Chapter 8: Cosilico - Infrastructure for the Future
GPT-4 gets tax questions wrong 33% of the time.
This is not an urban legend or an exaggerated claim. Blair-Stanek et al. (2023) developed SARA (StAtutory Reasoning Assessment)—a benchmark for evaluating large language models on US income tax calculations—and found that GPT-4 answered only 67% of true/false tax questions correctly. On scenario-based calculations, GPT-4 got tax liabilities exactly right only about a third of the time .[1]
The models confused marginal and effective rates. They misapplied filing status rules. They hallucinated phase-out thresholds that didn't exist.
"Today's LLMs cannot 'do taxes' on their own because tax calculations require 100% correctness. Today's models hallucinate."
This finding, from Column Tax's engineering blog in 2024, captures a fundamental problem: AI systems will increasingly mediate financial decisions, but they cannot reliably calculate the regulatory details that govern those decisions .[2]
PolicyEngine had demonstrated that deterministic tax-benefit calculations could be encoded as open-source code. The question was whether that approach could become infrastructure—not just for policy analysis, but for every fintech application, government agency, and AI assistant that needed accurate calculations.
That question led to Cosilico.
The Infrastructure Gap
Look at the landscape of tax and financial infrastructure in 2024:
Sales tax has Avalara—a company acquired for $8.4 billion in 2022 .[3] They provide APIs that calculate sales tax obligations for e-commerce transactions. But sales tax only.
Payroll tax has Symmetry, ADP, and others. They calculate employer and employee tax obligations. But payroll only.
Benefits screening has Benefit Kitchen and similar services. They estimate program eligibility. But coverage is limited—a handful of states, no tax integration.
Tax filing has TurboTax and emerging players like Column Tax. They help individuals file returns. But they're consumer-facing, not API-first infrastructure.
Policy simulation has academic models like PolicyEngine, EUROMOD, Tax-Calculator. They're rigorous but not production-ready for commercial applications.
No one provides the full stack: income tax + benefits eligibility + attribute prediction + population simulation in a single, production-ready API.
This gap matters because every application that involves money eventually runs into tax and benefit calculations. A lending app needs to estimate after-tax income. A benefits platform needs to determine eligibility across programs. A financial planning tool needs to project tax liability under different scenarios. An AI assistant asked about finances needs to call something to get accurate numbers.
Without unified infrastructure, each company builds fragmented, partial solutions—or gives wrong answers.
Why This Can't Be Trained Away
A natural response to the AI accuracy problem is: "Just train better models."
This won't work for tax and benefit calculations, for structural reasons:
Tax law changes annually. Every January brings new brackets, new thresholds, new credits. Training data from last year encodes rules that no longer apply. Models can't extrapolate to provisions they've never seen.
Fifty states have different rules. Each state has its own income tax (or no income tax), its own credits, its own benefit programs. The combinatorial explosion of jurisdiction-specific rules exceeds what pretraining can memorize reliably.
Eligibility depends on dozens of variables. SNAP eligibility involves income, household size, expenses, categorical eligibility, state supplements. One wrong variable produces wrong results. LLMs compress this complexity into statistical patterns that don't preserve exact logic.
Calculations must be 100% correct. A model that's 95% accurate sounds impressive until you realize that 1 in 20 tax returns would be wrong. Real-world tolerance for error in financial calculations is effectively zero.
The Column Tax engineers put it bluntly: "Today's LLMs cannot 'do taxes' on their own because tax calculations require 100% correctness" .[2]
The solution isn't better training. It's tools.
Deterministic + Auditable
Cosilico's thesis is that AI systems need deterministic, auditable tools for calculations—and that building those tools as open-source infrastructure creates value for everyone.
The architecture has three components:
Rules Engine: Every tax and benefit formula encoded as deterministic code, traceable to statute. The EITC calculation cites 26 USC § 32. The SNAP calculation cites 7 USC § 2014. Each computation includes a citation and the parameter values used.
Synthetic Populations: For aggregate analysis, you need representative data. Cosilico constructs synthetic populations by calibrating public microdata to known totals, imputing missing variables, and validating against administrative aggregates. The result is a privacy-preserving dataset of ~100 million synthetic households that produces correct aggregate tax revenue when run through the rules engine.
Scenario Simulation: With rules and population, you can answer counterfactual questions. What if the EITC expanded by 50%? Run the baseline, run the reform, take the difference. The output might show: cost estimates over ten years, households affected, poverty reduction in percentage points—all calculated from the underlying rules and population data.
The key properties:
- Deterministic: Same inputs → same outputs, always
- Auditable: Every calculation includes legal citation and parameter values
- Versioned: Git history tracks all rule changes
- Bi-temporal: Parameters track both effective date and knowledge date
When an AI agent calls Cosilico to answer a tax question, the calculation is provably correct, legally citable, and traceable. The AI explains; the infrastructure calculates.
Building on PolicyEngine
Cosilico doesn't start from scratch. It builds on the foundation PolicyEngine established.
PolicyEngine proved the core thesis: tax and benefit rules can be encoded accurately at scale. Over a million simulations have run on the platform. The UK Treasury has used the UK model for policy costing. US Congressional offices have used the analysis. The codebase covers US federal plus 50 states, the UK, and Canada—with 50+ open-source contributors maintaining and extending it .[4]
This is the starting point. But PolicyEngine was designed as a policy analysis tool—a nonprofit providing free research infrastructure. Cosilico is designed as commercial infrastructure—an open-core business providing production APIs.
The relationship is complementary:
- PolicyEngine continues as a nonprofit doing policy research and public-facing tools
- Cosilico builds on the same rule-encoding approach but serves commercial customers
- The open-source core is shared; the commercial layers are separate
This structure mirrors patterns in other successful open-source infrastructure: Linux and Red Hat, Kubernetes and cloud providers, PostgreSQL and managed database services.
The Open-Core Model
Cosilico is designed as open-core: the simulation engine would be open source (Apache 2.0 licensed), while hosted services and enterprise features would be commercial.
The envisioned model follows patterns proven by companies like GitLab, Elastic, and HashiCorp: give away the core technology, charge for convenience and enterprise features. Whether this model will work for policy simulation infrastructure remains to be tested.
The open-source foundation would serve multiple purposes:
Trust through transparency. When a fintech company integrates tax calculations into their product, they can inspect exactly how those calculations work. No black boxes.
Community contributions. Tax law is vast and constantly changing. Open source enables distributed maintenance—state-level experts can contribute state-specific rules.
Regulatory readiness. As AI regulation increases, audit trails matter. Calculations based on open-source, citable code are inherently more defensible than LLM outputs.
Competitive moat. Paradoxically, open-sourcing the core makes it harder for competitors to catch up. The comprehensive rule coverage becomes a network effect—each additional program encoded increases the value of the whole.
Why This Might Matter
The thesis is that accurate tax and benefit calculations will become essential infrastructure—not a niche academic tool but something every fintech app, government service, and AI assistant needs.
Several trends support this:
AI tool use is becoming standard. Function calling shipped in GPT-4 and Claude 3. Anthropic's Model Context Protocol is being adopted broadly. AI assistants need reliable tools to call—hallucinating tax calculations is unacceptable.
AI financial regulation is coming. The SEC, CFPB, and state regulators are examining AI in financial services. Audit trails and explainability will likely be required. Citation-based approaches are regulation-ready.
The precedent exists. Avalara built a large business ($8.4 billion acquisition) providing sales tax APIs alone. Someone will build the equivalent for income taxes and benefits. The question is whether it will be open or proprietary.
This is a bet, not a certainty. Many infrastructure plays fail. The market for policy calculation APIs is unproven. But the underlying need—accurate calculations that AI can call—seems real.
The Shared Substrate Vision
The deepest rationale for Cosilico connects to the book's larger thesis:
"Society is hard to optimize because nobody has a shared model to reason against. Congress debates with napkin math. Banks model risk without knowing policy changes. AI agents hallucinate eligibility rules.
Cosilico is the shared substrate—a simulation everyone can query, so decisions are grounded in the same reality."
When policymakers consider reforming the EITC, they should be able to query the same simulation that banks use to assess lending risk, that fintech apps use to estimate tax liability, that AI assistants use to answer questions. The calculations should be the same because the underlying reality is the same.
Today, fragmentation creates divergent answers. Official government estimates use proprietary models. Think tanks use different models. Companies build ad-hoc solutions. Individuals get inconsistent information. The policy debate suffers because participants aren't reasoning against the same facts.
Shared infrastructure could change this. A canonical, open-source, production-quality simulation that anyone can query—and anyone can verify—creates common ground. Disagreements can focus on values and priorities rather than on whose numbers are right.
This is infrastructure for democratic deliberation: transparent, accessible, shared.
From PolicyEngine to Cosilico
The transition represents an evolution in thinking about what policy simulation infrastructure should be.
PolicyEngine asked: "Can we make policy analysis accessible to everyone?" The answer was yes—through open-source models, web interfaces, and free public tools.
Cosilico asks: "Can accurate policy calculations become infrastructure that everything else builds on?" The answer requires commercial sustainability, production-quality engineering, and integration with the broader ecosystem of AI tools and financial applications.
PolicyEngine demonstrated the proof of concept. Cosilico is the attempt to scale it.
The nonprofit continues, focused on research and public goods. The company builds the commercial infrastructure layer. The open-source foundation supports both. And the vision—understanding policy through transparent, accurate simulation—remains the same.
What Success Looks Like
If Cosilico succeeds, the results would be visible throughout the financial ecosystem:
AI assistants that give accurate answers to tax and benefit questions—not because they've been trained better, but because they call reliable tools.
Fintech applications that correctly calculate after-tax income, benefit eligibility, and financial impacts—without each company reinventing the calculations.
Government agencies that use the same validated calculations as the private sector—reducing discrepancies and improving trust.
Policy debates grounded in shared numbers—where disagreements are about values, not whose model is right.
Individual citizens who can understand how policies affect them—accessing the same calculations that Congress uses.
This is the vision of "society in silico" applied practically: simulation as infrastructure for understanding and improving the systems that govern our lives.
The Work Ahead
As I write this in late 2024, Cosilico exists more as thesis than product. The company is incorporated. The rules engine is being built, drawing on PolicyEngine's foundation. Design partners are testing early versions. But there's no production API, no paying customers, no proof that the market exists.
I include this chapter not because Cosilico has succeeded, but because the problem it addresses is real regardless of whether this particular solution works. AI systems need deterministic tools for financial calculations. Someone will build this infrastructure. The question is whether it will be open or proprietary.
If Cosilico fails—and startups usually fail—the thesis remains: open-source policy simulation infrastructure, production-ready and commercially sustainable, would be valuable. Someone should build it. Maybe we will. Maybe someone else will do it better.
The honest framing is aspiration, not accomplishment. PolicyEngine demonstrated that open policy simulation is possible. Cosilico is an attempt to make it sustainable and scalable. Whether that attempt succeeds is unknown.
What I can say with confidence: the gap exists. GPT-4 gets tax questions wrong a third of the time. Every fintech company that needs accurate calculations today builds fragmented, partial solutions. The need for shared infrastructure is real, even if the path to building it remains uncertain.
Cosilico is a bet. You're reading this book partly to understand what the bet is on, and why it might matter even if the specific company doesn't survive.