\section{Mechanisms: what the evidence says about the theory}
\label{sec:mechanisms}

Table~\ref{tab:predictions} listed six predictions. Reading them against Section~\ref{sec:results} gives
a mixed scorecard that is more informative than a clean confirmation would have been.

\paragraph{H1 (informativeness falls in the shared-model share): not supported on levels, supported on dynamics.}
The stylized model predicts that a common error component in the signals of a growing share of traders
leaves a persistent gap between price and value. The simulator shows no such gap: the tracking error
is, if anything, lower with LLM agents at every share. The reason is visible in the prompts and the
decision logs. The LLM agents are given the public present value and told how it moves with the next
dividend; most of them hold no private signal (they inherit the information set of the slot they replace,
and only fundamentalists are informed); their modal decision is a limit order near the public value.
Their ``common error'' is therefore not a biased signal about the fundamental but a common
\emph{anchor} on the public value, and a population anchored on the public value prices the stock at the
public value---which is close to the latent value except in the window between a latent shock and its
revelation. In that window the prediction holds: the shock is absorbed more slowly as the share rises
(Section~\ref{sec:results_dose}), and the persistence of returns (variance ratio 2.1, autocorrelation
0.32 at 95\,\%) is the signature of a price that moves toward new information in many small steps against
a wall of stale limit orders. The misleading-news cell of the classical design shows what a genuinely
biased common signal does (Section~\ref{sec:results_classical}); the LLM cells did not inject one.

\paragraph{H2 (coordination rises with the share): supported for synchrony, rejected for flow correlation.}
Decision synchrony is near its ceiling at every share, consistent with a single model answering a single
snapshot. Pairwise flow correlation falls with the share because the agents trade with one another; the
stylized model, in which every trader submits a market order against a common market maker, has no room
for this. The classical analogue shows that rule-based agents of one type are more correlated than
LLM agents, which qualifies the common intuition that language models are the more homogeneous traders:
in this economy they are more homogeneous than a mixed classical population and less homogeneous than any
single classical rule.

\paragraph{H3 (bubbles and crashes rise with penetration): rejected; H3$'$ (leverage interaction): reversed.}
No bubble, crash or halt occurs in 112 live runs. With leverage present, the LLM agents reduce the
instability the leveraged traders create. Proposition~\ref{prop:spiral} is about an exposed cohort whose
positions are same-signed and leveraged; the LLM agents in this design are unleveraged and their
positions lean against the exposed cohort, so the proposition's premise is not met and its conclusion does
not apply. A design in which the LLM agents themselves carry leverage is the natural next test.

\paragraph{H4 (diversity restores informativeness): supported in sign, small in size.}
Diverse prompts and the two-model mix lower coordination and shorten the absorption of the latent shock
by about four ticks, as Proposition~\ref{prop:diversity} predicts; they also lower liquidity provision,
which the proposition does not consider because its traders do not supply liquidity.

\paragraph{H5 (communication raises synchrony): supported.}
The channel raises pairwise correlation to the highest level of all cells and the tracking error by a
small but precisely estimated amount; memory does the same, less strongly.

\paragraph{H6 (safeguards trade depth for delay): not testable at these shock sizes.}
The breaker never fires; the position limit binds and removes liquidity without removing any
instability, because there was none to remove.

\paragraph{The unifying mechanism.}
The one regularity that runs through every arm is liquidity provision by value-anchored agents who act
infrequently. It explains the tighter spreads and deeper books, the lower volume, the slower absorption of
latent information, the predictable returns, the trend followers' higher profits, the dampening of
leverage-driven fire sales, the liquidity cost of diversity (more market orders) and of position limits,
and the behaviour of the frozen library at $N = 1{,}000$. It is not the mechanism Section~\ref{sec:theory}
set out to test---correlated errors in a market-order population---but it is the mechanism the data
identify, and it suggests a different stylized model: a population of patient, public-value-anchored
limit-order traders whose aggregate order book slows price discovery of private information while
stabilizing prices against non-fundamental shocks. Whether that is good or bad for a market depends on the
mix of information and noise in its order flow, which the information-quality and misleading-news cells
begin to map.