spb/wp20_uqo

1% ============================================================================2\section{Discussion}3\label{sec:discussion}4% ============================================================================56\subsection{What the results mean for the cost approach}78The cost approach has always been limited by the cost of specifying the building. The results say that listing media relax that constraint for some inputs and not for others, and that the pattern is set by sellers rather than by models. Exterior materials, roof covering, garage and building type can be read from a fa\c{c}ade photograph with agreement against broker fields in the nineties; kitchens are usually shown and their counters and cabinets are read with high cross-extractor agreement even if their \emph{grade} is a judgment on which extractors differ; bathrooms are shown in one listing in five and basements in one in thirty, so that two of the three largest interior cost items must come from the broker's text. For a mass-appraisal office this means that the marginal information in photographs, relative to the broker fields it already has, is concentrated in condition and renovation evidence---the inputs to effective age---and in the finish grade of the principal rooms. That is where the cost indication responded: through depreciation, whose median falls from \DmedsonnetBZeroB{} to \DmedsonnetBThreeB{} between the roll-only and the multimodal baselines, with a modest gain in alignment with asking prices and a larger gain in level.910The second lesson is about the schedule rather than the images. The level of the indication depends more on the published unit cost than on anything extracted, and the two published anchors for Québec in 2026 differ by a third. A calibration against assessor values on five metropolitan areas points, under either anchor, to a standard hard cost near \$\calBBThreeCalibrationLevelImpliedSqft{}--\$\calABThreeCalibrationLevelImpliedSqft{} per square foot for a two-storey house and transfers to the Québec City hold-out; the Montréal hard-cost band understates what assessors and the market imply, and the Québec industry band is approximately right. The age gradient of the alignment ratios adds that a 60-year straight-line life depreciates old housing faster than either assessors or buyers do, in line with the net rates of one to two percent a year that the depreciation literature has long reported; a cost approach from listing media would do better with a slower, condition-driven schedule, and the condition information the images provide is exactly what such a schedule needs.1112\subsection{Transparency, human oversight and professional practice}1314The pipeline is interpretable in the plain sense that every number in the indication can be traced to a visible attribute, a declared rule or a sourced parameter; nothing is learned from prices. That is a property appraisers and assessors can audit, and it is the reason the paper insisted on a cost identity rather than a price model: a hedonic model with image embeddings would align better with prices and tell a reviewer nothing about why. The price of transparency is that the schedule's errors are the method's errors---the level and age-slope problems above are visible precisely because nothing was fitted to hide them---and that the indication is twice as dispersed around asking prices as a hedonic model trained on half a million sales. We see three uses in practice. As a \emph{screening} tool, the indication and its interval flag listings whose price is far from replacement cost plus land, or whose photographs contradict the broker's claims, for a reviewer's attention. As a \emph{specification} tool, the extracted ontology---with its evidence class and status flags---is a pre-filled inspection form that an appraiser verifies on site rather than fills from scratch. As a \emph{reconciliation} input, a cost indication whose assumptions are explicit is what valuation standards ask the appraiser to weigh against the comparison and income approaches; the paper supplies the indication and its uncertainty, and leaves the weighing to the professional.1516Human oversight is not optional in this design, and the results say where it should go. Confidence-based triage flags many listings without targeting the uncertain ones, because stated confidence is compressed and uninformative for judgment attributes. Review should instead be triggered by the status flags---text and image contradicting each other on basement, garage or storeys---by the absence of interior evidence, by visible defects, and by the age bands where the schedule is least reliable. The expert annotation protocol in Appendix~\ref{app:protocol} is designed to measure both the reliability of the judgment attributes and the gain from such targeted review.1718\subsection{Scalability and cost}1920At batch prices the three extractor tiers processed \nSample{} listings and \nImagesSample{} photographs for about US\$\apiCostUSD{}; the cheapest tier alone, which replicates the cost results (Section~\ref{sec:tiers}), costs a few dollars per thousand listings. Processing the \nFrame{}-listing frame, or the roughly 40,000 house listings active on the companion platform at any time, is therefore inexpensive; what does not scale is the evidence. Eight photographs capture what a listing will reveal, and no budget reveals the basement. The binding constraints on a deployment are rights and representativeness rather than compute: one large franchise network's media host disallows automated retrieval and was excluded, removing two fifths of the frame and its more urban, higher-land-value listings; terms of use vary by host; and listings are a selected sample of the stock. A production system would need licensed access to media and a sampling design that corrects for both selections.2122\subsection{Relation to the literature}2324Relative to the image-based valuation literature, the paper changes the target from price to specification and the model from a learned predictor to an audited chain, and it adds a measurement of what sellers photograph that the price literature has noted \citep{almaas2026giving} but not quantified by room and position. Relative to the depreciation literature, it shows that a condition score read from photographs moves effective age in the direction and with the age-dependence that theory predicts, and that its scale is extractor-dependent, which is a measurement-error statement of the kind \citet{schennach2016recent} would recognise. Relative to the construction-cost literature, it supplies the missing first step---obtaining the specification of an existing house from public evidence---and shows that the second step is still dominated by the level of the published schedule. And relative to the vision--language literature, it finds the well-known pattern: factual, visible attributes are reliable and abstention works, while judgment attributes are reliable only up to a scale that differs across models.25