spb/wp20_uqo

1% ============================================================================2\section{Conclusion}3\label{sec:conclusion}4% ============================================================================56This paper asked whether the photographs and text that sellers publish can supply the specification of the improvements that the cost approach to value requires, and what a transparent cost indication built on that specification is worth. It formalised an extraction-to-valuation chain with five evidence classes, deterministic aggregation and reconciliation rules, a fully sourced schedule and propagated uncertainty, and it ran the chain on \nSample{} Québec house listings matched to their assessment-roll unit, with \nImagesSample{} photographs read by three vision--language model tiers and every parameter either published or declared.78The answer has three parts. Visible exterior facts are read reliably and abstention works: roof covering, garage type, cladding and building type agree with broker fields in the mid-eighties to mid-nineties with coverage above 85\%, and disagreements concentrate on material look-alikes and roofs that cannot be seen from the street; storeys agree with the roll only loosely because the roll, the broker and the image count half-storeys differently. The binding constraint is what sellers show: in the first ten photographs, basements appear in one listing in thirty and bathrooms in one in five, so the largest interior cost items come from the broker's text, while condition and renovation evidence---the inputs to effective age---are what the images uniquely add. And the cost indication responds to that evidence mostly through depreciation: between a roll-only baseline and the multimodal baseline the median depreciation falls from \DmedsonnetBZeroB{} to \DmedsonnetBThreeB{}, the ratio of the indication to the asking price moves from \priceRatiosonnetBZeroB{} to \priceRatiosonnetBThreeB{} and its median absolute deviation from \priceMdapesonnetBZeroB\% to \priceMdapesonnetBThreeB\%---a real, replicated, but modest gain, next to a one-third difference between two published unit-cost anchors and an age gradient that says the assumed depreciation schedule is too steep for old houses and too lenient for new ones. A calibration against assessor values on five metropolitan areas transfers to a geographic hold-out and implies a standard hard cost near \$300 per square foot, consistent with 2026 Québec industry guides.910The paper is deliberately not an image-enhanced price model. No price enters the indication; the photographs specify the building, the schedule prices it, and the gap to the market is reported and interpreted rather than minimised. What the data cannot establish is stated as plainly as what they can: there are no expert labels, hidden defects are unobservable, cross-city cost levels are not published, assessor values are judgments, asking prices are not transactions, and the frame excludes a large network for rights reasons. The ontology, prompts, schedule, code, pre-registered plan and annotation protocol are released so that the next study---with experts in the loop and a depreciation schedule estimated rather than assumed---can start where this one stops.11