spb/wp20_uqo

1% ============================================================================2\section{Limitations, ethics and future research}3\label{sec:limitations}4% ============================================================================56\subsection{What the data cannot establish}78\paragraph{No human expert labels.} The study has no annotations by appraisers or building technologists. Agreement is measured against broker fields, which are claims, and against the roll, which records few of the attributes of interest; judgment attributes are evaluated by cross-extractor agreement only, which bounds their reliability from below but says nothing about their validity. Everything the paper says about finish grade and condition should be read as ``what three vision--language models report'', not as ``what the house is''. The protocol for the expert study that should follow is specified in Appendix~\ref{app:protocol} and pre-registered with the code.910\paragraph{Latent attributes.} Foundation integrity, hidden moisture or mould, structural soundness, permit status, code compliance, tenure and servitudes are not inferable from listing media and were never inferred. The Monte Carlo intervals are conditional on their absence, and their under-coverage of both prices and assessor values is partly the measure of that conditioning. A house that looks like new and is sitting on a failing foundation receives the same indication as its sound neighbour.1112\paragraph{No cross-city cost levels.} The Statistics Canada index gives temporal change by metropolitan area but not level differences; the two published anchors describe Montréal and ``Québec'' and were issued by an industry software vendor's secondary reporting and by builder-facing guides, not by a cost manual. The regional level is therefore a sensitivity parameter, and the Montréal alignment ratios, where land is a larger share of value, are the least reliable.1314\paragraph{Assessor values are not truth.} The roll's building value is the assessor's dated, cost-based indication reconciled with market evidence, equilibrated and subject to its own error; indexing it with a construction price index to the valuation date is an approximation that the roll-vintage split shows to be imperfect. Alignment with it is a benchmark comparison, not an accuracy measure.1516\paragraph{Asking prices are not transactions.} Market alignment is measured against asking prices for the main sample and against recorded sales for a small, self-selected set of quickly sold units. Asking prices exceeded recorded sales by a median of two to three percent in that set, and the cost indication was closer to the sale than to the ask; the main results would shift accordingly but the small sample does not allow more than that statement.1718\paragraph{Representativeness.} The frame is the set of house listings active on 67 connectors over ten weeks, matched within 15~m to a roll unit, with at least four photographs, excluding two classified-advertisement platforms and one franchise network whose media host disallows retrieval. The excluded network's listings are more urban and have higher land values; listings are not the housing stock; and the stratified sample over-represents the smaller metropolitan areas by design. Design weights are stored and the paper makes no claim about Québec housing as a whole.1920\paragraph{One provider family.} The three extractors are tiers of one provider's model family, chosen for comparability of prompts and schemas. Cross-extractor agreement therefore understates the disagreement that models of different families would show, and the results do not establish that any particular model is required.2122\paragraph{Production component not inspected.} The Vrai Prix production ``cost estimate'' used as a context benchmark has a generating code we could not inspect during the study; it is reported as an observed output of unknown formula.2324\subsection{Ethics and rights}2526The study handled copyrighted media by reference only, excluded hosts that disallow automated retrieval, masked all value information before any model saw a text, reproduced no photograph, and reports only aggregates. Residual uncertainties remain about the terms of use of individual broker sites and about the status of broker descriptions as copyrighted works; the data-rights document released with the code records them. Two further ethical questions concern deployment rather than research. A cost indication from listing media could be used to pressure sellers or to screen neighbourhoods; because the indication is explicit about its inputs, misuse is at least auditable, but the fairness analysis the paper could perform is limited to subgroups by price, age and region, not by the characteristics of residents, which the data do not and should not contain. And an extraction schema that asked about occupants or belongings would be a privacy violation; ours asks only about the building, and the one free-text field is instructed to name building cues only.2728\subsection{Future research}2930Four studies follow directly. First, the expert annotation study of Appendix~\ref{app:protocol}, which would convert cross-extractor agreement into validity statements and measure the gain from targeted review. Second, a depreciation schedule estimated, not assumed---condition-driven, slower than straight-line, and calibrated against assessor values and transactions jointly with an explicit model of their errors---which the age gradient of Section~\ref{sec:subgroups} shows to be the largest remaining lever. Third, access to the detailed cost schedules of the provincial manual, or to a licensed cost manual, which would replace the two industry anchors with a professional one and allow the regional level to be measured rather than varied. Fourth, a study of photograph selection as an economic choice: the room-by-position evidence of Section~\ref{sec:mnar} is the raw material for a model of what sellers choose to show, which would turn the missing-not-at-random structure from a nuisance into information.31