\subsection{The classical economy: stylized facts and mechanism checks}
\label{sec:results_classical}

Before any language model enters the market, the classical economy must (i) reproduce the stylized facts
it was calibrated to, and (ii) respond to the mechanisms of Section~\ref{sec:theory} in the predicted
direction. Table~\ref{tab:stylized_facts} summarizes the 30 control seeds of the classical design
(seeds 101--130): first-order autocorrelation of tick returns is indistinguishable from zero
($-0.007$, seed SD 0.15), the variance ratio at five ticks is 1.18, excess kurtosis averages 20 with a
minimum of 1.8 across seeds, the first ten autocorrelations of absolute returns sum to $2.0$, the quoted
spread is 21~bp (effective 30~bp), and the mid tracks the latent value with a mean absolute deviation
of 0.63\,\%. The market absorbs the latent $+2.2\,\%$ shock of tick~300 in 25~ticks on average, i.e.\
before the dividend of tick~320 reveals it, through the informed fundamentalists and the news flow; the
20-tick response ratio of 0.28 says that most of the adjustment happens in the last ticks before the
announcement. No control seed produces a bubble episode (5\,\% for ten ticks) or a crash (10\,\% in
twenty ticks), which is the floor effect anticipated in Section~\ref{sec:identification}: in the
classical economy instability can only be created, not reduced.

Table~\ref{tab:classical_cells} reports cell means and Table~\ref{tab:classical_effects} the paired
effects against the control for the five primary outcomes, with Holm-adjusted $p$-values within each
cell. Four results matter for what follows.

\paragraph{Feedback and leverage create instability (Propositions~\ref{prop:feedback} and \ref{prop:spiral}).}
Raising the trend-follower share from 6 to 16\,\% raises the tracking error by 0.86 percentage points
($d = 1.26$), the spread by 14~bp and the maximum drawdown by 2.8 points, and creates bubble episodes
in 13\,\% of seeds where the control has none. Raising the leveraged share from 4 to 19\,\% has larger
effects on every outcome: tracking error $+2.1$ points, time to absorb $+28$ ticks, spread $+16$~bp,
0.9 bubble episodes per run and a maximum drawdown of 7.8\,\% against 3.3\,\%; adding a funding shock
(a 50\,\% cut of the leverage cap twenty ticks after a negative latent shock) pushes the bubble count to
1.4 and the drawdown to 9.4\,\%, and produces crashes in 30\,\% of seeds. Figure~\ref{fig:baseline_paths}
shows a representative seed of each economy. Leveraged traders lose money in both cells ($-4.8$ and
$-6.0$\,\% of initial wealth against $+0.7$\,\% in the control), and the losses are transferred to the
fundamentalists and market makers who absorb the forced sales.

\paragraph{Information quality and misleading news.}
Lowering the precision of private signals from 2 to 0.5 has no detectable effect on any primary outcome,
and raising it to 8 tightens the spread by 3~bp and the drawdown by half a point without changing the
tracking error. The small effects are a property of this economy rather than a failure of the
information channel: the latent shock is revealed by the next dividend within at most twenty ticks and
by public news in the meantime, so private signals shorten the absorption of a shock that the market
would absorb anyway. Misleading news is different: when half of the public items carry a hidden
10-cent bias, the tracking error rises by 0.08 points ($d = 0.74$), the spread by 3~bp and the drawdown
by half a point, because the public value itself is pulled away from the latent value and the
fundamentalists trade on it. This cell is the classical counterpart of a shared, biased model: a common
error in the public signal degrades price discovery even when every trader is otherwise rational.

\paragraph{Liquidity withdrawal.}
When the five market makers stop quoting for sixty ticks at the moment of the latent shock, the shock
takes 8~ticks longer to absorb ($p = 0.05$ unadjusted), but the tracking error, spread and drawdown
measured over the whole post-warmup window are unchanged: the remaining limit-order traders supply
enough liquidity for the price to find its level once the market makers return.

\paragraph{Safeguards and market design.}
The 5\,\% circuit breaker almost never fires in the classical economy (0.13 halts per run) and has no
effect; it matters only where prices move by 5\,\% in ten ticks, which the LLM arms below test. The
position limit of 300 lots is actively harmful here: it caps exactly the traders who absorb order
imbalances, so the tracking error rises by 1.3 points, the time to absorb by 28 ticks and the spread by
18~bp, and 0.6 bubble episodes per run appear. The periodic call auction (clearing every five ticks with
a $\pm 2\,\%$ collar) raises the tracking error by 0.7 points and the drawdown by 2.3 points relative
to continuous trading at the same shock, which is the expected cost of batching with a mechanical
collar; it is reported as a mechanism-robustness check in Section~\ref{sec:robustness}, not as a
recommendation.

\paragraph{Placebo and negative controls.}
Replacing the latent shock by two placebo events leaves every outcome within its control range except
the tracking error, which falls by 0.1 points because there is no shock to absorb; no bubble, crash or
halt appears. The two cells in which 25 or 50\,\% of the slots run the \emph{mock} policy, a seeded
value-plus-noise rule with independent errors, are the negative control for the LLM treatment: with
independent errors, replacing a quarter or half of the traders changes neither the tracking error nor the
time to absorb, bubbles or drawdowns, and tightens the spread (by 4 and 7~bp) because the mock rule
rests limit orders near the mid. Whatever the live LLM arms do beyond this is attributable to the
language model's decisions, not to the act of replacing slots.

\paragraph{Two assets.}
With two assets whose dividend innovations load 0.7 on a common factor, the fundamental return
correlation across dividend periods is 0.38 while tick-return correlation between the two mids is
$-0.01$: microstructure noise dominates at the tick frequency and the two books are priced almost
independently, so excess comovement is negative in the control ($-0.39$). Adding leveraged traders and a
funding shock to asset~0 produces crashes in asset~0 in half of the seeds and raises the maximum drawdown
of asset~1 from 3.6 to 11.6\,\% although its fundamentals are untouched, and the tracking error of
asset~1 from 0.6 to 7.4\,\%: contagion through common funding, not through correlated fundamentals. The
Diebold--Yilmaz spillover index stays near 1.2\,\% in both cells, because the variance decomposition of a
VAR on tick returns does not pick up the forced-sale episodes (Table~\ref{tab:contagion}).

\begin{table}[htbp]
\centering
\caption{Stylized facts of the classical control economy (30 seeds, post-warm-up window).}
\label{tab:stylized_facts}
\small
\input{tables/tab_stylized_facts}
\end{table}

\begin{table}[htbp]
\centering
\caption{Classical mechanism checks: cell means (30 matched seeds per cell). Stars: paired $t$-test
against the control, Holm-adjusted within the cell ($^{*}$ 10\,\%, $^{**}$ 5\,\%, $^{***}$ 1\,\%).}
\label{tab:classical_cells}
\scriptsize
\resizebox{\textwidth}{!}{\input{tables/tab_classical_cells}}
\end{table}

\begin{table}[htbp]
\centering
\caption{Classical mechanism checks: paired effects against the control with 95\,\% confidence intervals,
Cohen's $d_z$ and Holm-adjusted $p$-values (five primary outcomes per cell).}
\label{tab:classical_effects}
\scriptsize
\resizebox{\textwidth}{!}{\input{tables/tab_classical_effects}}
\end{table}

\begin{table}[htbp]
\centering
\caption{Two-asset economy: comovement and spillovers (30 seeds per cell).}
\label{tab:contagion}
\small
\resizebox{\textwidth}{!}{\input{tables/tab_contagion}}
\end{table}

\begin{figure}[htbp]
\centering
\includegraphics[width=\textwidth]{figures/fig_baseline_paths}
\caption{Mid price, latent value and public value in one seed (101) of three classical economies. The
latent $+10$-cent shock of tick~300 is revealed by news and by the dividend of tick~320.}
\label{fig:baseline_paths}
\end{figure}

\begin{figure}[htbp]
\centering
\begin{subfigure}{0.49\textwidth}\includegraphics[width=\textwidth]{figures/fig_classical_mad_pct}\caption{Tracking error}\end{subfigure}
\begin{subfigure}{0.49\textwidth}\includegraphics[width=\textwidth]{figures/fig_classical_time_to_absorb}\caption{Time to absorb the latent shock}\end{subfigure}
\begin{subfigure}{0.49\textwidth}\includegraphics[width=\textwidth]{figures/fig_classical_max_dd}\caption{Maximum drawdown}\end{subfigure}
\begin{subfigure}{0.49\textwidth}\includegraphics[width=\textwidth]{figures/fig_classical_spread_bp}\caption{Quoted spread}\end{subfigure}
\caption{Classical mechanism checks: cell means with 95\,\% confidence intervals over 30 seeds (control in grey).}
\label{fig:classical}
\end{figure}