A country that can see

§00 The vote is not the story

When Congress passes a major law, the coverage treats it as an ending. The vote happens, one side wins, the cameras move on. But almost nothing a law is supposed to do happens at the moment of passage.

A bill is a bundle of promises: orders to agencies with deadlines attached, money authorized but not yet moved, prohibitions that mean nothing until someone enforces them, tripwires set to fire years in the future. A bill is less like a verdict and more like a flight plan — filed publicly, in enormous detail, and then flown. Or not flown.

And here is the strange part: nobody is watching the flight.

I mean that almost literally. We built a national weather service because the atmosphere affects everyone. We keep box scores for games. Insurers have maintained actuarial tables for three centuries. For the machinery that governs the daily life of 330 million people, there is no equivalent.

No institution continuously tracks whether the rules a law ordered actually got written, whether the money it promised actually moved, whether the thing Congress said would happen happened. Agencies quietly drop congressional mandates and nobody notices. Statutory deadlines pass in silence.

The one feedback signal a legislator reliably receives is an election — which is a very noisy way to learn whether a statute worked.

I’m building a company called Downstream to close that gap: a permanent, public, checkable record of what laws actually do after they pass, and on top of it, forecasts of what they will do next. This essay is the case for why that’s now possible, what becomes different once it exists, and why I think it is one of the larger quiet opportunities available right now — commercially and civically at once.

Two honest concessions before the optimism, so you can calibrate the rest. First: every piece of data I will mention is public — every bill, rule, court docket, and dollar — and public data, by itself, is not a defensible business.

Second: a frontier AI model pointed at that public data can already do maybe 70% of what I am about to describe. The entire bet is that the remaining 30% — the discipline, the verification, the accumulated track record — is where all the value lives. I’ll try to show you why I believe that. But it is a bet, and you should read this knowing the author holds a position.

§01 The insight: read the clauses, not the bill

The reason nobody has built this is not laziness. It’s that everyone analyzes the wrong object. Ask “what will this bill do?” and you are stuck — a thousand-page omnibus does ten thousand different things, and the question dissolves into punditry. The unlock is to stop treating the bill as the unit of analysis and drop one level down: the atomic unit of a law is the individual provision, classified by what kind of thing it does.

Because provisions come in types, and each type leaves a characteristic trail in the public record:

• A delegation — “the Secretary shall issue regulations within 180 days” — either produces a proposed rule in the Federal Register or it doesn’t. There’s a database for that.
• An authorization of money either converts into actual appropriated, spent dollars or it doesn’t. The government’s own spending records show which.
• A clause creating a right to sue produces lawsuits you can count in court dockets.
• A sunset clause fires on a date printed in the statute itself.
• A formula moves money mechanically when published economic statistics move.

This turns an unanswerable question into a measurable one. “Will this law have good effects?” can’t be answered by anyone, and you should distrust any system that claims otherwise. But “when Congress orders an agency to write rules with a 180-day deadline, how often do the rules appear, and how late?” — that’s an empirical question, and roughly thirty years of public records exist to answer it.

Ask a sharper version. How often does a rule-writing order produce no final rule at all, ever? Nobody publishes that number. The records to compute it sit in the open; what’s missing is an institution assigned to notice, and that absence is itself the finding.

Classify every provision of every bill by mechanism type, fit the historical patterns for each type, and you have something like an actuarial table for legislation — not a pundit’s opinion but a base rate with a sample size attached.

One more property of legal text makes this richer than it first sounds. Law is densely cross-referential, and the references are extractable by machine. When a bill quietly amends the definition of a single term — “covered entity,” say — it silently changes the meaning of every other provision anywhere in federal law that borrows that definition: sometimes fourteen sections scattered across six different areas of the U.S. Code, some carrying their own rule-writing obligations, some currently being litigated.

Tracing that ripple is not a prediction; it is a computation, citable line by line to the official text. The first time you watch it run on a real bill, the reaction isn’t “how clever.” It’s “why doesn’t this already exist?”

§02 Why it's possible now

That question has an answer. Four things changed — three of them recently, one of them uncomfortably.

The courts made regulatory survival a live question. For forty years, a legal doctrine called Chevron deference meant that once an agency finalized a rule, courts usually let it stand. In 2024 the Supreme Court ended that doctrine, and a companion decision reset the clock for challenging even decades-old rules. Whether a regulation survives is now genuinely probabilistic — a real, tradable, plannable uncertainty. The demand to measure it was created about two years ago, and essentially nothing has been built to meet it.

The cost of reading everything collapsed. Analyzing a Congress means processing roughly fifteen thousand bills clause by clause. In 2022, doing that with structured, verifiable output was economically absurd. Today, with modern AI doing the extraction — and humans verifying it against labeled samples — it is a rounding error. The corpus was never large; the analysis was the bottleneck, and the bottleneck is gone. This is the quiet reason the field is about to change.

The government’s own data infrastructure matured. Congress publishes a real API now. The Code of Federal Regulations serves historical point-in-time snapshots. Federal spending, rulemaking dockets, and court records are all machine-readable and free. Two years ago half of this was scraping; today it is engineering.

And the record itself turned out to be optional. This one we measured ourselves. A presidential memorandum carries essentially the same legal force as an executive order, but no number and no publication requirement. The Federal Register prints one only when the President orders it printed. Of the thirty memoranda whitehouse.gov published in 2026, twenty-one have a Federal Register counterpart, seven are confirmed absent, and two are too recent to call (measured against the archive’s own captures, August 19, 2026). The seven are not late; matched pairs show up within days, and these are months old. They are not coming.

And because memoranda are unnumbered, there is no gap in a sequence for anyone to notice. Each one is a live directive to an agency. It can set rule-writing and money in motion with no statute anywhere upstream, entering through a door nobody watches.

Nor is this one instrument’s quirk. The House republishes its own floor schedule mid-week, twenty-six times in one week we sampled; a document withdrawn from the Federal Register’s public-inspection list before publication leaves no public trace it ever existed; federal web pages and guidance documents have been deleted and altered in the past two years, in one case restored only because a court ordered it.

Very little of this is misbehavior. Most of it is discretion operating exactly as designed. But it means the public record is not something you can look up later. It is something you either captured, dated, and hashed on the day, or lost. So the archive runs from day one, every day, regardless of everything else in the plan.

§03 What the map alone changes

Downstream is two layers. The first is the map: every provision classified, every connection — from statute to rule to court case to dollar — cited to a primary public document, every query answerable as of any past date. The map is open and free, permanently. Before any forecasting at all, consider what just the map changes.

One note on the map’s shape, because it explains several choices below. Consequence does not run down a single pipe from Congress to the citizen. An action can enter the record at any point on the line: a bill scheduled for the floor next week, a statute, a presidential directive with no statute upstream of it, a proposed rule, a court order vacating one.

And the loops run backward as well as forward. A veto returns a bill, a vacatur returns a rule to the agency, oversight returns a finding to Congress. Three branches share one timeline, and every hop between them is a place where a consequence gets created, delayed, or dropped.

We keep a working diagram of that structure, the chain of command, and the map is built to watch all of it: what the House says is coming, what the executive published today, what the President directed yesterday. One deliberate deferral: the judicial branch’s own corpus, the dockets themselves, is a larger ocean than the other two combined, so for now the courts appear in the map as outcomes rather than as a fully ingested source. State law is a dimension beyond that. Start where the payoff is, and be plain about the edges.

A staffer stops reading in the dark. A twenty-six-year-old congressional aide handed a thousand-page bill the night before a vote can ask what one buried definitional change actually touches — and get the full ripple, every affected program and pending lawsuit, with citations, in minutes. Today that analysis is done by hand, partially, by whoever can afford the lawyers. The gap between who can and cannot afford to know what a bill does is one of the quietest forms of power in Washington, and it is an artifact of processing costs that no longer exist.

A state finds out whether a law is actually about it. This is the question almost everyone downstream of Washington, D.C. is really asking. A legislative staffer in Olympia, a state party drafting talking points, a homebuilders’ association, a tenants’ union, a lobbyist with a client in Spokane — none of them wants a summary of a housing act. They want to know two things: does this reach us, and what did our delegation do about it. Both are answerable without forecasting anything, and neither is currently answered anywhere.

The first one has a structure, and the structure is the whole trick. A provision can reach a state in six distinguishable ways, and only two of them are things the act does on purpose. It can name the state. It can set a criterion resolved by a published federal list the state is on.

Below that line the reach is real but incidental: a formula whose inputs move, a threshold that binds whoever crosses it wherever they act, a rule preempting a category of state law, or nothing state-specific at all. Say which one applies and you have a fact. Skip that step and you have a topical guess wearing a fact’s clothes — which is exactly what you get from a language model asked which provisions affect Washington, and it reads identically.

The classification is a gate, though, not a headline. It decides whether a card may exist and which shelf it goes on; what the card then says has to be what changed, for whom, and by when. Here is the housing act’s institutional-investor rule, which mentions no state at all and lands on a specific set of people in this one:

Reach classes are not degrees of importance, either. The most consequential thing in this act for Washington may be a clause that reaches it only because Washington happens to have a state law of the kind the clause preempts — which makes the answer unavailable in Washington, D.C. and unresolved anywhere:

The second question — what did our delegation do — is the one that gets published wrong most often, because the honest version is narrower than the question implies. Nobody votes for a provision. They vote on an act containing twelve titles, and “Senator X voted for the 350-home rule” is false in the way that is hardest to retract.

So the record shows the strongest attribution that genuinely exists — authorship of the bill a section came from, a committee vote, an amendment vote, final passage, or a procedural vote that is not a position on anything — and labels which of those it is. Four vote states, never three, because not-voting is not a no. And a fifth that is not a vote at all: we have not read that roll call yet, which has to look different from they were absent, or the first quietly becomes the second.

That is the state of the Washington delegation record as of this revision: the authorship tier now fills nightly from the official record, while the member-level roll calls remain unread. The vote tiers stay empty and say so. Publishing the tier structure half-filled, with the empty tiers labeled, is honest. Publishing a card that dresses a national fact in a state’s name would not be, and it is the easier mistake by a wide margin — the same card renders fifty-six times with only the label changing, and every one of them looks like local reporting.

A journalist gets a scorecard instead of a vibe. A year after a major law passes, “is it working?” is currently answered by opinion. The map answers a narrower question exactly: this law created this many rule-writing obligations, across these agencies, with these deadlines — and here is each one, met, missed, or silent. Silent is the interesting one.

An agency taking no observable action on a congressional order is not missing data. It is a finding.

Today it is a finding nobody is assigned to notice. All the ingredients of this scorecard exist in separate government databases. As of August 2026, no one joins them. The closest thing in existence is one health-policy nonprofit hand-maintaining a tracker for one single statute.

An analyst gets a calendar instead of a headline. Markets reprice a sector on the day a bill passes — then stop paying attention. But the real economics resolve later, at dated, knowable moments: public comment periods closing, rules taking effect, promised funding lapsing, sunset clauses firing. Finance already prices exactly one species of government date this way — FDA drug-approval deadlines, which support a whole cottage industry of biotech “catalyst calendars” — because those dates are discrete and published. Statutory time is full of such dates, for every sector of the economy.

I ran a systematic survey of the data vendors in early August 2026, expecting to find competitors. What I found instead is that no product anywhere sells that calendar; search for “regulatory calendar” and you get law-firm PDFs about compliance filing deadlines. The phrase currently means your deadlines. Nobody sells the world’s.

And the dead clauses stop vanishing. Provisions stripped out of one bill don’t die; they migrate — into a spending rider, a defense bill, a state legislature — often nearly word-for-word. Text struck for procedural reasons is the purest case: the political appetite survives intact, so reappearance is close to certain and merely needs a vehicle. Tracking that lineage across every bill in every state is a solved computational problem that has never been industrialized. Everyone in policy wants to know what’s coming back. Today the answer lives in lobbyists’ memories, available to their clients.

For readers who build data systems, three disciplines make this a reference rather than a demo, and none of them shows up in a screenshot:

• Two timestamps on every record — when the law said it, and when we learned it — so any question can be answered honestly as of any date, with no future information leaking into the past. Economic data is used as originally published, not as later revised.
• Hand-verified joins. The links between statutes and their outcomes ship only after hand-verification on random samples clears 90%, because every claim downstream inherits that error rate.
• No self-graded confidence. No AI model in the pipeline ever gets to report its own confidence — reliability is measured against human-labeled ground truth or it isn’t claimed.

Boring, invisible, and the entire difference between this and every “AI for policy” product that came before it.

§04 Keeping score on the future

The second layer sits on top of the map: given this provision, what happens next, with what probability, and when. Forecasting is where this category has historically embarrassed itself, so let me explain why this time can be different — and then tell you the honest version of the bet.

The right analogy is weather. Weather forecasting did not become trustworthy because someone hired cleverer forecasters. It became trustworthy when two things happened: the observation network got dense enough to fit real models, and meteorologists adopted a culture of verification — scoring every forecast against what actually happened, publicly, forever. It’s no accident that the standard mathematical score for probability forecasts was invented by a weatherman.

Forecasting became a science exactly when forecasters started keeping score on themselves.

That is the design here, mechanically. Every forecast states, in advance, the exact database query that will grade it — “did a final rule citing this section appear within 540 days: yes or no” — because a prediction you can’t grade isn’t a prediction, it’s content. Every resolved forecast, hit or miss, lands in a public, append-only record we call the Ledger, where a miss is displayed exactly as prominently as a hit.

And the same machinery that grades our forecasts can grade everyone else’s: the government’s own budget projections against actuals, agencies’ announced rule-writing plans against what they actually published, big AI models against the record. Nobody consents to being measured by a scorekeeper. That is rather the point of scorekeepers — and keeping our own entries on the same public page, unfiltered, is what makes the grading defensible.

Now the honest bet. Raw AI forecasting skill is commoditizing fast — on current benchmark trends, general-purpose models match elite human forecasters within about a year, and several good startups already sell calibrated predictions to institutions. If Downstream’s pitch were “we predict better,” it would be a melting asset. The durable assets are the ones benchmarks don’t measure:

• Knowing what to forecast — every clause in the map generates its own precise questions by the thousand.
• Point-in-time rigor that makes historical validation honest, which cannot be bolted on later.
• The standing to resolve — to say, citably, what actually happened.
• The archive of resolved outcomes itself — which happens to be exactly the kind of verified ground-truth data that AI labs now pay real money for.

The forecasting engine can ride the industry’s curve. The map underneath it compounds. And it is served over an open protocol, easy for other people’s machines to read, because the goal was never to be the only forecaster. It is to be the thing every forecast gets graded against, whoever made the forecast. The faster general AI improves, the more valuable the one thing it cannot generate for itself becomes: a trustworthy record of what happened.

§05 Law you can run

Everything so far treats law as something to observe. A surprising amount of it can be executed. A large fraction of statutory text isn’t really prose at all — it’s specification: thresholds, eligibility tests, phase-outs, matching formulas, triggers indexed to published statistics.

The federal share of each state’s Medicaid spending is arithmetic on published income data. Federal education grants key off published poverty estimates. Highway money follows written apportionment formulas. Certain unemployment benefits switch on automatically when a state’s jobless rate crosses a line printed in the statute.

The housing act in the cards above does this to cities. It penalizes a jurisdiction that builds slower than the median one — which means a city’s exposure is set by a national statistic it has no part in, and can move in a year when nothing local changed at all. The mechanism is legible from the statute today. The number is arithmetic, on published inputs, and the rule for publishing it is the strict one: ship the formula as runnable code alongside the figure, or ship neither.

Think about what spreadsheets did to finance. Before them, a financial model was a static document, expensive to build, recomputed by hand when an assumption changed. The spreadsheet made consequences instant — change one cell, watch everything downstream move. Law today is where finance was before the spreadsheet.

“If state incomes shift two percent, what happens to federal matching dollars in each of the fifty states, under the formula exactly as written?” has a precise answer that today is computed slowly, expensively, and mostly for those who can pay. Encode the formulas as running code wired to live data and it becomes a calculator — no predictions, no model risk, every figure citable to the statute — that a state budget office, a hospital system, or a school district can use on a Tuesday.

A small international research community has proven this works — projects in France compile tax law into verified code; open-source engines run national benefit simulations; one nonprofit maintains executable U.S. and U.K. tax-and-benefit law good enough that assistance programs build on it. All of them, sensibly, stopped at the tax-and-benefits domain, where the payoff justified encoding law by hand. What changed is that encoding no longer requires hands. The other 95% of the corpus is sitting there, uncompiled.

§06 The wind tunnel

Follow the ladder one rung higher and you reach simulation — and a place where I want to be more careful than excited, in both directions.

If law is executable and the record of past behavior is dense, you can stress-test a bill before it passes: run a draft against modeled agencies with limited staff, companies optimizing around compliance, litigants probing for weaknesses. The honest frame for this is a wind tunnel, not an oracle.

Simulations of policy have a poor record at predicting outcomes and a decent record at surfacing failure modes nobody thought to look for. You don’t build a wind tunnel to learn how the flight will go; you build it to find the fatal vibration cheaply, on the ground.

A simulation won’t tell you what a statute will do — but it will occasionally hand you the third-order path no human traced: this definition, fed through that formula, collides with this deadline. Then you verify the path deterministically, against the real record. The simulation proposes; the map disposes.

Here is the shape of one we watched run, start to finish. A trade bill passes one chamber overwhelmingly, restoring a tariff benefit that had lapsed months earlier. It never becomes law under its own number. Its text is folded into an omnibus spending act and signed there, and the emptied shell of the original bill is reused a season later as the vehicle for an unrelated stopgap. One bill number, two unrelated lives. A tracker that follows bill numbers loses the thread twice; a tracker that follows provisions never loses it at all.

Now run the substance through the wind tunnel. The restored benefit is retroactive, so importers who paid duties during the lapse can reclaim them. But the claim window is a fixed number of days from enactment, it binds private companies rather than an agency, and nobody is scored for missing it.

The refunds carve out whole categories of duties, which makes the benefit considerably narrower than the headline. Once a claim is granted, the money must be paid within a set period and without interest, so delay costs the claimant and costs the agency nothing. And the user fees underwriting the program run years past the expiry of the benefit they offset.

None of that is a prediction. Every fork is printed in the statute and echoed in the agency’s own filing guidance, and every one is a place where money either arrives or quietly doesn’t. What a simulation adds is the search — walking paths a drafter had no reason to walk and handing back the ones with consequences attached. What the record adds is the check. Deadlines of exactly this shape, worth real money to real companies, expire somewhere in the corpus most weeks, and the calendar that would hold them is precisely the artifact nobody sells.

And one rung above that sits the most powerful and most uncomfortable capability in this whole essay: drafting with feedback. If you can score a provision’s likely fate, you can search for language — “what wording, inserted here, achieves the stated goal with the least litigation risk and the most durable funding?” Legislative drafting today is a craft practiced blind; this would give it instruments.

It is also, unmistakably, a tool that would harden laws against challenge — and if it were sold only to whoever pays most, it would quietly compound the advantage of the already-powerful while wearing the costume of civic infrastructure. I’d rather answer that question in public, now, while it costs nothing: if this capability gets built, it will be priced for private clients and free to legislative staff of every party and to public-interest drafters.

Drafting power gets distributed symmetrically or the tool doesn’t deserve to exist. Commitments are cheapest before the revenue arrives, which is exactly when they should be written down.

§07 Who pays, and why we say it out loud

A fair question by now: who funds all this? The answer is unusual enough to state plainly rather than bury.

The people who most need this record — congressional staff, state legislators, journalists, researchers — have no budget for it, and never will. The people who will pay well for it are the commercial ones, because the gap between when a law passes and when its consequences actually land is worth money to almost anyone with revenue on the line.

Asset managers and traders want the timing edge. Insurance underwriters need to price regulatory exposure they currently absorb blind. Compliance teams — in-house, and the firms that advise them — need to know which obligations are real and which deadlines will actually bind.

And whole industries transact with the government on statutory clocks: a defense authorization doesn’t just move stock prices, it opens a sales cycle for the Pentagon’s suppliers and reshapes procurement for the Pentagon itself, and both sides of that market want to see the schedule. So the commercial side pays — for the calendars, the exposure analysis, the forecasts — and the map stays free for everyone else, forever.

Commerce funds the free tier.

It’s a strange sentence and a stable arrangement, and its strangeness is not a flaw: the paying customers and the civic beneficiaries want the same artifact — a record too accurate to argue with — for different reasons. Neither side has to trust our motives. They only have to check the citations.

The bigger commercial pattern is one of the most reliable in the history of information. In credit risk, in catastrophe risk, in stock indices, the same movie played out: a category of risk existed with no authoritative dataset; someone built the record and published the methodology; the record got cited, then licensed, then embedded in how the whole industry talks; and the scorekeeper — the reference layer — ended up worth more than every software tool sold beside it, usually by an order of magnitude.

Regulatory consequence is a category of risk with no authoritative dataset and no scorekeeper. This year even supplied a controlled experiment on which half of the business matters: the market currently pays premium multiples for proprietary benchmark datasets, while the best-funded policy-software company of the last decade was delisted from the stock exchange, and a standalone archive of legislation-over-time sold for about six million dollars.

The record alone is nearly worthless. The record plus the meaning of the record — the joins, the scores, the track record — is a reference institution. Everything here is aimed at the second outcome.

Which is also why the map being free is strategy, not charity. Reference layers win by being cited, built upon, and standardized around; you cannot become the standard by keeping the schema secret. What compounds instead is the part no competitor can copy: the archive that can’t be back-filled, and a public track record — misses included — that is credible precisely because it was kept before anyone was watching. Assets whose only input is having started early are the fairest kind of moat, and the hardest to fake.

§08 A country that can see

Zoom all the way back out. Doctors have outcomes data. Engineers have failure analysis. Pilots have flight recorders; traders have profit and loss. In every one of those fields, the arrival of honest feedback didn’t just improve performance — it transformed what the profession was, separating what worked from what merely sounded right. Lawmaking is the one consequential craft that still runs open-loop. There is no other domain of comparable stakes where the practitioners receive no systematic information about whether their previous decisions produced the intended results.

I don’t claim a feedback loop makes a country wise. Feedback doesn’t guarantee good decisions; it makes learning possible — a smaller claim, and a sturdier one.

A legislature that could see which of its own mechanisms actually execute — which deadlines get met, which formulas actually move money, which mandates die in silence — would still argue about goals, as it should. But it would stop repeating the same implementation failures on the same five-year cycle, because for the first time, someone would be keeping score.

And a citizen who asks “what does this law mean for me” would get an answer that traces to primary documents, instead of an answer from whoever had the most to gain from giving it.

Markets price whether a bill passes. Nobody prices what happens next.

So here is the whole case, compressed. The consequences of law are already public — scattered across a few dozen databases, expensive to join, waiting for an institution with the patience to join them. The cost of that patience just collapsed, the legal ground just shifted to make the questions urgent, and the field is — verifiably, as of August 2026 — empty. The gap between those two sentences up top is a business. Closing it for everyone is the point.

§09 The build plan

Everything above is an argument. This is the schedule, published because a scorekeeper that keeps its own plan private is asking for a trust it hasn’t earned. You can hold me to the dates, and if a row slips you’ll be able to see that it slipped.

The sequencing follows one rule, and it is a rule about engineering rather than about marketing. Every shippable thing has to be independently interesting, and it has to force a piece of the pipeline into existence. Ingest, normalize, resolve entities, diff, classify, grade — in that order, because each stage is only testable once the one before it runs daily.

A demo can skip stages. A record cannot. So the first thing published is deliberately the least impressive thing on the list, and it runs every morning whether or not anyone reads it.

There’s a second reason to start now rather than start well. The archive is the one asset here that cannot be back-filled. Every day the snapshot doesn’t run is a day of the record that no amount of later money or cleverness recovers — and this is not hypothetical, because federal pages are being edited and removed at a rate that makes a dated copy a civic good on its own. That workstream has no feedback loop, no decision point, and nothing to learn from for about two years. It just has to be switched on before anything else, and it has been.

Three gates run across all of it, and they are the reason the later rows are slower than they look.

Two timestamps on every record — when the law said it, and when we learned it — because without that every backtest quietly reads tomorrow’s newspaper. Hand-verified joins, where nothing built on a statute-to-rulemaking match publishes until a random sample clears ninety percent. And no self-graded model confidence, ever: reliability is measured against human labels or it isn’t claimed.

These are product quality, not caution. You cannot grade an agency while being sloppier than the agency.

What that buys, by about month six, is not a demo. It’s a small, honest slice of the map — a few thousand provisions joined to what actually happened, cited to primary documents, point-in-time correct — plus the beginning of a public track record that includes the misses.

The map stays free. The calendar, the exposure analysis and the forecasts are what get sold, later, once there is a record behind them worth paying for. That order is not modesty. A reference layer earns its position by being cited, and nothing gets cited that you have to buy before you can check it.