Float precision loss when aggregating multi-year budget-window totals

[MaxGhenis]

Summary

Budget-window impact fields are typed float and summed with += across up to 75 years. At trillion-dollar magnitudes this accumulates meaningful precision loss — the totals and the per-year impacts in the response will not be self-consistent to the cent (or even to the dollar).

Location

• projects/policyengine-api-simulation/src/modal/budget_window_results.py:54-72
• projects/policyengine-api-simulation/src/modal/gateway/models.py:113-128

What goes wrong

class BudgetWindowAnnualImpact(BaseModel):
    year: str
    taxRevenueImpact: float
    federalTaxRevenueImpact: float
    ...

def sum_annual_impacts(annual_impacts):
    totals = {"taxRevenueImpact": 0, ...}
    for annual_impact in annual_impacts:
        totals["taxRevenueImpact"] += annual_impact.taxRevenueImpact
        ...
    return BudgetWindowTotals(**totals)

Federal revenue impacts in the trillions sum over 75 years exceed 2^53 and quickly lose integer precision in IEEE-754 double. Even at sub-trillion magnitudes, the naive += accumulation introduces drift that makes the reported totals.budgetaryImpact disagree with sum(annualImpacts.budgetaryImpact) when clients re-check.

Suggested fix

• Model impacts as decimal.Decimal or integer cents at the boundary; convert to float only for display.
• Or use math.fsum(...) in sum_annual_impacts to reduce accumulation error.
• Add a test that sums 75 years of trillion-dollar impacts and asserts totals equal the exact sum.

Severity

Medium. Silent numerical drift in public-facing budget numbers.