\begin{appendices}
\section{Data-generating process and quality audit}\label{app:data}

\paragraph{Synchronisation and exit rule.} Each cycle re-reads every source; a listing absent from its source for two consecutive cycles is declared inactive by the platform. The research definition of exit (seven days of absence before the snapshot) is stricter: the per-source ``reactivation share'' (exit events later reversed, divided by exit events) is 0.44 for Otogo, 0.04 for AutoUsagée, below 0.1 for most dealer websites, and 0.52--0.94 for the four connectors excluded from the dynamic sample. Listings whose last observation falls in the final seven days of the window are right-censored.

\paragraph{Sustained price paths.} Of 428{,}484 price-log rows, 2{,}257 record a withdrawn price; 121{,}239 are price changes, of which 11{,}479 occur less than 24 hours after the previous change. The transient rule (a level shorter than 48 hours that reverts to the preceding level is deleted) removes the alternating-feed artefacts illustrated by a listing that alternated between \$37{,}840 and \$40{,}794 twice a day for four weeks; 70 listings with three or more transients are excluded from markdown analyses. Robustness to a 24-hour and a 72-hour rule is reported in Section~\ref{sec:robust}.

\paragraph{Identifier stability.} Within Otogo, 1{,}059 of 63{,}821 VINs (1.7 percent) appear under more than one listing identifier; inspection shows these are the same vehicle re-listed by a different dealer after a transfer, not identifier churn. Dealer websites assign stable identifiers.

\paragraph{Governance.} The snapshot excludes the platform's user table; VINs, URLs, images, free text and exact coordinates are retained only in the restricted raw and interim layers, never in released tables, which carry hashed listing and dealer identifiers and coordinates rounded to 0.01 degree.

\section{Preregistration and amendments}\label{app:prereg}

The research questions and the analysis plan (v1.0, 2026-11-01) were written after the data audit and before estimation. Amendments recorded during execution: (D-010) make--model and make--model--trim quality fixed effects instead of make--model--year, because model-year effects absorb vehicle age within an 81-day window; (D-011) linear-probability estimation of the competitor event study with high-dimensional fixed effects; (D-014) per-source left-censoring after the portal start-up stock (33{,}224 listings) was found to be treated as entries; (D-015) exclusion of panel weeks 0--1 from the liquidity index. The complete decision log, experiment log and claims ledger are distributed with the replication package.

\section{Additional estimates}\label{app:extra}

Additional coefficient tables (full spline bases, listing-age bins, powertrain and body effects, the log-logistic accelerated-failure-time model, the 10-km and 50-km competitor sets, the dealer-by-weekday specifications, LightGBM feature importances and calibration tables) are provided as machine-readable files in the replication package (\texttt{results/}).
\end{appendices}