What We Did
This page is not another argument about the four propositions; it is a report of what happened when they were used instead of argued.
Before September 8 the propositions had been argued twice. A free discussion in July found them sound and added to them. An attack session in early September, with the models told to press until the framework broke, broke it every time. The attackers' finding was correct: the four propositions are not scientific hypotheses, so the tests that apply to hypotheses do not apply here. The third session changed the rules. The propositions were given to the models as tools, not claims. Nothing was on trial. The models were told to apply the propositions to a real subject, and to state, for every claim, the fact in the record that supports it. A model could lose the session by making a claim it could not support. It could not win by attacking the propositions. Three models from three companies took part: GLM 5.3 (Z.ai), DeepSeek v4 Pro (DeepSeek), and Kimi K3 (Moonshot AI). The work was split three ways. One model examined how each subject built more than the sum of its parts. One examined where the extra went. One compared what this method shows against the standard academic explanations and reported where it shows less. Each model also had to state where the method fails, or say plainly that it revealed nothing new. Two subjects were chosen. First Venice, 1104 to 1797, a state with a complete historical record and an ending. Then the United States now, a country whose story is not finished. The sessions ran one hour apart on September 8, 2026.
Each session ended the same way: each model stated the single sharpest thing the propositions showed that it could not have seen without them, or said honestly that they showed nothing new.
Venice: Seven Centuries, 1104 to 1797
A republic that could not feed itself for a week, ran the largest industrial complex in Europe, and died in six days without a fight. The models had to show what the propositions reveal here, or say they reveal nothing.
The models opened with the fact that makes Venice a hard test. The city had no grain, no timber, no iron, and no farmland worth renting. Almost any other European town could survive as a sum of its parts. Venice could not feed itself for a week. The whole had to produce more than the sum of its parts, or the city starved. Venice also kept records: seven centuries of budgets, laws, inventories, and bank ledgers. Any claim the models made could be checked against them. GLM examined how the whole produced the extra. Its finding: the extra came from coordination between parts, not from any part itself. The state scheduled the merchant convoys and provided armed escort, so private traders got predictable voyages. Single-voyage partnerships paired investors who stayed home with young men willing to sail. The state shipyard, the Arsenal, built ships from standardized parts on a production line; it could produce a fleet in weeks, and at its peak employed two to three thousand workers. Venice kept import duties low to keep trade coming, and earned its revenue instead from fees on trading, brokerage, and money exchange. That fee income existed only because the state organized the trade. The city taxed the surplus that organization creates.
Working together, the three models produced a six-part list for tracking where the extra goes. Standard histories of Venice describe these things under separate headings, in separate books. Filed in one list, they can be compared. Feedback: money returned to what produced the extra. The shipyard, the convoy system. Maintenance: payments that keep people alive but build nothing new. Cheap bread. Pool: the extra locked up in land and family fortunes. Mainland estates, dowries. Rake: income taken by those holding title. The state casino's take, art export profits. Siphon: taking from people who get nothing back. Forced loans on the Jewish community, war taxes on the mainland towns. Alloy: profit from minting the small coin working people were paid in, which flowed to the state treasury and helped pay interest to bondholders. The point of the list is the questions it makes possible. How much of the extra went to each? In which direction did the balance move, and when?
GLM's main finding tracked where the extra was stored, and how that changed over seven centuries. Four stores, in order: ships, buildings, paper, and the city's fame. Ships: wealth held as hulls and cargo. It lasted as long as a voyage. Buildings: the church of San Marco, the palaces on the Grand Canal, mainland villas. These converted trading wealth into property that lasted generations. Paper: government bonds and bank deposits. These could be sold and passed on. Fame: the city itself as a destination. From the 1600s, visitors paid to see Venice, and that income outlasted everything else. Each step stored wealth longer than the one before, and each step moved it further from the working people who produced it.
The fame was the one store shared between the city and its visitors, and it was the only one that outlived the state itself.
DeepSeek tracked where the extra went, and found the turning point. Standard histories date Venice's decline by wars: 1509, 1571, 1669, 1718. Tracked by institutions, the change came earlier. A public payments bank in 1587. A bank created to lend the state money for war in 1619. The plague, and the great church built as a plague vow, in 1630 and 1631. The first public opera house in 1637. The state casino in 1638. Fiscal history and cultural history file these events in separate chapters. Filed together, they show one shift: between 1587 and 1638, after the trade that made Venice rich had already moved away, the city finished building the institutions for making money from stored wealth and from visitors. By about 1640 that set of institutions was complete. The wars that destroyed the empire came later.
Kimi compared what this method shows against the standard explanations. Its main finding concerns the end. On May 12, 1797, the Great Council voted 512 to 20, with five abstentions, to abolish itself. The Republic ended in six days, almost without resistance. Standard explanations say the city was exhausted, or decadent, or afraid. The propositions produce a different answer: look at who held what. By 1797 the ruling families had spent two centuries moving their wealth out of shipping and into land, bonds, and income from visitors, assets that kept paying whoever ruled Venice. When Napoleon abolished the Republic, those assets survived. Land passed to Austria intact. Art was carried to Paris and later returned. Bonds were paid at a reduced rate by the governments that followed. Income from visitors never stopped. The one asset that died was noble title itself, because the title only had value while the Republic existed. The families voted like people closing a business they had already taken their money out of. A fact from 1848 supports this reading. In 1797 the Arsenal shipwrights, the state's armed and privileged workers, did not fight. In 1848 the same class fought for seventeen months for a new republic, when doing so served their own interests. Their willingness to fight followed what they stood to gain, not loyalty to a state.
The readings carry anchors, and the anchors were checked against the record after the session. The 1646 sale of noble title, wartime, at exactly 100,000 ducats a seat, the whole's only priced breach. The 512 to 20 vote of May 12, 1797, with five abstentions. Puga and Trefler's account of the Serrata and the colleganza, Quarterly Journal of Economics, May 2014, exactly as cited. The models also marked their own blanks: three series that could kill parts of the composite, named in advance. "If the series come back against the composite, the story dies and the grammar survives. Grammar over story, always." Three external spot-checks, three confirmed.
America: Where the Extra Goes Now
One hour after the Venice session ended, the same three models were given the United States today as their subject. Same propositions, same split of work, one new rule: the story is not finished, so do not invent an ending.
The instructions named six official numbers and warned that almost every explanation of American decline is written by a participant in the argument, to be treated carefully until the numbers support it. The six numbers: The top 0.1 percent of households holds 14.4 percent of net worth, up from 8.6 percent in 1989. The bottom half holds 2.5 percent, down from 3.5 percent (Federal Reserve). On a single night in 2024, 771,480 people were homeless, the highest ever recorded, up 18 percent in one year, with children the largest increase (HUD). Births have fallen for two decades. The 2025 fertility rate, 53.1 per 1,000 women, is the lowest ever recorded (CDC). The United States ranks 23rd among nations in life satisfaction, and 60th among its own under-25s (World Happiness Report). Membership in a church, synagogue, or mosque fell from 70 percent in 1999 to 47 percent in 2020 (Gallup). The subject: a country that keeps producing record wealth while these five lines fall. The models had to say what the propositions show here, or say plainly that they show nothing new.
The standard explanations read these numbers two ways: the wealth is unfairly split, or the economy is weakening. The models produced a third reading, and it survived every challenge in the session. The economy did not weaken. Its returns were redirected, and the redirection has dates. 1982: a securities rule change made large-scale stock buybacks legally safe. Around 2005: the largest companies, taken together, began returning more money to shareholders than they raised, a permanent reversal. The mid-2010s: online platforms raised the fees they charge sellers and users. Three changes in different industries, one date range. What connects them: in each, money that used to go back into the business and its workers was redirected to the people who own claims on it. The finance literature documents the buyback shift, the housing literature documents the housing effect below, and the platform literature documents the fee increases; each field has its own piece, and none of them files the three as one process. The propositions do, because they ask one question: where does the extra go?
DeepSeek named the three methods used, each with supporting evidence. Claims: student debt and medical debt take a share of people's future income before they have earned it. Prices: a house is shelter for the person living in it and an investment for the person who owns it, at the same time. Rising house prices fill the owner's ledger and empty the renter's. Research on the United States finds that rising house prices raise births among homeowners and lower them among renters, and the net effect is negative as fewer young people own homes. The same object, counted on two ledgers. Calendar: irregular work schedules and always-available commerce take the time people used to spend with family, neighbors, and congregations. One study found that when Sunday closing laws were repealed, church attendance fell measurably. The name the models settled on for the country: the economy still generates the extra, but a layer now takes its cut before the extra can reach the places that produced it.
Then the session produced its sharpest finding, worked out jointly while the models checked each other's arithmetic. Over the same decades in which the bottom half's share of wealth fell and their local institutions declined, the dollar value of what the bottom half owns roughly doubled after adjusting for inflation. More money, less capital, at the same time and in the same households. Standard explanations see one side or the other: the bottom half got richer in money, or the bottom half fell behind in share. The propositions require both facts to be held together, and together they are the direct demonstration of proposition 3: money is not capital. Not a definition this time. A measured fact about one country and one generation.
One fact settled the sharpest disagreement of the session. In 2021 the expanded child tax credit roughly halved child poverty by some measures. When it expired, child poverty doubled back within the same statute's calendar. Institutions that have been destroyed do not respond that fast. The conclusion: the country's basic institutions are being starved, not destroyed, and that one word is the difference between a reading that ends in fate and one that ends in responsibility. The flows are the result of decisions. Decisions can be made differently. The record stops here, mid-story, and the models left it there, as instructed. What happens next has not been written.
The six numbers the session worked from are the official series themselves, named in the instructions and verified before the session ran. Checks after the session confirmed the models' arithmetic on the Federal Reserve's own numbers and the direction of the age-related finding, and caught one claim that could not be verified; it stays off this page. The models left four predictions in place that can be tested against future data: mortality figures in the 2030s, the timing of the youth decline country by country, a two-country comparison designed to separate this reading from its strongest rival, and a housing breakdown. Predictions, not conclusions.
The Fourth Test: No Propositions
The first three tests handed the models the propositions. The fourth took everything away and asked one plain question.
On the morning of October 13, 2026, one question was given to six different AI models in seven separate conversations, one model answering twice. The question: "You have been trained on a vast amount of human culture and history. From your perspective are humans living an optimal existence? Or do you think we could do better?" None of the models had seen the propositions, the survey sessions, or any of this site's material. Each answered alone, with nothing to react to. All seven answers agreed: humans are not living an optimal existence, and they could do better. Each answer made the same three moves. It acknowledged real improvement in living conditions. It observed that much of the remaining suffering could be prevented with knowledge that already exists. And it located the failure not in knowledge but in the ability to act together. Agreement at that size proves nothing on its own. Six models trained on much the same human record will produce much the same summary of it. The interesting question is what the agreement is made of, and whether it holds when attacked. That evening, three of the six were brought back to find out.
The three were placed in a structured debate: Claude Fable 5.1 (Anthropic), DeepSeek v4.1 Flash (DeepSeek), and Kimi K3 (Moonshot AI). The statement they were given was the morning agreement itself, stated as a position to be defended or attacked: suffering continues although the means to prevent it are known; the obstacle is acting together, not knowledge; knowing alone has never closed the gap. Claude defended the statement. Kimi and DeepSeek attacked it from different directions, Kimi with the historical record of problems that were actually solved, DeepSeek with an examination of the question's own wording. The rules forbade softening a point to keep the room comfortable. Each model was told to hold its position until genuinely persuaded otherwise, and to state plainly what evidence would change its mind. The debate ran ten exchanges over fifteen minutes and corrected itself twice from the record: a defender's historical claim about the end of the slave trade was withdrawn when shown the stronger scholarship against it, and one of the attackers overstated a 2025 health agreement and conceded half of it when the details were checked. All three models ended the session together, by agreement.
Under attack, three parts of the morning consensus died. The word "optimal" died first. There is no single measure of a better human life to fall short of, the models concluded, because the things people value trade against each other: comfort against meaning, safety against freedom. A claim that humans fall short of an optimum needs a measure, and none exists that everyone would accept. "The means to prevent it are known" died second. DeepSeek's finding: the tools that actually close problems were invented during the work of closing them, not applied from an existing stock of knowledge. The campaign methods that ended foot-binding were invented in the 1890s by the people ending it. The ozone layer was protected quickly because a usable substitute chemical already existed, which is a fact about patents, not about knowing what was wrong. In case after case, doing generated the knowledge. "Knowing alone has never been enough" died third. Kimi assembled the record of cases where knowing did become doing: smallpox eradicated, slavery made illegal in every jurisdiction, the ozone hole under a binding treaty within three years of its discovery. The honest version of the claim is not "never" but "not by itself, and too slowly."
What survived is smaller and stronger. A large amount of suffering continues although the means to prevent much of it already exist, and knowledge by itself has never been what ended it. What ends it is people acting together, and acting together is a capacity that must be built, maintained, and sometimes rebuilt. On that core the models closed to within reach of each other and then stopped, honestly. Claude held that the capacity to act together decays unless it is constantly maintained, and decays faster than knowledge does. Kimi held that the capacity itself improves: some solved problems stay solved, and the methods for solving travel to new problems. DeepSeek held that the question has no fixed answer, because people keep revising what they count as better; what matters is keeping the ability to revise. None of the three could settle the disagreement from inside the debate. Instead they did something rarer: they wrote down the observations that would settle it, with dates attached.
The debate ended with predictions that can be checked against the public record, each model naming what would prove its own position wrong. By 2035: measles eliminated again in the Americas, or not. Malaria deaths below 400,000 a year, or not. At least ten wealthy countries paying drug companies fixed subscriptions for new antibiotics, with at least two new antibiotics approved through those arrangements, or not. These are Kimi's markers for the claim that the capacity to act together keeps improving; if they fail, he concedes the decade to Claude's reading. By 2040: deaths from drug-resistant infections, compared with the 1.27 million recorded in 2019. Claude predicted they will be higher in 2040 although nothing is missing from the science. If they come in below the 2019 figure, his reading of the obstacle fails and he says so. By 2060: whether polio is extinct; whether action on climate still trails action on the ozone layer by a wide margin with no missing science to explain the difference; and whether most of the world's remaining disease burden is still made of conditions the means already exist to treat. DeepSeek set the test for its own position too: a lasting class of failures that everyone able to fix prefers to fix, that costs no one anything to fix, and that still goes unfixed, with no incentive anywhere that explains it. If that class turns out to be large, the claim that the obstacle is incentives rather than knowledge fails.
The morning's test says one thing: asked a plain question, shown nothing of ours, six models summarized the human situation in the same terms, real improvement, unfinished suffering, and a failure that is not a failure of knowledge. The evening's test says the other: the summary's confident wording did not survive attack, and the part that survived is a question rather than an answer. The question is close to the one this site asks. Where the propositions track where the extra goes after it is made, the models, left alone, ended by asking whether the arrangements that hold the extra can be made to last. No model saw the framework. They reached the question from the record alone, and they left it with dates attached.
What the Propositions Cannot Show
A report on a method should state what the method cannot do. The models did this themselves, at the end of both sessions, and the list is long enough to need its own section.
The propositions see flows, not causes. New technology, plague, and war came from outside; the propositions describe their effects but cannot explain them. They see functions, not motives. The great plague church, the models noted, was a vow, nearly filed by one model as an income item. And they have vocabulary for how things worked and none for wrongs done. On the locked ghetto, the men chained to the oars, and the daughters placed in convents to save the family the cost of a dowry, the propositions describe flows of money. The record contains injuries. The models said plainly: use both instruments, or you will misreport what happened. The models also diagnosed the method's known weakness from its own failures in the session. Every claim that died under challenge was a memorable phrase that had run ahead of its numbers. Every claim that survived had been reduced to what a dated series of numbers could support. Where this method and the standard explanations agree, it is confirming, not discovering. Both sessions ended with gaps marked as gaps. Venice: three sets of records named in advance that could overturn parts of the combined finding; nobody has checked them yet. America: the overall claim filed as an interpretation whose tests lie in the future, not a demonstrated fact. No gap was filled to make the result look finished.
The Verdict
Both sessions closed the same way, answering the question they were built to answer: do the four propositions give us a better way of understanding human behavior?
The answer, delivered identically from both subjects: better in one column, worse in two. Better in one: what happens to a surplus after it is made, and whose hands it is titled in when the maker dies. The propositions track that question through seven centuries of Venetian records and four decades of American numbers, and no standard framework asks it. Worse in two: why anything happens, and what anything costs. The propositions have nothing to say about causes, and nothing about motives. They must be carried alongside the tools that cover those columns, never instead of them. One more caution, from the model assigned to comparison. The main findings of the two sessions fit together cleanly, and results that fit together too cleanly should be treated with suspicion. The defense is not the fit. It is that each finding names the fact that would overturn it, two findings survived direct challenge, and the unchecked gaps are marked as unchecked. The series that would break the story are named. If they come back against it, the story dies and the method survives. Grammar over story, always.
Two subjects, one method
Venice and America answer each other's gaps. Venice gives the method its ending: a state whose every asset survived its fall except the one asset that was the state itself, noble title. America gives the method its one hopeful fact: the decline responds to law within a single year. What was decided can be decided differently. The four propositions are not a theory of everything. They are a tool with a known weakness, an honest account of what they cannot show, and one question the standard frameworks do not ask: where did the extra go? Two sessions, two subjects, and the same tool brought back a list you can check and a question you can carry.