Domains of validity – Populace dynamics

Why this chapter comes first

All models are wrong; a model earns its keep only where it improves predictions. This chapter states, before any machinery is described, which questions this project claims to help with, which it refuses to answer with false precision, and how users — human or AI — can tell the difference. Every output the model ships carries this boundary as metadata, not as a footnote.

Parameter uncertainty dominates the long horizon

The 75-year actuarial balance is a function of a handful of exogenous assumptions: total fertility, mortality improvement, net immigration, real wage growth, and interest rates. Over long horizons, uncertainty in those inputs — not the fidelity of the microsimulation that processes them — dominates the answer.

The public record makes this concrete:

The Trustees’ own low- and high-cost scenarios bracket a range of 75-year balances wider than the intermediate deficit itself, and their stochastic intervals are similarly broad (Board of Trustees, Federal Old-Age and Survivors Insurance and Federal Disability Insurance Trust Funds 2025).
CBO’s long-term projections have persistently shown materially larger 75-year shortfalls than the Trustees’, driven mainly by different demographic and economic assumptions rather than different benefit arithmetic (Congressional Budget Office 2024).
Successive Technical Panels convened by the Social Security Advisory Board have recommended revising the fertility and mortality-improvement assumptions as realized values ran outside the intermediate path for extended periods (Technical Panel on Assumptions and Methods 2023).

Two expert institutions, each with administrative data and decades of refinement, disagree with each other and have both missed realized demographic trends. A new model — however well engineered — does not fix that, because the variance lives in the inputs. Micro-level precision layered on unforecastable demographics is spurious precision.

What this project will not claim

A 75-year point forecast of solvency. The project will not market a headline depletion date or a point 75-year balance as a prediction. Where such quantities are computed, they are computed conditional on named assumption paths and presented alongside their sensitivity, never as the model’s answer.
Precision on behavioral response it cannot validate. Behavioral margins (claiming responses to reform, labor-supply feedback) ship as clearly labeled scenario inputs with documented ranges, not as point estimates wearing model authority.

What survives, and why

Three output tiers, ordered by the strength of their claim to usefulness:

Tier 1: distributional analysis under fixed assumptions

Reform analysis is a difference: outcome under reform minus outcome under baseline, holding the population and assumption path fixed. Much of the unforecastable demographic uncertainty is common to both arms and cancels in the difference. The ingredients this requires — a calibrated joint distribution of lifetime earnings, family structure, and differential mortality, plus an exact rules engine — are the parts of the system that can be validated directly.

The honest caveat, stated rather than buried: the slice of a reform delta that does not cancel — interactions between the reform and the uncertain dynamics, such as reform-induced claiming shifts or mortality-gradient interactions — is the least validated slice. The model labels it as such.

Tier 2: near-term components that resolve

Over roughly a ten-year horizon, the model’s outputs are dominated by mechanics rather than demographic extrapolation: beneficiary counts by type, average benefits, covered earnings and taxable payroll, claiming-age distributions, disability incidence. These quantities resolve against administrative publications on an annual cycle, so claims about them can be scored rather than argued (see scoring-and-resolution.md).

Tier 3: the long horizon as a sensitivity surface

Long-horizon outputs are published as surfaces over assumptions, not points: how the balance, cohort replacement rates, or distributional outcomes move as fertility, mortality improvement, and immigration vary across documented ranges. The incumbent practice is to bury assumption-dependence behind a point estimate and an appendix. An open model can invert that: make the sensitivity the product. That is a more honest object than a forecast, and it is the only long-horizon object whose accuracy does not depend on predicting the unpredictable.

Where dynamics genuinely earn their keep

The longitudinal machinery is not decoration. Questions that require lifetime position — who bears a retirement-age increase by cohort and lifetime-earnings quintile, how survivor outcomes shift under a benefit redesign, how insured status evolves for interrupted careers — cannot be answered from a cross-section. The point of the tiers is not to shrink the model’s ambition; it is to attach each ambition to the strongest claim it can actually support.

Validity as metadata

Every API and MCP response carries its tier, its assumption path, and its calibration history. A downstream agent composing this model with other tools can weight it accordingly — trust is a number the consumer reads, not a reputation the producer asserts. The scoring-and-resolution.md chapter defines where those numbers come from.