spb/wp20_uqo

1% ============================================================================2\section{Introduction}3\label{sec:introduction}4% ============================================================================56Of the three approaches to value, the cost approach is the one that depends least on the market and most on the building. Its logic is simple: a buyer will not pay more for an existing house than it would cost to acquire a comparable lot and build an equivalent house, less an allowance for the age and condition of the existing one. Its practice is demanding: it requires a specification of the improvements---their size, type, materials, finish grade, condition and renovation history---that only an inspection provides. Mass-appraisal systems, including the Québec assessment roll studied here, carry the size, the type and the year built but not the materials, the grade or the condition; appraisers obtain those on site. The question this paper asks is whether the photographs and text that sellers themselves publish when they list a house can supply that specification, how well, and what a transparent cost indication built on it is worth.78The question has become tractable for two reasons. Vision--language models now answer structured questions about an image---what room is this, what is the counter made of, how worn is the exterior---at a cost of a fraction of a cent per photograph, with an explicit option to say that the feature is not visible. And data infrastructures such as Vrai Prix, the author's research platform, link the millions of records of the provincial assessment roll to for-sale listings with their media, so that extracted attributes can be set against verified administrative fields, against the fields brokers enter, and against the values professional assessors have produced for the same buildings under a cost-based manual. The combination allows a study that is neither a demonstration on a few photographs nor a black-box price predictor, but an audit of an extraction-to-valuation chain at every link.910We make four contributions. First, we formalise the chain. Section~\ref{sec:framework} states the cost identity with explicit scope for each term, introduces an attribute ontology with five evidence classes---administrative, visual, proxy, text-claimed and latent---so that what is recorded, seen, judged, claimed and unknowable are never conflated, defines deterministic cross-image aggregation and text--image--record reconciliation rules, and maps the attributes into a cost schedule and a condition-adjusted age--life depreciation. Two propositions characterise the value of an image observation and the bias that misreported condition induces in depreciation. Second, we implement the chain on \nSample{} Québec house listings drawn by stratified sampling from a frame of \nFrame{} listings matched within 15~m to their roll unit, with \nImagesSample{} photographs analysed by three extractor tiers, every parameter of the cost schedule sourced or declared, and no parameter fitted on prices. Third, we evaluate each link: agreement of visual attributes with broker fields and the roll, information content of the image and text channels, cross-extractor reliability of judgment attributes, confidence calibration, the missing-not-at-random structure of what sellers photograph, and the alignment of the cost indication with assessor values, asking prices and recorded sales, under a geographic hold-out. Fourth, we are explicit about what cannot be established with these data---no expert labels, no view of hidden defects, no cross-city cost levels---and we release the ontology, prompts, schedule, code and a pre-registered plan for the expert annotation study that should follow.1112Three findings organise the results. Visible exterior facts are extracted well: against broker fields, the mid-size extractor agrees on roof covering in \accroofcovering\% of informative cases, on garage type in \accgaragetypeattacheddetached\% and on the primary cladding in \acccladdingprimarybrokersidings\%, with coverage above 85\% for each; storeys, by contrast, agree with the roll only loosely ($\kappa=\kappastoreysvisible$) because roll and image count half-storeys and walk-outs differently. The binding constraint is not the extractor but the seller: within the first ten photographs, only \shownBasement\% of listings show a basement and \shownBathroom\% a bathroom, while \shownKitchen\% show a kitchen and \shownExterior\% an exterior; basement finish, one of the largest add-ons in a cost schedule, is therefore essentially a text-claimed attribute in this setting, and interior condition is judged from living rooms and kitchens. And the cost indication responds to what is extracted mostly through depreciation: moving from a roll-only baseline to the multimodal baseline lowers the median depreciation from \DmedsonnetBZeroB{} to \DmedsonnetBThreeB{}, raises the median indication from \$\VsonnetBZeroB{} to \$\VsonnetBThreeB{}, and moves its ratio to the asking price from \priceRatiosonnetBZeroB{} to \priceRatiosonnetBThreeB{} with a median absolute deviation falling from \priceMdapesonnetBZeroB\% to \priceMdapesonnetBThreeB\%. The improvement is real, replicated across extractor tiers and robust to the depreciation assumptions we vary, but it is modest next to the effect of the published unit-cost anchor: switching from a Québec 2026 industry band to a Montréal hard-cost band moves the median indication by \sensAnchorAAltusMontralHardCostsDelta\%. A calibration of the schedule against assessor building values on five metropolitan areas transfers to the Québec City hold-out and, under the Montréal anchor, implies a standard hard cost of about \$\calABThreeCalibrationLevelImpliedSqft{} per square foot---inside the Québec industry band and well above the Montréal band.1314The paper is a cost-approach paper, not an image-enhanced hedonic paper, and the distinction matters for how the results should be read. A hedonic model would learn from prices how much a photographed granite counter is worth; it would be more accurate and less interpretable, and it would tell an appraiser nothing about replacement cost or depreciation. Here no price enters the construction of the indication; the photographs serve only to specify the building, and the gap between the indication and the price is a finding, interpretable in the light of the literature on prices and construction costs, not a residual to be minimised. Price prediction, cost estimation and appraisal judgment are three different things, and the paper keeps them apart.1516The rest of the paper proceeds as follows. Section~\ref{sec:setting} describes the Québec assessment framework and the Vrai Prix infrastructure, and states precisely which parts of the chain exist in production and which are new. Section~\ref{sec:literature} reviews six literatures and locates the gap. Section~\ref{sec:framework} develops the framework. Section~\ref{sec:data} describes the data, the ontology, the rights questions and the sample. Section~\ref{sec:methodology} describes the implementation, the five baselines and the validation protocol. Sections~\ref{sec:results_extraction} and~\ref{sec:results_cost} report the extraction and the valuation results; Section~\ref{sec:robustness} the robustness, uncertainty, missingness and failure analyses; Section~\ref{sec:discussion} the implications for practice; Section~\ref{sec:limitations} the limitations, ethical questions and the research agenda; Section~\ref{sec:conclusion} concludes.17