result assumes the tool was handed the correct guide. I built the proof of concept in two hours during the USA-Belgium World Cup match so some dedicated engineering with real test data will likely perform far better.

The reform: relocate the no, and make it produce receipts

If the disease is a rational officer resolving every doubt toward secrecy for their self-protection, the cure is to reprice the doubt. Today, clearing a document as unclassified is a personal career bet, while stamping it classified, or just sitting on the request, is free and the latter leaves no record. The reform does exactly two things about that. It takes the free act, the “no, this is classified” away from individuals and gives it to a system that keeps receipts. And it makes the careful act, the “yes, this is unclassified” safe to perform.

Here is the reformed process. A human security officer can still say yes to openness and operational necessity at every step: approve a release, grant a downgrade, clear a manuscript, agree a paragraph needs no stamp. Nothing about that changes, and nothing new gets released by machine. What no individual can do anymore is issue the authoritative “no”. Every denial, whether of a classification challenge, a downgrade or declassification request, or a drafter’s working-level “does this need a stamp” query, comes from one accredited automated adjudication system, and the system logs each denial with the specific guide rule it applied. The person who acted on the system’s concurrence gets a good-faith safe harbor tied to that logged event. If the call later proves wrong, the liability lands on the system, which can be measured and fixed, instead of on a career, which can only be ended. And silence stops working: a query nobody answers is automatically referred to the adjudicator, within five business days for a drafter’s question and never more than 120 days for the rest. The pocket veto does not die because security officers get braver. It dies because the security office does not have to subconsciously choose between protecting themselves and national security.

Notice the exhaust this throws off. Every “no” now flows through one chokepoint that records the question, the rule, and the outcome, and participating agencies report their denial, concurrence, escalation, and measured-accuracy rates every quarter. The decision-level record this entire assessment has been starved of becomes a routine byproduct of running the system. The reform is not just a fix. It is the collection platform.

What does the AI in the middle actually do? Three narrow things, none of them a magical oracle. It retrieves the governing Security Classification Guide(s) for the program at hand. It checks the content in front of it against that guide’s rules, the same lookup a diligent human reviewer is already supposed to perform. And it returns a determination that cites the rule it relied on, under an identifier that stays with the document. When the call is close or its models disagree, it does not guess. It escalates to the cleared humans inside the same accredited system, whose decisions are logged the same way. That is the arithmetic in section 8: escalate the hard 13.78 percent and the machine’s miss rate on what it decides alone falls to 2.16 percent on my benchmark. Give the humans the hard work, that takes real judgement to adjudicate ambiguous guidance or where world events have surpassed classification guides.

None of the structure depends on the AI being good. Accreditation is the gate: a trial against held-out, independently adjudicated documents, with guide retrieval certified as part of the trial, because my benchmark showed a model handed the wrong guide is worse than one handed nothing. Until an adjudicator passes that trial, the AI-only denier stays what section 8 says it is: a hypothesis with a published test attached.

Where would it live? My bias first: I know Department of War security far better than Intelligence Community security these days. You put the adjudicator on the systems that already store the classification guides and host security’s daily workflows. On the secret network that is the Defense Technical Information Center (DTIC), which holds most of the collateral SCGs. Above it, that is the Joint Access Database Environment (JADE), which is the front-end many of security’s existing classified workflows, and shows up in security-officer job listings as a daily tool. Put the adjudicator where the rules and the officers already are, and nobody builds a new server that holds the crown jewels.

The legal path is narrower than an enthusiast would like. Executive Order reaches roughly 90 to 95 percent of the decision mass, because the current allocation of denial authority exists only by executive order and what one order allocates a successor can reallocate. It cannot touch nuclear Restricted Data, cannot extinguish criminal exposure under the espionage statutes without legislation, and cannot displace the intelligence agencies’ independent statutory withholding authorities. Moderate confidence, pending the OGC review the memo itself demands.

The collection plan

Every judgment in this assessment that sits below high confidence comes with the data collection that would resolve it, and all of it is cheap for the one party that holds the data.

A representative decision-event audit with independent adjudication settles the rate, the ratchet, and the pocket-veto prevalence. A published accreditation trial on a held-out corpus of adjudicated documents settles the AI question. Deidentified sanction counts by error direction settle the asymmetry’s magnitude. A protected practitioner survey settles the chill. An institution that exercises coercive power through secrecy and declines to produce any of these measurements has made a choice, and the choice is itself informative.

Best arguments against this

The strongest counterargument: the whole cost case leans on a rate nobody measured, so I could be describing a system that over-classifies a little and calling it a crisis. I concede the premise about measurement and reject the conclusion. The direction is measured six independent ways. The floor is proven. The referee record, the elicitation, and the incentive census all point the same way, and the reform’s core is worth doing at the floor. If you believe the true rate sits at 3 percent, you are betting that every official who ever went on the record, and the appeals panel’s twenty-two-year docket, are all wrong in the same direction.

The second strongest: contested appeals are cherry-picked, so the 75 percent overturn rate overstates the problem. Partly true, and I keep the selection caveat welded to the number. But selection cuts both ways: the documents nobody can afford to appeal, inside CAPS and SAPS few ever see, face less scrutiny than the appealed ones, not more. If someone gets rejected for access to a SAP, they’re stuck on security’s slow approval process thereafter. If you’re not senate confirmed, then there are career consequences for rocking security’s boat.

The third: the AI cannot do the job. Correct, today, and this assessment says so at high confidence. That is why the reform’s core is accuracy-independent and the AI-only denier is a trial, not a deployment. The reality is that there is already a trust schism between the security apparatus, and the rest of the community. I expect security to distrust AI tools. That’s good, because the system collects data, make it better than what we have today. I hope to see the broader workforce and security band together to thread against the AI who is the bad robot saying “no.” That’s a healthy culture.

And a fourth, which deserves the most respect: some secrets are worth everything their protection costs, and a system tuned to never lose them will over-protect everything else as the price of the tail. I concede more than half of this. There is probably a line somewhere around the unacknowledged or waived access programs where the default is “no”. That’s a SAPCO call. What right-sizing cannot explain is zero sanctions, a correction channel nobody uses, secrecy orders on patents no longer get lifted even after the inventions ceases to be novel, and an enterprise that stopped measuring itself. A well-calibrated system would show a working feedback loop somewhere. This one shows none.

What would change my mind

A representative, independently adjudicated audit finding the true excess share near the 3 to 8 percent floor would collapse my cost estimate to the bottom of the slope, and I would publish that number with relief, because it would mean the measurement finally exists.

A single documented sanction for over-classification, in any channel, would bound the asymmetry and improve the paper. I searched for one and want to be sent one.

An accreditation trial in which a deployable adjudicator clears 98.6% with certified classified guide retrieval would move the AI-only denier from hypothesis to result.

A released internal audit showing agencies already track decision-level correctness with independent adjudication would retire the internal-measurement judgment outright.

The close

I went in believing over-classification was about people protecting a paycheck. It is not. It is a bet where one side ends careers and the other side has never once, in forty-seven years of public records, cost anyone anything, taken by rational people tens of millions of times a year.

Classification is not a judgment about a document. It is a bet about a career.

The intelligence community writes assessments about adversaries who hide their embarrassing and national secrets. The classification system now qualifies. Move the no somewhere it can be seen, and the government gets, for the first time, the one thing this assessment could not: the measurement of the secrets it keeps from itself.

China published its corruption verdicts, in bulk, until it did not. Fortunately, there is a leaked copy of the 2013 to 2021 court records floating around that lets us calculate the standard bribe per government seniority level, bribeflation per year, what gifts officials received, and even the most common watches used as a bribe (Omegas).

The chart shows both facts at once: the going rate by rank, year by year. How to read it: each line is one tier of official, from sub-section clerk to bureau chief, and each point is the median disclosed bribe for that rank in that year, in US dollars on a log scale. What to see: the lines never cross, so the rank ladder holds every single year, and every line tilts upward, so the whole schedule inflates together. The price of an official is both ordered by his rank and rising with the calendar. The dotted line at the floor is where the lowest tier would sit if bribes had merely tracked consumer prices. Nothing on the chart stays near it.

How to read it: each colored line is one rank of official, from sub-section at the bottom to bureau chief at the top; vertical position is the median bribe taken by that rank in that year, log scale in US dollars. Solid lines cover the bulk pre-deletion record, 2013 to 2021; the shaded band holds the directional survivor tail. What to see: the rank ladder never inverts, and every rank’s going rate rises across the span, with the lowest tier climbing roughly fivefold.

The findings

How to read it: one bar per rank, all bulk years pooled, median bribe with a 95 percent confidence interval, log scale in US dollars. What to see: the price of an official climbs steeply with rank, from about 13,000 dollars at the bottom to about 308,000 dollars for a bureau chief, a roughly twenty-two-fold spread.

The menu is real, and it is steep. The rank ordering is the sturdiest thing in this dataset. Sub-section, section, county, bureau: the median bribe climbs at every step and never reverses. That monotonicity matters more than any single number, because it survives every way the data is sliced. A county-level official commands roughly four times a section chief. A bureau chief commands more than three times a county official again, and more than thirteen times a section chief. The confidence bands on the four tiers do not overlap. It’s quite literally a priced Chinese menu of corruption.

The high end is confirmed from the outside. The court corpus is thick with low and middle ranks and thin at the top, because tigers are tried in a handful of courts and the corpus samples them sparsely. So the top rung was checked against a separate source: the anti-graft agency’s own public register of senior officials placed under investigation. That register runs the other way, heavy at the top, with about three quarters of its named officials at bureau level and roughly a fifth at provincial or ministerial rank. Where the two sources describe the charge, bribery is named in about ninety-six percent of them. The court record prices the bottom of the ladder with large numbers, the disciplinary register confirms the top of it, and both agree the dominant transaction is a bribe. One aside worth keeping: those disciplinary notices never state a sum. The number surfaces only at trial. Even the state’s public accounting of its own tigers withholds the price until the courtroom.

How to read it: two indexed lines, both set to 100 in 2013; blue is China’s consumer price index, red is the median bribe, solid through the bulk record and faded dashed for the survivor tail. What to see: consumer prices barely move while the median bribe rises almost ninefold by 2021.

Bribeflation. Between 2013 and 2021, the years the record is published in bulk, the median disclosed bribe grew about 790 percent while the official consumer price index grew about 16 percent. That is a gap of roughly fifty to one. And it is not one giant case dragging a mean. The ninetieth-percentile bribe climbed in step, from about 73,000 dollars in 2013 to about 520,000 in 2021, a sevenfold move. The whole distribution shifted up, year after year. Two honest readings compete. Either the bribes themselves got bigger, or a lengthening anti-corruption campaign moved up-market and kept landing bigger fish. The court record cannot fully separate those. But under either reading, the last near-decade the public docket shows before it goes dark is corruption getting steadily more expensive, not less.

How to read it: a stacked area over the bulk years showing the share of cases that mention each medium of payment, cash at the base with property, cars, gift cards, gold, liquor, watches and other forms above. What to see: cash dominates every year, with a steady minority in property, gifts, gold and luxury goods.

They take cash, mostly and stubbornly. The tradecraft would imply that sophisticated graft migrates from cash to property to crypto. The data refutes it. Cash is named in a majority of cases every year, ranging from roughly 59 percent to more than 90 percent, and it never yields the lead. Cryptocurrency appears in essentially zero percent of bribery convictions in every single year across the whole span. Property, cars, gift cards, gold, liquor, and watches all show up, but as the garnish on a cash meal, and the non-cash forms are undercounted here because judgments list them unevenly. The clean finding is the boring one, which is the most useful kind: prosecuted Chinese corruption is a cash business, start to finish.

There is a floor effect worth stating plainly. The classic low-end gift, a single bottle of Moutai or a carton of cigarettes, mostly sits below the roughly 4,000-dollar threshold for criminal prosecution and gets handled inside the Party as a discipline matter. It never reaches a court. So this record captures the going rate for prosecuted corruption, not the full gift culture. The watches and the liquor that do appear are riding inside bribes large enough to charge.

How to read it: the luxury-watch brands named in bribery and graft judgments, ranked by how often they appear. What to see: a short list of prestige Swiss brands accounts for most named watches, and the workhorse is the Omega, not the Rolex.

And when the in-kind items appear, they are a comedy of taste. The workhorse bribe watch is not the trophy piece you would guess. It is the Omega, which turns up more than three times as often as the Rolex, with Cartier, Longines, and Tudor filling out the wrist. The genuinely rich watches, the Patek Philippes and the Vacherons, sit in the long tail by count but at the top by value, a Patek running a median around 49,000 dollars. One state-bank investment chief took six watches in a single relationship, two Pateks, two Piagets, a Lange, and a TAG, appraised together at about 166,000 dollars. Liquor means Moutai in about 13 percent of the gift cases, and when the bottle is old the price climbs: fifteen-year is the common pour, but fifty-year and even eighty-year bottles surface in cases valued in the tens of thousands of dollars. The single fanciest object in the whole record is not a watch at all. It is 11.84 kilograms of gold bars, appraised at about 631,000 dollars, taken by an assistant general manager at a state steel group. Pine timber, an apartment, a villa, six ivory carvings, a tiger painted by a famous calligrapher: the in-kind ledger reads like an estate sale for a taste no one would admit to having. The cash is the meal. The Omega is the mint on the pillow. (This in-kind detail is drawn from the disclosed gift cases across the record and is illustrative, not a year-by-year series.)

How to read it: the curve plots prison months served per one million dollars taken against the size of the theft, log vertical scale, each point labeled with the number of sentenced cases behind it. What to see: the marginal punishment per stolen dollar falls sharply as the amount grows, so the largest thefts are, per dollar, the most lightly punished.

Does money buy leniency? Per dollar, yes. Ask it crudely and the answer is no: a bigger bribe draws a longer sentence and kills the odds of a suspended one. The suspension rate falls from about 44 percent at the small end toward almost nothing at the large end. Ask it per dollar and the answer flips. Sentence length does not scale with the bribe, because prison time hits a ceiling at life and death while money does not. Measured as months of prison per million dollars taken, the effective price of corruption collapses from roughly 3,700 months per million at the small end to around 60 months per million at the large end, a discount of nearly sixty to one. Steal a little and every dollar is expensive in years. Steal an empire and the marginal dollar is nearly free.

Geography has a price gradient too, faintly. With the usual small-sample caution, the priciest disclosed corruption sits in the far west and the capital, with Xinjiang around 93,000 dollars and Beijing around 69,000 at the median, while the cheapest sits inland, with Shanxi around 13,000, Guangxi around 16,000 and Shaanxi around 9,000. Do not hang a thesis on the provincial ranks. They earn a mention because the spread is wide enough to be interesting and narrow enough to be honest about.

How to read it: the largest single amounts on the disclosed record, one bar each, labeled by year, province and charge type, log scale in US dollars, no individuals named. What to see: the top cases reach into the tens and hundreds of millions of US dollars, clustered in the 2016 and 2020 filings.

The ceiling. After a plausibility screen that throws out impossibly large figures as source-document typos rather than launder them into a headline, the largest credible single cases run to roughly 118 million dollars for a public-funds misappropriation that drew a life sentence, and about 57 million dollars for an embezzlement. The largest pure bribery case is a city deputy mayor who took about 41 million dollars and drew a life sentence. The point of the ceiling is not the single number but that the disclosed distribution had a very long, very real tail, and the state was still publishing that tail, in detail, until it was not.

Best Arguments Against This

The strongest objection is selection, and it is real. A court record is not a census of corruption. It is a census of prosecuted, disclosed corruption, and both filters move over time. The whole 2013 to 2021 window sits inside a single, intensifying anti-corruption campaign, so the rising median could reflect enforcement that spent its early years on small fry and its later years on bigger targets, rather than any real inflation in the graft itself. I concede that the level of the trend is not clean. What I do not concede is the direction and the ladder. The direction now runs eight years without a single reversal, which is far harder to wave away than a two-year jump. The rank schedule is a cross-sectional fact, immune to the time-selection worry, and it is overwhelming, with non-overlapping confidence bands across four tiers and independent confirmation of the top rung from the disciplinary register.

The second objection used to be the salvage. One bulk year, 2019, was for a while held only as a partial salvage from a source-corrupt archive, which made its counts a lower bound. That year has since been recovered in full (and the article updated on 2026-7-16), and the complete record lands exactly where the salvage put it: the 2019 median sits on the staircase between 2018 and 2020, and its medium mix is unchanged, so nothing turns on it. And after 2021 the public court record thinned dramatically and non-randomly, so any figure drawn from the surviving cases is a groomed sample and is treated here as directional at most. That is why the load-bearing numbers live in 2013 through 2021, where the frozen copy is published in bulk.

What Would Change My Mind

• If a fuller or earlier slice of the same corpus showed the rank ladder inverting or flattening, the schedule claim weakens. The ordering is the spine of this piece.

• If procuratorate or court yearbook totals for those years showed disclosed bribe medians flat while mine rise, I would suspect an extraction artifact and re-audit the amount parsing. The current re-derivation audit agrees with the model on about 96 percent of stated amounts.

• If the non-cash forms, properly valued, turned out to carry a third or more of total bribe value in later years, the cash-business claim needs qualifying.

• If the discarded mega-cases turn out to be real rather than typos, the tail is even longer and the ceiling section understates the point.

• If independent reporting establishes that the low-end gift economy is being criminalized rather than handled as Party discipline, the threshold caveat changes.

Close

The price of bribing a Chinese official climbed almost ninefold while the price of groceries barely moved. Then the court records list went dark. And in Feb 2026, the U.S. suspended enforcement of bribing foreign officals. How long before we see the first GoFundMe?

Correction and update, 2026-07-20.

First, what does not change: everything this piece stands on. Bribeflation of +791 percent from 2013 to 2021 against consumer inflation of +16.4 percent, a gap of roughly fifty to one. The rank ladder that never inverts. The cash business. The per-dollar leniency discount. The ceiling. All unchanged, and for a simple reason: those numbers live at the 2013 and 2021 endpoints and in the pooled cross-section, and this update touches neither endpoint.

What happened is that the one weak year got fixed. When this ran, 2019 existed only as an 11.36 percent salvage of a corrupted archive, and I flagged it as such. The complete 2019 archive has since been recovered, all 10,750,409 records, and it turns out the salvage under-sampled large-amount corruption cases badly: 20 big-amount cases in the salvage, 1,105 in the full year. I re-ran the exact same pipeline over the full year. Four secondary printed figures moved:

The bureau-chief pooled going rate rises from about $286,000 to about $308,000, crossing the $300k mark. Xinjiang’s provincial median falls from about $120,000 to about $93,000. Hainan falls from about $75,000 to about $46,000, so Hainan is no longer the high-end standout; the high end is now Xinjiang at about $93k and Beijing at about $69k. Guangxi rises from about $13,000 to about $16,000. The rank ladder and the province paragraph’s caveats stand as written.

One caveat is now obsolete, and retiring it strengthens the piece. The Best Arguments Against This section conceded that 2019 survives only as a partial salvage, a lower bound. No longer. The complete year lands exactly where the salvage put it on the staircase: the 2019 representative-base median moves from about $54k to about $57k, still cleanly between 2018 at about $33k and 2020 at about $86k. The trend line held when the missing 89 percent of the data arrived.

These numbers are generated from the aggregated data, not hand-typed, so when the data improves, the numbers get re-run and the changes get dated and posted. Math

But I understand now. The financial minutiae is mind numbingly boring. It is really challenging to communicate enough economic complexity to refine the readers intuition, but make it interesting enough that people read it to walk away with a more intuitive understanding.

This capstone does three things. It grades China’s economic warfare machine honestly, cell by cell. It settles the series’ hardest case, semiconductors, at its true weight. And it converts eight posts of measurement into six operating rules a US strategist can carry into a meeting as the economic treaties with China expire this fall. I hope to now make thee knowledge actionable as new economic policy is enacted over the next few months.

Bottom Line Up Front

Here is the machine, graded objective by objective, each grade stated at its own strength.

The one thing China’s state money causally secured is survival. After a government guidance fund buys into a firm, going-concern distress drops measurably (a coefficient of -0.033, p = 0.003, with clean pre-trends). That is a rescue effect, not an improvement effect, and it is the strongest result in this entire body of work.

The economic chokepoints are real and held. In 2024, the latest full reporting year, China held 50 to 72 percent of world exports across five product families: photovoltaic modules at 71.5 percent, rare-earth magnets at 67.3, coated steel at 62.9, lithium-ion batteries at 55.0, cargo vessels at 50.2. Those are reporting-countries-only shares and run slightly high; no plausible correction pulls any of them below half the world market. Capacity and scale gains are real too, but they arrive as associations shaped by selection: the funds pick firms already ahead. Measured capability, productivity and patent quality, was not bought; three independent causal designs return the same null. And the one weapon China has repeatedly fired, mineral export controls, erodes with use, on every window long enough to read.

How to read it: five state objectives, each graded in plain words; the boxed row is the only grade with causal support. What to see: the machine won survival and holds chokepoints. Capability never came, and fired controls erode.

The caveats sit here, next to the claims. The capability result is a bounded null. I could detect no positive effect detected at my resolution but did not have enough data to certify Chinese money doesn’t lead to any innovative capabilities or more productive/efficient companies. The erosion grade rests on partner-mirror data because China stopped reporting its own trade after 2024, and the single completed erosion arc is a sample of one. The scorecard is a synthesis of what eight posts measured; it proves nothing new on its own.

What should a decision-maker now believe? That China built a mixed machine: a few dangerous wins, strong held positions, measurable failure modes, and a heaviest bet that is defensive and therefore durable. Plan against that machine. Not the monster, not the spreadsheet.

The findings

1. Walk the scorecard cell by cell, and keep the grades apart from each other. The separation is the whole product. This is what the system does, measured. It is not proof of a secret plan.

Survival is causal. China is deliberately keeping certain firms alive. Fund entry cuts a firm’s probability of going-concern distress, the pre-trends pass, and the effect is the cleanest number in the project. Its qualifiers travel with it: the money is a tourniquet, not a training program. Distress ran 26.4 percent among the delisted casualties against 0.7 percent for core champions and 7.1 percent for their listed peers, so the rescue is being bought for the sickest tier, and inside the state-owned cohort more subsidy tracks more distress, because the money runs toward the bleeding.

The economic chokepoints are held, not caused by the cash. The five families above cleared half of world exports on thin listed-firm money; the magnet position rests on roughly CNY 6.8 billion (about US$0.9 billion) of grants across a decade and a 2021 quota consolidation. Across six measured sectors, the cash and the leverage rank almost exactly inverse (a rank correlation of -0.94, a typology on six points, not an estimated law). The money poured hardest where dominance never came. They haven’t been able to create an economic chokepoint by lighting billions of dollars of money on fire.

Capacity and scale are associations. Subsidy intensity ranks with revenue growth and output scale, but the clean fund-entry design shows the funds entering firms that were already ahead, by about 2.4 points of R&D intensity and roughly 3 to 5 points of sector share, before the money arrived. The state picks winners more than it makes them.

I need to unpack a the apparent contradiction a bit between China using money to keep (distressed) firms alive, and using money to pick the winners from companies doing well. The state runs two tools with opposite selection signatures. Strategic equity funds buy equity in firms already ahead of their peers and causally lower their failure risk: they pick winners and insure them. Broad grant subsidy, especially to SOEs, pools in the weakest firms, a keep-alive tilt that shows up as correlation.

Capability is the null. Three causal designs on three data structures, a withdrawal difference-in-differences, a tax-threshold discontinuity, a staggered fund-entry study, and none finds a productivity or patent-quality gain. Among subsidized firms, more subsidy ranks with slightly lower patent quality, about -0.20. The money buys reported R&D and scale. It does not buy what those inputs are supposed to produce.

And the fired economic weapon erodes. The direction is consistent everywhere the window is long enough: the 2010 embargo cut Japan’s China-dependence from roughly 90 to roughly 60 percent over the following decade; gallium and germanium flows ran at about 0.60 of pre-control value across 31 months at a flat price; graphite at about 0.47 with no premium; antimony collapsed to about 0.15 of volume at 3.2 times the price, a pattern that is ambiguous by construction; the April 2025 magnet regime, eleven months in, sits near 0.94 and is not yet a verdict. All of it reads through the partner mirror, median-unbiased and product-level noisy. Erosion is where the evidence leans, stated as a lean.

Why belabor the grades? Because every bad China opinion poiece and policy that I have read this year comes from averaging them. Fuse the cells and you get either “it all works” or “it all fails.” Keep them apart and you get a machine you can actually plan against.

2. The semiconductor bet is defense, so expect persistence. The heaviest cash in the panel sits on chips: 7.3 percent of listed-firm revenue in support intensity, plus roughly CNY 368 billion (about US$51 billion) of state equity through the national chip funds. On the offensive scorecard that bet is a clean miss. China is the world’s largest importer of integrated circuits, the import bill rose from about US$250 billion to about US$413 billion across the buildout decade, and in January and February 2026, after more than three years of allied export controls, it was still rising: 550 billion yuan (about US$76 billion) of imports, up 36.8 percent year on year by value and 9 percent by volume. You will hear that China’s chip exports are surging too. They are, up nearly 70 percent in the same months, and they are overwhelmingly mature-node. The exports are the chips China can make. The imports are the ones it cannot.

Now put that miss next to the map from the previous post: the deepest feasible chokepoint anyone holds against China is the semiconductor toolchain. The heaviest cash landed on the deepest inbound dependence. That is not a failed bid for export dominance. It is a rational defensive bet on the one input whose denial China fears most, and a rational defensive bet does not get abandoned when the first decade looks ugly. It gets doubled.

How to read it: an annotated overhead of the flagship logic fab, marked as the defensive bet it is. What to see: the heaviest cash in the ledger guards the way in, not the way out.

Hold the wording precisely, because both halves are important. Rational defensive bet, not a measured escape. The capability null covers chips too, the import bill is growing in the wrong direction for the escape story, and nothing in my data shows the exposure closing. What the defensive read changes is the forecast: Washington should expect Chinese chip spending to persist regardless of measured returns, because the yardstick in Beijing is exposure, not profit. Budgeting for your rival to get discouraged is not a plan. The allied Asian and Western firms who are forgoing revenue to maintain the chip ban will be paying that opportunity cost for a long time.

3. What happens after the money moves on: four patterns, honestly bounded. The series also watched sectors after they stopped being the top priority, and four shapes emerged, split by one variable: supply discipline. Solar, demoted without a capacity cap, was buried by its own glut, net margins down about 5.4 points with three of the sector’s five listed-census delistings inside the post window. Steel and aluminum, demoted under caps, compressed about 3.6 points and survived, casualties near zero. Rare earths became a ward: margins actually rose about 1.6 points after demotion because quotas and licenses prop pricing in a way cash never did. Electric vehicles switched instruments, the visible purchase subsidy fading into tax and demand channels while the sector kept growing, which is why a falling grant line must never be read as abandonment. These are associations on a small set of dated triggers, casualties included, and they are shapes, not laws; agriculture is the standing warning, a sector whose sign flips from -0.47 to +0.86 points depending on whether you count the firms that died.

How to read it: each bar is one demoted sector’s median net-margin change after demotion, casualties included, labeled by its pattern. What to see: the uncapped sector is buried while the capped and warded sectors hold; the split runs on supply discipline, not on how much cash the sector once got.

A. Four post-demotion patterns generated/post8_body_gen_a.png

The pattern earns its place on the scorecard for one reason: Beijing read it the same way. Watching solar bury itself, the state reached for a capacity-retirement campaign, not a rescue check. The machine grades its own homework, and it looks like I graded it the same way.

4. The playbook. Everything prescriptive this series has to offer fits in six rules on a 4x6 index card:

1. Read Beijing’s subsidy ledger as a fear map, not an ambition map. Across the six measured sectors, cash and leverage rank inversely; the money marks where China feels exposed, and the intensity ranking is a free intelligence product on where the next yuan goes.

2. Do not build any plan whose load-bearing assumption is that China’s weak firms eventually die. Survival is the machine’s one causal win; the shakeout you are waiting for is the specific thing the money prevents.

3. Track caps, licenses, and ownership consolidation, not budgets. One capacity ceiling separated record aluminum profits from solar’s casualty list, and the only chokepoint China has repeatedly fired runs on quotas and licensing built on a rounding error of cash; budget telemetry misses every decisive instrument.

4. Force the fire-early-or-never dilemma, then spend every firing on diversification. Fired controls erode on every readable window, the 2010 embargo bought Japan a decade of measured de-risking, and closed routes respawn one border over within a year; a firing is a dated, free lesson in where dependence binds, and it expires.

5. Fight on the narrow allied front, not the deepest dependence. Vulnerability and feasibility are different axes; the workable levers sit in chip equipment, metrology, and high-bandwidth memory, where a US-Netherlands-Japan bloc holds roughly 56-66 percent of supply and China’s own category exports run 1.4-1.9 percent, while food, the deepest-looking dependence, is exposure, never a lever.

6. Assume the telemetry dies, and write mirror-based verification into every regime. China’s official trade feed has a documented off switch, dark since calendar 2024, and any deal, snap-back, or reporting requirement written against data the counterparty can switch off is written on sand; build the partner-mirror check in from day one.

That list is the whole series.

5. Why this lands now: every clock rings in fall 2026. The Busan truce of October 30, 2025 was a one-year pause, not a settlement, and its expiries are specific. China’s October 2025 rare-earth expansion, including the extraterritorial rule reaching any foreign magnet with 0.1 percent Chinese-origin content, is paused until November 10, 2026. Its US-directed gallium, germanium, and antimony ban is suspended until November 27, 2026. Washington’s Affiliates Rule is suspended until November 9, 2026. The April 2025 licensing regime on seven rare earths never paused at all, and controlled exports still ran roughly 50 percent below pre-control levels into the May 2026 Beijing summit. The two presidents meet again in September 2026, weeks before all three alarms ring. Rules 4 through 6 are not evergreen advice; they are dated homework.

The US instrument stack has meanwhile been stress-tested down to its durable core. The Supreme Court struck down the IEEPA tariffs in February 2026; their stopgap replacement, a 10 percent surcharge under Section 122, expires July 24, 2026 absent congressional action. What survived untouched is Section 301: 100 percent on Chinese EVs, 50 percent on solar, 50 percent on semiconductors, with graphite, magnets, and non-EV batteries added at 25 percent from January 1, 2026. Section 301 is the durable US core any next bargain gets built on. And Washington has begun answering the machine in its own grammar: a 9.9-10 percent federal stake in its flagship chipmaker converted from grant money in August 2025, a roughly 15 percent defense-department stake in its flagship rare-earth miner with a ten-year price floor at US$110 per kilogram, about double the miner’s 2024 realized price. The strongest tell that the channels are the weapon is that the other side has started copying them.

Best Arguments Against This

The strongest attack: the semiconductor shield relabels a miss. The cash went in, dominance never came, imports rose, and calling that “defense” rescues a failed program with vocabulary. I concede the entire offensive record, and the defensive read still earns its place, for two reasons. It is the only reading that holds three facts at once, the heaviest cash, the clean export miss, and the deepest inbound chokepoint sitting on the same sector. And it is falsifiable where a rhetorical rescue would not be: if subsidy simply chased vulnerability, chemicals, where China is 90 to 100 percent import-dependent on key feedstocks, should be drowning in state money, and it gets almost none. The defensive read is a per-sector judgment that happens to fit the biggest bet, not a law. The moment the import bill falls with capacity holding, the bet becomes an escape. It has not.

The second attack: both camps can cherry-pick this scorecard. The hawk quotes the held chokepoints; the skeptic quotes the capability null; my own matrix arms them both. Correct, and that is the finding, not a flaw in it. The camps are each half right about different cells, which is precisely why single-grade answers have produced a decade of bad China takes. The scorecard’s discipline is that no cell travels without its grade. Anyone quoting one row of it as the whole verdict is doing opinion with my footnotes.

What Would Change My Mind

1. A measured semiconductor escape: the IC import bill falling for four consecutive quarters while wafer capacity holds, checkable by mid-2027. As of January and February 2026 the bill was growing 36.8 percent year on year, the wrong direction for this trigger.

2. A fired control that holds: the April 2025 magnet regime printing four clean post-control quarters of sustained share and sustained volume at a price premium, readable in early-to-mid 2027. That would break the erosion grade where it matters most.

3. A citation-weighted Chinese, European, and world patent panel showing state support associated with rising per-patent quality. That panel is not public data, and it is the named falsifier for the capability null.

The close

Eight posts ago this project set out to measure a machine everyone had already made up their mind about. The measurements came back mixed, and the mix is the message: one causal win, real held positions, a defensive bet that will persist, a capability engine that never started, and a weapon that depreciates when fired.

The machine was not built to win. It was built to not lose. You do not out-wait a machine like that. You out-position it.

exeunt

An empty set is still data. This post is about a year of empty sets, and about eleven other places China’s official numbers changed form under my queries while I built the dataset behind this series. The public narrative is that China lies about its numbers. The ledger supports something narrower and more useful: China changes where you can see the numbers, exactly where seeing them would matter.

The narrower claim is worth more than the accusation. An accusation ends the analysis. A taxonomy tells you what each gap costs and how to route around it. You act based on that knowledge.

Bottom Line Up Front

Twelve documented reporting behaviors sit in the ledger behind this series, sorted into six categories: omitted, reclassified, basis-shifted, accounting-contaminated, delayed, and inaccessible. A seventh category exists for strictly false statements, meaning an official number contradicted by primary evidence. It’s empty. I didn’t find evidence that China is lying. This project has 29GB of raw Chinese government and corporate filings. There were mistakes yes. But I’m not seeing economic filing evidence that corroborates the widespread belief in China lying about its numbers. That makes the other twelve rows worth briefing.

China is omitting I sat between two lawyers at dinner tonight and they probably would have argued if lying by omission was lying. But that doesn’t make past data suspect. China filed its full 2024 trade data with UN Comtrade (the United Nations database where each country reports its own customs returns) in June 2025, on its normal cadence. It has filed nothing since. As of July 1, 2026, verified directly against the Comtrade API rather than by rumor, there is no China-reported 2025 trade data: all twelve monthly queries and the annual query come back empty. To be precise about what stopped: China’s customs agency still publishes its own headline releases at home. What stopped is the standardized UN filing, the one format that lets an outside analyst (me) audit product-level detail against everyone else’s books. The blackout covers exactly the period in which China’s 2023-2025 export controls need grading.

The workaround is a partner mirror: add up what every other country reports importing from China. The mirror is honest in aggregate and noisy per product, and both halves of that sentence matter. In the years China still reported its own trade, the mirror is unbiased at the median, a mirror-to-reported ratio of about 1.019 across 131 validated product-years. But only about a third of product lines land within plus or minus 10 percent of China’s own figure, and roughly one in six is off by more than a factor of two. Yuan figures below convert at CNY 7.2 per US dollar, the single round rate used across this series. I assuming that I couldn’t find all of the partner data because many of those partner countries aren’t sophisticated data reporters (and there is probably a lot of graft going on at the receiver end in some countries). The research data was current as of June 19, 2026; the Comtrade absence was re-verified live on July 1, 2026.

The strategic translation fits on a PostIt note. Never write a verification regime, a sanctions damage assessment, or a deal-compliance clause against data the counterparty can switch off. The off switch has now been demonstrated by China on the record.

The findings

1. They don’t lie, they omit. Each of the twelve rows records one statistic, what happened to it, the innocent explanation, and what would falsify the entry. The categories sort by what the analyst loses. Omitted, three rows: a series stops or never appears, the trade blackout being the live case. Reclassified, four rows: the line survives but folds into a broader one, so the specific figure stops being legible. Basis-shifted, one row: the series continues on a changed methodology, so new numbers do not compare with old. Accounting-contaminated, one row: the figure is accurate as filed but mixes in items an outside reader does not expect. Delayed, one row: the number arrives years after the question. Inaccessible, two rows: paywalled or never published.

Then there is the seventh category, false, and it is empty. That emptiness is the credibility engine of this whole post. Anyone can shout fraud at an adversary’s statistics; the shout costs nothing and proves nothing. A ledger that hunted for a year and still reports zero strictly false statements is a ledger you can trust when it reports everything else. And every row carries its boring explanation printed beside it: funds get consolidated when they wind down, methodologies change, notifications lag, accounting standards mix items. I dispute none of those. The taxonomy exists to price the gaps, not to prosecute them.

How to read it: twelve documented reporting changes sorted into six categories, with the column for strictly false statements standing empty.

What to see: zero lies, six kinds of dark.

2. The flagship omission. The tracked basket is thirteen products: the controlled minerals, gallium and germanium, graphite, antimony, rare-earth metals and compounds, and the magnets made from them, plus the dominant subsidized exports, solar modules, lithium-ion batteries, electric vehicles, and coated steel. Through calendar 2024, China reported its own exports of all of them. Its 2024 batch landed at the UN in June 2025, roughly six months after year end, the normal rhythm. For 2025, China has reported zero of the thirteen. Not late. Absent, as of July 1, 2026.

Now put the control dates next to the blackout. Gallium and germanium went under export licensing in August 2023, graphite in December 2023, antimony in September 2024, rare-earth magnets in April 2025. The single question every Western planner asks about those controls, did the flows actually stop, turns on 2025 data. That is the year the feed went dark.

How to read it: the red line is China’s own reported gallium and germanium exports to the UN database; the blue line is the sum of what every other country reports importing from China.

What to see: the red line stops dead after 2024; the blue line keeps reporting straight through the blackout year, which is the only reason the year can be graded at all.

The mirror routes around it, at a measured price. Every export that left China arrived somewhere, and that somewhere filed a customs return. Summing those returns rebuilds China’s exports from the receiving end. Against 2014-2024, when both sides of the ledger exist, the reconstruction is unbiased at the median. Product by product it is rough. About a third of validated product-years land within plus or minus 10 percent. Roughly one in six misses by more than a factor of two. And the misses concentrate exactly where the stakes are: on the thin mineral lines, gallium and germanium ran 1.6 to 5.9 times China’s own figure in the early validation years, antimony ore has had as few as four reporting partners in a year, and the worst thin-reporter mineral year is off by roughly twenty-four times. The fog is thickest directly over the chokepoint products.

How to read it: each count is one validated product-year, 2014-2024, comparing the partner-mirror reconstruction against China’s own report for the same product; the horizontal scale is logarithmic.

What to see: the distribution centers almost exactly on 1.0, a third of product-years land within ten percent, and a sixth miss by more than a factor of two; the mirror is honest in aggregate and rough per product.

Partner customs returns are logs from the far side of the connection. Useful, but it’s lossy. The mirror supports annual, directional grading of fired controls, which is what the scorecard in Post 5 ran on. It does not support quoting any single small-mineral level as truth, just truth with error bands.

3. Complying with a WTO ruling by changing methodology. The feed-in-tariff fund. A feed-in tariff (FiT) is the above-market price a government guarantees renewable generators; China paid its FiT through a surcharge on electricity bills, collected into a dedicated central fund with its own named line in the Ministry of Finance accounts. From budget-year 2023 that named line vanishes, folded into a catch-all “other government funds” line. The fold-in is provable from the table itself: the 2023 other-funds revenue figure of CNY 111.4 billion (about USD 15 billion) is stated in the table to be 102.8 percent of the prior year’s renewable line. The identity is legible. The line is not. The innocent explanation is genuinely strong: the FiT was winding down, and consolidating a dying fund is ordinary accounting. What keeps it in the ledger is the timing. The fold came immediately after a roughly CNY 394.5 billion (about USD 55 billion) two-year clearing of the fund’s payment arrears, a magnitude reconstructed from explanatory notes, so treat it as approximate. The decade’s largest green-support instrument became non-itemized right after finishing its largest job.

The WTO basis shift. DS511 was the World Trade Organization dispute the United States brought, and won in February 2019, over how China measured the price support it pays wheat and rice farmers. Before the ruling, China’s notifications counted eligible production as only the grain actually procured at the support price, which printed support ratios of 7.25 percent of value for wheat and 6.13 percent for rice, comfortably under the allowed ceiling. In the first post-ruling batch, filed in December 2022 and covering 2017-2020, the basis flips to the total production of the major producing regions. Eligible wheat quantity rises about 4.3 times, rice about 7.7 times, the ratios print at 19 to 25 percent of value, and China notifies its first positive total support figure, CNY 240.6 billion (about USD 33 billion) for 2017. Yes, this level of “eligible-production basis” bores me too. But it moved a printed national statistic by a factor of eight which has geopolitical implications I wanted to understand. The innocent explanation is again real: changing methodology after losing a ruling is called compliance. The analytical consequence is also real: the same support channel reads at wildly different magnitudes depending on the basis, which is why this series carries it as a band, roughly CNY 141-248 billion (USD 20-34 billion) per year, and never as a point.

And one clause covers the quietest category: the grant-carrier lines in listed-company filings are accurate as filed but mix value-added tax (VAT) mechanics into what a Western reader books as subsidy income, overstating true grants by 3.2 to 4.6 times in heavy industrial classes, so the number is right and the naive read is wrong.

4. The changes cluster exactly where the economic question becomes important. Run the timing test across the ledger. The national trade feed stops in the first year that would have graded the export controls. The FiT line folds directly after the arrears clearing. The agriculture notification basis flips in the first filing after the courtroom loss. The named industrial-upgrade transfer line leaves the central tables after 2019, just as central industrial support migrates into guidance funds and capital-budget specials that are far less legible, though that row may prove to be a renaming rather than a burial. The agriculture notifications themselves run about two years late, so the latest covered year is 2020 and the recent support wedge has nothing official to check against.

Every one of those rows keeps its innocent alternative, and I am not claiming intent. Clustering is consistent with coordinated design, and it is equally consistent with many separate agencies each independently preferring dim light when a number turns sensitive. The data distinguishes behavior, not motive, and the behavior is the finding. One honest clause on the final category: inaccessible is not suppressed. The pre-2017 capacity-utilization back-history and the province-level split of roughly CNY 453 billion (about USD 63 billion) in central transfers are unpublished or paywalled, not deleted, and calling them suppressed would make this post the mirror image of the accusation it declines to make.

5. Write the verification before the deal, and point it at the receiving end. This is the finding for the staffer with a drafting pen. The next negotiation with Beijing will produce commitments: purchase volumes, export ceilings, licensing rates, something. Whoever drafts the compliance clause needs a data source to score it against. If that source is the counterparty’s official feed, the clause ships with an off switch installed, and the off switch has a documented history of being used at the analytically decisive moment.

Three design rules follow from the numbers above. First, score against receiving-end data from day one: partner customs returns, port throughput, disclosures filed in the buyer’s jurisdiction, things the counterparty cannot un-publish. Build the mirror before the crisis, validate it while both feeds still exist, and record the bands. Second, budget those bands into every threshold. A compliance trigger set at plus or minus 10 percent, running on a sensor where only a third of product lines resolve to 10 percent, is not a trigger. It is a coin flip with legal consequences. Third, treat cadence as a term of the deal. Data the counterparty files annually, at its own discretion, six months in arrears, enforces nothing on any timeline that matters.

The deadline is not hypothetical. The control pauses Beijing granted under the October 2025 truce expire in November 2026. If controls refire, every Western damage assessment will run on partner mirrors with exactly the error bars measured here, because China’s own feed will show nothing at all. Better instruments exist only if someone commissions them now, while the calibration data still flows.

Yes officer, I was driving the speed limit.

Best Arguments Against This

The strongest objection: everything in this ledger is normal statistical practice somewhere. Funds get consolidated when their purpose ends. Methodologies change after legal rulings, which is compliance working as designed. Notifications lag in most WTO members. Accounting carriers mix items because standards say so. Some series were never collected. All of that is correct, and I concede it row by row rather than argue with it; the innocent alternative is printed beside every entry because most entries, taken alone, deserve it.

A sharper version of the objection notes selection: a ledger compiled while researching sensitive topics will find its gaps in sensitive places, because that is where the researcher was looking. Partly right, and worth saying out loud. But the flagship rows are not obscure corners I went digging in. They are the country’s headline feeds: the national customs return to the UN, the largest green fund in the central budget, the WTO agriculture notification. Normal practice scatters its gaps randomly. These cluster on schedule, at the decision points. The concession covers the rows. It does not cover the clustering.

What Would Change My Mind

1. China posts its 2025 trade batch to Comtrade. On the 2024 cadence, a batch could land any week in the second half of 2026. If people would find it useful then I will re-run the availability query. If it lands, the row reclassifies from omitted to delayed, the taxonomy absorbs it, and the mirror gets a free validation year. If China’s numbers then diverge materially from the mirror on the major products, the reconstruction, and the control grades built on it, would need to be redone. Both outcomes are progress. If people wouldn’t find it useful then I’m going to put down financial projects for a while to recover from country-level financial filings analysis.

2. A benign published reconciliation. A Ministry of Finance table re-itemizing the renewable fund line, or evidence the industrial-upgrade transfer line was renamed and remained itemized, converts those rows into fully explained consolidations and shortens the ledger.

3. A strictly false statement. One official number contradicted by primary evidence would fill the empty column, and I would have to retire the zero-false headline and reframe the post from visibility to fabrication. Through the data freeze, none exists.

4. A consistent-basis re-notification of China’s 2011-2016 agriculture support, or a WTO finding blessing the original procured basis, would collapse the CNY 141-248 billion band toward a point and downgrade the basis-shift row.

The close

Statistics are not a window. They are a valve. Beijing has never needed to falsify the flow; it adjusts the aperture instead, which lines stay itemized, which basis applies, which year gets filed. A valve, unlike a window, closes without breaking, and nothing in this ledger broke. The defense is not louder accusation. It is plumbing of your own: sensors on the receiving end, error bands you measured while both feeds still ran, and the standing assumption that the official feed fails precisely when you need it most. The data going dark is not noise. Dark, on schedule, is signal.

Next, Post 8: the scorecard and the playbook. What the machine actually delivered, what it cost Beijing, and what the data says the West should do about it.

Common wisdom assumes China’s economic vulnerabilities and allied Economic weapon feasibility are the same thing. Squeeze China where they are most dependent and they feel the most economic pain. But the data says vulnerability and economic weapon feasibility barely co-move. This post scores the difference, because the previous five posts mapped where China holds leverage over everyone else, and the obvious next question is the mirror one: where does anyone hold leverage over China, and which of those holds could actually be used?

The answer is not the intuitive one. The deepest dependences are the worst levers. The usable economic warfare front is narrow, technical, and already in allied hands.

Bottom Line Up Front

A target list sorted by China’s depth of dependence is mis-sorted. We took fifteen candidate inputs that China imports at scale and scored each on four external-control feasibility tests, two points apiece, eight points maximum: does an allied bloc actually control the supply, can China substitute or onshore it, are the relevant stockpiles thin, and is the collateral abroad tolerable. Raw dependence and feasibility turn out to be nearly orthogonal. Knowing how badly China needs an input tells you almost nothing about whether you can withhold it.

The extremes make the point. Lithium carbonate is China’s deepest single dependence, roughly 98 percent import-dependent, and it scores 5 of 8, because no allied bloc controls the ore and China owns the conversion midstream. Soybeans and maize, the deepest raw exposures after lithium, score 0 of 8. The only inputs that pass all four tests sit in the semiconductor toolchain: lithography, etch, and deposition equipment at 8 of 8, metrology and inspection instruments at 8 of 8, and high-bandwidth memory (HBM, the stacked memory that binds AI accelerators) at 7 of 8. A United States, Netherlands, and Japan bloc holds roughly 56 to 66 percent of the equipment market. China’s own world exports in those categories run 1.4 to 1.9 percent.

How to read it: fifteen candidate levers ranked twice, by dependence depth on the left and by feasibility on the right.

What to see: the lines cross hard. The deepest dependences make the worst levers, and the usable front is three chip inputs wide.

The strategic translation is a mirror. Beijing’s own playbook against its bottlenecks is visible in this data: stockpile deep, diversify origins, own the midstream, subsidize the residual chokepoint. The West spent late 2025 and 2026 building exactly that machine for minerals, with names and dates. The feasible offensive complement is narrower than most target lists assume: hold the chip toolchain front, with allies, for years, and do not confuse the other thirteen rows with leverage.

The findings

1. Vulnerability and feasibility are different axes, and lithium is the proof. Walk the deepest dependence through the four tests. China imports roughly 98 percent of its lithium carbonate, the deepest import dependence on every index in the panel. First test, allied control: raw lithium supply splits across Australia at about 37 percent of world mine output, Chile at about 20 percent, and Argentina at about 8 percent. Australia is allied. Chile and Argentina are friendly and non-aligned. There is no single controller, so an embargo needs a cartel first; and building a cartel was not on my life’s bingo card. Second test, substitution: China already owns the step after the mine, roughly 65 to 70 percent of the world’s conversion midstream, the step that turns ore into battery-grade chemical. Its route around a squeeze is mine equity abroad, which Chinese firms keep buying, so the controllable point is the ore and the window on it is closing. Third test, stockpiles: China’s strategic lithium reserve is undisclosed, which is not the same as thin, and commercial inventories sat in a glut at the data freeze. Fourth test, blowback: the one test lithium half-passes, since the collateral falls mostly on China’s own battery build-out rather than on civilians abroad.

Score: 5 of 8. The deepest dependence on every index is a medium lever at best. That is the whole sorting insight in one row.

One thing must be said plainly, here, where the map appears. This map is analysis, not a targeting recommendation. It scores dependences; it names no firms and recommends pulling nothing. The food rows in particular are flagged as exposures with severe civilian collateral, never as targets.

How to read it: each dot is one of fifteen inputs China imports at scale; right means China is more import-dependent, higher means the input passes more of the four external-control feasibility tests.

What to see: the deepest dependences, food and lithium on the right, sit low; the usable band at the top holds only chip equipment, metrology, and high-bandwidth memory.

2. The feasible front is narrow, allied, and already fortified. Three inputs pass the tests. Lithography, etch, and deposition equipment: 8 of 8. Metrology and inspection instruments, the tools that tell a fab whether its wafers are any good: 8 of 8. HBM: 7 of 8. Concentration does the work. Five toolmakers, Applied Materials, Lam Research, and KLA in the United States, ASML in the Netherlands, and Tokyo Electron in Japan, hold roughly 56 to 66 percent of the equipment market among them. China’s own world-export share in these categories runs 1.4 to 1.9 percent, which is what low frontier substitutability looks like in trade data. HBM comes with its own caveat: the 7 of 8 holds only at the HBM sub-segment, supplied by SK Hynix, Samsung, and Micron, where Chinese buyers have been drawing down pre-ban stock. The broad memory category is not a chokepoint at all; China exports about 27 percent of it. Precision matters more than breadth here.

Two features make this front different from every other row. Stockpiles fail China quietly: Chinese fabs stockpiled tools ahead of each control round, but the binding consumables are advanced-tool spares and service, and those are thin. And the blowback is low, because the collateral falls on chipmakers, not on civilians. The control architecture already exists: the 2023 United States, Japan, and Netherlands trilateral understanding on chip equipment, Japan’s 23 equipment categories licensed from July 2023, the first United States controls on HBM in the December 2, 2024 package, and Dutch national licensing of metrology and inspection equipment from April 1, 2025. This is the one place on the map where the deepest feasible lever and the operating policy already coincide. The erosion risks are mature-node leakage and indigenization, and they are risks, not facts, as of the data freeze.

How to read it: allied-bloc control of each toolchain input beside China’s own exports in the category.

What to see: the feasible front is narrow, and it is allied by construction.

3. Food is an exposure. But we have a soul. Soybeans and maize are China’s largest inbound dependences after lithium, and they score 0 of 8. Run the tests and watch every one fail. Origins are substitutable: Brazil and the United States together supplied 92 percent of China’s 2024 soybean imports, Brazil is non-aligned and gaining share, and China already rehearsed the reroute once, after 2018. Stockpiles are deep: state soybean reserves run around 40 to 45 million tonnes, roughly five months of demand, and China holds about 60 percent of the entire world’s corn stocks, around 166 million tonnes. And the blowback test is not close: a food squeeze lands on feed, meat prices, and nutrition, which is to say on civilians.

So the line gets drawn explicitly. Food is a hardening target for Beijing and a collateral catastrophe as an instrument of a Western Democracy. It would not coerce, because the reserves and the reroute absorb it, and it would injure civilians while failing as an economic weapon. This analysis scores the dependence. Read Beijing’s behavior and you see the same map: the May 2026 Beijing summit’s touted agriculture purchases, about US$17 billion a year through 2028, are reassurance procurement for an exposure Beijing intends never to feel. Food on this map tells you what China fears. It does not tell you what you can use.

4. The China subsidy follows economic vulnerability law dies on contact; the per-sector stories survive. There is a seductive statistic in this data. Across nine sectors, the rank correlation between how vulnerable China is to an inbound cutoff and how much subsidy the sector receives runs about 0.67. That looks like doctrine: Beijing pays where it bleeds. It looked like doctrine to me for the first week of this project. Then the robustness checks came back. The correlation holds only with agriculture in the sample and only on the grant channel; drop agriculture and it falls to roughly 0.52 and loses significance. Agriculture’s cash is income and margin support for a politically vital constituency, not import substitution, so the one sector carrying the law does not even belong in it. Chemicals then break it from below: China is roughly 90 to 100 percent import-dependent on p-xylene and the polyethylenes, exactly the profile a targeting law says should draw defensive money, and the listed subsidy response is near zero. The feasibility map agrees those rows are safe to neglect, at 2 of 8, since the supply is a diffuse world commodity market with easy route-arounds. A blind spot, but a cheap one. One reconciliation note, because an earlier post leaned on a similar-sounding number: the six-sector intensity ranking the fear-map post used as corroboration is the cell that survives these checks. The nine-sector cross-channel law is what dies here.

What survives is the strongest single story, semiconductors as a defensive bet. Chips draw the heaviest support intensity of any sector in this panel, 7.3 percent of listed-firm revenue, plus roughly CNY 368 billion (about US$51 billion) in state equity, and chips sit precisely where the feasible external levers concentrate. The money points at the map. What the money has not bought is escape: China’s integrated-circuit import bill rose from about US$250 billion to about US$413 billion across the buildout, and China remains the world’s largest chip importer. A rational defensive bet, not a measured escape, and the distinction is significant. Print the stories. Do not print it as a law.

5. The counter-doctrine is already on the map: play China’s hardening game back, on the narrow front. Look at how Beijing treats its own limiting factors, its LIMFACs (limiting factor, the military term for the bottleneck that constrains everything else). It stockpiles deep, five months of soybeans and 60 percent of world corn stocks. It diversifies origins, shifting soy from the United States to Brazil. It buys the midstream, 65 to 70 percent of lithium conversion. And it subsidizes the residual chokepoint it cannot buy or reroute, the chip toolchain. That is a coherent doctrine, and it is a doctrine anyone can run.

The West, mostly without saying so, spent 2026 standing up the same machine. The Comprehensive Outbound Investment National Security Act (COINS Act), signed December 18, 2025, made outbound investment screening statutory instead of an executive order. Project Vault, announced February 2, 2026, put up to US$10 billion of Export-Import Bank financing behind a public-private strategic critical-minerals reserve, the stockpile leg. FORGE (Forum on Resource Geostrategic Engagement), launched February 4, 2026 as the successor to the Minerals Security Partnership, is designing coordinated reference prices and price floors backed by adjustable tariffs, the midstream-economics leg. And the June 17, 2026 G7 platform agreed at Evian added monitoring, lithium and nickel stockpile pilots, and a dated target: below 60 percent dependence on any single non-G7 supplier of rare earths and magnets by 2030. Diversify, stockpile, own the economics, screen the capital. Each mechanism is a named, funded rhyme with a line of Beijing’s own playbook.

That is the defensive half. The offensive half is what the map licenses: hold the equipment, metrology, and HBM front, with the allies who own it, and hold it for years. The previous post’s arithmetic applies with full force here, because held leverage beats fired leverage, and this front only stays feasible while the three-country bloc stays aligned and indigenization stays slow. The doctrine is not “hit them where they are most dependent.” It is “hold what they cannot substitute, and harden what you cannot.”

Best Arguments Against This

The strongest attack: this feasibility scoring is a structured framework on public data, not an economic war plan supported by the full weight of the intelligence community and people who actually enjoy reading financial filings. Correct, and I will state it harder than the critic would. Four tests at two points each is a sorting rubric. The stockpile axis is mostly qualitative, because China does not publish strategic-reserve volumes for lithium or chemicals; undisclosed got scored, not measured. The framework carries no substitution elasticities, so “China substitutes poorly” is a reading of trade shares, not a model. And public trade data cannot separate leading-edge chips from mature ones inside a customs category. What survives the concession is ordinal, not cardinal: the claim is that dependence-sorted target lists are mis-sorted, and every one of those limitations would have to break in the same direction, food becoming usable and the chip toolchain becoming substitutable, to un-sort the map. I have never met an OPLAN that fit on a one page chart. This is the chart you check the OPLAN against.

The second attack: the vulnerability-targeting result is carried by one sector. Conceded, by design. The 0.67 correlation was printed above with its own obituary attached, because it dies without agriculture and chemicals falsifies it from below. Nothing downstream leans on the law; the findings lean on lithium’s supplier map, food’s reserve depths, and the toolchain’s concentration, which stand or fall row by row.

What Would Change My Mind

1. An omitted lever. If someone names an input China depends on that is allied-controlled, hard to substitute, thin-stockpiled, and low-blowback, and it is not in my fifteen, the map is wrong and the narrow front may be wider. Then the thread we pull on is why my analysis of Chinese financial filings identify that input. There would be gold at the end of that rainbow.

2. Indigenization moving the trade data. China’s world-export share in chip equipment and metrology climbing well past the 1.4 to 1.9 percent band, or frontier tools shipping from Chinese vendors at scale, degrades the 8 of 8 rows on a schedule of quarters, not decades.

3. A measured semiconductor escape. If China’s integrated-circuit import bill falls while its wafer capacity holds or rises, the defensive bet becomes a closing chokepoint and the front narrows from the other side.

4. The correlation resurrecting. A multi-specification vulnerability-targeting result that survives dropping agriculture and holds on all-channel cash would upgrade Beijing’s defensive arm from per-sector stories to doctrine, and make the mirror strategy more urgent, not less.

5. The architecture failing to ship. If the Evian stockpile pilots and FORGE’s price-floor design produce nothing bankable by the 2027 G7 summit, the hardening half of the mirror strategy is a communique, and this post’s recommendation loses its instruments.

The close

A dependence table is not a target list. It is the other side’s hardening schedule. Vulnerability is what hurts them; feasibility is what you can hold; and in China’s case the two coincide in exactly one place, a toolchain three inputs wide that the United States, the Netherlands, and Japan already control. Sort by dependence and your list starts with food that a Democracy can never touch and lithium you cannot corner without forming a Cartel. Sort by feasibility and the list is short, allied, boring, and real. Or become the godfather of a lithium cartel; I hear cartel chiefs get hippopotamus or is it hippopotami?

Next, Post 7: every number in this series that rests on China’s own reporting stops in 2024, because Beijing stopped filing. The data went dark on schedule, and that is a finding, not an accident.

That distinction, between what China holds and what happens when China fires, is the most consequential thing the trade data has to say about economic coercion. Washington mostly argues about the first chart, the dominance shares, and treats the second question as a matter of opinion. It is not a matter of opinion. Beijing has now fired the weapon enough times to grade it.

So this post grades it. Held first, fired second, and the seams between them are where the strategy lives.

Bottom Line Up Front

China’s export chokepoints are real, and the moment Beijing fires one they start losing it. Those are two different findings, and the policy debate keeps fusing them into one. As of 2024, the latest full reporting year, China held roughly 71.5 percent of world photovoltaic (PV aka solar panel) module exports, 67.3 percent of neodymium-iron-boron (NdFeB) magnets, 62.9 percent of coated steel, 55 percent of lithium-ion batteries, and 50 percent of cargo vessels. These are reporting from receiving countries, so they run slightly high, but no plausible correction drags any of them below half the world market. In strategic parlance, China is holding their adversary’s economic infrastructure at risk.

Firing these economic weapons behaves differently. The one completed case is the 2010 rare-earth embargo on Japan: Japan’s dependence on Chinese supply fell from roughly 90 percent to roughly 60 percent over the following decade. The 2023-2025 controls are printing the same signature faster, on partner-mirror reconstructions of trade China no longer reports. Gallium and germanium: exercised flow down about 40 percent in 31 months at a roughly flat price. Graphite: flow roughly halved with no price premium. Antimony: volume at about 0.15 times pre-control and price at about 3.2 times, a pattern that is ambiguous by construction. The April 2025 magnet controls are eleven months in, a dip and then a recovery, but too early to declare a verdict.

The honesty note sits up front: China stopped reporting its own trade after calendar year 2024, so every post-2024 number is rebuilt from what the rest of the world says it imported from China. That mirror is unbiased at the median and wide product by product. Erosion-on-firing is therefore the hypothesis the evidence leans toward, not a measured law; the single completed arc is a sample of one. The research data was current as of June 19, 2026.

How to read it: Japan’s dependence on Chinese rare earths after the 2010 embargo, with an inset showing the 2023-2025 controls printing the same shape faster.

What to see: fired leverage starts a clock, and every firing since runs the same decay, quicker.

The strategic translation: expect threats more often than embargoes. Held beats fired, and Beijing’s own behavior, pausing its broadest controls for exactly one year at the October 2025 Busan truce, reads like a government that has seen its own scorecard. Every one of those pauses expires in fall 2026. The clock only pays the defender if the warning is spent on diversification rather than relief.

The findings

1. The held inventory is wide, one-sided, and stamped 2024. Beyond the five families above, electric buses sit at 50.8 percent of world exports and lithium hydroxide, the battery-grade midstream chemical, at about 80.1 percent. The deeper feature is asymmetry: the dependence mostly runs one way. On a reverse-dependence index built from the same trade data, coated steel scores about 77.6, magnets 68.8, PV modules 59.2, and batteries 52.9. The rest of the world buys from China; China buys little back. Cargo vessels post a 99.9 on that index but we need to be very careful with that number: China imports almost no ships, so the denominator is tiny and the ratio explodes. The 50.2 percent export share is the useful ship number.

How to read it: green bars are China’s share of world exports in 2024, the latest full reporting year; purple bars are a single-year index of whether the dependence runs one way.

What to see: five product families clear 50 percent of world exports with one-sided dependence; the cargo-vessel 99.9 is a thin-denominator artifact, not the story.

Two caveats belong next to the inventory, not after it. First, asymmetry is not substitutability. The index reads one year of trade. It cannot see stockpiles, inventories, or how fast a buyer can qualify a second source, and those are what decide whether a held position can be fired without self-harm. Second, the mineral positions, rare earths above all, run on refining share and licensing machinery rather than visible subsidy cash; that structural machinery was Post 4. And the stamp matters: if you saw January 2026 headlines about China’s shipbuilding share falling, those describe 2025 new orders, roughly 63 percent of tonnage ordered and the first decline in five years, a different metric from the 2024 delivered-vessel export share, and both are true. One position is already sliding on its own: coated steel peaked at 68.6 percent in 2022 and sits at 62.9, roughly a sixteen-year half-life at that pace. A held position is not a permanent one. It is just the strongest form the leverage will ever take.

2. Caveat lector: the data went dark. China filed its 2024 trade with the United Nations trade database (UN Comtrade) in June 2025 and has filed nothing since; calendar 2025 onward is blank. The blackout sits exactly where the fired-control question lives. The workaround is a partner mirror: add up what every other country reports importing from China. Validated against China’s own 2014-2024 filings, the mirror is unbiased at the median, a ratio of about 1.02, but only about a third of product-years land within ten percent of truth, and the loosest lines are exactly the small mineral flows these controls touch. Worse, the mirror is annualized: the within-year halt, the weeks when licenses simply do not issue and the weapon actually bites, is invisible in available data. This is a bad place to bluff. So every grade below is a run-rate the mirror can support.

3. Every completed economic weapon firing shows the same arc: bite, then erosion. Walk the one finished case concretely. In 2010, during a territorial dispute, China halted rare-earth shipments to Japan. The embargo bit immediately; prices spiked; Tokyo had no near-term substitute. Then Japan spent the decade on stockpiles, recycling, substitution, and an Australian supplier, and its China-dependence slid from roughly 90 percent to roughly 60 percent. Some series put the slide at about 82 to about 54 percent by 2015; the band does not change the shape. The diversification is partial, heavy rare earths and processing dependence remain far higher than the headline, and Beijing is now re-applying pressure to the same target, tightening magnet-technology licenses aimed at Japan in January 2026.

How to read it: each row is one fired export control; the multiples compare the trailing-twelve-month run-rate after the control against the twelve months before it, for value, volume, and price.

What to see: every value multiple sits at or below 1.0; no firing since 2023 shows sustained volume at a price premium, the signature durable delivery would leave.

Since 2023 Beijing has refired the weapon four more times, helpful for my analysis but not for anyone else. All four regimes are licensing systems, and all four remain in force today. They are not the ban you read about: the separate December 2024 outright ban on gallium, germanium, antimony, and superhard-material sales to the US is not part of this scorecard and is not currently in force, suspended in November 2025 until November 27, 2026 under the truce. The licensing regimes graded here never paused.

Gallium and germanium, licensed in August 2023, are the strongest case for Beijing, and even they are not a win. China’s share of reported world exports held around 37 percent. But the exercised flow eroded underneath the share: trailing-twelve-month export value ran at about 0.60 of pre-control across 31 months, at a roughly flat price. Share durability with a shrinking, unpriced flow is not delivered leverage.

Graphite, licensed from December 2023, is self-eroding. Natural graphite ran at about 0.47 of pre-control value with no price premium, artificial graphite similar. A control that cuts your flow without raising your price is not coercion. It is a subsidy to your competitors.

Antimony, licensed from September 2024, is the ambiguous one, and the ambiguity is structural. Volume collapsed to about 0.15 of pre-control while price rose about 3.2 times. That pattern can mean the weapon bit, or that buyers fled a collapsing base; price-up plus volume-down does not self-interpret. Two outside checks help. The United States Geological Survey (USGS) shows China still holding about 60 percent of world antimony production while its export share fell, a genuine withheld-exports signal. And rest-of-world antimony supply rose about 57 percent in the year after the control. The adaptation half of the story shows up at the buyers, in data China cannot black out.

Magnets, licensed April 2025, are eleven months in: a dip in May and June 2025, then recovery toward baseline, value around 0.94. That is fewer than four clean quarters and it is not a durability verdict in either direction. The bite was real; Ford idled a Chicago assembly line for a week in May 2025 for want of magnets. Whether the bite lasts is precisely what the 2027 data will answer.

How to read it: left panel, post-control run-rate multiples by product, where 1.0 is the pre-control pace; right panel, the bound on how much Chinese value sits inside connector-country exports.

What to see: fired controls run at or below pre-control value everywhere; the connector bound is wide, which is why relocation gets graded by behavior, not by a point estimate.

4. Tariffs as economic warfare defenses was whack-a-mole. Export controls are China’s half of the exchange. The other half is what Western tariff walls do to China’s exported gluts, and the answer is not “stop them.” In 2024, roughly USD 5.5 billion of solar modules flowed out of Vietnam to the US and EU, built substantially on Chinese cells and Chinese-owned capacity. On April 21, 2025, Commerce finalized anti-dumping and countervailing duties (AD/CVD) on Cambodia, Malaysia, Thailand, and Vietnam at company rates from about 41 percent to over 3,500 percent. That route is now substantially closed. Within a year the flow respawned in India, Indonesia, and Laos, and the enforcement followed it: preliminary countervailing duties on February 24, 2026, preliminary anti-dumping on April 23, 2026, combined deposit rates roughly 100-234 percent.

Whack-a-mole is the right frame, and the moving mole is the point. Capacity that re-emerges one border over in under a year is relocation, not paperwork. Three markers say so. The solar cell-in to module-out ratios on the main routes collapsed from 2022 to 2024, Vietnam from about 0.081 to 0.008 and Thailand from about 0.260 to 0.008, which is rising local content, not rising pass-through. The disclosure trail agrees: cell and upstream plants abroad trigger billion-class board filings in their native currencies, while US module-assembly sites trigger no material filing at all. And the Chinese value embedded in connector exports is a bound, not a number: a 2-22 percent floor on 2022 input-output tables, with a mirror ceiling up to about 85 percent only on the transshipment-leaning solar lanes. Relocation and pass-through coexist, sorted by product; do not average them.

Europe runs the same play one act behind. About USD 4.0 billion of batteries flowed west out of Hungary in 2024, and Mexico’s EV flow to the same walled markets ran about USD 8.2 billion. The Hungarian route is open and producing, with the flagship battery and EV plants at Debrecen and Szeged reaching production start in early 2026, but it is newly contested: Hungary’s April 2026 election changed the government, the incoming coalition has pledged review of Chinese investments, and the EU’s revised investment-screening regime enters into force in summer 2026. The escalation ladder is more of a dance: what couples the two sides is imitation. When Washington writes an ownership rule, Beijing answers with an ownership rule; when one side reaches for extraterritorial licensing, the other copies the form within months. There is no tested tit-for-tat clock in this data. There is mirrored paperwork.

5. All of the economic warfare truces run out in fall 2026. The Busan framework of October 30, 2025 is a one-year truce, not a settlement, and its clocks are specific. China paused its October 9, 2025 rare-earth expansion, the five added elements and the extraterritorial rule reaching any foreign magnet with 0.1 percent Chinese-origin content, until November 10, 2026. It suspended the US-directed gallium, germanium, and antimony ban until November 27, 2026. Washington suspended the Bureau of Industry and Security (BIS) Affiliates Rule until November 9, 2026. Meanwhile the April 2025 licensing regime on seven rare earths was never suspended, and controlled rare-earth exports were still running about 50 percent below pre-control levels going into the May 2026 Beijing summit. The leaders meet again in the US in September 2026, weeks before all three alarms ring. It’s all right around the US mid-term elections and potential for a number of lame duck congress members and senators. And the yearly US continuing resolution season as political capital is expended on budgets.

Read Beijing’s posture against its own scorecard and it makes sense. A state that believed firing economic weapons was free would not holster this carefully. The 2025 magnet squeeze extracted real concessions, relaxed US positions on jet-engine parts, chip-design software, and ethane within about two months, and Beijing then sold the pause itself as a concession, twice. That is a threats-first doctrine: hold the inventory, fire briefly, monetize the fear, reholster before the erosion compounds. The Western counterpart obligation is just as specific. Each firing is a dated, free lesson in where dependence actually binds, and the truce year is the cheapest diversification runway anyone will be offered. Fall 2026 is the next scheduled lesson.

Best Arguments Against This

The strongest attack is the simplest: the exercise evidence is too recent, and China’s own trade data is missing after 2024. That attack is largely correct, and I concede the point. The mirror approximates China’s exports at the median; it never recovers them, and it cannot see the within-year license halt where the weapon actually bites. The 2010 arc is a sample of one. The magnet regime, the most strategically important firing, has eleven months of reconstructed data and could still print durable delivery in 2027. What survives the concession a weak claim that is strong enough to plan on: no export control China has actually used has ever paid off cleanly and kept paying.

The second-best attack: durability is the wrong test, because a weapon that hurts for two quarters can still coerce if the demand is small and fast. Also correct, and it is not a rebuttal. The 2025 squeeze got paid in concessions within two months, before any erosion mattered. That is exactly the held-beats-fired doctrine this post describes: the coercive yield front-loads, the depreciation compounds, and Beijing behaves accordingly.

What Would Change My Mind

1. The magnet window, readable in early-to-mid 2027: four clean post-control quarters showing sustained share and sustained volume at a price premium would kill the erosion hypothesis where it matters most.

2. China resuming its UN Comtrade reporting with a 2025 batch (on the 2024 cadence, that could land any time in the second half of 2026) whose real numbers materially diverge from the partner mirror. That would force this scorecard to be regraded.

3. Any fired control, on any product, holding share and volume at a premium through two full years. None has.

4. Firm-level value-added data showing connector exports are mostly pure transshipment. That would flip the relocation finding to pass-through and shrink the whack-a-mole story to a customs-fraud story.

5. A fall 2026 snap-back firing with no rest-of-world supply response within a year, against the antimony and graphite precedents. That would mean the adaptation machinery has stalled and fired leverage keeps more value than this record shows.

The close

China’s economic sector dominance charts are real, and they are the strongest the leverage will ever be. Every economic weapon firing converts a little more of it into someone else’s supply chain. If they truly fire all of their economic weapons then they will be out of economic ordinance within a few years. Beijing prices its weapon this way; the one-year pauses are the tell. Price it the same way, and spend the warning.

Next, Post 6: the West’s counter-target list, where the data says something uncomfortable. The deepest dependence is not the most feasible lever. Your target list is wrong.

This post is about the instruments that decide outcomes in China’s economic-warfare machine, and why almost none of them cost the state money. The evidence comes from a decade of audited annual filings of listed Chinese firms, with the delisted casualties deliberately kept in the panel, pinned to dated policy documents.

Bottom Line Up Front

You will miss the weapons if you track Beijing’s economic arsenal by the money. The instruments that decided the outcomes in this dataset are administrative: capacity ceilings, export licenses, production quotas, ownership consolidation. They cost the state almost nothing to hold. Budget telemetry, the thing Western analysis is built to watch, misses every one of them.

The core evidence is a natural experiment that Beijing itself. In 2022 the state withdrew the same subsidy instrument, cheap industrial electricity, from two power-hungry sectors in the same year. Aluminum smelting, which sat under a hard capacity ceiling, came out ahead of its own-sector peers by 9.4 percentage points of net margin and 10.1 points of return on invested capital (ROIC), with zero delistings in the capped cohort. Solar, which had no ceiling, saw margins fall 5.4 points once the dead firms are counted, with three of the sector’s five listed-census casualties landing in the post-shock window. Same lever, same year. The cap separated the survivors from the corpses. The caveat sits here, not in a footnote: the contrast rests on five treated smelters, a strong directional result rather than a decimal-precision law.

The second pillar is the one economic chokepoint weapon China has actually fired: critical minerals. The measured subsidy cash to listed rare-earth firms over the decade is CNY 6.8 billion, roughly USD 0.9 billion, an intensity of 0.07 percent of revenue. A rounding error. The leverage that idled a Ford assembly line in 2025 rests on the 2021 quota consolidation, export licensing, and state ownership. The cash bought nothing. The paperwork bought everything.

And Beijing itself just confirmed the model. Watching solar bury itself, the state reached for the cap, not the checkbook: a capacity-retirement campaign, not a rescue subsidy. More on that below.

So the watch instruction for a policy audience is uncomfortable but simple. Track caps, licenses, and consolidation. The money is the loudest and least informative signal the system emits.

How to read it: two annotated overheads, the capped Weiqiao aluminum complex and the uncapped Daqo polysilicon base, with their post-withdrawal outcomes.

What to see: same lever, same year, opposite fates. The cap decided. Imagery shows footprint and scale, not ownership or cash.

1. Beijing ran the experiment for me

One lever, two sectors, opposite fates. In 2022 the Chinese state cancelled the preferential electricity pricing that had built two industries. Aluminum smelting lost its cheap power. Polysilicon, the raw feedstock at the top of the solar chain, lost its cheap power. Both sectors are essentially electricity converted into product. Both shocks trace to dated policy documents.

Two champions make the contrast concrete before any statistic. One polysilicon maker went from a 48.7 percent gross margin in 2022 to negative 16.6 percent by 2024, selling below the cost of making the material. One aluminum smelter posted a record net profit of about CNY 24.6 billion, roughly USD 3.4 billion, in the same window. Same shock. Opposite sign.

Now the measurement. The tool is a difference-in-differences (DiD) design: take the firms hit by the policy, take same-sector peers who were not, measure how each group changed across the event, and subtract. The subtraction removes the period’s general weather, because the peers lived through it too. The design’s honesty check is parallel pre-trends, meaning the two groups moved together before the shock. On margin and ROIC, they pass for both sectors.

Treated aluminum smelters beat their peers by 9.4 percentage points of net margin and 10.1 points of ROIC after the withdrawal. Read that again. The state removed a cost advantage from an industry whose entire cost structure is electricity, and the industry pulled ahead. Not a paradox. A supply story.

The difference was a ceiling. China had written a hard cap of roughly 45 million tonnes of annual smelting capacity into policy in 2017. When cheap power appeared, smelters could not respond the way commodity producers usually do, by building more and racing to the bottom. The cost increase passed through into the aluminum price, and the incumbents captured it. Solar had no ceiling. Its response to the same shock was more capacity into a glut: listed solar capacity rose about 326 percent across the window while the sector’s median net margin went from 12.3 percent to negative 9.5 percent.

Two disciplines on this claim, stated next to it. First, I report aluminum alone. Blend aluminum with steel into one heavy-industry cohort and the signal washes out to a meaningless 1.1 points, an artifact of mixing two experiments. Second, the attribution to the 2017 ceiling stays modest: studied as its own event, its margin interval touches zero on eight firms. The DiD carries the claim. The ceiling is the mechanism candidate, not a second proof.

The cap is not a relic of the past, either. China’s 2025 primary aluminum output came in at 45.02 million tonnes, a record, and essentially at the 2017 ceiling even without cheap electricity.

2. Bring out your dead

Survivorship bias would have hidden the entire harm. This is the finding I would teach first, because I caught it pretty late in the project.

Run the solar DiD on the firms that still exist and the withdrawal looks benign: survivors beat their peers by 2.4 points of margin. Positive. Fine. Now ask who is in that number. The survivors, by construction. A survivors-only panel quietly deletes the firms the policy may have killed. Put the casualties back in, coded to the year they died, and the solar margin effect turns negative, minus 0.7 points. Nothing changed except whether I counted the dead. The exact size of the flip depends on the control group, so I state the control set every time; the direction never changes. Counting the corpses always makes solar look worse.

The delisted-inclusive accounting of the full post-2022 window is the cleaner summary: solar net margins down 5.4 points, ROIC down 3.6, and three of the sector’s five listed-census delistings inside the four-year post window. The capped cohort’s count in the same window: zero. Not one capped smelter delisted.

Where did the uncapped glut go? Abroad. Export unit values, total value divided by quantity, are a signal and not a price; they blend product mix, quality, and price. On that signal, lithium-ion batteries recently sold at roughly 0.80 of the world-average unit value on exploding volume, and solar modules and coated steel show the same signature. That is dumping pressure as a warning light, not a measured below-cost margin. Finished electric vehicles are the one exception, volume up at a stable relative unit value near 0.81, genuine cost-quality competition, and they stay out of this bucket.

For a strategist the survivorship correction is not a statistics footnote. The weapon does not need every producer to live. It needs enough output to survive long enough to deny everyone else a return. The casualties are not missing data. They are friendly fire in an economic weapon intended to kill (foreign) companies.

3. The lever works in reverse

Cut capacity and margins rise, on command. If a ceiling turns a subsidy withdrawal into a margin recovery, then removing capacity outright should do the same thing more directly. It does, and China ran that experiment too.

The 2016 steel de-capacity campaign, 去产能, “remove production capacity,” was a dated, enforced reduction. Run as a DiD against a machinery-matched control, it raised steel net margins by about 4.64 percentage points, with an interval of 2.46 to 6.82 that excludes zero, and gross margins by about 7.29 points. The causal weight sits on the margin legs only; I label the estimate conservative because the pre-trend bias runs against the effect. Utilization corroborates from an independent source: ferrous capacity utilization rose from about 71.7 percent in 2016 to 78.0 percent by 2018.

How to read it: real margin paths for the capped and uncapped cohorts around the 2022 power withdrawal, with the delistings marked on the falling branch.

What to see: one shared shock, a fork. Capped rises; uncapped falls with a body count.

The 2022 window supplies the negative control. Steel took the same power-price shock as aluminum but received no new supply discipline, and it walked into the property slump: minus 1.0 points of margin and minus 2.2 of ROIC against peers. The slump is a genuine confounder there, so I lean on steel only as the uncapped contrast, not as a second clean experiment.

State the claim precisely: supply discipline separated profitable survivors from casualty sectors under the observed 2016 and 2022 episodes. Cap forward, cut backward, same lever. Beijing can raise or lower an industry’s margins by adjusting how much of it is allowed to exist. No cash changes hands. Nothing shows up in a budget.

4. China’s only fired economic chokepoint runs on licenses, not cash

The rare-earth position was built with paperwork. Across this whole research project, critical minerals are the one place where China has repeatedly exercised chokepoint leverage against foreign targets, not merely held it. So look at what that leverage is made of.

It is not made of subsidy. The measured decade of cash to listed rare-earth firms is CNY 6.8 billion, about USD 0.9 billion, an intensity of 0.07 percent of revenue. For calibration, that is among the smallest sector figures in the dataset. The leverage rests on three administrative structures instead: production quotas, consolidated in 2021 into a handful of state groups; export licensing under the dual-use control regime; and state ownership of the consolidated groups. After the sector’s subsidy cash faded in the late 2010s, listed rare-earth margins actually rose, about 1.6 points, because the quota and license regime props pricing in a way cash never did. The state does not pay this sector to be a weapon. It permits it to be scarce.

The leverage is not hypothetical. China holds 67.3 percent of world rare-earth-magnet exports in the latest full reporting year, a reporting-countries-only figure that slightly overstates the share. When Beijing put seven medium and heavy rare earths under export licensing in April 2025, the queue for licenses idled Ford’s Explorer line in Chicago for a week within a month. At mid-2026, partner-reported trade data still had exports of the controlled elements running roughly 50 percent below pre-control levels, and that April regime stayed in force through every subsequent truce. The precedent cuts both ways, though: after the 2010 embargo episode, Japan cut its rare-earth dependence on China from roughly 90 percent to roughly 60 percent over the following decade. Fired leverage starts a diversification clock. Hold that thought for the next post.

Now the mirror, because it sharpens the lesson. Washington’s 2025 answer to this position was cash: an equity stake in its flagship rare-earth miner and a ten-year magnet-input price floor at USD 110 per kilogram, roughly double the miner’s 2024 realized price. A price floor is a rational answer to an adversary who can crash prices by adjusting a quota. But notice the asymmetry. The US instrument is budgetary, visible, and appropriated. The Chinese instrument is administrative, free, and invisible until the moment it binds. One side’s weapon shows up in telemetry. The other side’s does not.

How to read it: the left bars are a decade of cumulative listed-firm grant cash by sector, with rare earths highlighted; the right stack is the quota, licensing, and ownership structure that actually carries the mineral leverage.

What to see: the one chokepoint China has repeatedly fired sits on the smallest cash bar in the panel; the leverage was built by permission, not payment.

5. Beijing reached for the cap, not the checkbook

The thesis is confirming itself in real time. Since mid-2025, Beijing has been trying to retrofit exactly this discipline onto its solar casualty. The “anti-involution” campaign against ruinous competition was elevated to a central policy priority in July 2025. The instrument set that followed is the tell: a proposed CNY 50 billion fund, about USD 7 billion, to retire over 1 million tonnes of the roughly 3.5 million tonne polysilicon capacity base; a December 2025 consolidation joint venture spanning six producers; and 2025 polysilicon output down 26.4 percent. Execution is visibly hard, and the antitrust regulator’s January 2026 warning complicated the cartel-style mechanics. But the direction is unambiguous. The state watched an uncapped sector bury itself for three years, held the checkbook it has used for decades, and reached for the cap instead. Beijing read its own experiment the same way I did.

That is the lesson compressed. A state that has learned the cap works holds a cheaper, quieter, more deniable toolkit. A rival that audits budgets will systematically underestimate it.

Best Arguments Against This

The strongest surviving counterargument: these are sector cycles, not supply discipline. Aluminum is a global commodity that had a good post-2022 run; solar was in a worldwide glut that punished producers everywhere; any cap-versus-no-cap story is riding coincidental cycles and calling itself causal.

The objection has real force, and two features of the design answer most of it. First, the DiD nets out sector weather by construction: treated firms are measured against same-sector peers living through the identical cycle. Second, the steel mirror breaks the coincidence reading. The 2016 cut and the 2022 withdrawal are different years, different cycle positions, and different lever directions, and the margin response follows the capacity discipline both times, up when capacity is cut, down when the field is open. One cycle can mimic one result. It is much harder for two cycles, six years apart, to mimic opposite results on cue.

What I concede: five treated smelters is a small cohort; the steel 2022 contrast carries a property-slump confounder; and the 2017 ceiling attribution is a mechanism candidate, not a proven cause. There is also a fair version of the objection I have not fully answered: caps only mint winners where earlier cash built the capacity, so the instruments are sequenced, not independent. True. The telemetry point survives it. The cash phase is loud and mostly over; the control phase is silent and current.

What Would Change My Mind

1. A capped sector that busts. If aluminum, sitting at its literal 45 million tonne ceiling, posts sector-wide negative margins with delistings in fiscal 2026-2027 filings absent a demand collapse, the cap thesis weakens badly.

2. An uncapped sector that shrugs. Any uncapped Chinese sector that absorbs a dated support withdrawal in 2026 or 2027 and holds peer-beating margins would be the counterexample this dataset lacks.

3. The anti-involution test fails. If the polysilicon retirement fund and consolidation land as designed and solar margins still have not recovered by fiscal 2027 filings, the cap lever is weaker than the steel and aluminum episodes suggest.

4. An input-cost decomposition of the smelter cohort showing the post-2022 gains came mostly from falling alumina and power costs, independent of the ceiling.

5. A firm-level transaction-price series showing batteries, solar modules, or coated steel selling at or above comparable world prices, which would retire the unit-value warning light.

The Close

A subsidy is a payment. A cap is a decision about who is allowed to exist. Payments show up in budgets, get counted by analysts, and buy capacity that anyone can also buy. Decisions about existence show up nowhere, cost nothing, and cannot be replicated by a treasury.

The money is not the weapon. The permission is.

Next in this series: what actually happens when Beijing fires one of these chokepoints, and why the record says a chokepoint is a wasting asset with a clock that starts at the shot.

Bottom Line Up Front

Grade China’s industrial subsidies the way an economist would and the machine fails. I tested capability three independent causal ways: a difference-in-differences (DiD) design around dated subsidy withdrawals, a regression discontinuity at a corporate tax threshold, and an event study of state fund entries. All three agree with the raw association: state money buys research and development (R&D) spending and output scale, and it buys no measured productivity gain and no patent-quality gain. Among subsidized firms, more subsidy ranks with slightly lower patent quality (a rank correlation of about -0.20). Read that as a bounded null: no positive effect detected at my resolution, not a certified zero.

Grade it the way the state does and one effect is causally secured: survival. After a government guidance fund buys into a firm, the firm becomes measurably less likely to slide into distress. The coefficient on the distress flag is -0.033 with p = 0.003, and the pre-trends pass. That is the cleanest causal result in this entire project. It is a rescue effect, not an improvement effect.

Hold both results at once and the strategic picture resolves. The product of this machine is not excellence. It is endurance: excess capacity, retained engineers, intact supplier networks, and production routines that refuse to exit. A rival cannot wait this machine out. Any plan that assumes China’s weak firms will eventually die is a plan against the wrong machine.

How to read it: one square per listed firm in the sector census; red squares are firms that delisted.

What to see: the casualties are in the data. Every result in this post counts the dead.

One perimeter note before the details: this is the listed-company layer, so the biggest rescue instruments, unlisted funds and off-book recapitalizations, are systematically undercounted, which biases the survival finding down, not up.

1. The money tracks inputs, not results

The association is already telling before any causal design touches it. Take the roughly 7,900 firm-years across some 520 listed Chinese firms, 2010 through 2025, with the delisted casualties deliberately kept in the sample. Rank the subsidized firms by how much they get. More subsidy comes with more R&D spending and more output scale. It does not come with more capability. Patent quality ranks slightly negative against subsidy intensity (about -0.20), and the total factor productivity (TFP) residual sits at roughly zero. The state is buying inputs. The outputs those inputs are supposed to produce do not show up.

How to read it: four cards, one per research design; the three muted cards are the capability nulls, the green-bordered card is the survival effect.

What to see: three ways of asking whether state money buys excellence come back empty. The one causal yes is survival.

The US-citation slice deserves its own sentence, because it is corroboration and not proof. I scored 41 firms on roughly 2,000 US-granted patents using forward citations, the standard quality proxy. More subsidy associates with lower per-patent citations in all three specifications, rank correlations of about -0.15 to -0.27, none individually significant. That is a narrow, survivor-skewed subsample. It corroborates the null from independent data. It does not power a test. The panel that would settle the question, citation-weighted Chinese, European, and world patent filings, is paid data I did not buy. I will come back to that, because it is the strongest argument against this post.

2. Three causal designs, one shape

Associations have confounds, so I asked the question three causal ways, on three different data structures. Their agreement is the part I trust most.

Walk the cleanest one fully, because it shows the mechanism. A firm certified as a High and New Technology Enterprise (HNTE) pays corporate income tax at 15 percent instead of the standard 25 percent. Certification hinges partly on an R&D-intensity threshold: 3 percent of revenue. A hard threshold is a natural experiment. Firms just above the line and firms just below it are close to identical, except one group gets the tax break. That is a fuzzy regression discontinuity design (RDD), and the first stage is strong: crossing the 3 percent R&D line raises the probability of paying the low rate by 13.9 percentage points, with a confidence interval of 8.4 to 19.3. The tax door really opens at the R&D line.

Capability does not walk through it. Every capability outcome behind the threshold is null, and the one statistically significant sign runs the wrong way: sector share falls at the cutoff (p = 0.021). A 2021 study in the American Economic Review of this exact tax notch explains why. Firms bunch right at the 3 percent line and relabel spending to clear it; about 24.2 percent of reported R&D at the notch is relabeled from other budgets, not new science. The subsidy bought reported R&D. It did not buy what reported R&D is supposed to represent. And a caution that generalizes: HNTE-certified firms are the healthiest cohort in the panel, and that is selection, not treatment. Taxable income is the precondition for a tax break. The tax door only opens for firms that already make money.

The other two designs stay at altitude, and they rhyme. The withdrawal DiD studies dated subsidy removals: if the money had built capability, pulling it should reveal capability moving. The 2022 power-support withdrawal moved sector margins by roughly nine percentage points, a story Post 4 owns, and the capability needle barely moved with it: about +0.8 points of R&D intensity and +0.04 of the productivity residual, both inside the noise. The fund-entry design studies government guidance funds as staggered events, with a sector-mandate instrument confirming entry is real (first stage +9.6 percentage points, p below 1e-10). The balance check fails in the revealing direction: the firms the funds picked were already ahead before the fund arrived, by about 2.4 points of R&D intensity and roughly 3 to 5 points of sector share. Post-entry capability: null. The state picks winners more than it creates them, and the clean causal result is survival, not improvement.

The null also survives the obvious measurement attacks. Recomputing productivity with sector value-added deflators moves the estimate toward zero, never positive (about -0.01 across all firms, -0.15 among subsidized firms). The grant channel shows the relabeling signature beyond the tax notch: grant intensity ranks +0.36 with R&D-spend growth and negative with real productivity, while the patent-quality reading straddles zero. Better measurement does not rescue the claim. If you came here for proof that a state can purchase innovation on command, this is the wrong post.

3. The one thing the money causally buys is survival

Fund entry keeps firms alive, and that result is clean. After a guidance fund buys in, the probability of going-concern distress drops: coefficient -0.033 on the distress flag, p = 0.003, pre-trends pass. Weeks of failed experiments, then that. I did not expect the most strategically important number in this project to come out of an accounting flag.

Do not soften this until it disappears. Survival preserves capacity, employment, supplier relationships, and engineering teams. It preserves the option to surge later. If your objective is strategic capacity rather than shareholder return, keeping a weak firm alive is rational even when it ruins a productivity chart.

Do not harden it either. Survival is not capability. A tourniquet is not a training program. Inside state-owned enterprises (SOEs) the sign even flips: within the SOE cohort, more subsidy tracks more distress, which is what you expect when the money is running toward the bleeding. And the firms outside the rescue perimeter show what the rescue is worth: going-concern distress hit 26.4 percent of the delisted casualties, against 0.7 percent of the core champions and 7.1 percent of their listed peers.

4. The survivors lean, and the leaners die harder

Endurance has a texture, and the texture is fragility that does not exit. Across the decade, 11.6 percent of profitable firm-years clear zero only after counting one-off items: asset sales, fair-value swings, impairment reversals, and government grants. Call that non-recurring-dependent, never subsidy-dependent; grants are the largest single component but only about 32 percent of the profit-flipping gap at the median. The share is macro-cyclical, not a trend: 18.9 percent in 2014, down to a 6.6 percent trough in the 2021 boom, back to roughly 10 to 12 percent in the 2023 to 2025 downturn. The state’s flattery of the books peaks exactly when the cycle turns against the firm.

Now the survivorship contrast, which is why the dead firms stay in my sample. Among the delisted casualties, about 33.7 percent of their profitable firm-years leaned on one-off items, measured across 172 profitable firm-years, against roughly 9 to 12 percent for survivors. The firms that died were leaning about three and a half times harder. Read a survivor-only panel and you would never see it.

Airlines are the case that makes rescue visible as a shape. From 2020 through 2025 the core state carriers absorbed roughly CNY 117 billion (about USD 16 billion) in recapitalizations against roughly CNY 208 billion (about USD 29 billion) in cumulative losses, with COVID as the named confounder behind the worst years and only five carriers in the post-COVID window, so treat it as an illustration rather than an estimate. The shape is the point. The peak recapitalization landed in 2023, a full year after the worst loss. The carriers were still loss-making after the money went in. Capital went in; returns did not come out. That is equity buying solvency, not recovery. It is the survival result drawn in a single sector.

5. Endurance is the strategic product

Translate this to the state level and the machine snaps into focus. China does not need every subsidized firm to become excellent in order to change another government’s choices. It needs enough capacity, engineers, suppliers, and production routines to stay alive that a rival cannot simply wait for the weak firms to die. Market logic says the shakeout comes. The rescue machine says it does not.

That forces the rival state into an unpleasant menu: protect your own industrial base indefinitely, accept dependence, or pay to build a parallel one. Every option is expensive. None of them is “wait.”

In military language, this is not force improvement. It is force preservation. An economist looks at a subsidized Chinese firm with flat productivity and sees waste. A planner should look at the same firm and see a unit that has been kept in the field for fifteen years without ever winning an engagement, and has not gone home either. Capacity and scale gains are real in the data too, but they arrive through selection more than treatment, so I hold them at association strength. The load-bearing causal fact is simpler. The firms do not die.

Best Arguments Against This

The strongest surviving attack: I measured the wrong quality. My patent yardsticks lean on US-granted filings and an invention-share proxy. Maybe Chinese-language patents, process learning, or tacit engineering know-how would show the capability gain the US filings miss.

It is a fair objection, and it is the residual I could not close on free data. Here is the honest answer. The invention-share proxy on the full panel is negative. The US forward-citation slice is negative in all three specifications, independent data pointing the same way. The null survives recomputing productivity on deflated output rather than nominal revenue. And three causal designs on three separate data structures all fail to find a capability gain. For the objection to win, the unpurchased citation-weighted CN/EP/WO patent panel would have to reverse a sign that holds in every free design I can run. It might. That panel is the named falsifier, and until someone buys it, the residual stands: the capability counterfactual for firms that applied and were refused is unobtainable on free data, and I log that rather than guess it.

The second-best attack is that survival gains from fund entry are just selection too, funds picking firms that were never going to die. The pre-trends say otherwise: treated and untreated firms tracked each other before entry and diverged after. Selection explains who the funds picked. It does not explain the timing of the divergence.

What Would Change My Mind

As of the June 19, 2026 data freeze, none of these has fired.

1. A citation-weighted CN/EP/WO patent panel showing state support associated with higher per-patent quality would break the capability null.

2. A clean productivity gain surviving pre-trends in either the withdrawal DiD or the fund-entry design would mean the money buys improvement, not just inputs.

3. A balanced fund-entry sample showing post-entry capability creation, rather than funds entering already-ahead firms, would flip picking winners toward making them.

4. A secular, monotone decline in the non-recurring-dependence share, replacing the 2014 to 2025 cycle, would mean the system is weaning its firms rather than sustaining them.

The mental model to carry forward

Stop asking whether China’s subsidies work. Ask what they bought. They bought R&D budgets, output scale, and above all the refusal of subsidized firms to die. By the economist’s yardstick that is failure. By the state’s yardstick it is the mission.

How to read it: the left panel shows the three causal designs, with the real first-stage and selection effects in green and the capability reduced forms in gold; the right panel grades the same money on the economist’s yardstick and on the state’s.

What to see: the tax door and the fund selection are real, capability barely moves, and survival is the one causal win.

China’s subsidy machine is not a capability forge. It is a life-support ward. And a rival that plans for the patients to die is planning against the wrong machine.

Next in this series: the checkbook turns out to be the loudest and least informative instrument Beijing owns. The deadliest instruments are free.

That inversion is the subject of this post. It is also the quiet death of Washington’s favorite one-liner about China: “they subsidize their way to global dominance.” I spent a few weeks building the sector ledger that could test that line. The line fails. And the way it fails is more useful than the line ever was.

Bottom Line Up Front

Stop reading Beijing’s subsidy ledger as an ambition map. Read it as a fear map. Across the six strategic sectors I measured (solar, electric-vehicle (EV) batteries, rare-earth magnets, steel and aluminum, shipbuilding, and legacy semiconductors), the ranking is almost exactly inverted: the more intensely China subsidizes a sector, the less export leverage it holds there.

Semiconductors take the heaviest subsidy intensity of any sector, 7.3 percent of listed-firm revenue, plus roughly CNY 368 billion (about US$51 billion) in state equity. China remains the world’s largest chip importer, and its chip import bill rose from about US$250 billion to about US$413 billion over the decade of the buildout. The deepest one-way chokepoints China holds, cargo ships, coated steel, and rare-earth magnets, sit on the thinnest listed-firm cash in the panel. Six sectors is a pattern, not a law. But the pattern is consistent under every measure I tried, and it survives removing any single sector.

Here is what a decision-maker should now believe. Subsidy flows mark where Beijing feels exposed, not where it plans to dominate. Dominance formed where the product was exportable and the technology achievable, often on thin cash. The heavy cash sits where China is losing.

The caveats belong next to the claim, not in a footnote. The panel is six sectors. The mechanism is selection, not strategy; nobody in Beijing wrote “fund our failures” on a whiteboard. And shipbuilding’s near-zero measured intensity partly reflects support the listed ledger cannot see. All three get unpacked below.

The Findings

1. Rank the sectors both ways and the ladders cross. Subsidy intensity here means each listed firm’s disclosed state support as a share of its revenue, averaged across firms and years within the sector. Line the six sectors up by it: shipbuilding reads 0.0 percent (an artifact, fixed in finding three), rare-earth magnets 0.07, steel and aluminum 0.24, solar 1.4, EV batteries 4.0, semiconductors 7.3. Now line the same sectors up by one-way trade dependence, an asymmetry score built as rest-of-world dependence on China minus China’s dependence on the rest of the world, computed on each sector’s flagship product in 2024 trade data, reporting countries only: ships 99.9 (a thin-denominator artifact, also flagged below), coated steel 77.6, magnets 68.8, solar 59.2, batteries 52.9, chips minus 9.

How to read it: each line is one of the six measured sectors, connecting its rank by listed-firm subsidy intensity (left) to its rank by one-way export leverage (right); these are ranks, and no line is fitted. What to see: the lines cross almost perfectly; the heaviest-subsidized sector, chips, lands last on leverage, and the deepest chokepoints climb from the bottom of the cash ranking.

The first ladder climbs. The second falls. Sector for sector, they cross.

If you want the one statistic, here it is, once: a rank correlation of -0.94 across the six sectors (exact permutation p = 0.017, computed on subsidy intensity versus trade-asymmetry rank; descriptive, not causal). Six sectors is a typology, not an estimated law. Remove any single sector and the inversion stands. Swap the measures, grant cash for intensity, world export share for asymmetry, and every pairing still comes out negative, though none of the alternates clears significance on its own. Six points cannot estimate a relationship. They can only exhibit a pattern.

2. The heaviest bet sits on the weakest position. Semiconductors absorb the heaviest firm-level support intensity in the panel, that 7.3 percent of revenue, plus roughly CNY 368 billion (about US$51 billion) of state equity through the national chip funds and their provincial siblings. The cash did not buy export leverage. China’s world export share in integrated circuits (ICs) runs 8 to 27 percent depending on the chip category, with the strongest slice, memory, at the top of the band. Chips are the only sector in the panel where the dependence runs against China: it is the world’s largest importer of them. The import bill did not shrink under the buildout. It rose, from about US$250 billion to about US$413 billion over the decade.

It is still rising. In January and February 2026, after more than three years of allied export controls, China imported 550 billion yuan (about US$76 billion) of ICs, up 36.8 percent year on year by value and 9 percent by volume. You will hear that China’s chip exports are surging too, up nearly 70 percent over the same months. They are, and they are overwhelmingly mature-node. The exports are the chips China can make. The imports are the ones it cannot.

One perimeter note, stated plainly: my intensity figure sees listed firms. Several of the biggest chip bets, the national memory champions and the tool builders, sit substantially off the listed books in state-held vehicles. Count them and the chip bet grows. The inversion deepens; it does not soften.

3. The chokepoints formed on thin cash. Now walk the other end of the ladder. On 2024 trade data, the latest full reporting year, China holds at least half of world exports in a set of products that includes photovoltaic (PV) modules at 71.5 percent, neodymium-iron-boron (NdFeB) magnets at 67.3, coated steel at 62.9, lithium-ion batteries at 55.0, and cargo vessels at 50.2. Those shares are computed from reporting countries only, so they run slightly high. They are still enormous.

How to read it: green bars are China’s world export share per product (2024, reporting countries only, so slight over-estimates); purple bars are the one-way-dependence score for the same product (how much the world depends on China for it, minus how much China depends on the world). What to see: every product here clears half of world exports on thin sector cash; the cargo-vessel 99.9 is a thin-denominator artifact, flagged on the chart itself, because China barely imports ships.

Look at what built the deepest of those positions. The magnet chokepoint rests on about CNY 6.8 billion (under US$1 billion) of listed grants across a decade, an intensity of 0.07 percent of revenue. The instrument was structural: mining quotas, licensing, and a 2021 consolidation of the producers. Structure, not checks. Shipbuilding’s measured intensity literally reads zero, and that zero is an accounting artifact. The state’s support ran through the balance sheet, roughly CNY 210 billion (about US$29 billion) of cumulative support, mostly debt-to-equity recapitalizations that a flow-over-revenue metric cannot see. Heavy rescue, thin flow. Either way it is a fraction of the chip bet, and it sits beneath half of the world’s cargo-vessel exports. Coated steel holds 62.9 percent of world exports on 0.24 percent intensity.

One time-stamp so the headlines do not bite you: in January 2026 you may have read that China’s shipbuilding share fell for the first time in five years. That was share of new orders, roughly 63 percent for full-year 2025, dipping as Korea rebounded under US port-fee pressure. It is a different metric from the 2024 delivered-vessel export share printed here, and a useful reminder that held positions erode.

4. The inversion is selection, not strategy. Do not read finding one as “subsidies prevent dominance.” The mechanism is simpler and colder. The state pours money where the frontier is hard and the exposure hurts: chips above all. It spends lightly where dominance came cheap, from an achievable manufacturing base, or from resource position and quota, because dominance that comes from structure never needed a check. Where the product was exportable and the technology achievable, solar, batteries, ships, coated steel, capacity converted into world share, sometimes with cash helping. Where the frontier was hard, the same machine bought capacity, engineers, and a rising import bill instead.

So the think-tank chart that plots subsidy totals next to export dominance and draws an arrow is plotting a coincidence backwards. The cash did not buy the chokepoints. And the chokepoints did not need the cash.

5. The same ledger tracks China’s own vulnerabilities. There is independent corroboration for the fear-map read. Rank the six sectors by how exposed China itself is to inbound cutoffs, and subsidy intensity tracks that ranking positively, a rank correlation between 0.83 and 0.93 depending on the vulnerability measure. Same six sectors, same small-panel limits, and semiconductors do much of the work, so treat it as corroboration, not proof. But both rankings agree on the read: the cash flows toward the exposure.

How to read it: each dot is a sector; the horizontal axis is China’s own inbound vulnerability, the vertical axis is listed-firm subsidy intensity. What to see: the money climbs the fear gradient. Corroboration, not proof.

That makes the ledger an intelligence product. If you want to know where Beijing believes it can be hurt, do not read its white papers. Rank its subsidy intensities. The allied chip controls that began in October 2022 landed on the exact sector the ledger marks as the deepest fear: 7.3 percent intensity, US$413 billion of imports. Washington read China’s chip push as an ambition and answered with a US$52.7 billion chip program of its own. The ledger says the push was always defense. Both readings end with chips mattering most. Only one of them predicts where the next yuan goes.

Best Arguments Against This

Three attacks survive contact, and the strongest deserves full volume.

The strongest: with six sectors, a determined skeptic can call the exact ranking luck. Six points is a coffee-table dataset, and the headline pairing was found by exploring the measured panel, not pre-registered. Both true. My response is not a better p-value; it is that the coefficient was never the load-bearing part. The per-sector facts carry the weight: the chip bet is the largest and chips are the import hole; the magnet chokepoint was built by quota and consolidation on trivial cash; ships were rescued through the balance sheet, not the income statement. Delete the statistic entirely and those three facts still invert the story.

Second: shipbuilding’s zero flatters the inversion. Correct. Restore shipbuilding’s equity support to the ranking and the inversion flattens but does not flip. The direction survives.

Third: the mid-ranking sectors cut the other way. Solar, at 1.4 percent intensity, holds 71.5 percent of world module exports; batteries, at 4.0 percent, hold 55.0. Heavy state cash and real dominance coexist there, and the cash plausibly helped build both positions. Also correct, and it bounds the claim. Money can help build dominance where the product is exportable and the technology achievable. What the ranking kills is the general law, subsidy in, dominance out. The heaviest money went where dominance never came.

What Would Change My Mind

Four triggers, dated and falsifiable.

1. China’s 2025 trade data lands on UN Comtrade (on the cadence of the 2024 batch, plausibly in the second half of 2026). If re-ranking on 2025 flows turns the chip asymmetry positive or flips the intensity-versus-leverage inversion, the fear-map read dies.

2. Four consecutive quarters of falling Chinese IC import value, checkable by mid-2027. That would mean the defensive bet is finally closing the exposure, and the “losing” in my title needs retiring. As of January and February 2026 the bill was growing 36.8 percent year on year, which is the wrong direction for this trigger.

3. Fiscal-2026 annual reports, filed by spring 2027, showing subsidy intensity surging into the held chokepoints, magnets or shipbuilding climbing toward battery-level intensity. Money chasing strength would break the inversion going forward.

4. A replication on a broader panel, ten sectors or more, in which intensity rises with export leverage. Six sectors exhibit a pattern; a bigger panel could overturn it.

The Close

A subsidy ledger is not a statement of ambition. It is a map of fear. China’s map is specific: the state is calm about the sectors where the world depends on China, and frightened about the one where China depends on the world. Read the ledger the way a poker player reads a nervous bet. The money is not the plan. The money is the tell.

Next in the series: follow the money down to the firms and ask what it buys where it lands. The answer is not excellence. It is refusal to die.

The military has a term for a strategy that does not try to beat you in the open field but makes it prohibitively expensive for you to show up at all: anti-access/area-denial, A2AD. China built the economic version. I started this trying to understand PRC industrial subsidies, the data has the size and shape of an economic warfare arsenal.

Bottom Line Up Front

China runs an economic-warfare machine with three measurable arms: market denial pointed outward through exported overcapacity, access denial held in reserve through export chokepoints, and a defensive arm that pours the heaviest money into hardening China’s own bottlenecks. Keep one thing in mind through this whole series: this is what the system does, as measured. I have no proof of two PLA Colonels writing a secret economic warfare plan.

What makes the machine unusual is not the size of the checks. It is breadth. China runs material programs across all nine support channels in my taxonomy: profit-and-loss grants, deferred grants, state equity, tax relief, demand-side purchase support, administered power prices, below-market inputs, directed credit, and in-kind transfers of land and infrastructure. The peer economies run three or four. The European Union scores four of nine, South Korea four, the United States three, Japan three. The caveat sits next to the claim: that count rides my own taxonomy and a judgment about what counts as material, so treat the nine-versus-three as a shape, not precision. Spending intensity points the same direction. An IMF-derived estimate puts China’s industrial support near 4.4 percent of GDP against 0.9 for the EU and 0.4 for the United States; the direction is robust, the exact multiple is not, because the accounting bases differ.

Why now? The Chinese/American trade clocks run out in the fall 2026. The Busan truce of October 30, 2025 was a one-year pause, not a settlement, and its expiries are specific. China's October 2025 rare-earth expansion, including the extraterritorial rule reaching any foreign magnet with 0.1 percent Chinese-origin content, is paused until November 10, 2026. Its US-directed gallium, germanium, and antimony ban is suspended until November 27, 2026. Washington's Affiliates Rule is suspended until November 9, 2026. The April 2025 licensing regime on seven rare earths never paused at all, and controlled exports still ran roughly 50 percent below pre-control levels into the May 2026 Beijing summit. The two presidents meet again in September 2026, weeks before all three alarms ring.

Here is what a policy-maker should now believe. China’s economic statecraft is not a larger version of everyone else’s subsidies. It is a different class of system, running more instruments at greater intensity, and its behavior, measured sector by sector, follows a loop: build capacity, overshoot, export the glut, convert export share into chokepoints, and spend the biggest money defending the economic chokepoints held against China. Whether anyone in Beijing drew that loop on a whiteboard is a question the data cannot answer. Whether the loop exists is answered by the data. The research data behind this series was current as of June 19, 2026. Yuan figures convert at roughly CNY 7.2 per dollar throughout.

1. The loop is measured behavior, not a blueprint

Let’s start with the concession. A large share of the machine’s supply side is not central strategy at all. It is bureaucratic competition. Chinese cities fund factories the way American universities fund football programs: because the neighboring city has one, and because careers are scored on it. The payer records show the split plainly. Cities and provinces fund supply, the capacity race. The central government funds demand and security, the purchase subsidies and the strategic bets. Nobody architected the machine in one document that I can find.

That is precisely why the Economic A2AD frame survives scrutiny. I am not claiming intent. I am claiming mechanism. A rainstorm has no plan, and you still bring an umbrella. Local tournaments build the capacity. The uncapped sectors overshoot and bust. The glut exits through the ports. The export share accumulates into one-way dependencies. And the central government, watching which dependencies point back at China, aims the heaviest cash at exactly those. Every link in that chain is measured somewhere in this series, at its own strength, with its own failure modes printed next to it.

How to read it: seven nodes trace the loop from instruments through capacity, overcapacity, export glut, and world dominance to a latent chokepoint, with a dashed defensive arm pointing back at China’s own vulnerabilities. What to see: the machine is a loop, not a list of subsidies, and one arm of it aims home.

2. Breadth is the anomaly

Nine channels but never one total. The first discipline that we enforced on the project is that the nine subsidization channels cannot be added together. A deferred grant is a stock; a power-price discount is a flow; state equity is a balance-sheet position. Anyone who hands you one all-in “China subsidy number” has already made an accounting error you should not forgive. The channel mix is not a measurement nuisance. It is the message. Tax relief rewards firms that already have profits. Demand subsidies conjure markets. Equity keeps strategic bleeders alive. Reading which channel a sector gets tells you what the state’s intent, and that decoder ring does more work in this series than any yuan figure.

How to read it: rows are governments, columns are the nine support/subsidization channels under one collapsed taxonomy; a filled cell is a channel run at material or primary scale, a small circle is a minor program below the bar. What to see: China’s row is fully lit. No peer row lights more than four of nine.

The intensity ranking points the same way, held loosely. China near 4.4 percent of GDP, the EU at 0.9, Korea at 0.67, Japan at 0.50, the United States at 0.39, on an IMF-derived basis. The one study that measured everyone on a single yardstick put China lower in level but still far ahead of every peer. Direction: settled. Multiple: honest range, not a point.

3. The method is the product

Three disciplines run through every post, and they are where this series earns the right to its conclusions. First: count the dead. Most subsidy studies grade the survivors, which is like grading a casino by interviewing the people still at the tables. Every sector-level result in this series includes the delisted firms, and in at least one sector that choice flips the sign of the answer. Second: when China’s own data goes dark, and it does, the numbers get rebuilt from the other side of the transaction, from what the rest of the world reports importing from China, with the error bands printed rather than waved at. Third: causal claims, associations, and hypotheses are kept in separate sentences. When a post says the state causally bought survival, there is a research design behind the word, and when a post says a pattern is suggestive, it means exactly that and no more.

The policy debate about China runs almost entirely on opinions with artistic infographics. The wager here is that Chinese financial ledgers, read with discipline, will beat opinions. And hard-conclusions become more actionable when soft-suggestions are called out as deriving from spotty data.

4. Washington started copying it, which is the tell

The strongest objection to this series used to be “every country subsidizes.” In 2025 the United States answered it for me. In August 2025 the federal government converted CHIPS Act grants into an equity stake of roughly 10 percent in Intel. A month earlier the Department of War took a 15 percent position in MP Materials, the rare-earth producer, and wrapped it in a ten-year price floor and an offtake guarantee. By October the federal government held stakes in five public companies. I am not grading those moves here. I am observing that when Washington starts buying golden shares and writing price floors, it has conceded that grants-versus-tariffs was never the real instrument set. The channels are the economic weapon. The United States just started acquiring them.

5. What the next seven posts will show

Each post kills or sharpens one piece of Washington conventional wisdom, stands alone, and carries its own chart. In order:

Post 2. The Money Goes Where China Is Losing. Across the six sectors measured, subsidy intensity and export leverage rank almost exactly inverted. Semiconductors absorb the heaviest support intensity, 7.3 percent of listed-firm revenue plus roughly CNY 368 billion (about US$51 billion) in state equity, and China remains the world’s largest chip importer. Read Beijing’s ledger as a fear map, not an ambition map.

Post 3. It just so happens that your friend here is only MOSTLY dead. Subsidies don’t buy excellence, they buy refusal to die. Three independent causal designs find no measured productivity or patent-quality gain. One effect is causally secured: survival. The product is endurance, and you cannot wait it out.

Post 4. The Deadliest Economic Instruments Are Free. The same 2022 support withdrawal made capped aluminum more profitable by 9.4 points of margin and buried uncapped solar. China’s one repeatedly fired economic weapon, critical minerals, runs on quotas and licenses built on a rounding error of cash. Track caps and licenses, not budgets.

Post 5. Weapons Depreciates When Fired. As of 2024 China held 50 to 72 percent of world exports across five product families. Held leverage is real. Fired leverage erodes: the 2010 rare-earth embargo cut Japan’s China-dependence from roughly 90 to roughly 60 percent over a decade, and the 2023-2025 controls are printing the same signature faster.

Post 6. The Target List Is Wrong. Vulnerability and feasibility are nearly orthogonal. China’s deepest dependences, food and lithium, are the West’s worst levers. The usable front is three chip inputs wide and requires allies.

Post 7. Never Verify Against a Sensor the Adversary Owns. Twelve documented Chinese reporting changes, zero strictly false statements, and no 2025 trade data posted to the UN as of July 2026. Never write a verification regime against a sensor the adversary owns.

Post 8. So, Is it Working? There is one causal win, and it is not the one Washington argues about. The honest grades, cell by cell, and six operating rules for US strategy, each tied to a measured finding.

Best Arguments Against This

“This is local overinvestment and bureaucratic competition, not warfare.” Largely true on the supply side. The answer is that the frame never required intent. The loop’s outputs, chokepoints held, gluts exported, defenses funded, are real regardless of whether the inputs were ordered or emergent. Post 2 will show the strongest evidence for the emergent read: the money flows toward China’s fears, not its ambitions, which is exactly what a fear-driven allocator looks like and exactly what a master plan does not.

“Your nine-of-nine is your own taxonomy.” It is. A different taxonomy would move the count. I have not found one under which the peers reach China’s breadth, and the graph above states the judgment it rides on. If someone produces a defensible taxonomy where the EU runs nine of nine then please send it my way.

What Would Change My Mind

1. A primary Chinese central-planning document that specifies the loop as doctrine. That would upgrade “operating model” to “intent,” and this series would say so.

2. A pre-specified, cross-sector test in which subsidy cash positively predicts later export dominance. Post 2 stakes the opposite claim.

3. Peer economies scoring seven-plus of nine channels under this project’s own taxonomy.

4. China resuming trade reporting to the UN and reconciling the flagged gaps benignly. Post 7 carries the dated version of that trigger.

5. A fired Chinese export control that holds market share and volume at a price premium for four consecutive quarters. Post 5 explains why none has yet.

Economic A2AD is not a doctrine China published. It is a behavior China exhibits. Doctrine is an argument about documents. Behavior is an argument about data, and data is what the next seven posts are made of.

Post 2 runs next: the money goes where China is losing.

How to read it: each bar is a level of AI adoption. Blue (above the line) is the annual national-security gain from productivity using the Torres defense production-function model. Red (below) is deliberately conservative: it charges every bar the full expected annual cost of all ten feared security-event categories, as if AI adoption owned the entire security ledger. The white tick is the net. No adoption escapes none of that cost, because the same threat categories (insider risk, spillage, foreign cyber activity, supply-chain exposure, and today’s unmanaged workarounds) persist under the status quo as part of the workforce reverts to unmanaged AI usage; it is all cost and no gain. By medium adoption the gains win, reaching +8.9% per year improved national security impact at full adoption. The gold zone is where the U.S. government started in early 2026 in chatbots on older generation AI models: the only zone where the net is negative. Yes, everyone is moving as fast as they can. This article is about turning the Security organization into the accelerator.

Security professionals see real adversaries, real classified information, real CUI, real insider risk, real supply-chain risk, and real accreditation obligations. They are right to reject any AI adoption plan that treats those risks as paperwork.

The broader workforce sees something else that is also real: modern AI can already compress writing, coding, research, discovery, review, triage, summarization, training, and administrative work. On classified and CUI systems, those are not office conveniences. They are mission throughput.

The mistake is turning those truths into a binary fight between AI and security. That is not the decision. The decision is whether the government should:

1. delay while conducting narrow pilots;

2. tolerate unmanaged workarounds as users seek capability outside the approved path; or

3. field managed AI inside accredited CUI and classified environments with measurable controls, logging, provenance, no external egress, model supply-chain controls, and human verification for high-impact work.

That third path is the policy this model supports. Not because AI is safe. Because managed adoption appears safer than either unmanaged adoption or delay once productivity, baseline threats, AI-incremental risk, and defensive control gains are scored on the same national-security ledger.

Security should not have to prove AI is risk-free before fielding it. The burden of proof for delay should be that a proposed delay reduces more national-security harm than it causes.

Bottom Line Up Almost-Front

This is the shortest version of the argument.

Managed AI adoption on classified and CUI systems appears to increase national security more than the feared AI-incremental security risks decrease it. A categorical delayed-adoption default does not remove most of the threat landscape. It mainly gives up the productivity term while baseline insider risk, spillage risk, foreign cyber activity, supply-chain exposure, flawed analysis, aggregation risk, and user workarounds remain.

The load-bearing numbers are straightforward. By 2035, Low AI adoption adds about 1.3% of annual national-security output, Medium adds about 5.1%, and High adds about 10.5%. The equivalent cleared-labor figures are shown below because they are more relatable to the national-security enterprise than percentages.

Those “effective worker” numbers are not a force-reduction argument. They are a mission-capacity argument. They mean fewer analytic backlogs, faster software delivery, faster acquisition review, faster issue triage, faster security processing, better document review, more red-team coverage, and more time for expert judgment. Converting the gain into mission output also buys time to retrain people toward emerging mission and a more rapid response to societal needs.

The central managed-enclave policy is about +3.5% of annual national-security output after implementation cost drag and counterfactual AI risk. Unmanaged adoption is about -2.1%, which is why the recommendation is not “let everyone use whatever AI they want.” The recommendation is the opposite: give the workforce a sanctioned path that is strong enough to pull real work into the managed control surface.

Now the vocabulary those results assume.

Low adoption means standalone chatbot use with the latest model: ask, copy, paste, and clean up. It is useful, but outside the main workflow. Medium adoption means an assistant (Codex, Claude, Copilot, and peers) embedded in the workflow plus retrieval-augmented generation (RAG), grounded on approved government corpora. That is the recommended near-term target. High adoption means human-supervised agentic workflow: multi-step, tool-using orchestration over mission tools, gated by least privilege, logging, sandboxing, and human approval for high-impact actions.

Managed adoption means current-generation AI inside an accredited enclave with approved corpora, no external egress, access control, provenance, logging, monitoring, red teaming, and model supply-chain review. “Current-generation” model is critical, for a reason the model section spells out: the national-security enterprise should not be operating on knowledge months older than the rest of the world. Unmanaged adoption means AI use outside the sanctioned control surface: uncontrolled contractor-side use, transcribing material to and from the low side, personal accounts and devices, or other routes with weak telemetry.

AI-incremental risk is the counterfactual difference between risk with managed AI and risk under the status quo. It does not mean the risk is linear or small. It means the model charges AI only for the part AI causes or worsens. When a chart says “full expected event toll,” it is using a harsh accounting test that charges every adoption scenario the same gross expected annual toll from the ten modeled security categories. It is not the formal counterfactual risk model, and it is not a claim that every imaginable incident happens every year.

The first chart is the executive sanity check.

Read this as a burden-of-proof chart. It charges every adoption scenario the same expected annual toll, about 3.4% of output, from the ten security-fear categories, as if AI owned that entire ledger. That is deliberately harsh to the AI side. Even then, full adoption lands around +7.1% after the toll. Medium adoption lands around +1.8%. No adoption does not become safe. It sits at zero before the same security-event toll and around -3.4% after it. One accounting note so the two charts reconcile: this toll is larger than the roughly 1.6% per year charged in the opening chart because the two come from different risk priors. This one comes from a standalone elicitation of the ten fears; the opening chart uses the counterfactual world model. Both charge the full toll against AI, and the scenario ordering is the same under either.

That chart is not the formal risk model. It is the “stop making the accounting error” model. The accounting error is treating all security risk as AI-created risk. Insiders exist without AI. Spillage exists without AI. Foreign cyber operations exist without AI. Supply-chain compromise exists without AI. Flawed analytic assessments exist without AI. Aggregation risk exists without AI. A categorical non-adoption policy can avert only the AI-caused or AI-worsened part of those risks. It cannot make the status quo clean.

The second chart runs the decision over time. It uses 40,000 simulated futures from 2026 to 2035.

The conservative move in this chart is easy to miss: AI capability is held fixed, and only adoption changes. That is almost certainly false in the real world. AI capability is improving quickly. The chart freezes it anyway because the point is not to argue with assumptions. The point is to measure and argue with data. More adoption means more national-security output. Faster adoption means the advantage compounds sooner. “Delay while we pilot” has a measured cost.

One convention before the rest of the argument: every claim here is time-boxed to the public model as of June 2026. This is not a classification decision. It is an unclassified, public-source policy model. Where the model projects, it labels the projection. Where the evidence is transferred from commercial work into classified work, it says so.

How To Grade An AI Security Policy

The usual question is too vague: “Is AI risky?” Of course AI is risky.

The better risk question is: which policy produces less total national-security damage?

A risk policy works if a causal chain runs from the policy to measured national-security gain or measured risk reduction. A risk policy fails if the risk appears anyway through another path: baseline insider activity, existing supply-chain compromise, public-SaaS workarounds, low-side cut-and-paste, contractor-side tools, degraded official tools, or adversary adoption. A policy fails if it gives up more national-security output than the risk it can actually avert.

Graded that way, the scorecard splits in a way the current debate often misses.

Security professionals often agree on many of the controls. The bottleneck is usually not the concept of controls; it is the evidence burden and the review process around them. Bodies of evidence take weeks or months to assemble, move, interpret, and rework, and there is often no clear threshold at which the answer becomes “yes.” That failure mode turns accreditation into a queue instead of a risk decision. Queuing up more advanced AI models and tools causes delay, and delay also has a cost.

The answer is not to bypass Security. The answer is to make accreditation evidence-producing: predefined test suites, accepted control metrics, continuous telemetry, red-team reports, signed model provenance, DLP results, marking-propagation accuracy, and human-review gates. If the evidence passes, the workflow should move. If it fails, the workflow should pause at that boundary and the control owner should know exactly what to fix.

That is the doctrine: no categorical non-adoption default, no unmanaged adoption, and managed adoption with Security as the accountable control owner.

The National Security Model

There is no perfect public equation for national security. If someone tells you they can precisely measure national security in one number, be skeptical. This decision does not require a perfect number. It requires a defensible way to compare a labor-productivity gain against an AI-incremental risk term.

The project uses a citable defense production-function model from Jose L. Torres’ 2020 PLOS ONE paper, “The production of national defense and the macroeconomy.” It models defense output as a function of capital and labor:

D = B * [mu*K^rho + (1-mu)*L^rho]^(1/rho)
rho = (sigma - 1) / sigma

The important part is not the Greek letters. The important part is that the model decomposes national-security output into labor and capital. AI is a labor-productivity shock, so AI enters as effective labor:

L_AI = L * (1 + g * alpha * e)

Here, g is the productivity gain on tasks AI can help with, alpha is the share of cleared knowledge labor exposed to AI-augmentable tasks, and e is realized adoption, not mere access.

The e term matters. Access is not adoption. A chatbot in another browser is not the same thing as an assistant embedded in the analyst tool, the IDE, the document system, the acquisition workflow, or the security-processing queue. And a model that cannot retrieve approved mission data is not the same thing as RAG inside an accredited enclave. This deserves a footstomp, because it is also why the math assumes any adoption path keeps its AI models current. An AI model has a knowledge cutoff; it knows nothing after that date. People adapt to a changing world faster than a static model does. If the nation is at war tomorrow, the people will know, but a stale model needs constant reminders. If peace breaks out the day after, the people will know, but the stale model still infers war. The world changes, policy changes, regulations change, laws change. Current models plus RAG over approved corpora are what give an AI current, pertinent information.

The Low, Medium, and High adoption cases are work modalities, not arbitrary labels.

Commercial analogies help:

Low adoption looks like a ChatGPT-style side chat. In government, that maps to a standalone assistant for unclassified or tightly bounded CUI tasks, with copy/paste risk, weak provenance, limited mission data, and limited workflow integration.

Medium adoption looks like GitHub Copilot, Microsoft 365-style copilot, Claude Code/Cowork, OpenAI Codex, etc. In government, that maps to an inline assistant in coding, analytic, document, acquisition, legal, contracting, or security-review tools, grounded on approved corpora. The government-specific burden is need-to-know, classification markings, source lineage, no external egress, audit logs, DLP, and accreditation. Much of this already exists on the high side; it just needs integration with the AI tools.

High adoption looks like tool-using agents in software, ticketing, data, or workflow systems. In government, that maps to human-supervised orchestration over mission tools, with least privilege, sandboxed execution, command logging, human approval for high-impact actions, and narrow authority boundaries.

That matters for the recommendation. This is not an argument that analysts should become spectators while autonomous agents run the Intelligence Community. The recommendation is more disciplined: field assistants plus RAG inside accredited enclaves, then gate agentic workflows until the enclave, workflow instrumentation, and controls are mature.

The Productivity Term

The cleared knowledge workforce modeled here is about 1.85 million people: Intelligence Community (IC) and Department of War (DoW) government knowledge workers plus cleared defense and intelligence contractors. The model estimates the AI-exposed share at about 57%, or roughly 1.06 million AI-augmentable cleared knowledge workers.

The model then applies conservative task-level productivity assumptions. Low adoption uses a 10% uplift, below the larger writing, customer-support, and coding RCT effects because standalone chat is intermittent and outside the workflow. Medium adoption uses a 25% uplift, aligned with peer-reviewed 2026 evidence around integrated copilot/RAG-style work, including software and consulting tasks, but then discounted by exposure and realized adoption. High adoption uses a 40% uplift, below favorable agentic-workflow results and treated as phase two, not immediate whole-job autonomy.

Those are not the largest numbers in the literature. Noy and Zhang report large gains for professional writing. Brynjolfsson, Li, and Raymond find material gains in customer support, especially for less experienced workers. Peng et al. report large coding-task speedups in an early Copilot experiment. Cui, Demirer, Peng, and coauthors report about a 26% completed-task gain across three software-development RCTs in peer-reviewed 2026 work. Dell’Acqua et al. report strong in-frontier gains for consultants and worse performance off the frontier. Ju and Aral report favorable agentic-teamwork gains, while the Remote Labor Index suggests fully autonomous whole-job completion remains small. METR’s open-source developer result is a warning that AI can slow experts on complex, mature codebases, and Humlum and Vestergaard’s near-null population-level result is a reason the model discounts capability numbers by realized adoption instead of taking them at face value.

That is why the model does not assume every cleared worker instantly becomes 40% more productive. It discounts by exposed labor share and realized adoption. It also holds AI capability fixed through 2035 in the central time-series chart, even though the capability frontier has already measurably improved since this project started. This project is asking how much effective labor increases once AI is used on the work it can actually help with.

Under Medium adoption, the 2035 result is equivalent to adding roughly 185,000 effective cleared workers. Even a rough version is strategically large. The national-security enterprise argues constantly about recruiting, clearances, retention, billets, contractor ceilings, analytic coverage, cyber backlogs, software queues, acquisition speed, and security-processing timelines. Leaving a six-figure equivalent mission-capacity gain on the table is not a neutral act. It is an opportunity cost.

The Fears Are Real

A serious AI adoption argument has to steelman the security fears. This project uses ten categories and maps them onto the government’s own EO 13526 damage vocabulary.

The “some damage” concerns are CUI mishandling, privacy or civil-liberties harm, and loss of provenance, auditability, or classification lineage. The “serious damage” concerns are data spillage, cross-domain contamination, hallucination corrupting analysis, poisoning, prompt injection, automation bias, and tradecraft erosion. The “exceptionally grave damage” concerns are aggregation or mosaic inference, AI-amplified insider threat, model or data exfiltration, and supply-chain or foreign-model compromise.

This is the right list to worry about. It is also the wrong list to use as a categorical stop sign unless the counterfactual question is answered:

1. How much of each risk is caused by managed AI adoption?

2. How much already exists?

3. How much gets worse under unmanaged adoption?

4. How much gets better because AI improves monitoring, classification support, provenance, anomaly detection, review consistency, and defensive triage?

That is where the “delay while we pilot AI” case weakens.

The counterfactual model estimates three worlds (status-quo delay, managed accredited adoption, and unmanaged adoption) using 60,000 Monte Carlo draws.

One disclosure for skeptics, because it changes how these numbers should be read: the event probabilities and losses behind this chart are a synthetic expert panel: four independent AI models, each answering as six security personas, pooled into twenty-four elicitations. The four models disagree on absolute magnitudes but agree on the ordering: managed below status quo, unmanaged far above both. The pooled panel is an explicitly labeled prior, not a substitute for cleared human expert judgment, and the falsifier section below says exactly how human elicitation should update the model.

The managed result is not “AI is risk-free.” At the 95th percentile, managed AI-incremental loss is about +0.85%. At CVaR99 (the average loss in the worst 1% of modeled outcomes), it is about +1.91%. That is real. But it is smaller than the Medium productivity benefit to national security.

If the goal is to improve national security, the burden of proof on delay arguments is that the specific risks impact national security by more than the cost-drag-adjusted productivity benefit.

AI Risk Can Be Nonlinear

One objection to the phrase “AI-incremental risk” is that AI may not add a small linear increment. It could create nonlinear or compounding risk. That objection is legitimate.

AI can make aggregation risk worse because retrieval, summarization, and synthesis make it easier to combine individually innocuous facts. AI can amplify insider threat because one person can query, summarize, transform, and stage more information faster. AI can accelerate data exfiltration if egress controls fail. AI can create correlated analytic error if many offices rely on the same flawed model output. AI can scale prompt injection, poisoning, and automation bias in ways that do not look like ordinary software risk.

The word “incremental” is not meant to deny any of that. It means counterfactual. The policy question is not whether AI can worsen a risk category. It can. The policy question is how much managed AI changes that risk compared with delay and compared with unmanaged adoption.

That distinction is exactly why managed adoption matters. No egress, approved corpora, retrieval provenance, least privilege, user and entity behavior analytics, DLP, red-team testing, model provenance, human verification, and classification-aware review are not cosmetic controls. They are the tools that keep nonlinear risk from becoming uncontrolled risk. If those controls fail measurable tests, the workflow should pause. If they pass, indefinite delay needs its own quantified justification.

“Delay While We Pilot” becomes Unmanaged Adoption

The strongest form of the non-adoption argument assumes no official AI other than pilot programs means no AI risk. That assumption does not survive contact with human incentives.

The official system can say no, but the surrounding ecosystem still contains public AI systems, home computers, contractor-side tooling, low-side summaries, personal productivity accounts, and colleagues who can access stronger models. Even ordinary google search now returns AI-generated answers by default, so much of the workforce already uses AI through unmanaged paths every day without ever deciding to adopt it. Then the Information Assurance arm of Security restricts access to google and then the workforce has a choice between being trapped in a previous decade, or doing work on their unmanaged personal devices.

This is not a moral claim about users. It is an institutional claim about incentives. If the sanctioned path is too slow, too weak, or unavailable where the work happens, some work will route around the system. Some of that will be careful. Some will not. The organization will see less telemetry, not more.

That is the worst of both worlds: much of the risk and far less of the productivity gain. The model sees that branch clearly. Unmanaged adoption is negative. A categorical delay posture does not guarantee no AI. It can produce the worst AI.

Delay also undercounts the defensive side. AI can make some baseline security problems worse, but it can also make some baseline security problems easier to detect, measure, and control. AI-assisted user and entity behavior analytics can detect anomalous bulk access faster. Classification-aware tools can catch marking mistakes and spillage attempts. Provenance systems can make derivative outputs easier to audit. Retrieval-grounded assistants can force source display and citation checks. Models can serve as tireless reviewers for missing caveats, inconsistent markings, and analytic hygiene.

None of that happens automatically. It happens only if Security owns the control package instead of keeping the capability outside the system.

Security’s Own Scale Cuts Both Ways

Security has a powerful vocabulary for harm. Confidential means unauthorized disclosure could cause damage. Secret means serious damage. Top Secret means exceptionally grave damage. That vocabulary is usually pointed at disclosure risk, but this project uses it as a magnitude yardstick for both sides of the scale.

This is not a legal reclassification claim. EO 13526 does not assign percentages to its tiers. This is a conditional burden-of-proof argument: if feared AI incidents are being scored on a national-security damage scale, then the national-security output lost to AI restraint should be scored on the same scale for an apples-to-apples comparison.

By 2035, every less-than-full-rapid adoption path imposes an annual opportunity cost in the model’s exceptionally-grave magnitude band. Medium adoption versus full rapid adoption forgoes about 5.4% per year. A conservative rollout forgoes about 6.6% per year. Low adoption forgoes about 9.2% per year. No adoption forgoes about 10.6% per year.

Again, this is not a literal classification determination. It is a symmetry test. If a low-probability AI incident deserves grave language because it could damage national security, then a recurring, near-certain productivity loss of 5% to 10% of annual national-security output also deserves grave language. Security policy should not use the damage scale only on the risk side of the ledger.

The China Argument Cuts Both Ways

China is a real adversary, and not just a stalking horse in national security debates. China might compromise a model, exploit AI spillage, poison data, use foreign-origin models as collection surfaces, or accelerate cyber and intelligence operations with AI. Those are valid threat paths.

But China also benefits if the United States makes its cleared workforce slower. China benefits if American analysts, engineers, operators, acquisition officers, program managers, and security processors are kept off frontier-grade tools while Chinese operators, companies, and state organs adopt aggressively under their own risk calculus.

“Delay AI while we pilot” is not a shield against China. It is a tax collected on ourselves. “Behind” is not the same as “blocked.” China does not need perfect AI to use AI. It needs enough capability, enough urgency, and enough institutional permission to keep moving. The United States national-security enterprise has capability. The control question is whether it can give itself permission in a way that is secure enough and fast enough.

Recommended Doctrine

The doctrine should be:

1. no categorical non-adoption default;

2. no unmanaged adoption;

3. managed accredited adoption with Security as the accountable control owner.

That changes Security’s mission from “prevent AI use” to “field the controls that make AI usable at scale.”

The control missions are concrete.

The implementation order should be boring and fast:

1. Reject categorical non-adoption as the default.

2. Prohibit unmanaged adoption by giving the workforce a sanctioned path.

3. Field a government-hosted CUI enclave first. The Department of War and much of the Intelligence Community are effectively here.

4. Field classified RAG enclaves with approved corpora and no external egress. That means the AI has access to the classified intel the user has a need-to-know and authorization for.

5. Treat agentic classified workflows as phase two, gated by least privilege, command logging, sandboxed execution, and human approval.

6. Make accreditation evidence-producing. Authority to Operate should not mean queue time. It should mean measured controls.

7. Require a quantified burden of proof for further delay.

A security office should be able to delay a workflow only by showing one of four things:

1. the workflow cannot meet the required controls, where the controls themselves have a demonstrated, measured national-security benefit;

2. managed AI-incremental risk exceeds the cost-drag-adjusted productivity benefit;

3. specific threat intelligence changes the supply-chain, insider, or exfiltration risk enough to cross the break-even line; or

4. the proposed use is actually unmanaged or agentic beyond approved authority.

“We are concerned about China” is not a risk determination. It is the beginning of one.

What Would Change The Model

This should not be an argument that lives forever in abstraction. The model should change when measured evidence changes.

None of this has to be argued abstractly. It can be measured. That is the point.

The Mental Model

AI adoption is not automatically safe. Unmanaged AI has real risks. Managed AI is the highest-security policy in the model.

Default delay does not remove most baseline threats. Default delay does remove the productivity gain, and that gain is strategically large. The mission is not to block AI. The mission is to field the controls that let cleared personnel use AI faster than adversaries can exploit either our data or our hesitation.

That is the job Security should want.

Citations

Main public sources behind the model:

• Jose L. Torres, “The production of national defense and the macroeconomy,” PLOS ONE, 2020: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0240299

• Correlates of War National Material Capabilities data: https://correlatesofwar.org/data-sets/national-material-capabilities/

• OpenAI GDPval: https://openai.com/index/gdpval/

• Noy and Zhang, “Experimental evidence on the productivity effects of generative artificial intelligence,” Science, 2023: https://doi.org/10.1126/science.adh2586

• Brynjolfsson, Li, and Raymond, “Generative AI at Work,” NBER, 2023: https://www.nber.org/papers/w31161

• Cui, Demirer, Jaffe, Musolff, Peng, and Salz, “The Effects of Generative AI on High-Skilled Work,” Management Science, 2026: https://doi.org/10.1287/mnsc.2025.00535

• Dell’Acqua et al., “Navigating the Jagged Technological Frontier,” Organization Science, 2026: https://doi.org/10.1287/orsc.2025.21838

• Peng et al., “The Impact of AI on Developer Productivity,” arXiv, 2023: https://arxiv.org/abs/2302.06590

• Ju and Aral, “Collaborating with AI Agents,” arXiv (revised February 2026): https://arxiv.org/abs/2503.18238

• METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity,” arXiv, 2025, with METR’s February 2026 experiment update at metr.org: https://arxiv.org/abs/2507.09089

• Humlum and Vestergaard, “Large Language Models, Small Labor Market Effects,” NBER Working Paper 33777 (revised 2026): https://www.nber.org/papers/w33777

• Remote Labor Index, arXiv, 2025: https://arxiv.org/abs/2510.26787

• EO 13526 and ISOO FAQ: https://www.archives.gov/isoo/faqs/e-o-13526-and-32-cfr-part-2001

• NIST AI Risk Management Framework 1.0: https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10

• NIST Generative AI Profile, NIST AI 600-1: https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence

• NSA, CISA, FBI, and partners, “Deploying AI Systems Securely,” 2024: https://media.defense.gov/2024/apr/15/2003439257/-1/-1/0/csi-deploying-ai-systems-securely.pdf

• ODNI 2026 Annual Threat Assessment: https://www.dni.gov/files/ODNI/documents/assessments/ATA-2026-Unclassified-Report.pdf

Methodology note: every model output in this piece regenerates deterministically from the project’s code and data (fixed random seeds; figures render directly from model outputs with no hand-entered numbers). The risk priors are a labeled synthetic elicitation pending cleared human expert elicitation and classified workflow pilots, for which prepared instruments exist.

I spent a weekend thought exercise thinking about the unthinkable, and what would escalated conflict in North America look like.

The 2000s had a signature weapon: the improvised explosive device. Cheap, ugly, lethal, hard to attribute, and strategically out of proportion to its cost. The 2010s were the decade of anti-access and area denial, of long-range missiles like the DF-21 and DF-26, and of navies relearning what distance meant. For the first third of this decade I expected the 2020s to be characterized by hypersonics.

That was wrong. The wars in Ukraine and around Iran have been defined by something cheaper and less elegant: one-way attack UAVs. A Shahed-class drone is, functionally, an IED with range. It takes the core tactical fact of the roadside bomb and adds hundreds or thousands of kilometers.

That changes the way old disputes look.

For my entire life, the geography of North America has been one of America’s defense assets. Oceans on two sides. Canada and Mexico on the other two. Friendly neighbors, integrated economies, long institutional habits, and an interior that distance keeps safe.

But distance only protects you when range is expensive. And both the Iran and Ukraine wars showed how a non-great power can match their will against a great power; without nuclear weapons.

So I built the map I needed to see but didn’t really want to see. Where are the unsettled US territorial, maritime, EEZ, water, and border-adjacent disputes? Which are real conflicts, which are symbolic flashpoints, which are old paper claims, and which parts of CONUS fall inside the geometry of cheap one-way UAVs if those regional disputes ever stopped being managed over a pitcher of McNallys ale?

Again, this is intended as a “what if” thought exercise. I would prefer our regional disputes keep being negotiated over a pitchers of beer.

Bottom line up front

I cataloged 39 American backyard disputes: 16 active, 3 dormant, 2 latent, 8 largely resolved, and 10 resolved; and another 8 perception-asymmetry cases, where another country counts or experiences a conflict with the United States that Washington does not track as a territorial, border, or sovereignty dispute.

The strangest case is Serranilla Bank, a western Caribbean reef where Colombia keeps a naval garrison on a cay tied to a US claim that Washington never formally renounced and no longer presses. That makes Serranilla a potential flash point. It is an example of the category: a dormant legal fact, a physical outpost, a resource squeeze, and a deteriorated diplomatic relationship all stacked in one place.

The Mexico cases need a different vocabulary. The Rio Grande water debt, the Colorado River shortage-sharing fight, and the Tijuana sewage crisis are real conflicts. They are not classic territorial disputes, but they are real interstate stressors with legal instruments, deadlines, money, public health, and retaliation tools attached. The new Mexico cartel-strike case is different again: a sovereignty conflict Mexico sees in threatened unilateral US force, while Washington frames it as counter-narcotics.

This is a thought exercise in the unthinkable: regional armed conflict in North America and the Caribbean. It is worth doing precisely because the premise has been unthinkable.

There are more potential flash points within UAV range of the United States than I expected, certainly more than I wanted, and enough that maybe it isn’t unthinkable.

Conflict is up, but mostly not by shooting

The first result is not a border result. It is a temperature result.

2025 was a structural break in new US disputes. Count by counterparty and you get roughly 43 new disputes. Count by episode, where one tariff wave is one event rather than twenty-six country rows, and you get roughly 4.

Severity weighting does not make it disappear. In the severity-weighted series, 2025 is roughly 4.5 times the next-highest year. But the stack matters. This is not primarily a shooting-war spike. It is trade, sanctions, diplomatic coercion, and pressure tools. The militarized slice grew, but it did not define the tower.

That is the right frame for the rest of the article.

The geopolitical temperature is up. Most of the heat is economic and diplomatic. The risk is not that a tariff order becomes a drone strike. The risk is that an environment where economic coercion becomes normal is also an environment where old boundary disputes are less insulated from national politics.

Some disputes have nowhere physical to go. A name change does not move the Gulf. Others have a cay, a reef, a fishery, a river, a base, a strait, or a continental shelf. Those are the ones where political heat can become the flashpoint. So let’s map those.

Chasing waterfalls

First, we built a registry of the disputes. Not just islands. Maritime boundaries, EEZ overlaps, river-boundary residues, water treaty conflicts, sewage infrastructure, and symbolic sovereignty fights. Then we separated the old “is this a dispute?” question from a more useful one: what is the current US posture on paper? Then we pushed on the blind spot that mattered most: what if the US does not count the conflict at all?

Those two axes do not always line up.

A case can be active in daily politics even though the underlying line is settled by treaty. The Colorado River is a real conflict even though the legal entitlement is not vague. A case can be dormant in public and still unrenounced on paper. Serranilla and Bajo Nuevo are like that. A case can be loud and symbolic but not a real territorial conflict at all. Gulf of America is like that. And a case can be experienced abroad as a sovereignty conflict while Washington files it under national security, sanctions, counter-narcotics, trade, or old litigation.

Then we pulled filings from the other side too to help identify potential blind spots. Where the dispute had a physical surface, we looked at imagery. A cay, a lighthouse, a garrison, a sewage outfall, a reservoir bathtub ring, a runway, a reef. Where a 2025 or 2026 fact was load-bearing, we separated official acts from rhetoric.

The loudest sentence is often not the most important fact. Sometimes the most important fact is the thing a government did not say.

Then we drew the UAV envelopes from these disputed areas.

What cheap UAV range adds

A Shahed-131-class envelope is not a CONUS-wide blanket if you only ring islands and maritime edges. From the Gulf of Maine and Machias Seal Island, it touches New England. From the Juan de Fuca and Dixon Entrance region, it touches the Pacific Northwest. From Caribbean dispute zones, the cheap envelope reaches parts of Florida and the Gulf edge, while the longer Shahed-136-class envelope reaches much more of the Gulf Coast and Atlantic seaboard.

The water disputes changed my concern and the geometry because they sit on the border itself. Thirsty people can blow up dams. Once Rio Grande, Colorado, and Tijuana are included as stress zones, the cheap ring covers much more of the populated lower 48, and the longer Shahed-136-class ring puts every flagged CONUS economic target inside the footprint.

That is important. It keeps the piece from becoming fantasy, but it also keeps the border in view. Cheap drones do not erase distance everywhere. They erode it at the margins, and the margins are exactly where maritime, territorial, water, and sovereignty disputes live.

The map includes the Rio Grande, Colorado River, and Tijuana sewage conflicts because they are real. They involve treaty obligations, deadlines, crop losses, public-health closures, project delivery, and pressure tools. They can raise temperature. They can sour a relationship. They could produce retaliation.

The perception-asymmetry rings work the same way. Mexico, Cuba, Colombia, Panama, and Nicaragua are ringed because their 2,000 km geometry reaches CONUS. Venezuela’s mainland, Greenland, and Brazil are not. The rings show proximity to the US homeland. They do not assert capability, staging, sponsorship, or intent.

So the map asks two different questions:

Where is range cheap enough to matter?

Where is political temperature high enough that range might someday matter?

The inventory: most of this works

Before getting to the strange cases, it is worth saying the reassuring part clearly.

Most of the diplomacy machinery works.

Boundary commissions work. IBWC minutes work. Treaty settlements work. Arbitration worked in the San Juan Islands. The US gave up dozens of old Pacific guano claims by treaty. The US and Mexico settled the Chamizal. The Western Gulf gap is delimited. Many Caribbean lines are applied even where Senate ratification lags.

The live residue is smaller than the complete catalogue.

But the residue has a pattern. It clusters at the edges: Arctic shelves, lobster waters, open straits, tiny islands, reefs, banks, old guano claims, and water systems under climate stress. Some are active and managed. Some are dormant and unrenounced. Some are symbolic. Some are real conflicts that are not territorial in the classic sense.

The most dangerous analytical mistake would be to flatten them. Machias is not the Colorado River. Serranilla is not Gulf of America. Tijuana sewage is not a maritime boundary. Bering is not Wake. The categories matter because the escalation pathways differ.

The conflicts only one side is counting

There is a class of conflicts where the other side counts a dispute with the United States and the United States does not. That does not mean the other side is correct on the law. It means the political fact exists abroad while the US bureaucracy files the same episode somewhere else.

Panama is the cleanest territorial example. A US threat to “take back the Canal” did not become a US territorial filing. Washington treats the issue as China leverage and strategic messaging. Panama treated it as a sovereignty threat serious enough to raise at the United Nations under the Charter’s prohibition on force. The Canal is not an old paper claim like Serranilla. It is the reverse: the US is the actor, Panama is the state logging the conflict, and the US does not call it a dispute.

Greenland is the other territorial example. Denmark and Greenland hear acquisition talk as a sovereignty threat. Washington calls it strategic interest. Again, there is no formal US claim to map. The point is not that a Danish island suddenly becomes a conventional border dispute. The point is that one side hears coercion while the other side refuses the category.

The counter-narcotics cases are sharper because they involve force. Venezuela frames the 2025-2026 Caribbean campaign as aggression and regime change. Colombia frames the strike campaign as a sovereignty violation, separate from Serranilla. Mexico responded to cartel-strike threats by hardening its anti-intervention posture. The US language is different in all three: counter-narcotics, counter-terrorism, sanctions, interdiction.

Then there are the economic and legal cases. Cuba treats the embargo as aggression and bundles Guantanamo into the same narrative. Nicaragua still treats the unpaid 1986 ICJ judgment as live. Brazil frames sanctions and tariffs tied to its domestic judiciary as sovereignty interference. Those are not border disputes. They are not launch geography. But they are part of the same pattern: the US acts, the other government counts a conflict, and Washington does not count a dispute.

That matters because Serranilla is not just weird on its own. It is one direction of a larger asymmetry. At Serranilla, the US holds a dormant claim and does not press it while Colombia occupies the rock. In the newer cases, the US is the actor and the other country is the one saying, in effect: you are in a conflict with us whether you admit it or not.

The common thread is silence. Not absence of action. Silence about category.

Serranilla Bank: the weirdest case

Serranilla Bank is the case that makes the whole project feel less like a catalogue and more like a warning label.

On Beacon Cay, Colombia maintains a compact naval outpost: roughly fifteen marines, a lighthouse, barracks, solar power, cisterns, tracks, and a cleared pad whose function should be described cautiously. It may be aviation-related. It may be a utility or turning pad. The imagery does not prove a helipad (there is no windsock or big H).

The cay is small. The legal problem is not.

The United States claimed Serranilla and Bajo Nuevo in the nineteenth century under the Guano Islands Act. In 1972, the US and Colombia signed the Vasquez-Saccio Treaty to clean up a set of Caribbean cay claims. The US renounced Quita Sueno, Roncador, and Serrana by name. It did not name Serranilla or Bajo Nuevo. Article 7 preserved matters the treaty did not address.

That means the US claim was never formally extinguished.

It also means the easy conspiracy version is wrong. A dedicated archival dig found no evidence that Washington deliberately kept Serranilla and Bajo Nuevo as strategic leverage. The record leans against it. Across 1890, 1928, and 1972, neither side appears to have put the two banks at the center of the bilateral negotiation. The safer line is the truer one: shelved by inattention, not preserved as a weapon.

The modern US posture is even stranger. The US has not formally renounced the claim. But it has stopped acting like an owner. Its December 2023 extended-continental-shelf package named no Caribbean region. NOAA charts no US maritime line at the banks. Current maritime databases attribute the waters to Colombia, while still flagging the US claim at Navassa. The silence is not a footnote. The silence is the fact.

Colombians have called this litigio silencioso, silent litigation.

One side occupies the cay. The other retains a paper title it does not use.

The resource squeeze is not hypothetical. It is fish.

After the 2012 International Court of Justice ruling shifted rich fishing grounds toward Nicaragua, the Raizal fishing community of San Andres lost access to waters that were not abstract legal space to them. They were livelihood, food security, and identity. The lobster collapse is not a metaphor. It is measurable.

The oil story is different. There is unproven hydrocarbon potential in the broader region. There is no public reserve estimate for the disputed zone. The one frontier well came up dry.

The diplomacy is a potential accelerant. The US-Colombia relationship ruptured in 2025: decertification on 15 September 2025, an OFAC designation of President Gustavo Petro personally on 24 October 2025, and a lethal Caribbean and eastern Pacific boat-strike campaign that became Operation Southern Spear. Colombia’s strike grievance is its own perception-asymmetry case, distinct from Serranilla. A February 2026 Oval Office meeting improved tone but it did not reverse the instruments.

Tone is not policy.

There is no evidence Washington intends to revive the Serranilla claim. I assume no one in the US Government remembers there is a claim but it’s easy to find for anyone looking. The risk is that in a resource or diplomatic rupture, an unrenounced paper claim gives an administration a legal vocabulary to light a fire.

That is how dormant claims become dangerous. Not because they predict war. Because they remain available.

Two guardrails keep the story credible.

First, the 2012 ICJ case did not extinguish the US claim. The US was not a party. It also did not “enclave” Serranilla and Bajo Nuevo the way careless summaries sometimes imply.

Second, Colombia is not “arming an island against the United States.” Colombia is garrisoning territory it administers as its own, backed by a World Court judgment as against Nicaragua. The drama is a latent clash of titles, not an active US-Colombia military standoff.

The contrast with Bajo Nuevo is useful because it reduces the story to one sentence: Colombia mans one and lights the other.

Canada: allied management and Arctic reopening

The Canada disputes are the best evidence that old conflicts do not automatically become military problems.

The Beaufort Sea wedge is active. The United States published extended continental shelf limits in December 2023. The US and Canada created a Beaufort boundary task force in September 2024. US Arctic offshore-leasing planning advanced through 2025 and into 2026. There is real resource value there, and climate change is making the shelf more legible as a commercial and strategic space.

But there is still cooperation.

The Northwest Passage is even clearer. The US rejects Canada’s internal-waters position and treats the Passage as an international strait on paper. But it has never run a freedom-of-navigation operation against Canada there. Canada is absent from recent US operational-assertions reports. Washington challenges Russia and China with ships. It challenges Canada with law, notes, restraint, and probably a few drinking contests.

That is what the system working looks like.

Machias Seal Island is the human-scale exception: a lighthouse, puffins, lobster, and two flags. The temptation is to overheat it because a president restated the US claim in June 2025. But the operational incident people remember was in 2018 where US Border Patrol vessels stopped Canadian fishing vessels, and there was a collision between a US Coast Guard vessel and a Canadian lobster boat. The 2025 statement was messaging. The distinction matters.

The sentiment shift in Canada is not just about any single boundary. It is about sovereignty talk. Northern and Indigenous voices have hardened around “nothing about us without us,” while broader Canadian opinion sharpened after a year of annexation and “51st state” rhetoric. That does not make a boundary fight likely. It makes boundary management something no longer doable at the bar.

Mexico: three real water conflicts, one sovereignty case, and one symbolic flashpoint

The Mexico section needs careful sorting.

There are three real water or public-health conflicts.

Rio Grande water debt is a real conflict. Mexico owes water under the 1944 treaty. Cycle 36 closed in deficit in late 2025, with Mexico having delivered roughly half the required amount. The US denied a Tijuana water request on 20 March 2025. Trump threatened a 5 percent tariff on Mexican goods on 9 December 2025, but the tariff was not actually levied. A deal followed on 12 December 2025, and a technical plan was announced in February 2026. The pressure worked tactically. The structural fight remains.

Colorado River shortage-sharing is a real conflict. Here the obligation runs the other way. The US owes Mexico a treaty floor, and the shortage-sharing architecture around Minute 323 (meeting minutes from the International Boundary and Water Commission of the United States and Mexico 323rd meeting) expires on 31 December 2026. A successor is not yet signed. 2026 is a Tier 1 shortage year. This is not going to end well.

Tijuana sewage is a real conflict. It is not a sovereignty dispute, but it is a cross-border public-health crisis. Imperial Beach closures, polluted surf zones, sickened military trainees, and a binational infrastructure chain running through Minute 328, a July 2025 EPA-SEMARNAT memorandum, and Minute 333 in December 2025. This is not symbolism. It is waste in the water.

Then there is a sovereignty case.

The cartel-strike conflict is real, but it is not a water dispute. The US designated cartels as foreign terrorist organizations in February 2025 and floated unilateral strike options. Mexico heard that as a threat to act on Mexican soil, water, sea, or airspace without consent. President Claudia Sheinbaum’s government moved to harden Article 40 of the Mexican constitution with anti-intervention language. Washington frames the same pressure as counter-narcotics. Mexico frames it as sovereignty. Cartels are known to respond with force and they are probably better resourced than some western hemisphere countries.

Caribbean: old claims and modern stress

The Caribbean is where the legal ghosts live.

Navassa is the precedent. The US administers it as a wildlife refuge. Haiti claims it. The US reactivated a dormant Guano Act claim there in 1916 and never left. That is why Navassa matters to Serranilla’s Columbian garrison. It proves the mechanism exists.

But Navassa also shows why analogies can mislead. At Navassa, the US holds possession. At Serranilla, Colombia does. At Serranilla, Colombia also has the 2012 ICJ judgment against Nicaragua behind it. A US move there would be weaker and more provocative than Navassa was.

Guantanamo is a live lease standoff. It belongs in the conflicts catalogue and on the appendix map. The base is politically radioactive, the lease history is its own universe, and there already are far better writeups of GTMO than I could do.

The newer Caribbean stress is not an island claim. It is the strike campaign. Venezuela frames US operations as aggression. Colombia frames at least one strike as murder and a violation of its waters. Cuba frames the embargo as ongoing aggression and connects that story to Guantanamo. Washington does not track these as disputes. It tracks them as counter-narcotics, sanctions, terrorism, and policy.

The Bahamas, Dominican Republic, and Haiti maritime cases are quieter. They are EEZ geometry, not mass-politics. But they matter to the map because they sit close to Florida, Puerto Rico, and the Windward Passage. The quiet lines are often the ones that only become visible when something valuable is discovered or when enforcement changes.

The Caribbean lesson is simple: the strangest claims are not always the loudest ones.

Pacific and Bering: farther out, still real

The Pacific cases are not central to CONUS cheap-UAV range, but they matter strategically.

Wake Island is the most open US Pacific sovereignty dispute. The US administers it and has invested heavily in military infrastructure. The Marshall Islands claim it as Anen Kio. That claim is not the fake micronation story. It is a serious cultural and historical claim on the record.

The Bering Sea line is another one to watch. The US ratified the 1990 Baker-Shevardnadze maritime boundary in 1991. Russia never ratified it. Both sides have applied it provisionally, but Russia formally objected to the US Arctic continental shelf filing on 18 March 2024. A provisional line between nuclear powers in a warming Arctic is not an academic footnote.

The Mariana ECS overlap with Japan is the opposite case: a real seabed overlap handled by allies as a technical issue and probably settleable with a karaoke competition. Swains/Olohega is state-to-state settled but still alive as a Tokelauan cultural grievance. The Pacific is mostly the good-news control group: treaties work when both sides choose closure.

What the map does not say

The entire piece depends on restraint. A map is not a prediction. Range is not capability. Capability is not intent.

Rhetoric is not policy unless it becomes an official act: a filing, sanction, tariff, treaty minute, operational move, published boundary, lease program, or administrative change.

For most of American history, geography did part of homeland defense for free. Two oceans and two friendly borders put the interior beyond the reach of most regional conflict. Cheap UAVs do not erase that advantage. They soften it at the margins.

The margins are where these disputes live.

Appendix: The Conflict Zones Behind The Map

This is the field guide to the conflict zones: what they are, why they matter, what the imagery shows, what sentiment has changed, and what would make the risk move.

How to read this appendix

The main article uses a simple distinction.

Conflict zones are places where a territorial, maritime, EEZ, island, strait, or base dispute has geography. They can appear on a map. They can anchor a range ring.

Conflict flashpoints are stressors that can raise the political temperature without being launch geography. The three US-Mexico water and public-health conflicts are real conflicts. Gulf of America is not. It is a symbolic naming flashpoint.

Perception-asymmetry conflicts are cases where another government counts or experiences a conflict with the United States that Washington does not track as a territorial, border, or sovereignty dispute. Some are territorial in another country’s eyes. Some are military, legal, or economic. They are grouped separately because the escalation pathway is different.

That is why this appendix groups the cases by escalation pathway rather than by database row.

1. The northern maritime edge: Canada and the Arctic

Beaufort Sea

The Beaufort Sea is a friendly dispute with serious material stakes.

Canada’s position follows the 141st meridian north from the land boundary. The US position uses an equidistance line that bends eastward. The wedge between them is roughly Lake Ontario-sized and sits over an oil and gas province. The broader Arctic shelf context got much sharper after the US published its extended continental shelf limits in December 2023.

This is not a hostile file. The US and Canada created a joint task force in September 2024 to negotiate the boundary. The important fact is that the file is active again. Climate, seabed value, and continental shelf law turned an old managed line into a current negotiation.

Sentiment is not anti-American in the ordinary sense. Northern anger often aims at Ottawa. The recurring line is that Arctic decisions cannot be made over the heads of Inuvialuit, Inuit, territorial governments, and northern communities. The politics harden when sovereignty becomes a southern slogan and northern people are treated as scenery.

Risk moves if the task force stalls while leasing or shelf assertions advance. The current file is cooperative. The danger is a gap between negotiation tempo and resource tempo.

Central Arctic and Beaufort extended shelf

The Central Arctic extended-shelf overlap is the quieter, deeper version of the Beaufort problem.

In December 2023, the United States published outer limits for its extended continental shelf. Canada already had an Arctic submission in the Law of the Sea process. Where those asserted seabed entitlements overlap beyond 200 nautical miles, the issue is not a coastline visible from shore. It is a deep seabed claim, tied to oil, gas, minerals, mapping, and the long future of Arctic access.

This is not a popular flashpoint yet. There is little local public voice around the exact deep-shelf geometry. The better local frame is Inuit Nunangat: Arctic sovereignty cannot be treated as an empty federal chessboard. The people who live with the consequences do not want to be treated as scenery.

Risk moves if the Beaufort task force stalls, if leasing or seabed claims advance faster than diplomacy, or if Russia and China pressure the Arctic enough that Washington and Ottawa stop treating the overlap as a technical file.

Northwest Passage

The Northwest Passage is a legal-status dispute, not an areal boundary fight.

Canada treats the Passage as internal waters enclosed by Arctic straight baselines. The US treats it as an international strait. That sounds confrontational until you look at the instruments. The US has protested diplomatically and maintained its position on paper. It has not run a freedom-of-navigation operation against Canada (yet?).

That contrast matters. The US uses operational assertions against rivals. It has not done that to Canada. The mechanism is legal disagreement plus alliance management.

Sentiment changed less because of the Passage itself than because of broader sovereignty rhetoric. Inuit and northern leaders increasingly frame Arctic security as housing, food, infrastructure, search and rescue, and self-determination. After a year of annexation-style rhetoric from US politicians, the Canadian public mood hardened around sovereignty. That makes quiet management politically harder even if policy does not change.

Risk moves if a US or allied ship transits without Canadian consent and turns a paper disagreement into an operational one. That has not happened.

Machias Seal Island and the Gulf of Maine Grey Zone

Machias Seal Island is tiny, photogenic, and unusually concrete. The US and Canada both claim the same dry land. Around it sits the Gulf of Maine Grey Zone, where lobster makes the abstract line economically real.

This is not a new crisis. The dramatic boarding event often cited in the press is from June 2018, not 2025 or 2026. The 2025 development was rhetorical: a US presidential restatement of the claim. That distinction is essential. Rhetoric can raise temperature, but it is not the same as a new enforcement action.

Sentiment is divided in a way that makes the case interesting. National and provincial politicians can be sharp. Local fishers often live with the ambiguity because they have to. The Grey Zone is a worksite before it is a symbol.

Risk moves through enforcement, not speeches: arrests, gear conflict, patrol refusals, or USMCA-linked pressure around fishery regulation.

Dixon Entrance and Juan de Fuca

Dixon Entrance and the seaward Strait of Juan de Fuca are quieter than Machias and Beaufort. They matter because they show how long a line can remain unsettled without becoming central politics.

Dixon is a Canada-US disagreement over the A-B line and maritime delimitation. It flares around fisheries. Juan de Fuca is smaller and more technical, frozen partly because boundary method precedents matter.

These are dormant or low-salience disputes with real geography. Their risk is not ideology. It is a fisheries or enforcement incident that forces a government to defend a line it usually prefers to ignore.

2. The US-Mexico water and infrastructure flashpoints

The Mexico section has to be kept clean.

There are three real water or public-health conflicts: Rio Grande water debt, Colorado River shortage-sharing, and Tijuana sewage. They are not territorial launch zones, but they are real interstate disputes with legal instruments and material harm.

Gulf of America is not a real territorial conflict. It is a symbolic naming flashpoint.

The separate Mexico cartel-strike sovereignty case is handled in the perception-asymmetry section below. It is a real sovereignty conflict, but it is not a water dispute and not a naming fight.

Rio Grande water debt

The Rio Grande conflict is the sharpest US-Mexico water file because Mexico owes the United States.

Under the 1944 treaty, Mexico is supposed to deliver an average of 350,000 acre-feet per year from named tributaries over five-year cycles. The cycle ending in late 2025 closed with Mexico far short, roughly half of the required amount delivered. The US denied a Tijuana water request on 20 March 2025. The administration threatened a 5 percent tariff on Mexican goods on 9 December 2025, but it was not actually levied. A deal followed on 12 December, and a technical plan was announced in February 2026.

That sequence matters because it shows escalation. A water dispute became a trade threat, then a delivery plan. The tariff did not happen, but the threat had negotiating value.

Sentiment is already hot. In northern Mexico, water politics have produced protests, barricades, and violence. The 2020 La Boquilla confrontation, where Yessica Silva was killed, sits behind the modern grievance. On the Texas side, the conflict is framed in agricultural survival terms: sugar, citrus, irrigation districts, and jobs.

Risk moves if the next cycle inherits the deficit without predictable deliveries, or if the “extraordinary drought” clause becomes a permanent excuse rather than an emergency exception.

Colorado River shortage-sharing

The Colorado is the mirror image of the Rio Grande.

Here the US owes Mexico. The 1944 treaty gives Mexico a baseline entitlement, modified in practice by shortage-sharing minutes. Minute 323 expires on 31 December 2026. A successor is not yet signed. 2026 is a Tier 1 shortage year, and the broader post-2026 Colorado operating rules are still unsettled.

The conflict is real because the deadline is real. It is a treaty architecture reaching the end of its current operating framework while the reservoirs remain stressed.

Sentiment on the Mexican side is local and existential. Mexicali Valley farmers and Cucapa voices talk about a river that no longer reaches the sea, about dying fields, and about an ecological system turned into an accounting problem.

Risk moves if the successor minute is not signed in time, if basin-state politics freeze the US position, or if shortage reductions become harder to sell to Mexican users who already feel they are carrying another country’s drought.

Tijuana River sewage

The Tijuana River sewage crisis is a real conflict because the damage crosses the border every day.

Untreated flows have closed beaches, sickened people, and made South San Diego’s coastline a public-health zone. This is not a line on a map. It is human waste in the water. The governing chain is technical but important: Minute 328 in 2022, the EPA-SEMARNAT memorandum in July 2025, and Minute 333 in December 2025. The question is whether the infrastructure build-out meets its schedule.

Sentiment is local, angry, and practical. Imperial Beach wants the beach open and the smell gone. Tijuana needs wastewater capacity. National governments announce milestones; border communities live with the outflow.

Risk moves if project deadlines slip, funding breaks, or the promised wastewater capacity fails before the next major storm season.

Gulf of America

Gulf of America is not a real border conflict.

Executive Order 14172 changed US federal usage. Federal agencies updated names. Some maps and platforms displayed different labels by user location. Mexico objected and litigated. Politicians performed sovereignty.

No maritime boundary moved. No EEZ moved. No lease block moved. No fishery moved. No continental shelf entitlement moved.

That is why it belongs in the article as a flashpoint but not as a conflict. It measures political temperature, not territorial reality. If the article treats it as equivalent to Rio Grande, Colorado, or Tijuana, it is wrong.

Risk moves only if symbolic nationalism spills into real instruments: trade retaliation, energy permitting, maritime enforcement, or treaty cooperation. The name itself does not change the map.

3. Other Mexico and Gulf technical flashpoints

These cases are not the heart of the article, but they close important gaps in the map.

Transboundary hydrocarbons

The US-Mexico transboundary hydrocarbons agreement is a settled-boundary, unsettled-implementation case.

The 2014 framework exists so reservoirs that cross the maritime boundary can be developed by unitization rather than by a race to drain the same pocket from opposite sides. That is the good-news part. The flashpoint is not a line. It is whether political energy nationalism, Pemex finances, and US-operated Gulf projects can make the framework work when a shared reservoir matters.

Risk moves if a commercially meaningful cross-border reservoir is found and the two sides disagree over operatorship, revenue, data sharing, or unitization terms. The boundary is settled. The resource politics are not guaranteed to be boring.

Eastern and Western Gulf gaps

The Western Gap is mostly closure. The in-force 2001 treaty split the deepwater doughnut hole in the Perdido Fold Belt, with the United States getting the smaller share and Mexico the larger share. It belongs in the appendix because it is proof that a large offshore resource boundary can be settled cleanly.

The Eastern Gap is cleaner on paper than in institutional process. The United States, Mexico, and Cuba have treaty architecture that largely settles the deepwater eastern Gulf, but US Senate consent and the three-party Cuba tripoint remain the loose ends. That does not make it a hot dispute. It makes it a technical file that could become politically relevant if offshore activity or US-Cuba relations change.

Risk moves if Senate consent remains stalled while leasing, seabed activity, or Cuba policy changes create a reason to care about a line most voters have never heard of.

Rio Grande bancos and shifting river islands

The Rio Grande bancos are the older, quieter version of the border-water story.

Rivers move. Boundaries do not always move with them. The United States and Mexico built a treaty system to handle bancos, islands, channel changes, and rectification, including the 1905 and 1970 rules. Hundreds of small pieces of land have changed sovereignty under that machinery.

The case is largely resolved, but not dead. Border barriers, levees, flood-control work, and channel maintenance can still create accusations that one side is prejudging an island, impeding flow, or violating the free-flow and maintenance logic of the boundary regime.

Risk moves through engineering, not ideology: a wall segment, levee project, flood event, or IBWC disagreement that forces governments to answer which side of a moving river a piece of land belongs on.

The 1978 and 1997 US-Mexico maritime boundary

The US-Mexico Pacific and Gulf maritime boundary is backbone context.

The 1978 treaty and later entry into force give the two countries the basic territorial sea and EEZ framework within 200 nautical miles. That is why Gulf of America does not change anything, and why the real modern Gulf questions are the deeper shelf gaps, hydrocarbons implementation, and resource administration rather than the basic maritime line.

This is not a flashpoint, and it keeps other flashpoints from being worse.

4. The conflicts only one side is counting

The project started by asking where the United States has unsettled territorial, maritime, EEZ, water, and border-adjacent disputes. That question naturally privileges what the US government says or does not say. We also asked the inverse question: where does the other side think it is in a conflict with the United States even though Washington does not track the case as a dispute?

That distinction matters because Serranilla is only one direction of asymmetry. At Serranilla, the US holds a dormant claim it does not press while Colombia occupies the cay. In the new cases, the direction often flips. The US is the actor, the other government experiences a sovereignty problem, and Washington calls the same episode national security, counter-narcotics, sanctions policy, trade policy, or closed litigation.

Panama Canal

Panama is the cleanest territorial example because the legal geography is not obscure.

The Canal is Panamanian. The modern sovereignty settlement is the 1977 Torrijos-Carter treaty architecture. A US threat to “take back the Canal” therefore does not create a normal boundary dispute. It creates a perception-asymmetry case. Washington can frame the issue as messaging, China leverage, or strategic concern. Panama can treat it as a sovereignty threat.

The official act is Panama’s response. Panama filed a United Nations complaint invoking Article 2(4) of the UN Charter after the US threat. That does not mean the US filed a claim. It did not. That is the point. The US does not need to log a dispute for Panama to experience one.

The Canal case should not be written as a settled boundary suddenly becoming unsettled. It should be written as a sovereignty system reacting to coercive language. The risk pathway is diplomatic and economic first: port concessions, audits, treaty language, pressure on Chinese-linked infrastructure, and the possibility that rhetoric gets translated into administrative or military planning.

Greenland and Denmark

Greenland is the other territorial case, but it is stranger because geography pulls it into North America while sovereignty remains Danish and Greenlandic.

The US does not have a formal claim to Greenland. It has strategic interest, acquisition talk, and messaging that would not rule out force. Denmark and Greenland hear that differently. Danish intelligence treated the United States as a national-security threat for the first time that we could find, and Denmark summoned the US ambassador over Greenland-related concerns.

This is not a drone-range case for CONUS. It is out of range on the article map. It belongs because it shows the same category problem. Washington can say there is no dispute. Denmark can still treat US pressure as a sovereignty problem.

Venezuela

Venezuela is the marquee military-sovereignty case.

The United States frames the 2025-2026 Caribbean campaign through counter-narcotics and counter-terrorism. Venezuela frames it as aggression, regime change, and violation of sovereignty. That difference is not just rhetoric layered on nothing. The underlying US acts are official: strikes, designations, operations, and force.

The article map does not ring Venezuela’s mainland because the CONUS geometry does not reach cleanly from the mainland case under the working assumptions. But the case is still central to the appendix because it shows the category gap at its highest intensity. One side says interdiction. The other side says war.

That does not make Venezuelan claims true in every legal detail. It means the article should not let US category language erase the fact that the other side is counting a conflict.

Colombia strikes

Colombia now appears in the article twice, and the two appearances must not be collapsed.

Serranilla is a dormant paper-claim problem. The US claim is unrenounced but silent, and Colombia physically administers the cay. The 2025 strike grievance is different. It is a live sovereignty and force problem created by US counter-narcotics operations, decertification, sanctions, diplomatic rupture, and Colombian allegations that Colombians were killed and Colombian waters were violated.

That makes Colombia the bridge case. It is both the home of the article’s weirdest dormant territorial claim and one of the new perception-asymmetry cases. The same bilateral relationship carries both forms of stress, but they should not be treated as one dispute.

The writing line should be: Serranilla is the dormant legal vocabulary; the strike campaign is the active diplomatic accelerant.

Mexico cartel-strike sovereignty

Mexico has three real water or public-health conflicts in this article. The cartel-strike issue is not one of them. I almost treated the cartels as nation-state for the purposes of this thought exercise.

The US designated cartels as foreign terrorist organizations and floated unilateral strike options. Mexico responded by hardening anti-intervention language around Article 40 of its constitution. The substance is not a boundary line. It is whether the United States may use force on Mexican territory, water, sea, or airspace without Mexican consent.

That makes it more serious than Gulf of America and different from Rio Grande, Colorado, and Tijuana. It is not symbolic naming. It is not water accounting. It is the old non-intervention principle meeting a new US counter-narcotics theory.

The map rings Mexico because the geography is obvious: it is the US border. That ring is not a claim that Mexico would launch anything. It is a reminder that an unfriendly border changes the defense geometry immediately. And the cartels are a well resourced military-capable actor in and of themselves.

Cuba embargo

Cuba is partly already in the catalogue through Guantanamo Bay. The embargo is different.

The US treats the embargo as sanctions policy. Cuba treats “el bloqueo” as economic aggression and often folds Guantanamo into the same anti-sovereignty narrative. The UN General Assembly vote is useful here not because it decides the legal question, but because it shows the persistence and scale of the foreign-perceived grievance.

Nicaragua 1986 ICJ judgment

Nicaragua is a legal-memory case.

In 1986, the International Court of Justice ruled against the United States in the Nicaragua case over mining harbors and support to the Contras. The US never paid the reparations Nicaragua says it is owed. Washington treats the matter as closed. Nicaragua periodically treats it as unresolved, including through later UN correspondence.

This is not a territorial dispute. The Bryan-Chamorro and Corn Islands residue belongs in the resolved-history bin. But the unpaid ICJ judgment is still useful in the article because it shows how legal time differs between capitals. A case the US treats as dead can remain politically alive for the other side.

Brazil sanctions and tariffs

Brazil is the weakest of the new live cases for this article.

The issue is not land, water, or maritime space. It is economic and judicial sovereignty: US sanctions and tariffs tied to Brazil’s domestic judicial politics and the Bolsonaro prosecution. Brazil frames that as interference. The US frames it as sanctions and trade policy.

Brazil is out of range on the CONUS map and should not be over-weighted. It belongs as a category marker, not as a central scene. Its value is that it shows the perception-asymmetry pattern can appear outside the Caribbean and outside classic territorial claims.

Negative controls

The negative controls are important because they keep the appendix from becoming a grievance collector.

Ecuador and Bolivia are reversed cases. Ecuador’s current government has courted US basing rather than treating it as a live government-level sovereignty conflict. Bolivia’s new government readmitted the DEA in February 2026, which retires the old anti-empire framing at the state level for now.

Guyana, Suriname, Peru, Paraguay, and Uruguay do not fit the pattern. The current government posture is partnership, not a foreign-perceived US conflict.

Argentina and Chile in Antarctica are latent only. Their claims run into the US non-recognition and reservation posture, but the Antarctic Treaty freezes the problem and their sharper friction is with the United Kingdom, not the United States.

Honduras and the Swan Islands are the resolved foil. The US had a Guano Act claim and gave it up by treaty in 1971-1972. That matters because it proves the US can let one go. It is the clean Navassa comparison, and a useful reminder that not every old guano claim has to stay weird forever.

5. The Caribbean: old claims, bases, and EEZ geometry

Serranilla and Bajo Nuevo

Serranilla is the centerpiece because it combines every ingredient the article is about; and quite honestly I found it really interesting.

There is a physical outpost. Colombia mans Beacon Cay. There is a dormant paper claim. The US never formally renounced Serranilla or Bajo Nuevo. There is resource stress. The Raizal fishery around San Andres has collapsed. There is diplomatic stress. US-Colombia relations deteriorated sharply in 2025. And there is silence. Washington has not reasserted the claim.

That silence is what makes the case unusual. Active claims are easy to see. Dormant claims are not. Serranilla is a title that remains available without being used.

We’ll be careful writing down the visual evidence. The garrison is compact and persistent. It is not a sprawling base. The 2018-2025 imagery shows entrenchment, not a build-out.

The resource story should also be disciplined. Fish is fact. Hydrocarbons are unproven. A seabed discovery could change the risk environment, but the current human conflict is about fisheries, livelihood, and the Raizal sense of sea-territory.

Bajo Nuevo is the contrast: same dormant US claim family, much less physical presence. Colombia lights it with a beacon. It does not man it like Serranilla.

Navassa

Navassa is the precedent.

The US administers Navassa as a wildlife refuge. Haiti claims it. The US reactivated a dormant Guano Act claim there in 1916 and still holds the island. That is why Navassa matters in a Serranilla article. It proves that dormant guano claims can be revived as state claims.

But the analogy has limits. At Navassa, the US holds possession. At Serranilla, Colombia does. Haiti also lacks the practical state capacity to press the claim the way Colombia can defend Serranilla. Navassa is the mechanism, not the prediction.

Sentiment is anti-colonial and diasporic. Haitian voices frame the island as stolen territory and resource loss. The state claim survives, but the state itself has bigger problems.

Risk moves if a stabilized Haitian government, diaspora legal campaign, or international forum makes Navassa a live diplomatic issue again.

Guantanamo Bay

Guantanamo is a real sovereignty and lease standoff, but it should not dominate this article.

The US treats the 1903 and 1934 instruments as preserving its right to occupy the base until mutual agreement or abandonment. Cuba calls the base illegal occupation and refuses to cash the rent checks. The base sits at a strategic point near the Windward Passage and carries all the political weight of detention, migration, and US-Cuba history.

It belongs in the appendix because it is a live territorial standoff in the region. It belongs outside the center because it would swallow the article.

Risk moves through base mission changes, US-Cuba diplomatic rupture, or a new use that makes the lease issue politically unavoidable.

Bahamas, Dominican Republic, Haiti, and Cuba maritime lines

The Caribbean also has quiet EEZ files that look dull until you draw range or find resources.

The US-Bahamas boundary is undelimited in several segments and became more salient after the 2023 US extended shelf filing. The Dominican Republic archipelagic-baseline dispute affects the Mona Passage near Puerto Rico. The US-Haiti maritime boundary is tangled with Navassa. The US-Cuba line is functionally applied but historically provisional because the 1977 agreement has not entered into force.

These are not popular mass conflicts. Sentiment is often thin, elite, or technocratic. Their risk is not street politics. Their risk is that an enforcement action, seabed discovery, migrant-route pressure, or formal court move makes a quiet geometry problem visible.

6. Pacific and Bering: farther from CONUS, still strategically real

Bering Sea

The Bering Sea line is the most strategically serious “farther out” case.

The US and Soviet Union signed the Baker-Shevardnadze line in 1990. The US ratified it in 1991. Russia never did. Both sides have applied it provisionally, but the line remains a political target for Russian nationalist resentment. Russia’s 18 March 2024 objection to the US Arctic extended shelf filing made the broader shelf dispute current again.

This is beyond the CONUS cheap-UAV frame, but it is not beyond strategic relevance. It is a provisional line between nuclear powers in a warming ocean.

Risk moves if Russia formally denounces provisional application, if fisheries enforcement becomes confrontational, or if Arctic resource bargaining turns the line into leverage.

Wake Island

Wake is the most open US Pacific sovereignty dispute.

The US administers it and invests in airfield and missile-defense infrastructure. The Marshall Islands claim it as Anen Kio. That claim is serious and cultural. It should not be confused with the bogus “Kingdom of EnenKio” micronation story.

The risk is not cheap-UAV range to CONUS. The risk is that a heavily militarized atoll with a live cultural claim becomes part of a broader US-Pacific Island negotiation over defense, compacts, and sovereignty.

Mariana ECS and Swains/Olohega

The Mariana extended-shelf overlap with Japan is the cooperative version of a real issue. Both sides are allies. The overlap is technical. It requires bilateral handling, but it is not a confrontation.

Swains Island, or Olohega, is the cultural version of a settled state-to-state issue. The US holds Swains as part of American Samoa under the Treaty of Tokehega. Tokelauan voices still treat it as a stolen fourth atoll. The state claim is settled. The grievance is not.

These cases are useful because they show that settlement and sentiment are different things. A treaty can close a state claim without closing a people’s memory.

7. Resolved controls: what closure looks like

The resolved cases matter because they gave me more hope.

The San Juan Islands were settled by arbitration. Chamizal was settled by treaty and land transfer. Rio Rico and the Horcon Tract were resolved after the river moved. The Western Gulf gap was delimited. The US renounced many Pacific guano claims by treaty with Cook Islands, Kiribati, Tuvalu, and others. The US-UK Caribbean boundaries are settled. The US-Venezuela Aves line is settled between those two states.

The water controls matter too. The All-American Canal lining case shows how a US infrastructure project can solve one legal fight while leaving a transboundary-groundwater gap in political memory. The Colorado salinity crisis, resolved through Minute 242 in 1973, shows the opposite lesson: a water-quality conflict that damaged Mexican agriculture could be converted into a technical binational settlement.

The small Pacific controls are boring in the best way. The Federated States of Micronesia, Niue, Cook Islands, Kiribati, and Tuvalu boundaries show what closure looks like when guano-era residue and EEZ lines are cleaned up by treaty. Not every old claim becomes Serranilla.

The system can work. The US does renounce useless claims. Boundaries do get drawn. Old disputes can become boring.

Not because I need another AI model to feed my agents. I miss it because there is a specific kind of innovative tasking where the frontier model matters. You have an “I wonder if” hypothesis. You ask the model to help you build the code, analyze the data, test the bounds, and characterize the result. If the thing you are discovering was not already publicly known in the model’s training data, weaker models keep trying to pull you back toward what they already know. The best models do not just answer. They help you hold onto the new thing long enough to prove it.

That is the productivity gain people miss when they talk about frontier AI as if it is just a better chat box.

So I wondered what the cost is of delaying the next frontier-model generation in the United States.

The answer: about $170 million in gross US GDP per working day. That’s about $43 billion per year at today’s adoption in a $31.9 trillion US GDP. That’s only 0.13% of GDP if delayed a full year.

This is not a net-welfare claim. I am not pricing national security, safety, adversary access, diplomatic leverage, or the political value of slowing a technology that scares Washington. This is one side of the ledger: the economic output the country gives up when it holds back the next frontier generation.

The public story is not the operative story

If you read the press straight, you will get the wrong model.

The press story will always be professionally spun. Companies preserve optionality. Public companies and pre-IPO companies avoid saying they are in a political fight with the US government because share prices and offering windows are real constraints. The legal theory is not the relevant. The point is leverage.

When the government signals that broad rollout of a frontier model will create a conflict, the lab does not need to receive a clean, durable, court-tested order before it pauses. Boards understand the risk. Investors understand the risk. A CEO who wants an IPO, a stable market cap, government contracts, and regulatory goodwill understands the risk.

That is the mechanism I model here.

Anthropic, OpenAI, and Google are all operating under the same political constraint: Washington currently wants control over the release of frontier models. The public explanations differ. Export control, limited preview, commercial delay, safety review, customer-by-customer access. Those are the surface forms. The economic effect is the same: the US economy keeps using the current generation while the next generation is withheld from broad productive use.

That delay has a meter running.

How $31.9 trillion becomes $170 million a day

The model is deliberately simple:

Daily cost = GDP x AI productivity uplift x adoption today x restricted frontier share
             x one-generation gain / working days

Start with US gross domestic product: about $31.9 trillion annualized.

Then ask how much frontier AI could add at full adoption. The literature is all over the place. A skeptical read is low-single-digit GDP impact. Goldman/McKinsey/Merali-style central estimates land around 7%. More aggressive views go higher. I use 7% central, with a wide band.

Then apply today’s adoption. Not everyone uses AI all day. A lot of workers never touch it. A lot of companies have “AI adoption” that doesn’t go further than a powerpoint deck. I use 10% effective labor augmentation today, with sensitivity from 6% to 20%.

Then apply the restricted frontier share. In the central case, Anthropic, OpenAI, and Google account for about 83% restricted-weighted frontier use after accounting for OpenAI’s current limited preview.

Then apply the value of one model generation. The clean Anthropic Fable benchmark delta is about +12% on the strict within-harness coding comparisons. The central transfer method maps that to about a 23% increase in the value of the AI-using slice, not a 23% productivity gain for the whole workforce.

Divide by 250 working days.

That gets you to $170 million per working day.

The important thing is not false precision. The important thing is scale.

Even if you reject the central capability-to-productivity bridge and use the conservative benchmark-proportional method, the all-three estimate is about $89 million per working day. If you only count the measured Anthropic case and ignore OpenAI and Google, it is still about $62 million per working day.

That is the floor I would want a policy team to look at before pretending delay is free.

The uncertainty is not where people think it is

The easiest critique is to argue about benchmarks. That is not where the model mostly lives or dies.

I re-audited the benchmark data and used a strict subset for the headline capability delta: SWE-bench Pro and SWE-bench Verified. That gives a clean +11.5% Fable-over-Opus gain, which supports the rounded +12% input. Other benchmark sets move around, and some are noisy, but the central result is not driven by cherry-picking a leaderboard.

The real uncertainty is macro:

• How much productivity frontier AI ultimately adds at full adoption.

• How much of the economy is actually using it today.

Those two numbers dominate the sensitivity analysis.

If today’s AI adoption is only the St. Louis Fed’s hours-share floor, the estimate falls. If adoption is closer to worker breadth with meaningful intensity, it rises. If AI’s full-adoption GDP impact is Acemoglu-low, the number is smaller. If the Goldman/McKinsey/IMF-style estimates are closer, the number gets large quickly.

But across the plausible range, the daily cost is not zero. It is tens to hundreds of millions of dollars per working day. You’re welcome to pick the assumptions you believe, and pick one of the nine boxes above to get the GDP cost which fits your personal mental baseline.

What I would tell the government

If you are going to control frontier model rollout, price the control honestly.

Maybe you think the safety case is worth it. Maybe you think the national-security case is worth it. Maybe you think the United States should slow the commercial frontier until the government has a better grip on evaluation, exportability, cyber risk, bio risk, or strategic alignment with allies.

Fine. Make that argument. But do not make it while implicitly assuming there is no economic cost.

The current generation of models is not a static consumer product. It is an input into software development, research, analysis, operations, finance, engineering, security, medicine, law, and management. A better model does not merely answer faster. It expands the set of tasks where a person can successfully delegate part of the work. It changes how many tasks someone can run in a day. It changes how quickly a small team can test an idea. It changes the throughput of knowledge work.

Holding back a generation means the country runs those workflows on worse tools.

That is a GDP cost.

And because adoption is rising, the daily cost rises too. If the freeze persists for months, the linear bill is already large. If the US keeps missing successive frontier generations while other paths advance, the compounding scenario becomes ugly fast.

I am not saying every model should be released the moment a lab wants to release it. I am saying the policy conversation should make sure the ledger shows a net gain to the country. My central estimate is that delaying Frontier AI rollout costs the US GDP $170 million per working day of delay. The benefit of delays should outweigh the cost of delay.

How to read it: left is faithful translation quality against human-curated gold; right is technical-detail extraction accuracy. Error bars are 95% confidence intervals. What to see: DeepSeek-V4-Flash joins GPT-5.5’s tier on both tracks, Qwen3.6-35B-A3B bf16 thinking reaches the GPT-5.5 extraction tier, and the practical recommendation still depends on serving reliability.

Results up front

I have spent the last few weeks paying for frontier model tokens to read Chinese financial and technical filings.

Not for literary translation.

For the data that actually matters to the questions I want to answer: semiconductor disclosures, subsidy language, data-center capacity, R&D spend, government grants, risk factors, customer concentration, and the data’s units that become load-bearing when you publish.

Those LLM tokens got more expensive than I wanted to achieve the fidelity I needed.

More importantly, it risked changing how I research answers. When every Chinese excerpt feels like a token bill, you start sampling carefully. That is the wrong instinct. For this work I want to read broadly, translate everything that might matter, extract the numbers, and reserve expensive review for the passages that carry a claim. I was searching for needles in the haystack so I needed to be able to automate going through it haystack by haystack, hay strand by hay strand, until I found the needles or be relatively confident those haystacks didn’t bear needles.

So I ran the experiment I should have run earlier.

The question was narrow:

Can a locally hosted open-weight model replace frontier LLM for technical Chinese filing work?

The answer, as of mid June 2026, is yes.

With one important qualification: this does not replace skilled Chinese translators where judgment matters. It replaces the broad, expensive Frontier AI pass. Human translation effort is scarce and valuable. I would not spend it grading a model bake-off unless the review itself had direct strategic value. I would spend it on the handful of passages where a real research claim depends on nuance.

That is the whole point.

The benchmark result is real. DeepSeek-V4-Flash tied GPT-5.5 on human-gold translation chrF++ (65.1 vs 65.0) and on technical-detail extraction (85.2% vs 85.7%), while significantly beating GPT-4.1 on both tracks. It also produced zero hallucinations in the judge run.

The workflow changes from “pay a Frontier AI to read everything” to:

Bulk technical translation: use Qwen3.6-35B-A3B bf16 direct. Do not turn on thinking mode by default.

Unit-heavy financial extraction: use Qwen3.6-35B-A3B bf16 thinking, Google Gemma 4 31B IT bf16, or NVIDIA Nemotron-3 Super 120B-A12B-NVFP4. DeepSeek-V4-Flash also passes the stress test, but I would not make it the default until I can serve the model without layers of kludges. Do not use Qwen direct for this job.

Structured JSON extraction: use Qwen3.6-35B-A3B bf16 thinking with strict unit instructions.

Publication-critical material: use GPT-5.5 audit (I had dedicated my Anthropic token budget to Fable while it was enabled), and human review if the Chinese excerpts are strategically important to geopolitical decision making.

Sensitive working notes and large source corpora: run local first. Do not upload unnecessarily to an API.

Benchmark curiosity: DeepSeek-V4-Flash is the strongest all-around local result in this run, including the unit stress test. I am not making it the default until vllm “just works”.

The salient point here is that I was becoming reticent to spend a few hundred dollars on frontier LLM tokens to achieve the necessary fidelity for each high-confidence assessment. I already have a sunk personal cost of high-end AI GPUs. By reframing my mental workflow from “is it worth spending a few hundred bucks on something that will probably fail” to “these tokens are almost free” I could take on more project risk which led to some very interesting findings.

Put through the entrepreneur’s lens, converting a recurring operating cost to a higher sunk cost let me optimize the human aspects of a workflow. Money well spent.

The result that changed the recommendation

The most important finding was not the headline leaderboard. It was the unit stress test.

Chinese filing work is often not hard because the sentence is hard. It is hard because 亿元 means hundred-million yuan, 万元 means ten-thousand yuan, 千元 means thousand yuan, and a missed table heading can move a number by a power of ten.

The model does not need haiku.

It needs unit discipline.

So I added a stress test: 77 objective targets from my work with Chinese annual reports and prospectuses. The past projects had already cross checked these transcribed numbers and unit conversion against other (audited) filings. That’s a validated test suite.

How to read it: 77 objective prospectus/annual-report targets; 95% confidence intervals; dashed line is GPT-5.5. What to see: Qwen3.6-35B-A3B bf16 direct collapses. Gemma 4 31B IT bf16 and Nemotron-3 Super 120B-A12B-NVFP4 lead at 96.1%. GPT-5.5, Qwen3.6-35B-A3B bf16 thinking, MiniMax-M2.7-NVFP4, and DeepSeek-V4-Flash all land at 94.8% with overlapping confidence intervals.

The result:

1. Google Gemma 4 31B IT bf16, 32.7B checkpoint parameters: 96.1%

2. NVIDIA Nemotron-3 Super 120B-A12B-NVFP4, MoE, 120B total / 12B active: 96.1%

3. OpenAI GPT-5.5: 94.8%

4. Qwen3.6-35B-A3B bf16 thinking, MoE, 35B total / 3B active: 94.8%

5. MiniMax-M2.7-NVFP4, MoE, about 116B checkpoint parameters: 94.8%

6. DeepSeek-V4-Flash mixed fp4/fp8, MoE, 284B-class: 94.8%

7. OpenAI gpt-oss-120B MXFP4, MoE, 120B, 128 experts / top-4: 93.5%

8. OpenAI GPT-4.1: 92.2%

9. Z.ai GLM-4.7-Flash bf16 thinking, MoE-lite, 64 experts / top-4: 89.6%

10. Z.ai GLM-4.7-Flash bf16 direct, MoE-lite, 64 experts / top-4: 83.1%

11. Mistral Small 4 119B 2603 NVFP4: 77.9%

12. Qwen3.6-35B-A3B bf16 direct, MoE, 35B total / 3B active: 64.9%

DeepSeek’s late stress-test run matters because it did not merely win the broad leaderboard and then fail the unit trap. On this 77-target prospectus stress set, it scored 94.8%, with a 95% CI of [89.6, 98.7], matching GPT-5.5, Qwen3.6-35B-A3B bf16 thinking, and MiniMax-M2.7-NVFP4. It also had 0.0% scale errors. That makes DeepSeek part of the same unit-discipline tier, even though I still would not make it the default workflow model until the serving stack just works.

That changed the operating doctrine.

Qwen3.6-35B-A3B bf16 direct (non-thinking) is my default translator.

It is not my extractor for unit-heavy filings.

In one example, Hygon disclosed R&D spending as:

研发投入 280,977.56 万元

The requested answer was in 亿元. That’s a unit which means hundred-million yuan. The right answer is 28.10.

Qwen3.6-35B-A3B bf16 direct returned 2,809.78.

Qwen3.6-35B-A3B bf16 thinking returned 28.10.

Same model. Same excerpt. One setting changed.

That is not a style problem. That is a hundred-fold error. And if you’ve read my 10 part series on China as an AI Great Power, you might remember that one of the hardest parts of analyzing Chinese data is keeping the numbers’ units correct. The smallest error is at least a 10x error.

So the rule is simple: translate direct; extract unit-heavy numbers with thinking mode or with a model that already has strong unit discipline.

What I tested

This was not a generic translation benchmark. I used the artifacts from my recent research projects: Hunting Huawei’s Hidden Fab, Mapping China’s AI Fab Industry, Mapping China’s AI Model Builders, Mapping Chinese Data Centers, and an exploration of Chinese state subsidies which I haven’t published yet. These are the kinds of Chinese excerpts I was already paying models to process while doing real work. And more importantly, I have been cross-checking numbers from alternative sources and modalities so I find myself with a really effective translation evaluation set for the problems I care about.

The shorthand matters, so here is what the local names mean in this article:

Qwen direct / Qwen thinking means Qwen3.6-35B-A3B bf16, MoE, 35B total / 3B active parameters, run in direct or thinking mode.

Gemma means Google Gemma 4 31B IT bf16, 31B class / 32.7B checkpoint parameters.

Nemotron means NVIDIA Nemotron-3 Super 120B-A12B-NVFP4, MoE, 120B total / 12B active parameters, 512 routed experts, top-22 routing, mixed NVFP4/FP8.

MiniMax means MiniMax-M2.7-NVFP4, MoE, about 116B checkpoint parameters, 256 experts, top-8 routing, NVFP4.

GPT-OSS means OpenAI gpt-oss-120B MXFP4, MoE, 120B class, 128 experts, top-4 routing.

GLM means Z.ai GLM-4.7-Flash bf16, MoE-lite, 64 routed experts, top-4 routing; substitute for the larger GLM-4.7-NVFP4 target.

DeepSeek means DeepSeek-V4-Flash base mixed fp4/fp8, a 284B-class MoE, served through the prebuilt vLLM-lucifer docker stack. It is included in both the main leaderboard and the 77-target prospectus stress test, where it matches GPT-5.5 at 94.8% with zero scale errors.

Mistral means Mistral Small 4 119B 2603 NVFP4, 119B class, NVFP4A16 group-16 quantization.

Two tracks:

Track A: faithful translation. Chinese quote in, faithful English out. No summary, no commentary, no invented connective tissue. Preserve the numbers, units, names, and technical language. The track has 180 translation pairs: 75 human-curated gold pairs and 105 GPT-4.1-referenced pairs.

Track B: technical-detail extraction. Chinese filing excerpt in, structured English values out: revenue, grants, percentages, PUE, megawatts, cabinet counts, subsidy categories, R&D spend. The track has 439 records and 1,534 scoreable targets.

Then came the 77-target unit stress test above.

That last test is why the article is about Chinese filing work, not just translation.

The leaderboard, corrected

Here are the core results.

How to read it: translation is chrF++ (a translation similarity metric) on human-gold references; extraction is target-level accuracy; judge is GPT-5.5 scored on a 1 to 5 scale; latency is mean seconds per item. What to see: DeepSeek reaches GPT-5.5’s tier on both objective tracks, Qwen3.6 thinking reaches the extraction tier with much smaller operational complexity, and GPT-5.5 remains the best judged publication translator.

The raw point estimate says Qwen3.6-35B-A3B bf16 thinking scores 86.5% on extraction and GPT-5.5 scores 85.7%. The first draft would have said Qwen3.6-35B-A3B “beats” GPT-5.5.

That is a little too strong. I care most about data extraction fidelity, but I do care about translation especially for aggregating sentiment analysis.

After bootstrapping the comparisons, Qwen3.6-35B-A3B bf16 thinking vs GPT-5.5 on extraction is a statistical tie: +0.8 percentage points, 95% confidence interval [-0.7, +2.6]. MiniMax-M2.7-NVFP4 vs GPT-5.5 is also a tie.

DeepSeek-V4-Flash is the new wrinkle. On human-gold chrF++, DeepSeek scores 65.1 and GPT-5.5 scores 65.0: a tie, +0.09, 95% CI [-1.3, +1.5]. On extraction, DeepSeek scores 85.2% and GPT-5.5 scores 85.7%: also a tie, GPT-5.5 +0.46 points, 95% CI [-1.1, +2.0]. DeepSeek significantly beats GPT-4.1 on both human-gold chrF++ (+2.9, 95% CI [+1.3, +4.5]) and extraction (+2.5 points, 95% CI [+0.4, +4.9]).

The durable claims are these:

• DeepSeek-V4-Flash ties GPT-5.5 on both main objective tracks and significantly beats GPT-4.1 on both. That is the strongest all-around local benchmark result.

• Qwen3.6-35B-A3B bf16 thinking significantly beats GPT-4.1 on data extraction: +3.8 points, 95% CI [+1.7, +6.2].

• NVIDIA Nemotron-3 Super 120B-A12B-NVFP4 significantly beats GPT-4.1 on human-gold chrF++ translation: +2.4, 95% CI [+0.9, +4.0].

• Thinking genuinely improves Qwen3.6-35B-A3B data extraction: +2.9 points, 95% CI [+1.8, +4.0].

• GPT-5.5 remains the best judged publication translator: it leads the GPT-5.5 judge overall, and significantly beats Qwen3.6-35B-A3B bf16 thinking on that judge, +0.11, 95% CI [+0.04, +0.19]. DeepSeek’s objective translation score is excellent, but its fluency score is lower than GPT-5.5’s.

The honest headline is not “DeepSeek replaces GPT-5.5.”

It is: local open-weight models reach GPT-5.5’s tier on the main objective tracks, reach or clear GPT-4.1 for this workflow, and leave GPT-5.5 as the final prose-audit ceiling.

That is enough to change how I work. This surprises many people: I don’t codify Claude or Codex skills very often (a skill is a specific technique you encode into your AI agent). I typically manage hierarchies of agents like I manage organizations of people. I give them objectives, requirements, metrics, and expectations; I provide rigorous oversight; and I don’t care how they tactically or technically achieve my objectives. My workflow results in the agents building one-time-use tools bespoke to the problem we are addressing. But in this case, I had Claude and Codex both codify a translation skill for each agent. Pro tip: you can just tell your agent this URL and tell it to create a translation skill based on the blog post.

What the examples show

For ordinary technical disclosure, Qwen3.6-35B-A3B bf16 direct is good enough and far cheaper at scale if you have suitable GPUs already.

Chinese disclosure:

如果我们因出口管制而无法采购足够数量的先进半导体芯片，我们可能面临重大挑战，包括产能受限、成本上升及效能下降。

Reference:

If we are unable to procure a sufficient quantity of advanced semiconductor chips due to export controls, we may face significant challenges, including capacity constraints, rising costs, and degraded performance.

Qwen3.6-35B-A3B bf16 direct:

If we are unable to procure sufficient quantities of advanced semiconductor chips due to export controls, we may face significant challenges, including constrained production capacity, increased costs, and reduced performance.

The load-bearing terms are there: export controls, advanced semiconductor chips, constrained capacity, costs, performance. I would not spend frontier tokens on this.

For close prose, GPT-5.5 still has the edge. It tends to preserve disclosure register more precisely, while local models sometimes paraphrase into a phrase that is accurate but less filing-native. That matters when a sentence is going into an article as evidence.

This is why GPT-5.5 stays in the workflow.

Not everywhere. Just where the language needs to be precise. That’s precisely where I would have spent favor tokens on a human translator when I was working national security mission.

Thinking mode is not a religion

Thinking mode is a knob. Use it where it buys accuracy.

How to read it: direct vs thinking for Qwen3.6-35B-A3B bf16 and Z.ai GLM-4.7-Flash bf16. Left: extraction accuracy. Middle: translation chrF++ against human gold. Right: output tokens per item on a log scale. What to see: thinking helps extraction, barely moves translation, and costs roughly 25x the tokens.

The direct/thinking trade-off in numbers:

Qwen3.6-35B-A3B bf16 direct: extraction 83.6%; translation chrF++ 62.9; 96 output tokens per item; 2.5 s mean latency.

Qwen3.6-35B-A3B bf16 thinking: extraction 86.5%; translation chrF++ 63.6; 2,606 output tokens per item; 63.3 s mean latency.

GLM-4.7-Flash bf16 direct: extraction 61.5%; translation chrF++ 62.0; 85 output tokens per item; 2.7 s mean latency.

GLM-4.7-Flash bf16 thinking: extraction 84.4%; translation chrF++ 61.8; 2,167 output tokens per item; 134.0 s mean latency.

For translation, thinking is mostly wasted latency.

For extraction, thinking can be decisive. GLM-4.7-Flash bf16 direct is bad at extraction; GLM-4.7-Flash bf16 thinking is competitive. Qwen3.6-35B-A3B bf16 direct looks acceptable on broad extraction, then fails the unit stress test. Qwen3.6-35B-A3B bf16 thinking fixes the exact problem.

So the rule is simple: Translate direct. Extract with thinking when units matter.

How each model fits the workflow

The leaderboard tells you who won. The failure modes tell you what to do with the model.

DeepSeek-V4-Flash base mixed fp4/fp8, 284B-class MoE: strongest all-around benchmark result; ties GPT-5.5 on human-gold chrF++ and extraction; matches GPT-5.5 on the 77-target unit stress test at 94.8%; zero hallucination in the judge run. Do not make it the default yet unless you are deliberately doing serving-stack research or the stack stabelizes.

Qwen3.6-35B-A3B bf16 direct, MoE 35B / 3B active: good translation; unsafe on unit-heavy extraction. Use it as the default bulk translator.

Qwen3.6-35B-A3B bf16 thinking, MoE 35B / 3B active: strong extraction; 0.0% hallucination in the judge run; slow. Use it as the default structured extractor.

Google Gemma 4 31B IT bf16, 32.7B checkpoint parameters: strong all-around; 96.1% on the unit stress test. Use it as the best fallback and possible direct-mode extractor.

NVIDIA Nemotron-3 Super 120B-A12B-NVFP4, MoE 120B / 12B active: high chrF++; literal/verbose prose. Use it for analyst drafts; it is weaker as publication prose.

NVIDIA MiniMax-M2.7-NVFP4, MoE about 116B checkpoint parameters: strong extraction; slow and two-GPU. Useful if already serving, but not the default.

OpenAI gpt-oss-120B MXFP4, MoE 120B, 128 experts / top-4: solid but not the best at anything here. Viable backup.

Z.ai GLM-4.7-Flash bf16 direct, MoE-lite 64 experts / top-4: very weak extraction. Avoid for extraction.

Z.ai GLM-4.7-Flash bf16 thinking, MoE-lite 64 experts / top-4: extraction rescued, but very slow. Use only if already committed to GLM.

Mistral Small 4 119B 2603 NVFP4: too many wrong values and scale errors. Avoid for structured data.

GPT-5.5: best translator, strong extractor. Use for final audit.

How I would keep this from lying to me

This workflow is not “let a local model write the research.”

It is a controlled pipeline:

1. Translate broadly. Use Qwen3.6-35B-A3B bf16 direct for the first pass over Chinese filings.

2. Extract with schemas. Use Qwen3.6-35B-A3B bf16 thinking, Google Gemma 4 31B IT bf16, or NVIDIA Nemotron-3 Super 120B-A12B-NVFP4 for structured JSON extraction.

3. Validate units deterministically. Any value in 亿元, 万元, 千元, 元, or % gets checked before it enters a claim ledger.

4. Detect scale errors. Values off by powers of ten get flagged automatically.

5. Audit only what matters. GPT-5.5 checks the passages that carry the argument.

6. Use human translation where it is valuable. If a passage is strategically or geopolitically important and nuance matters, that is where scarce human review belongs.

The point is not to remove human judgment. The point is to spend human judgment where it matters.

The cost answer

The cost comparison uses a pricing snapshot: GPT-5.5 at $5 per 1M input tokens and $30 per 1M output tokens, local electricity at roughly $0.15/kWh on one RTX PRO 6000-class GPU (which is a sunk cost). Token prices change. Recalculate this based on your power costs and today’s token costs if you’re making decisions based on it.

The exact dollars are less important than the order of magnitude:

This benchmark, 619 excerpts: frontier API proxy $6.75; local electricity proxy $0.02; about 340x difference.

10,000 excerpts: frontier API proxy about $109; local electricity proxy about $0.31; about 350x difference.

100,000 excerpts: frontier API proxy about $1,090; local electricity proxy about $3.12; about 350x difference.

Do not read that as a cloud economics paper. Hardware is not free. Engineering time is not free. Local serving breaks in annoying ways.

DeepSeek is the warning label on that sentence. I am glad I got the score. I am not pretending a model is cheap to operate when it only runs on a pile of infrastructure exceptions (today).

Read it as the workflow answer.

At frontier-token prices, you ask “which excerpts deserve translation?”

With a local model, you ask “what did I miss?” and “did I remember to turn the air conditioner on in my office before it gets uncomfortably warm?”

“What did I miss?” is the better research question.

There is also a control benefit. Running locally lets me process large working corpora, intermediate notes, and sensitive leads without sending everything to an API. Cost is the obvious reason to do this. Control is the quieter reason for those projects I choose not to publish.

Reproduce the practical workflow

Use Claude or Codex or your coding agent of choice:

> Download the largest quant of Qwen 3.6 that will fit on my GPU. Download and configure the latest version of vLLM to serve up Qwen 3.6. Test it. Then write and install a user-level skill to run and use Qwen for Chinese translation and data extraction based on the blog post at https://www.mfrantzen.com/p/the-translation-bill-for-bulk-chinese

Notice the wording: I would ask the agent to stand up Qwen, not DeepSeek. DeepSeek is the most interesting benchmark result. Qwen is the model I would currently put into a repeatable translation workflow.

The two prompts for translation and data extraction deliberately boring.

Translation: expert Chinese-to-English translator for technical and financial disclosure; preserve every term, number, unit, and named entity; do not omit, summarize, or add; output only English.

Extraction: meticulous bilingual financial-filing extractor; extract only values explicitly present; use null when absent; fields ending _yi are in 亿元; convert 千元, 万元, 元, and percentages explicitly; return only JSON.

The boring prompt is part of the result. Fancy prompting was not the point. Unit discipline was.

Caveats that matter

First, the translation judge is GPT-5.5. That creates a possible family bias toward GPT-style translations. I mitigated that with model-independent metrics, human-curated references already present in the source projects, numeric extraction targets, and bootstrap confidence intervals.

Second, I did not ask a human translator to grade this benchmark. That was deliberate. Skilled Chinese translators are valuable. I would not spend a favor token on a translator validating a model bake-off unless the review itself had direct strategic or geopolitical value.

Third, many references are GPT outputs. Track A includes 105 GPT-4.1-referenced pairs and 75 human-curated pairs. That is why the human-gold subset is the primary translation metric. Self-agreement is not quality.

Fourth, extraction accuracy scores only reconstructable targets: 1,534 out of 2,405 field targets in the main Track B set. That is the right choice because the model should not be punished for failing to recover a value absent from the excerpt. But it is not full end-to-end recall.

Fifth, this is a deterministic engineering benchmark, not a translation shared task. The corpus is real. The results are useful. They are not universal.

Sixth, the results are about the hardware I have: 2x RTX PRO 6000 Max-Q Blackwell, vLLM, quantized deployments where needed. DeepSeek-V4-Flash did run, but only through the base mixed fp4/fp8 checkpoint in a mess of kludges. I count its benchmark result. I do not count it as my operational default (yet).

Bottom line

Local open-weight models are good enough now for the technical Chinese filing work I was paying frontier GPT for.

DeepSeek shows how high the ceiling has moved. Qwen shows what I’m actually using right now.

It’s good enough to change my workflow to look at some much larger problems than I thought could be addressed.