\section{Limitations and future research}
\label{sec:limitations}

The synthesis is bounded by its retrieval process. Only \retrievalRate{} percent of the stage-one hedonic studies could be
obtained as legal open-access full texts, so the extracted sample over-represents recent, open-access and non-US work; the
paywalled core of the 1980s--2000s North American journal literature, which the Sirmans programme summarised, is
under-represented, and the pooled estimates should be read as a synthesis of the accessible literature. The discovery itself
relied on three machine-readable sources; SSRN, Google Scholar and institutional repositories were not harvested, and the
multilingual queries retrieved mostly English-language records. Screening and extraction were performed by language-model
agents following written protocols; the validation exercise measures their agreement on eligibility and on the numbers they
both recorded, but the \valCoefRecall{} percent recall against an independent extraction means that some printed coefficients
were missed, and the systematic review community's standard of two human reviewers was not met. Harmonisation required
choices: construction-year variables, standardised coefficients, interactions and un-harmonisable units were excluded, which
removes a quarter of the extracted rows; the Kennedy correction assumes normal errors; semi-elasticities per square metre mix
houses and apartments whose marginal square metre is not the same good. The moderators are coarse (continent, income group,
estimator class) and the meta-regressions have limited power for families with fewer than twenty studies. Publication-selection
diagnostics are difficult to interpret under extreme heterogeneity and dependence, and no correction was applied. Finally, the
AI synthesis rests on forty studies and cannot support claims about the general accuracy of machine-learning valuation. Future
work should extend retrieval through institutional access, add the human double-screening that the protocol envisaged, and
collect the reporting elements (units, transformations, standard errors, test-set definitions) that this exercise shows to be the
binding constraint on cumulative knowledge in hedonic research. Table~\ref{tab:gaps} lists the characteristics with fewer than
ten studies: interior amenities, market and transaction characteristics, several environmental risks and almost all AI-derived
characteristics remain understudied, as do Africa, South America and Oceania.

{\scriptsize\input{tables/tab15_gaps_long}}