Every major bill that has passed in your lifetime has been scored by the same handful of institutions. The Congressional Budget Office. The Joint Committee on Taxation. A few think tanks. A few hedge funds with private microsimulation models that never see daylight.
Inequality has hit historic highs anyway.
This is not a coincidence.
The analytical infrastructure that models federal policy has been excellent at modeling federal policy. It has been atrocious at communicating what that policy does to the people who live under it. CBO publishes a $3.4 trillion deficit number for the One Big Beautiful Bill Act and the conversation becomes a fight over whether the President is good or bad.
The distributional tables stop at the top quintile, which contains households earning $250K and households worth $2.5 billion in the same bucket.
The talking-head economists explain why the bottom 80% should be patient. The same bottom 80% gets poorer.
Maya is 31, a single mom in suburban Ohio, working a 36-hour week on the floor of a national retail chain. She makes about 32K a year, plus SNAP and the EITC. She rents. She doesn’t have a 401(k). She carries a few thousand in revolving credit. Her largest financial exposure isn’t an asset price — it’s the cost of groceries and the price of the gas she puts in a car she doesn’t quite own outright.
Maya has never read a CBO report. She would not recognize OBBBA by name. But she has lived under it since July 4, 2025. She has lived under every major bill that has passed since the day she was born. And not once, in thirty-one years, has anyone shown her what those bills actually did to her.
Charles is 58. Founding partner at a private equity shop. Net worth around $120M, mostly carried interest. Charles’s tax exposure on OBBBA is roughly two orders of magnitude larger than Maya’s. Charles’s situation under the bill was modeled in detail by every law firm he hires, every wealth manager he employs, and the lobbyists his firm contributes to. When OBBBA was being marked up, the people who care what it does to Charles knew before the vote.
Maya didn’t.
This is what fifty years of “trust the experts” has bought us. The experts trusted the wrong people. Or more precisely… the experts were for the wrong people.
The analytical infrastructure exists. It produces excellent work. It has been almost exclusively in the hands of the class that already wins the policy.
I built Civitas1 to end that.
What Civitas actually is
To fix this, I didn’t just write an essay… I built an engine.
Civitas is a desktop macroeconomic flight simulator.
It runs a stylized macro model — fiscal, monetary, tax, trade, labor, energy, foreign, regulation, technology, social, housing, defense — monthly across a ten-year horizon, tracking sixty-plus metrics including real GDP, headline and core inflation, unemployment, the long rate, debt-to-GDP, federal debt, the USD reserve share, bank stress, and currency basket levels.
It ships with eighteen historical bills modeled as compositions of factual policy provisions, from ERTA 1981 through Volcker, ZIRP/QE, Dodd-Frank, NAFTA, the 2018 Section 301 tariffs, TCJA 2017, the CARES Act, the IRA, and OBBBA 2025. Eighty-plus policy provisions across thirteen categories. Seven tax schedules ranging from current law to the 1950s 91% top marginal rate.
It carries twenty-eight named personas spanning the full US wealth distribution — Maya at the bottom, Charles at the top, the median couple and the dental partner and the leveraged real-estate investor in between — plus international comparators for the UK, Germany, Japan, and China.
And it does not stop at tax policy.
It models thirty-five hazard-event presets: thirteen conflict scenarios from Gulf War 1991 through Iraq, Russia-Ukraine, Israel-Gaza, and hypothetical Iran, Taiwan, and Korean peninsula contingencies; twelve climate-hazard presets from the 1988 drought through Katrina, the California wildfires, the Florida insurance market collapse, and the +1.5–2°C chronic shift; five pandemic presets from H1N1 through COVID acute, COVID long tail, H5N1 spillover, and antibiotic-resistance crisis; five domestic-crisis presets including the fentanyl peak, the next-generation synthetic, ACA repeal, and Medicaid full repeal. Each one fires per-persona hazard channels distinct from generic macro pass-through, so the defense engineer in Fort Worth, the reservist EMT in southern Ohio, and the long-haul trucker on the Oklahoma corridor resolve to very different outcomes from the same Iran-war bill.
You pick a scenario. You pick the personas you want to see it land on. You hit Run.
In five seconds you get a distributional verdict, six KPI tiles, a per-persona impact table with lived-impact labels — Underwater, Below cushion, Working-class, Comfortable, Top 5%, Top 1%, Top 0.1% — and a ten-year trajectory line per persona. Then you ask the in-app AI why is Maya hit hardest? and it gives you a causal explanation citing the exact levers that drove it. Not a hallucination. Not a plausible-sounding paragraph. A chain of multiplications grounded in the simulator’s actual state.
You can fork any bill, modify any provision, author entirely new provisions. The AI can author them for you. The AI can design entire bills for you on the Design a Bill page, every edit logged and reversible. If you run Claude Code, Civitas exposes itself as an MCP server so a coding agent can simulate a hundred bill variants overnight and write you a report.
The same infrastructure used by hedge funds to position macro books and by lobbyists to draft talking points now runs on a laptop, for free, for anyone who wants it.
Who the existing tools are actually for
The standard policy-information stack — the news article, the CBO report, the LLM — each does one thing well. Each of those things serves a different class of person.
None of them serves Maya.
The news article is for the political class. It tells you who voted how, what the floor speeches said, which constituency reactions were loudest. It gives you the framing that will dominate the news cycle. It is excellent at that, and nothing else does it. The illustrative household it sometimes includes — for a family of four earning $75,000, this means $1,200 less in taxes per year — is a prop. A number chosen to make the framing concrete, not to model the variance across actual households. Your household is not that prop. Maya’s household isn’t either.
The CBO and JCT reports are for the analytical class. CBO has about 275 staff, mostly economists and policy analysts with advanced degrees. JCT runs the revenue-side scoring with its own bench of PhD economists, attorneys, and accountants, and by statute the CBO has to use JCT’s revenue estimates for tax legislation. Between them, the two agencies represent more analytical infrastructure than any private institution in the United States can match. They produce gold-standard work. They produce it for committee staff and policy professionals, in the language of committee staff and policy professionals, with distributional tables that stop at the top quintile because that is the granularity Congress traditionally asks for. CBO is a fortress. Most citizens couldn’t enter the fortress if they wanted to, and the fortress was never built to invite them.
LLMs are for the curious literate. You can ask Claude or ChatGPT to explain OBBBA in plain language and get a fluent answer in twenty seconds. The popular critique is that they can’t follow tweaks. That critique is wrong — modern frontier models handle multi-turn refinement just fine. The real problem is there’s no engine underneath. When you ask an LLM what OBBBA does to a household earning $32,000 with two kids on SNAP, it gives you a plausible-sounding answer composed from its training data and its general reasoning. It is not running a microsimulation. It is not tracking how the deficit impulse flows through long rates into credit-card APRs. It is reasoning about what the answer probably looks like, based on what similar bills have done in the documents it was trained on. Sometimes that reasoning is excellent. Sometimes it is confidently wrong about a specific dollar amount in a way you cannot easily check. Every “what if” tweak is the model’s best plausibilizing of what such a tweak might do. The conversation feels productive while quietly drifting away from any underlying numbers.
None of these tools was built for Maya. None of them resolves to her household. None of them lets her ask what happens if she changes a provision. None of them shows her what Charles got in the same bill she got nothing from.
What Civitas isn’t
Before going further… Civitas is not a CBO replacement.
It is a stylized macro simulator calibrated against historical bills, federal-only, not predictive, and wrong about specific bills sometimes. The validation library publishes a per-line discrepancy report inside the app so you know exactly where the model breaks down — and so you can argue with it.
CBO will continue producing the official numbers for committee staff. Civitas produces the household-resolved numbers for everyone else.
What this changes
Five things shift when policy analysis runs on the user’s laptop instead of in the fortress.
Distribution beyond the top 20%. I cannot say this strongly enough… most distributional analyses stop at the top quintile, and the top quintile is meaningless. It contains a household with $250K in household income and a household with $2.5 billion in net worth, in the same bucket.
Most modern tax-policy capture happens inside the top 1% and the top 0.1%, not at the boundary between the fourth and fifth quintile. Civitas ships Anne (top 1%) and Charles (top 0.1%) as standard personas in the default comparison view. When you score OBBBA, the bill’s progressivity verdict against the median household is one thing. Its verdict against Charles is something else. Both numbers are true. The first is the framing the news article will use.
The second is closer to the actual fight.
Lived-impact framing. Percent changes are technically informative and practically useless. A 90% loss on $18M still leaves a household with $1.8M in net worth — still top 5% of US wealth. A 100% loss on $200K leaves a household at zero. The same percent delta means radically different things at different starting points. Civitas shows ending net worth in dollars alongside a lived-impact label — Underwater, Below cushion, Working-class, Comfortable, Top 5%, Top 1%, Top 0.1%.
When Maya finishes the ten-year window under OBBBA, the report doesn’t tell you she lost 4%. It tells you her cushion is gone.
Provenance. Every persona-level outcome traces back to a specific lever in the simulator’s EconomyLevers object. When the AI says Maya’s grocery costs rose 8%, it can tell you that came from EnergyShock plus TariffRate pass-through plus TradeDisruptionRate. The deficit-to-long-rates-to-consumer-credit channel is a specific lever stack, not a hand-wave — you can read the multiplications in the source. The number isn’t a vibe. You don’t have to take anyone’s word for it.
Interactivity. Five seconds to score a bill against a comparison group. Modify any provision, hit re-score, watch the outcome reshape. Ask the in-app AI what if I lower the SALT cap and it edits the bill, re-scores it, and explains what changed. CBO produces one score per bill version. Civitas produces however many scores you have the patience to try.
Authoring. The Design a Bill page lets the AI compose bills directly — adding provisions, setting metadata, authoring entirely new provisions — every edit logged and reversible by Undo. Through the MCP server, a coding agent can do the same work programmatically, simulating a hundred variants overnight. Bills export as JSON or compressed URL embeds so anyone with the app can replay your scenario exactly.
Policy modeling becomes something you compose, not something you receive.
The right mental model
When a major bill drops, I open Civitas first.
I load the bill if it’s pre-modeled, or compose it from the provision library if it isn’t. I pick a comparison group that puts Maya, Sarah and James, and Charles in the same view. I hit Run. In five seconds I have a per-persona impact table, a distributional verdict, sixty-plus tracked metrics, a ten-year trajectory. I can see what the bill does to my household, to my parents’ household, to the household down the street, and to the household whose lobbyists wrote the thing. Five minutes total.
Then I read the news article. Five minutes, mostly for political temperature and floor-speech tone.
I look at the CBO summary when it appears, days or weeks later, mainly to sanity-check my own model against the official aggregate. If there’s a major divergence, the Civitas validation library tells me where the calibration drifts. Twenty minutes.
I ask Claude or ChatGPT to translate any specific provision I don’t understand. Ten minutes if the bill has unusual language. The built in AI will actually build it for you in the app.
Forty minutes total instead of ninety, and the household-resolved view leads instead of trailing. The other three tools become supplementary context — political background, official cross-check, language translation. They are not the analysis. The analysis is the simulation that ran on my laptop in five seconds.
This is the inversion. The existing stack put the official aggregate at the top and the household resolution nowhere. Civitas puts the household resolution at the top and the official aggregate in its proper place — as one number among many that any honest model should be able to recover from first principles.
Policy belongs to all of us
Maya does not have a wealth manager. Charles has six. Maya does not have a tax attorney. Charles’s firm employs three. Maya does not have access to a private microsimulation model. Every fund Charles invests in does.
Until now, the asymmetry in policy understanding was a function of who could afford to be analyzed. That asymmetry is over. The same modeling infrastructure that the fortress kept for fifty years now runs, free, on Maya’s laptop. She does not have to take anyone’s word for what the bill does to her. She can score it herself.
The next time a major bill drops — and another one will — every person in the country with a laptop and an opinion can ask the question the experts have been answering privately for half a century: what does this actually do to me?
That question is the entire point of self-government. Civitas is the answer.
Try it
Civitas is free for Windows and macOS at civitas.marketally.eu. The free tier includes the full simulator, all eighteen built-in historical bills, all eighty-plus provisions, all thirty-five hazard presets, all twenty-eight built-in personas, and AI chat with your own API key. Paid tiers add managed AI through a bundled token allowance, full bill-authoring tools with unlimited storage, and MCP integration with Claude Code for agentic workflows.
To run the experiment that prompted this essay: open the Dashboard, set the scenario to OBBBA 2025, set the comparison group to the wealth-distribution default — Maya at Q1, Anne at top-1%, Charles at top-0.1%. Hit Run. Watch the per-persona impact table. Then open the Bills page, fork OBBBA, remove or modify one provision, re-score it. That second step is what fifty years of policy commentary never gave you. It’s why I built this.
Then try a Florida insurance collapse with Maya, an Iowa farmer, and a coastal homeowner in the same view. Try the Iran-war preset with the defense engineer and the reservist EMT. Try Medicaid repeal with Maya and Marcus. The same tool runs every scenario the country has run through, every scenario it might run through, and shows you who actually wears the cost.
Real-time FRED, Treasury, and BLS integration is shipping soon. Today the macro baseline is a snapshot; the next release scores against current rate curves and last week’s CPI print.
If you find a discrepancy with the official CBO score that matters to you, file an issue from inside the app — it lands directly against the repo. That’s how the model gets better. The point of publishing the validation library inside the app, rather than hiding it, is so you can argue with it.
Civitas is built by MarketAlly OÜ. It is a stylized macroeconomic simulator, not a replacement for official analysis. The source is private; the simulator’s calibration patterns and per-bill validation results are documented in the app’s “Validate Library” report.