\section{Empirical discipline and a falsifiable short-run design} \label{sec:empirical} The model is calibrated to 2026 and simulated conditionally; nothing in it is a forecast, and the historical sample cannot identify the post-AGI parameters ($\alpha^\infty_i$, the floors $\underline p$ and $\underline\varepsilon$, $\sigma_T$ at compute intensities never observed). What the data can do is (i) pin the base point and the pre-AI elasticities, (ii) reject the model's description of the \emph{current} phase of the transition, and (iii) discipline the sign of the mechanisms. This section states what would falsify the model in the next decade and what we have checked against the historical record. \subsection{Validation against pre-AGI patterns} The frozen-technology balanced growth path reproduces by construction the 2026 shares of Table~\ref{tab:calibration}. Three untargeted comparisons are informative. First, the model's aggregate wealth-to-GDP ratio in 2026 ($\WoverGDPbase$) compares with household net worth of $6.0$ times GDP in 2026Q2 \citep{FedZ1_2026}: the model's balance sheet, which has no government debt and no foreign assets, is of the right size, with the value of land ($\LandOverGDPbase$ GDP) within the range implied by \citet{DavisHeathcote2007} and \citet{KnollSchularickSteger2017}. Second, the model's price--earnings ratio of structures, $1/(r+\delta_N)=\PEstructuresbase$, is of the order of the historical earnings yield of equities \citep{Shiller2026}; the compute price--earnings ratio of $\PEcomputebase$ is what rapid obsolescence and falling prices imply and is not observed separately in national accounts. Third, the heterogeneous economy's wealth Gini ($\GiniBase$) and top-10\% share ($\TopTenBase$) compare with $0.85$ and $0.69$ in the Distributional Financial Accounts \citep{FedDFA2026}; the top-1\% share ($\TopOneBase$ against $0.325$) is under-predicted, as expected without return heterogeneity. The historical decline of the labor share \citep{KarabarbounisNeiman2014,Autor2020} is not evidence for or against the model: the model's labor share is flat on the frozen path by construction and falls only through the compute-equivalent channel. What the 2015--2026 record does show, and the model matches at the base point, is that compute income is still a small share of GDP (information-processing-equipment investment of 2.6\% of GDP, data-center electricity of 4.4\%): the economy is at the beginning of the transition, where the source's Proposition~7 predicts gradual wage effects. \subsection{Short-run predictions that can be checked} Along the baseline path the model predicts, for 2026--2040, a set of co-movements that distinguish it from competing accounts: \begin{enumerate} \item \emph{Wages by task exposure.} Wages of the skill whose task is being automated fall relative to the other skill around the automation date and recover afterwards (Restrepo's Proposition~7; Figure~\ref{fig:baseline}). A design that regresses occupation-level wage growth on the interaction of task exposure \citep{Eloundou2024,FeltenRajSeamans2021,Webb2019} with the time path of compute costs, with occupation and year fixed effects and pre-2022 trends as a placebo, tests the sign and the \emph{timing}: the model predicts that wages fall \emph{after} the compute-equivalent cost of the task crosses the wage, not when the technology becomes feasible. \item \emph{Investment and margins by industry exposure.} Industries with high task exposure should show rising compute investment and \emph{falling} margins on the labor-intensive part of their business (compute is competitively supplied), while margins rise only where a scarce complement (sites, grid connections, proprietary data) is owned. The model's distinction between $R^Q$ (competitive) and the markup wedge of the AI-monopoly scenario gives the two hypotheses. \item \emph{Valuations.} The model predicts that the market value of compute capital per unit of compute income is low ($\PEcomputebase$) and that the value of claims to scarce complements rises relative to GDP. A firm-level design would sort firms by the share of their capital that is compute (servers, accelerators; from property, plant and equipment footnotes in annual reports) versus land and long-lived structures, and test whether the price--earnings ratios of the former are lower and whether their earnings yields track the user cost $(1+r)(1+\pi_Q)-(1-\delta_Q)$. Data: Compustat if licensed, or the SEC XBRL financial statements, which are public. \item \emph{Energy prices.} The energy-constrained scenarios predict a rising relative price of electricity in compute-intensive regions and a rising share of electricity going to data centers (4.4\% in 2023, 6.7--12\% projected for 2028 by \citealp{Shehabi2024}). A region-level design using interconnection-queue data and wholesale prices would test the compute--energy price link that drives utilization in the model. \end{enumerate} \subsection{Identification, confounding and measurement} Each design has the same threats. \emph{Pre-trends}: exposed occupations and industries were already on different trends (offshoring, routine-biased change); the 2000--2021 window provides placebo tests. \emph{Confounding}: AI exposure is correlated with education, tradability and firm size; the model implies that the relevant treatment is the compute-equivalent \emph{cost} crossing the wage, which varies over time within exposure groups and allows occupation-by-year designs. \emph{Measurement}: exposure indices measure feasibility ($\alpha$ finite), not realized automation ($A\alpha