spb/wp20_uqo

1% ============================================================================2\section{Related literature and research gap}3\label{sec:literature}4% ============================================================================56The paper sits at the intersection of six literatures. We organise the review around the questions the framework must answer rather than around disciplines, and we close with the gap the study addresses. All references were verified against primary bibliographic records; the review characterises each work by its contribution and does not attribute to it figures we did not re-derive.78\subsection{Cost approach, depreciation and the land--structure split}910The cost approach rests on the principle of substitution and on the decomposition of a property into land and depreciated improvements. The economics of depreciation has been studied largely through prices rather than costs: \citet{hulten1981estimation} estimate economic depreciation from vintage asset prices, and a sequence of housing studies estimates net depreciation rates of the order of one to two percent a year, with large variation across markets, tenures and maintenance levels \citep{malpezzi1987microeconomic,shilling1991measuring,knight1996depreciation,harding2007depreciation,wilhelmsson2008house}. Two lessons from this work shape our depreciation model. First, chronological age is a poor proxy for the loss of value because it confounds deterioration, vintage and expected redevelopment \citep{clapp1998residential,coulson2008estimating,chau2005adjusting}; the age--life method must therefore use an \emph{effective} age. Second, observed maintenance and condition offset age substantially \citep{knight1996depreciation,harding2007depreciation,wilhelmsson2008house}, which is exactly the information a photograph can carry and a roll cannot. The decomposition of value into land and structure has been pursued with replacement-cost residuals at the aggregate level \citep{davis2007price,davis2008price} and with hierarchical repeat-sales models at the property level \citep{francke2017land}; the ``builder's model'' of \citet{diewert2017alternative} is the econometric mirror of the cost identity we implement. These studies also document that house prices can exceed replacement cost plus land for long periods where supply is constrained \citep{glaeser2005why,gyourko2006construction}, which is why we treat the ratio of a cost indication to a price as a market-alignment statistic, not an error. Published construction cost indices are known to be imperfect measures of cost \citep{somerville1999residential}; we use the Statistics Canada index only for temporal updating and treat levels as uncertain. Depreciation has also been characterised for commercial and multifamily property \citep{bokhari2018characteristics} and across regions \citep{yoshida2020economic}; the regional heterogeneity it reveals motivates our metropolitan-area diagnostics.1112\subsection{Automated valuation, assessment and valuation standards}1314Automated valuation models have progressed from hedonic regressions \citep{rosen1974hedonic,malpezzi2003hedonic,hill2013hedonic} through spatial and machine-learning estimators \citep{case2004modeling,bourassa2010predicting,zurada2011comparison,mayer2019estimation,pace2020examining} to large-scale commercial systems whose accuracy depends above all on the breadth of the characteristics they observe \citep{kok2017big}. The information in listing text has been shown to add to structured fields \citep{nowak2017textual,shen2021information,baur2023automated}, as has condition information from inspection reports \citep{oust2025assessing}. Unobserved renovations bias both indices and valuations \citep{bogin2019property}. On the appraisal side, valuation accuracy and variation have long been recognised as inherent \citep{crosby2000valuation}, and assessed values exhibit systematic error that ratio studies measure \citep{paglin1972equity,mcmillen2020assessment}. Interval reporting for AVMs is an active topic, with model-based intervals outperforming error-based ones \citep{krause2020uncertainty}. Québec-specific hedonic work documents the capitalisation of landscaping, schools, transit and location-specific attribute values \citep{desrosiers2001size,desrosiers2002landscaping,theriault2003modelling,dube2014spatial}, and the author's companion working papers describe the Vrai Prix data and the empirical performance of comparable-based methods on them. None of this work implements a cost approach whose inputs are extracted from listing media.1516\subsection{Computer vision for real estate}1718A decade of work has used images to predict house prices or to measure what buyers see. Early contributions combined visual and textual features \citep{ahmed2016house}, graded the luxury of interiors \citep{poursaeed2018vision}, and exploited satellite and street-view imagery \citep{bency2017beyond,law2019take,you2017image,kang2021understanding,kucklick2023tackling}. Exterior appearance has been shown to carry price premia, with the important caveat that looks are partly strategic \citep{glaeser2018computer,johnson2020valuing,lindenthal2021machine,nordstrom2024using}, and photographs in listings affect price and marketing time \citep{benefield2011relationship,seiler2012toward}. Closer to our object, building condition has been estimated from fa\c{c}ade patches \citep{koch2018visual}, construction epoch and energy demand from photographs \citep{zeppelzauer2018automatic,despotovic2019prediction}, condition classes have been fed into pricing models \citep{jerer2024platforms}, and, most recently, interior condition from listing images has been shown to improve automated valuation for Oslo condominiums \citep{almaas2026giving}. Multi-view approaches combine interior, exterior and satellite images \citep{nouriani2022vision}. Across this literature the target is the \emph{price}, the model is a supervised predictor trained on prices, and the visual signal is a learned embedding or a condition class whose link to any cost quantity is implicit. Perception-scaling methods such as Streetscore \citep{naik2014streetscore} show how human judgments can be scaled by vision models but also that the judgments scaled are perceptual, not technical.1920\subsection{Vision--language models, structured extraction and calibration}2122Contrastive pre-training \citep{radford2021learning}, few-shot multimodal models \citep{alayrac2022flamingo} and visual instruction tuning \citep{liu2023visual} have produced general-purpose vision--language models that answer schema-constrained questions about an image, a setting that descends from visual question answering \citep{antol2015vqa}. Schema-constrained extraction with large language models has been validated in scientific text \citep{dagdelen2024structured} and applied to real-estate documents \citep{zhao2024utilizing,shehbaz2026structured}, and language models have been tested as appraisers, where they prove sensitive to prompting and to anchors \citep{geerts2025performance}. Two risks are documented. Models hallucinate---in language \citep{ji2023survey} and, for vision--language models, by reporting frequent objects that are not in the image \citep{li2023evaluating}---which is why an abstention value is a design requirement rather than an option. And stated confidence is only partially calibrated \citep{guo2017calibration,kadavath2022language}; ensembles and Bayesian approximations provide alternative uncertainty signals \citep{lakshminarayanan2017simple,gal2016dropout}, and conformal methods offer distribution-free intervals once a calibration set exists \citep{angelopoulos2023conformal}. Human-in-the-loop designs organise the interaction between such models and expert review \citep{mosqueirarey2023human}.2324\subsection{Measurement error, missingness, leakage, interpretability and fairness}2526Attribute extraction errors enter the cost model as measurement error in categorical and continuous inputs, with the attenuation and misclassification consequences surveyed by \citet{hausman2001mismeasured}, \citet{bound2001measurement} and \citet{schennach2016recent}. Which rooms a seller photographs is a missing-data mechanism in the sense of \citet{rubin1976inference}, and there is every reason to expect it to be missing not at random; the selection problem is the one \citet{heckman1979sample} formalised, here at the level of images within a listing. Leakage---the model seeing the target through a side channel---is a documented cause of over-optimistic machine-learning results \citep{kapoor2023leakage}, and in listings the asking price is literally written in some descriptions, which motivates our masking. Because the cost approach is used for high-stakes decisions, interpretability is a requirement, in the sense of \citet{lipton2018mythos} and the argument of \citet{rudin2019stop} for inherently interpretable models; and because valuation inputs can encode neighbourhood composition, fairness across housing segments must be examined \citep{mehrabi2021survey,howell2021increasing,kermani2021racial}. Reliability of human coding, which our proposed expert protocol would require, has standard measures and conventions \citep{cohen1960coefficient,cohen1968weighted,landis1977measurement,hayes2007answering,artstein2008inter,hallgren2012computing,mchugh2012interrater,koo2016guideline}, and non-expert annotation can approach expert quality when aggregated \citep{snow2008cheap}.2728\subsection{Construction cost estimation}2930Parametric and learning-based estimation of construction cost has a long history in construction management, with regression, neural-network and case-based models reaching single-digit percentage errors at the early design stage when the quantity and the grade are known \citep{kim2004comparison,gunaydin2004neural,elfaki2014using}, and with dynamic models for the indices that update them \citep{hwang2009dynamic}. This literature takes the building's specification as given; the question of \emph{how to obtain the specification of an existing house from the evidence a listing provides} is not asked.3132\subsection{The gap}3334Putting the strands together, the state of knowledge is this. Images carry price-relevant information, and condition extracted from images improves price prediction. Depreciation depends on condition and maintenance, which rolls do not record. Vision--language models can answer structured questions about images with known failure modes. Cost estimation is accurate once the specification is known. What is missing is a study that (i) extracts an \emph{appraisal ontology}---materials, grades, condition, renovation evidence---rather than a price embedding; (ii) separates what is seen from what is claimed and from what is recorded; (iii) maps the extracted attributes into a transparent cost identity with every parameter sourced; (iv) measures the missing-not-at-random structure of listing photographs that governs what can be extracted at all; and (v) evaluates the result against administrative records and against professionally produced assessor values rather than against prices alone. The study closest to ours, \citet{almaas2026giving}, does (part of) the condition extraction for condominiums and feeds it to a price model; ours is, to our knowledge, the first to work on the cost approach, on detached houses with a verified administrative link, and with an explicit evidence-class ontology. We state this as a positioning claim, not a priority claim: the field is moving quickly and we searched publisher databases, Crossref and preprint servers but cannot exclude concurrent work.35