Organizations restructure for ordinary reasons — a merger, a funding round, a founder's exit, a board that wants cleaner reporting lines. It is rare for one to restructure around a claim about knowledge.
This one did.
Building the engine had taught its builders something they did not set out to learn: that rules, data, and prediction are separable concerns. For most of a decade the three had lived together inside one tool — one set of repositories, one brand, one habit of mind — and the joints between them stayed invisible because the bundle worked and nothing ever pulled on them. Verification pulled on them.
Once you require each part of a policy estimate to prove itself against ground truth, the parts stop resembling one another. A rule proves itself against the statute it encodes and against a reference calculator — TAXSIM, UKMOD, a SOUTHMOD model — that has read the same statute and either reproduces the encoding's answer to the last cent or forces the difference to be explained. It does this now, at the moment the rule is written, before it touches a single household. A population estimate proves itself differently: against surveys, and against the administrative totals the government already publishes — a calibration you can check the day you build it. A prediction can prove itself in neither way, because the thing it describes has not happened yet. It waits. It is scored only when reality arrives and the official number lands.
Three kinds of claim, then, on three different clocks. What is true of a rule is settled today, against a document. What is true of a population is settled today, against a count. What is true of a forecast is not settled until some future Thursday when a statistical agency posts a release. And the clocks turn out to imply the failure modes. A rule fails by contradicting the law — a bug against a fixed target, findable and fixable. A calibration fails quietly, by drifting away from a population it used to resemble. A forecast fails loudly and unarguably, by being wrong about the world — a miss you cannot litigate away and can only record. You do not manage these three failures with the same reflexes, or the same people, or — this is the chapter's claim — inside the same institution.
Take a single tax credit. Whether it has been encoded correctly is answerable this afternoon, against the section of the code that defines it and an oracle that computes it independently. Whether the households it reaches look like the country is answerable this afternoon too, against totals the tax agency already reports. But how many families will actually claim it next filing season is answerable only next filing season, when the agency counts them. The first two are audits. The third is a bet.
Axiom states, Thesis predicts
The reshaping that followed took its cue from that observation rather than from the forces that usually move an org chart. PolicyEngine divided along the seam that verification had exposed, into two poles that are opposite in the most elementary way two claims can be.
An axiom is an accepted truth. It is not on trial; it is the thing you reason from. The question what does the law say? has exactly that shape — the answer was fixed by a statute someone already wrote, and the only live question is whether an encoding faithfully matches it. A thesis is the mirror image: a proposition advanced in order to be tested, granted no standing until the evidence rules on it. The question what will happen? has that shape instead — no document settles it, and any answer stays provisional until the world supplies the verdict. Two words, two epistemic tempers, and the entire reorganization is an attempt to stop housing them in one body.
The Axiom Foundation is the determinism pole .[1] It encodes law exactly — federal statutes, state codes, agency manuals, the council-tax reduction scheme of a single London borough — and it asks its agents to state, never to guess. Its public launch is set for July 28, 2026. Everything it ships is meant to be right the way a proof is right: checkable line by line against a source, and wrong only if the source is.
The Thesis Institute is the prediction pole .[2] Its product is a public docket of forecasts about government metrics — the first print of the unemployment rate, a program's caseload, a wait time — including forecasts made under explicit policy counterfactuals. One live page asks what Medicaid call-center wait times will be in March 2027 if the work-requirement deadline is delayed: a number that does not exist yet, about a world that has not happened yet, staked out in advance. Its public launch is set for around September 15, 2026. Everything it ships is meant to be provisional — a value with an interval around it, entered into the record early enough that reality can grade it.
Opposite as they are, the two poles are built to compose. The Medicaid question is really two questions stacked. What the rule would be if the deadline slipped is an axiom — encoded exactly, by the determinism pole. What the caseload and the wait times would then do is a thesis — forecast, and graded, by the prediction pole. A policy counterfactual is precisely that: an encoded change to the law, run forward against a population, scored against the world that results. The determinism pole supplies the exact if; the prediction pole supplies the uncertain then. Separating them into two institutions does not cut the pipeline between them. It makes plain which half of every counterfactual is a matter of settled fact and which half is a matter of open forecast.
Here the discipline this book has been pressing becomes a literal feature of the product. Every forecast on the docket carries a published interval and will be graded when the official number lands. As of this writing, none has resolved. The scoreboard is live; the grades arrive with reality.
The engine, recomposed
Where does the original tool go, in a world divided between an axiom pole and a thesis pole? It goes to both, because it was always both, bundled together tightly enough that no one had to notice. For years one team shipped one release, and the rules, the data, and the projections rode out together; nothing in the daily work ever forced the question of which of them was a proof and which was a guess. The oracles and the scoreboard forced it, each part suddenly made to answer for itself. PolicyEngine is being recomposed into the two layers that were underneath it the whole time. Its rules — the thousands of encoded provisions that answer what a household owes and is owed — become part of Axiom. The calibrated microdata beneath every population estimate becomes populace, the shared substrate both poles stand on. The model code that used to live in PolicyEngine's own repositories retires downward into those two layers, and the product on top keeps its name, its interface, and its users.
That last point is the one to hold onto, because it is where the recomposition stops being an internal diagram and starts mattering to anyone outside. The teacher checking a family's benefit cliff, the journalist scoring a bill, the congressional staffer running a distributional table — none of them sees a seam. The thing they use answers the same questions it always did. What changed is underneath: it now stands on two legs that can each be checked, a rules layer verified against statute and a data layer verified against administrative totals, rather than on one monolith whose internals you had to trust as a whole.
Populace deserves one plain sentence and no more fanfare than that. It is the calibrated-microdata commons — households synthesized and reweighted until the totals they imply line up with the counts the government already publishes — and it is the ground the entire stack rests on, because an exactly encoded rule and an honestly scored forecast both still have to be applied to a representative population before they mean anything about the country. The methods matter; the brand does not. Name it once, and lean on the work.
What the recomposition actually buys is not neatness. It is accountability, of a kind the field has never really had. Route the population layer into a forecast, post that forecast on the Thesis docket with an interval and a resolution date, and a policy estimate stops being a number a model asserts and becomes a claim the world will eventually grade. "This reform costs fifty billion dollars" turns into "this reform costs fifty billion dollars, here is the interval, and here is the date we will find out." That closes a loop the institutions producing official estimates leave open. The Congressional Budget Office publishes a score and moves on. The Tax Policy Center publishes a distributional table and moves on. EUROMOD publishes and moves on. Nothing circles back to keep the tally. The recomposition wires the tally in — not as a report card imposed from outside, but as the native output format of the prediction pole.
Wrongness, isolated
Splitting the poles into separate institutions is not brand hygiene. It is a firewall around one specific hazard: being wrong.
The Thesis Institute has to be wrong in public, and often. A docket that never missed would be a docket with its intervals padded wide enough to be useless, or one quietly declining to post the questions that are actually hard. The worth of a graded forecast comes precisely from the fact that some of the grades will be bad — visibly, on the record, with the interval printed beside the miss — because that is the only thing that makes the good grades worth anything. Public wrongness is not a risk the prediction pole runs despite its mission; it is the mission. A forecaster who cannot be caught out is not being honest. He is being vague.
The determinism pole cannot live that way for a moment. A calculator that a benefits screener consults, that a state agency wires into its workflow, that a tax product ships to customers, has exactly one job — to be right about what the law says — and its cardinal virtue is that it is boring. Same input, same answer, and the answer matches the statute every time. Housed alongside the forecasting work, that calculator would inherit a reputation it has no business carrying. Every published miss on the scoreboard would raise the wrong question about it: is the engine that just under-forecast a caseload the same engine a caseworker leans on to check whether this family qualifies for SNAP? Technically the two share code and a population. Epistemically they make opposite kinds of claim, and the surest way to keep opposite claims from contaminating each other is to keep them under different roofs, with different promises attached to their names. One roof promises to be interesting and sometimes wrong. The other promises to be dull and always faithful. Wrongness isolation is a property of the design, not a vanity of the org chart.
A commons and its operators
There is a second reason to make the structure explicit, and it predates the two-pole split. An earlier chapter argued that a layer this fundamental — the encoding of the law itself, the calibrated portrait of the population — is not well served by a private gatekeeper; it is closer to a dictionary or a map projection than to a product, and it should be open, forkable, and free to audit for the same reasons those are. That principle survives intact. The 2026 structure gives it a sharper, two-sided form.
On one side stand the nonprofit institutes, stewarding the commons: the rules, the data, the scoreboard, all public. On the other stand commercial operators, building on that commons the things production actually demands — managed runtimes, uptime guarantees, support contracts, applied products for the banks and agencies and firms that need someone to answer the pager at two in the morning. The distinction the whole structure turns on is a single preposition. Operators build on the commons. They never own it. The foundation states the commitment plainly: the commercial service tier will always be a separate entity, so that the institution stewarding law-as-code never finds itself competing with the builders who depend on it. A hospital system or a state agency cannot run its eligibility screening on a best-effort open endpoint; it needs a contract, a service level, someone accountable when the pager goes off — and providing that is honest work, worth paying for. The structure's only insistence is that whoever provides it be constitutionally unable to fence off the commons underneath. A steward that also sells is a steward with a thumb on the scale.
One such operator already exists, in formation, organized around graded simulation offered as a service. It is not yet public, and — this is the point — nothing about the commons depends on it. The rules and the data would stay open whether or not any company ever billed for a runtime built on them. That is the test of whether a commons is real rather than rhetorical: it has to survive its operators, outlast any one of them, and owe none of them anything.
The pattern underneath
Pull back from the org chart and a repeatable pattern comes into focus, one with nothing intrinsically to do with taxes or benefits. Every project here has the same three-part shape. Find a corpus that is complete, public, and — the surprising part — un-computed, sitting in plain sight in a form no one has turned into anything queryable. Point agents at it. Ship the verified, open version of what was there all along. Axiom encodes the world's policy corpus. Populace integrates the world's microdata. Thesis forecasts the government's metrics. Three instances of a single move: the analytic stack of government, rebuilt in the open, one verifiable layer at a time.
The move comes with a gate, and the gate is the whole discipline of this book compressed into a test. A corpus qualifies only if its output can be checked, per unit, against ground truth — a rule against its statute and an oracle, a household against administrative totals, a forecast against the number reality eventually prints. Where that per-unit check exists, agents can work at a scale no institution could ever staff, precisely because each of their millions of outputs can be caught the moment it goes wrong; the check is what converts raw generation into trustworthy production. Where the check does not exist, the same agents at the same scale produce the exact opposite of knowledge — fluent, voluminous, and confidently false. That is the line separating encoding the law from generating plausible law-shaped prose, and it does not move for enthusiasm or for compute. It is the difference between a corpus you can verify and one you can only admire.
For most of history the corpus of public life sat un-computed not because it was hidden but because computing it did not pay. Encoding a single statute faithfully took a trained specialist days, and there are millions of provisions spread across thousands of jurisdictions; forecasting the budget of every county took analysts no one was ever going to fund. So the law stayed on paper, the microdata stayed in survey files, and the metrics stayed unforecast — public in principle, computed almost nowhere. What changed was the unit cost. An agent can encode a country's tax-benefit system in days rather than years, as the five-country week showed; it can hold a corpus of tens of thousands of provisions in view at once; it can draft a forecast for every metric on a docket and never tire. The corpus was always open. The missing piece was a way to compute it that is at once cheap enough to attempt at scale and disciplined enough to trust — and the gate is what supplies the second half.
The gate admits more than tax and benefits, which is where this goes next. Local zoning codes are public, largely un-encoded, and checkable line by line against the text of the ordinance — a national map of what may legally be built where, assembled the way the tax code was. The fiscal futures of tens of thousands of county and municipal governments are forecastable and, in time, gradable against the budgets those governments actually adopt. Bills introduced in a legislature could be scored the day they are dropped, run through the same engine that scores enacted law, so that a proposal arrives with its distributional consequences already attached. None of these is a promise; each is a direction, and each earns its place only if it clears the same gate — per-unit verifiability against something real. But the list runs long, because most of the corpus of public life shares the one property that matters here: it is already written down, and no one has computed it yet.
The third question
Two of the three questions now have homes. What does the law say belongs to Axiom; what will happen belongs to Thesis. Between them they cover a remarkable amount of what a policy argument is actually made of — the facts of the statute and the facts of the forecast. But they do not cover the part that decides the argument once the numbers are agreed. The oldest question is still open: what do we want?
That one has no institute, and it may be the hardest of the three to anchor in ground truth, because its truth is not a statute or a first-print statistic. It is something latent in a population's own preferences — harder to observe, easier to distort, and quick to punish anyone who claims to speak for it. Yet the embryo is already here: a survey-simulation tool that tries to elicit and forecast values the way the other layers elicit rules and metrics, and that the next chapters take up in earnest. Rules, predictions, values — the third primitive is at once the frontier of the whole program and its least finished piece, the place where a book about computing the government has to become a book about what the government is for. There is a version of this stack whose deepest payoff is not any single number but a shared, verifiable substrate everyone can reason against, so that public disagreement can finally be about ends rather than about whose figures are right.
That is the horizon. It is not, however, the next step. Before values, the book owes a debt it has been quietly running up — a question every estimate in every preceding chapter has managed to dodge. Each of those numbers arrived without an honest account of its own uncertainty. This reform costs fifty billion dollars: plus or minus what, and how would we know? The prediction pole exists to answer exactly that — to turn a bare point estimate into a graded interval with a date on it. And so Part IV begins not with what we want, but with how sure we can be of what we already claim to know.